f3c40898fc7d305a9815acccc1e6e366c9f3dd5b
[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                 blk_end_request_all(req, -EIO);
901                 scsi_next_command(cmd);
902                 break;
903         case ACTION_REPREP:
904                 /* Unprep the request and put it back at the head of the queue.
905                  * A new command will be prepared and issued.
906                  */
907                 scsi_release_buffers(cmd);
908                 scsi_requeue_command(q, cmd);
909                 break;
910         case ACTION_RETRY:
911                 /* Retry the same command immediately */
912                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY, 0);
913                 break;
914         case ACTION_DELAYED_RETRY:
915                 /* Retry the same command after a delay */
916                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY, 0);
917                 break;
918         }
919 }
920
921 static int scsi_init_sgtable(struct request *req, struct scsi_data_buffer *sdb,
922                              gfp_t gfp_mask)
923 {
924         int count;
925
926         /*
927          * If sg table allocation fails, requeue request later.
928          */
929         if (unlikely(scsi_alloc_sgtable(sdb, req->nr_phys_segments,
930                                         gfp_mask))) {
931                 return BLKPREP_DEFER;
932         }
933
934         req->buffer = NULL;
935
936         /* 
937          * Next, walk the list, and fill in the addresses and sizes of
938          * each segment.
939          */
940         count = blk_rq_map_sg(req->q, req, sdb->table.sgl);
941         BUG_ON(count > sdb->table.nents);
942         sdb->table.nents = count;
943         sdb->length = blk_rq_bytes(req);
944         return BLKPREP_OK;
945 }
946
947 /*
948  * Function:    scsi_init_io()
949  *
950  * Purpose:     SCSI I/O initialize function.
951  *
952  * Arguments:   cmd   - Command descriptor we wish to initialize
953  *
954  * Returns:     0 on success
955  *              BLKPREP_DEFER if the failure is retryable
956  *              BLKPREP_KILL if the failure is fatal
957  */
958 int scsi_init_io(struct scsi_cmnd *cmd, gfp_t gfp_mask)
959 {
960         int error = scsi_init_sgtable(cmd->request, &cmd->sdb, gfp_mask);
961         if (error)
962                 goto err_exit;
963
964         if (blk_bidi_rq(cmd->request)) {
965                 struct scsi_data_buffer *bidi_sdb = kmem_cache_zalloc(
966                         scsi_sdb_cache, GFP_ATOMIC);
967                 if (!bidi_sdb) {
968                         error = BLKPREP_DEFER;
969                         goto err_exit;
970                 }
971
972                 cmd->request->next_rq->special = bidi_sdb;
973                 error = scsi_init_sgtable(cmd->request->next_rq, bidi_sdb,
974                                                                     GFP_ATOMIC);
975                 if (error)
976                         goto err_exit;
977         }
978
979         if (blk_integrity_rq(cmd->request)) {
980                 struct scsi_data_buffer *prot_sdb = cmd->prot_sdb;
981                 int ivecs, count;
982
983                 BUG_ON(prot_sdb == NULL);
984                 ivecs = blk_rq_count_integrity_sg(cmd->request);
985
986                 if (scsi_alloc_sgtable(prot_sdb, ivecs, gfp_mask)) {
987                         error = BLKPREP_DEFER;
988                         goto err_exit;
989                 }
990
991                 count = blk_rq_map_integrity_sg(cmd->request,
992                                                 prot_sdb->table.sgl);
993                 BUG_ON(unlikely(count > ivecs));
994
995                 cmd->prot_sdb = prot_sdb;
996                 cmd->prot_sdb->table.nents = count;
997         }
998
999         return BLKPREP_OK ;
1000
1001 err_exit:
1002         scsi_release_buffers(cmd);
1003         if (error == BLKPREP_KILL)
1004                 scsi_put_command(cmd);
1005         else /* BLKPREP_DEFER */
1006                 scsi_unprep_request(cmd->request);
1007
1008         return error;
1009 }
1010 EXPORT_SYMBOL(scsi_init_io);
1011
1012 static struct scsi_cmnd *scsi_get_cmd_from_req(struct scsi_device *sdev,
1013                 struct request *req)
1014 {
1015         struct scsi_cmnd *cmd;
1016
1017         if (!req->special) {
1018                 cmd = scsi_get_command(sdev, GFP_ATOMIC);
1019                 if (unlikely(!cmd))
1020                         return NULL;
1021                 req->special = cmd;
1022         } else {
1023                 cmd = req->special;
1024         }
1025
1026         /* pull a tag out of the request if we have one */
1027         cmd->tag = req->tag;
1028         cmd->request = req;
1029
1030         cmd->cmnd = req->cmd;
1031
1032         return cmd;
1033 }
1034
1035 int scsi_setup_blk_pc_cmnd(struct scsi_device *sdev, struct request *req)
1036 {
1037         struct scsi_cmnd *cmd;
1038         int ret = scsi_prep_state_check(sdev, req);
1039
1040         if (ret != BLKPREP_OK)
1041                 return ret;
1042
1043         cmd = scsi_get_cmd_from_req(sdev, req);
1044         if (unlikely(!cmd))
1045                 return BLKPREP_DEFER;
1046
1047         /*
1048          * BLOCK_PC requests may transfer data, in which case they must
1049          * a bio attached to them.  Or they might contain a SCSI command
1050          * that does not transfer data, in which case they may optionally
1051          * submit a request without an attached bio.
1052          */
1053         if (req->bio) {
1054                 int ret;
1055
1056                 BUG_ON(!req->nr_phys_segments);
1057
1058                 ret = scsi_init_io(cmd, GFP_ATOMIC);
1059                 if (unlikely(ret))
1060                         return ret;
1061         } else {
1062                 BUG_ON(blk_rq_bytes(req));
1063
1064                 memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1065                 req->buffer = NULL;
1066         }
1067
1068         cmd->cmd_len = req->cmd_len;
1069         if (!blk_rq_bytes(req))
1070                 cmd->sc_data_direction = DMA_NONE;
1071         else if (rq_data_dir(req) == WRITE)
1072                 cmd->sc_data_direction = DMA_TO_DEVICE;
1073         else
1074                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1075         
1076         cmd->transfersize = blk_rq_bytes(req);
1077         cmd->allowed = req->retries;
1078         return BLKPREP_OK;
1079 }
1080 EXPORT_SYMBOL(scsi_setup_blk_pc_cmnd);
1081
1082 /*
1083  * Setup a REQ_TYPE_FS command.  These are simple read/write request
1084  * from filesystems that still need to be translated to SCSI CDBs from
1085  * the ULD.
1086  */
1087 int scsi_setup_fs_cmnd(struct scsi_device *sdev, struct request *req)
1088 {
1089         struct scsi_cmnd *cmd;
1090         int ret = scsi_prep_state_check(sdev, req);
1091
1092         if (ret != BLKPREP_OK)
1093                 return ret;
1094
1095         if (unlikely(sdev->scsi_dh_data && sdev->scsi_dh_data->scsi_dh
1096                          && sdev->scsi_dh_data->scsi_dh->prep_fn)) {
1097                 ret = sdev->scsi_dh_data->scsi_dh->prep_fn(sdev, req);
1098                 if (ret != BLKPREP_OK)
1099                         return ret;
1100         }
1101
1102         /*
1103          * Filesystem requests must transfer data.
1104          */
1105         BUG_ON(!req->nr_phys_segments);
1106
1107         cmd = scsi_get_cmd_from_req(sdev, req);
1108         if (unlikely(!cmd))
1109                 return BLKPREP_DEFER;
1110
1111         memset(cmd->cmnd, 0, BLK_MAX_CDB);
1112         return scsi_init_io(cmd, GFP_ATOMIC);
1113 }
1114 EXPORT_SYMBOL(scsi_setup_fs_cmnd);
1115
1116 int scsi_prep_state_check(struct scsi_device *sdev, struct request *req)
1117 {
1118         int ret = BLKPREP_OK;
1119
1120         /*
1121          * If the device is not in running state we will reject some
1122          * or all commands.
1123          */
1124         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1125                 switch (sdev->sdev_state) {
1126                 case SDEV_OFFLINE:
1127                         /*
1128                          * If the device is offline we refuse to process any
1129                          * commands.  The device must be brought online
1130                          * before trying any recovery commands.
1131                          */
1132                         sdev_printk(KERN_ERR, sdev,
1133                                     "rejecting I/O to offline device\n");
1134                         ret = BLKPREP_KILL;
1135                         break;
1136                 case SDEV_DEL:
1137                         /*
1138                          * If the device is fully deleted, we refuse to
1139                          * process any commands as well.
1140                          */
1141                         sdev_printk(KERN_ERR, sdev,
1142                                     "rejecting I/O to dead device\n");
1143                         ret = BLKPREP_KILL;
1144                         break;
1145                 case SDEV_QUIESCE:
1146                 case SDEV_BLOCK:
1147                 case SDEV_CREATED_BLOCK:
1148                         /*
1149                          * If the devices is blocked we defer normal commands.
1150                          */
1151                         if (!(req->cmd_flags & REQ_PREEMPT))
1152                                 ret = BLKPREP_DEFER;
1153                         break;
1154                 default:
1155                         /*
1156                          * For any other not fully online state we only allow
1157                          * special commands.  In particular any user initiated
1158                          * command is not allowed.
1159                          */
1160                         if (!(req->cmd_flags & REQ_PREEMPT))
1161                                 ret = BLKPREP_KILL;
1162                         break;
1163                 }
1164         }
1165         return ret;
1166 }
1167 EXPORT_SYMBOL(scsi_prep_state_check);
1168
1169 int scsi_prep_return(struct request_queue *q, struct request *req, int ret)
1170 {
1171         struct scsi_device *sdev = q->queuedata;
1172
1173         switch (ret) {
1174         case BLKPREP_KILL:
1175                 req->errors = DID_NO_CONNECT << 16;
1176                 /* release the command and kill it */
1177                 if (req->special) {
1178                         struct scsi_cmnd *cmd = req->special;
1179                         scsi_release_buffers(cmd);
1180                         scsi_put_command(cmd);
1181                         req->special = NULL;
1182                 }
1183                 break;
1184         case BLKPREP_DEFER:
1185                 /*
1186                  * If we defer, the blk_peek_request() returns NULL, but the
1187                  * queue must be restarted, so we plug here if no returning
1188                  * command will automatically do that.
1189                  */
1190                 if (sdev->device_busy == 0)
1191                         blk_plug_device(q);
1192                 break;
1193         default:
1194                 req->cmd_flags |= REQ_DONTPREP;
1195         }
1196
1197         return ret;
1198 }
1199 EXPORT_SYMBOL(scsi_prep_return);
1200
1201 int scsi_prep_fn(struct request_queue *q, struct request *req)
1202 {
1203         struct scsi_device *sdev = q->queuedata;
1204         int ret = BLKPREP_KILL;
1205
1206         if (req->cmd_type == REQ_TYPE_BLOCK_PC)
1207                 ret = scsi_setup_blk_pc_cmnd(sdev, req);
1208         return scsi_prep_return(q, req, ret);
1209 }
1210 EXPORT_SYMBOL(scsi_prep_fn);
1211
1212 /*
1213  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1214  * return 0.
1215  *
1216  * Called with the queue_lock held.
1217  */
1218 static inline int scsi_dev_queue_ready(struct request_queue *q,
1219                                   struct scsi_device *sdev)
1220 {
1221         if (sdev->device_busy == 0 && sdev->device_blocked) {
1222                 /*
1223                  * unblock after device_blocked iterates to zero
1224                  */
1225                 if (--sdev->device_blocked == 0) {
1226                         SCSI_LOG_MLQUEUE(3,
1227                                    sdev_printk(KERN_INFO, sdev,
1228                                    "unblocking device at zero depth\n"));
1229                 } else {
1230                         blk_plug_device(q);
1231                         return 0;
1232                 }
1233         }
1234         if (scsi_device_is_busy(sdev))
1235                 return 0;
1236
1237         return 1;
1238 }
1239
1240
1241 /*
1242  * scsi_target_queue_ready: checks if there we can send commands to target
1243  * @sdev: scsi device on starget to check.
1244  *
1245  * Called with the host lock held.
1246  */
1247 static inline int scsi_target_queue_ready(struct Scsi_Host *shost,
1248                                            struct scsi_device *sdev)
1249 {
1250         struct scsi_target *starget = scsi_target(sdev);
1251
1252         if (starget->single_lun) {
1253                 if (starget->starget_sdev_user &&
1254                     starget->starget_sdev_user != sdev)
1255                         return 0;
1256                 starget->starget_sdev_user = sdev;
1257         }
1258
1259         if (starget->target_busy == 0 && starget->target_blocked) {
1260                 /*
1261                  * unblock after target_blocked iterates to zero
1262                  */
1263                 if (--starget->target_blocked == 0) {
1264                         SCSI_LOG_MLQUEUE(3, starget_printk(KERN_INFO, starget,
1265                                          "unblocking target at zero depth\n"));
1266                 } else
1267                         return 0;
1268         }
1269
1270         if (scsi_target_is_busy(starget)) {
1271                 if (list_empty(&sdev->starved_entry)) {
1272                         list_add_tail(&sdev->starved_entry,
1273                                       &shost->starved_list);
1274                         return 0;
1275                 }
1276         }
1277
1278         /* We're OK to process the command, so we can't be starved */
1279         if (!list_empty(&sdev->starved_entry))
1280                 list_del_init(&sdev->starved_entry);
1281         return 1;
1282 }
1283
1284 /*
1285  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1286  * return 0. We must end up running the queue again whenever 0 is
1287  * returned, else IO can hang.
1288  *
1289  * Called with host_lock held.
1290  */
1291 static inline int scsi_host_queue_ready(struct request_queue *q,
1292                                    struct Scsi_Host *shost,
1293                                    struct scsi_device *sdev)
1294 {
1295         if (scsi_host_in_recovery(shost))
1296                 return 0;
1297         if (shost->host_busy == 0 && shost->host_blocked) {
1298                 /*
1299                  * unblock after host_blocked iterates to zero
1300                  */
1301                 if (--shost->host_blocked == 0) {
1302                         SCSI_LOG_MLQUEUE(3,
1303                                 printk("scsi%d unblocking host at zero depth\n",
1304                                         shost->host_no));
1305                 } else {
1306                         return 0;
1307                 }
1308         }
1309         if (scsi_host_is_busy(shost)) {
1310                 if (list_empty(&sdev->starved_entry))
1311                         list_add_tail(&sdev->starved_entry, &shost->starved_list);
1312                 return 0;
1313         }
1314
1315         /* We're OK to process the command, so we can't be starved */
1316         if (!list_empty(&sdev->starved_entry))
1317                 list_del_init(&sdev->starved_entry);
1318
1319         return 1;
1320 }
1321
1322 /*
1323  * Busy state exporting function for request stacking drivers.
1324  *
1325  * For efficiency, no lock is taken to check the busy state of
1326  * shost/starget/sdev, since the returned value is not guaranteed and
1327  * may be changed after request stacking drivers call the function,
1328  * regardless of taking lock or not.
1329  *
1330  * When scsi can't dispatch I/Os anymore and needs to kill I/Os
1331  * (e.g. !sdev), scsi needs to return 'not busy'.
1332  * Otherwise, request stacking drivers may hold requests forever.
1333  */
1334 static int scsi_lld_busy(struct request_queue *q)
1335 {
1336         struct scsi_device *sdev = q->queuedata;
1337         struct Scsi_Host *shost;
1338         struct scsi_target *starget;
1339
1340         if (!sdev)
1341                 return 0;
1342
1343         shost = sdev->host;
1344         starget = scsi_target(sdev);
1345
1346         if (scsi_host_in_recovery(shost) || scsi_host_is_busy(shost) ||
1347             scsi_target_is_busy(starget) || scsi_device_is_busy(sdev))
1348                 return 1;
1349
1350         return 0;
1351 }
1352
1353 /*
1354  * Kill a request for a dead device
1355  */
1356 static void scsi_kill_request(struct request *req, struct request_queue *q)
1357 {
1358         struct scsi_cmnd *cmd = req->special;
1359         struct scsi_device *sdev = cmd->device;
1360         struct scsi_target *starget = scsi_target(sdev);
1361         struct Scsi_Host *shost = sdev->host;
1362
1363         blk_start_request(req);
1364
1365         if (unlikely(cmd == NULL)) {
1366                 printk(KERN_CRIT "impossible request in %s.\n",
1367                                  __func__);
1368                 BUG();
1369         }
1370
1371         scsi_init_cmd_errh(cmd);
1372         cmd->result = DID_NO_CONNECT << 16;
1373         atomic_inc(&cmd->device->iorequest_cnt);
1374
1375         /*
1376          * SCSI request completion path will do scsi_device_unbusy(),
1377          * bump busy counts.  To bump the counters, we need to dance
1378          * with the locks as normal issue path does.
1379          */
1380         sdev->device_busy++;
1381         spin_unlock(sdev->request_queue->queue_lock);
1382         spin_lock(shost->host_lock);
1383         shost->host_busy++;
1384         starget->target_busy++;
1385         spin_unlock(shost->host_lock);
1386         spin_lock(sdev->request_queue->queue_lock);
1387
1388         blk_complete_request(req);
1389 }
1390
1391 static void scsi_softirq_done(struct request *rq)
1392 {
1393         struct scsi_cmnd *cmd = rq->special;
1394         unsigned long wait_for = (cmd->allowed + 1) * rq->timeout;
1395         int disposition;
1396
1397         INIT_LIST_HEAD(&cmd->eh_entry);
1398
1399         /*
1400          * Set the serial numbers back to zero
1401          */
1402         cmd->serial_number = 0;
1403
1404         atomic_inc(&cmd->device->iodone_cnt);
1405         if (cmd->result)
1406                 atomic_inc(&cmd->device->ioerr_cnt);
1407
1408         disposition = scsi_decide_disposition(cmd);
1409         if (disposition != SUCCESS &&
1410             time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1411                 sdev_printk(KERN_ERR, cmd->device,
1412                             "timing out command, waited %lus\n",
1413                             wait_for/HZ);
1414                 disposition = SUCCESS;
1415         }
1416                         
1417         scsi_log_completion(cmd, disposition);
1418
1419         switch (disposition) {
1420                 case SUCCESS:
1421                         scsi_finish_command(cmd);
1422                         break;
1423                 case NEEDS_RETRY:
1424                         scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1425                         break;
1426                 case ADD_TO_MLQUEUE:
1427                         scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1428                         break;
1429                 default:
1430                         if (!scsi_eh_scmd_add(cmd, 0))
1431                                 scsi_finish_command(cmd);
1432         }
1433 }
1434
1435 /*
1436  * Function:    scsi_request_fn()
1437  *
1438  * Purpose:     Main strategy routine for SCSI.
1439  *
1440  * Arguments:   q       - Pointer to actual queue.
1441  *
1442  * Returns:     Nothing
1443  *
1444  * Lock status: IO request lock assumed to be held when called.
1445  */
1446 static void scsi_request_fn(struct request_queue *q)
1447 {
1448         struct scsi_device *sdev = q->queuedata;
1449         struct Scsi_Host *shost;
1450         struct scsi_cmnd *cmd;
1451         struct request *req;
1452
1453         if (!sdev) {
1454                 printk("scsi: killing requests for dead queue\n");
1455                 while ((req = blk_peek_request(q)) != NULL)
1456                         scsi_kill_request(req, q);
1457                 return;
1458         }
1459
1460         if(!get_device(&sdev->sdev_gendev))
1461                 /* We must be tearing the block queue down already */
1462                 return;
1463
1464         /*
1465          * To start with, we keep looping until the queue is empty, or until
1466          * the host is no longer able to accept any more requests.
1467          */
1468         shost = sdev->host;
1469         while (!blk_queue_plugged(q)) {
1470                 int rtn;
1471                 /*
1472                  * get next queueable request.  We do this early to make sure
1473                  * that the request is fully prepared even if we cannot 
1474                  * accept it.
1475                  */
1476                 req = blk_peek_request(q);
1477                 if (!req || !scsi_dev_queue_ready(q, sdev))
1478                         break;
1479
1480                 if (unlikely(!scsi_device_online(sdev))) {
1481                         sdev_printk(KERN_ERR, sdev,
1482                                     "rejecting I/O to offline device\n");
1483                         scsi_kill_request(req, q);
1484                         continue;
1485                 }
1486
1487
1488                 /*
1489                  * Remove the request from the request list.
1490                  */
1491                 if (!(blk_queue_tagged(q) && !blk_queue_start_tag(q, req)))
1492                         blk_start_request(req);
1493                 sdev->device_busy++;
1494
1495                 spin_unlock(q->queue_lock);
1496                 cmd = req->special;
1497                 if (unlikely(cmd == NULL)) {
1498                         printk(KERN_CRIT "impossible request in %s.\n"
1499                                          "please mail a stack trace to "
1500                                          "linux-scsi@vger.kernel.org\n",
1501                                          __func__);
1502                         blk_dump_rq_flags(req, "foo");
1503                         BUG();
1504                 }
1505                 spin_lock(shost->host_lock);
1506
1507                 /*
1508                  * We hit this when the driver is using a host wide
1509                  * tag map. For device level tag maps the queue_depth check
1510                  * in the device ready fn would prevent us from trying
1511                  * to allocate a tag. Since the map is a shared host resource
1512                  * we add the dev to the starved list so it eventually gets
1513                  * a run when a tag is freed.
1514                  */
1515                 if (blk_queue_tagged(q) && !blk_rq_tagged(req)) {
1516                         if (list_empty(&sdev->starved_entry))
1517                                 list_add_tail(&sdev->starved_entry,
1518                                               &shost->starved_list);
1519                         goto not_ready;
1520                 }
1521
1522                 if (!scsi_target_queue_ready(shost, sdev))
1523                         goto not_ready;
1524
1525                 if (!scsi_host_queue_ready(q, shost, sdev))
1526                         goto not_ready;
1527
1528                 scsi_target(sdev)->target_busy++;
1529                 shost->host_busy++;
1530
1531                 /*
1532                  * XXX(hch): This is rather suboptimal, scsi_dispatch_cmd will
1533                  *              take the lock again.
1534                  */
1535                 spin_unlock_irq(shost->host_lock);
1536
1537                 /*
1538                  * Finally, initialize any error handling parameters, and set up
1539                  * the timers for timeouts.
1540                  */
1541                 scsi_init_cmd_errh(cmd);
1542
1543                 /*
1544                  * Dispatch the command to the low-level driver.
1545                  */
1546                 rtn = scsi_dispatch_cmd(cmd);
1547                 spin_lock_irq(q->queue_lock);
1548                 if(rtn) {
1549                         /* we're refusing the command; because of
1550                          * the way locks get dropped, we need to 
1551                          * check here if plugging is required */
1552                         if(sdev->device_busy == 0)
1553                                 blk_plug_device(q);
1554
1555                         break;
1556                 }
1557         }
1558
1559         goto out;
1560
1561  not_ready:
1562         spin_unlock_irq(shost->host_lock);
1563
1564         /*
1565          * lock q, handle tag, requeue req, and decrement device_busy. We
1566          * must return with queue_lock held.
1567          *
1568          * Decrementing device_busy without checking it is OK, as all such
1569          * cases (host limits or settings) should run the queue at some
1570          * later time.
1571          */
1572         spin_lock_irq(q->queue_lock);
1573         blk_requeue_request(q, req);
1574         sdev->device_busy--;
1575         if(sdev->device_busy == 0)
1576                 blk_plug_device(q);
1577  out:
1578         /* must be careful here...if we trigger the ->remove() function
1579          * we cannot be holding the q lock */
1580         spin_unlock_irq(q->queue_lock);
1581         put_device(&sdev->sdev_gendev);
1582         spin_lock_irq(q->queue_lock);
1583 }
1584
1585 u64 scsi_calculate_bounce_limit(struct Scsi_Host *shost)
1586 {
1587         struct device *host_dev;
1588         u64 bounce_limit = 0xffffffff;
1589
1590         if (shost->unchecked_isa_dma)
1591                 return BLK_BOUNCE_ISA;
1592         /*
1593          * Platforms with virtual-DMA translation
1594          * hardware have no practical limit.
1595          */
1596         if (!PCI_DMA_BUS_IS_PHYS)
1597                 return BLK_BOUNCE_ANY;
1598
1599         host_dev = scsi_get_device(shost);
1600         if (host_dev && host_dev->dma_mask)
1601                 bounce_limit = *host_dev->dma_mask;
1602
1603         return bounce_limit;
1604 }
1605 EXPORT_SYMBOL(scsi_calculate_bounce_limit);
1606
1607 struct request_queue *__scsi_alloc_queue(struct Scsi_Host *shost,
1608                                          request_fn_proc *request_fn)
1609 {
1610         struct request_queue *q;
1611         struct device *dev = shost->shost_gendev.parent;
1612
1613         q = blk_init_queue(request_fn, NULL);
1614         if (!q)
1615                 return NULL;
1616
1617         /*
1618          * this limit is imposed by hardware restrictions
1619          */
1620         blk_queue_max_hw_segments(q, shost->sg_tablesize);
1621         blk_queue_max_phys_segments(q, SCSI_MAX_SG_CHAIN_SEGMENTS);
1622
1623         blk_queue_max_sectors(q, shost->max_sectors);
1624         blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
1625         blk_queue_segment_boundary(q, shost->dma_boundary);
1626         dma_set_seg_boundary(dev, shost->dma_boundary);
1627
1628         blk_queue_max_segment_size(q, dma_get_max_seg_size(dev));
1629
1630         /* New queue, no concurrency on queue_flags */
1631         if (!shost->use_clustering)
1632                 queue_flag_clear_unlocked(QUEUE_FLAG_CLUSTER, q);
1633
1634         /*
1635          * set a reasonable default alignment on word boundaries: the
1636          * host and device may alter it using
1637          * blk_queue_update_dma_alignment() later.
1638          */
1639         blk_queue_dma_alignment(q, 0x03);
1640
1641         return q;
1642 }
1643 EXPORT_SYMBOL(__scsi_alloc_queue);
1644
1645 struct request_queue *scsi_alloc_queue(struct scsi_device *sdev)
1646 {
1647         struct request_queue *q;
1648
1649         q = __scsi_alloc_queue(sdev->host, scsi_request_fn);
1650         if (!q)
1651                 return NULL;
1652
1653         blk_queue_prep_rq(q, scsi_prep_fn);
1654         blk_queue_softirq_done(q, scsi_softirq_done);
1655         blk_queue_rq_timed_out(q, scsi_times_out);
1656         blk_queue_lld_busy(q, scsi_lld_busy);
1657         return q;
1658 }
1659
1660 void scsi_free_queue(struct request_queue *q)
1661 {
1662         blk_cleanup_queue(q);
1663 }
1664
1665 /*
1666  * Function:    scsi_block_requests()
1667  *
1668  * Purpose:     Utility function used by low-level drivers to prevent further
1669  *              commands from being queued to the device.
1670  *
1671  * Arguments:   shost       - Host in question
1672  *
1673  * Returns:     Nothing
1674  *
1675  * Lock status: No locks are assumed held.
1676  *
1677  * Notes:       There is no timer nor any other means by which the requests
1678  *              get unblocked other than the low-level driver calling
1679  *              scsi_unblock_requests().
1680  */
1681 void scsi_block_requests(struct Scsi_Host *shost)
1682 {
1683         shost->host_self_blocked = 1;
1684 }
1685 EXPORT_SYMBOL(scsi_block_requests);
1686
1687 /*
1688  * Function:    scsi_unblock_requests()
1689  *
1690  * Purpose:     Utility function used by low-level drivers to allow further
1691  *              commands from being queued to the device.
1692  *
1693  * Arguments:   shost       - Host in question
1694  *
1695  * Returns:     Nothing
1696  *
1697  * Lock status: No locks are assumed held.
1698  *
1699  * Notes:       There is no timer nor any other means by which the requests
1700  *              get unblocked other than the low-level driver calling
1701  *              scsi_unblock_requests().
1702  *
1703  *              This is done as an API function so that changes to the
1704  *              internals of the scsi mid-layer won't require wholesale
1705  *              changes to drivers that use this feature.
1706  */
1707 void scsi_unblock_requests(struct Scsi_Host *shost)
1708 {
1709         shost->host_self_blocked = 0;
1710         scsi_run_host_queues(shost);
1711 }
1712 EXPORT_SYMBOL(scsi_unblock_requests);
1713
1714 int __init scsi_init_queue(void)
1715 {
1716         int i;
1717
1718         scsi_sdb_cache = kmem_cache_create("scsi_data_buffer",
1719                                            sizeof(struct scsi_data_buffer),
1720                                            0, 0, NULL);
1721         if (!scsi_sdb_cache) {
1722                 printk(KERN_ERR "SCSI: can't init scsi sdb cache\n");
1723                 return -ENOMEM;
1724         }
1725
1726         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1727                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1728                 int size = sgp->size * sizeof(struct scatterlist);
1729
1730                 sgp->slab = kmem_cache_create(sgp->name, size, 0,
1731                                 SLAB_HWCACHE_ALIGN, NULL);
1732                 if (!sgp->slab) {
1733                         printk(KERN_ERR "SCSI: can't init sg slab %s\n",
1734                                         sgp->name);
1735                         goto cleanup_sdb;
1736                 }
1737
1738                 sgp->pool = mempool_create_slab_pool(SG_MEMPOOL_SIZE,
1739                                                      sgp->slab);
1740                 if (!sgp->pool) {
1741                         printk(KERN_ERR "SCSI: can't init sg mempool %s\n",
1742                                         sgp->name);
1743                         goto cleanup_sdb;
1744                 }
1745         }
1746
1747         return 0;
1748
1749 cleanup_sdb:
1750         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1751                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1752                 if (sgp->pool)
1753                         mempool_destroy(sgp->pool);
1754                 if (sgp->slab)
1755                         kmem_cache_destroy(sgp->slab);
1756         }
1757         kmem_cache_destroy(scsi_sdb_cache);
1758
1759         return -ENOMEM;
1760 }
1761
1762 void scsi_exit_queue(void)
1763 {
1764         int i;
1765
1766         kmem_cache_destroy(scsi_sdb_cache);
1767
1768         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1769                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1770                 mempool_destroy(sgp->pool);
1771                 kmem_cache_destroy(sgp->slab);
1772         }
1773 }
1774
1775 /**
1776  *      scsi_mode_select - issue a mode select
1777  *      @sdev:  SCSI device to be queried
1778  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
1779  *      @sp:    Save page bit (0 == don't save, 1 == save)
1780  *      @modepage: mode page being requested
1781  *      @buffer: request buffer (may not be smaller than eight bytes)
1782  *      @len:   length of request buffer.
1783  *      @timeout: command timeout
1784  *      @retries: number of retries before failing
1785  *      @data: returns a structure abstracting the mode header data
1786  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
1787  *              must be SCSI_SENSE_BUFFERSIZE big.
1788  *
1789  *      Returns zero if successful; negative error number or scsi
1790  *      status on error
1791  *
1792  */
1793 int
1794 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
1795                  unsigned char *buffer, int len, int timeout, int retries,
1796                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1797 {
1798         unsigned char cmd[10];
1799         unsigned char *real_buffer;
1800         int ret;
1801
1802         memset(cmd, 0, sizeof(cmd));
1803         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
1804
1805         if (sdev->use_10_for_ms) {
1806                 if (len > 65535)
1807                         return -EINVAL;
1808                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
1809                 if (!real_buffer)
1810                         return -ENOMEM;
1811                 memcpy(real_buffer + 8, buffer, len);
1812                 len += 8;
1813                 real_buffer[0] = 0;
1814                 real_buffer[1] = 0;
1815                 real_buffer[2] = data->medium_type;
1816                 real_buffer[3] = data->device_specific;
1817                 real_buffer[4] = data->longlba ? 0x01 : 0;
1818                 real_buffer[5] = 0;
1819                 real_buffer[6] = data->block_descriptor_length >> 8;
1820                 real_buffer[7] = data->block_descriptor_length;
1821
1822                 cmd[0] = MODE_SELECT_10;
1823                 cmd[7] = len >> 8;
1824                 cmd[8] = len;
1825         } else {
1826                 if (len > 255 || data->block_descriptor_length > 255 ||
1827                     data->longlba)
1828                         return -EINVAL;
1829
1830                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
1831                 if (!real_buffer)
1832                         return -ENOMEM;
1833                 memcpy(real_buffer + 4, buffer, len);
1834                 len += 4;
1835                 real_buffer[0] = 0;
1836                 real_buffer[1] = data->medium_type;
1837                 real_buffer[2] = data->device_specific;
1838                 real_buffer[3] = data->block_descriptor_length;
1839                 
1840
1841                 cmd[0] = MODE_SELECT;
1842                 cmd[4] = len;
1843         }
1844
1845         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
1846                                sshdr, timeout, retries, NULL);
1847         kfree(real_buffer);
1848         return ret;
1849 }
1850 EXPORT_SYMBOL_GPL(scsi_mode_select);
1851
1852 /**
1853  *      scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
1854  *      @sdev:  SCSI device to be queried
1855  *      @dbd:   set if mode sense will allow block descriptors to be returned
1856  *      @modepage: mode page being requested
1857  *      @buffer: request buffer (may not be smaller than eight bytes)
1858  *      @len:   length of request buffer.
1859  *      @timeout: command timeout
1860  *      @retries: number of retries before failing
1861  *      @data: returns a structure abstracting the mode header data
1862  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
1863  *              must be SCSI_SENSE_BUFFERSIZE big.
1864  *
1865  *      Returns zero if unsuccessful, or the header offset (either 4
1866  *      or 8 depending on whether a six or ten byte command was
1867  *      issued) if successful.
1868  */
1869 int
1870 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
1871                   unsigned char *buffer, int len, int timeout, int retries,
1872                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1873 {
1874         unsigned char cmd[12];
1875         int use_10_for_ms;
1876         int header_length;
1877         int result;
1878         struct scsi_sense_hdr my_sshdr;
1879
1880         memset(data, 0, sizeof(*data));
1881         memset(&cmd[0], 0, 12);
1882         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
1883         cmd[2] = modepage;
1884
1885         /* caller might not be interested in sense, but we need it */
1886         if (!sshdr)
1887                 sshdr = &my_sshdr;
1888
1889  retry:
1890         use_10_for_ms = sdev->use_10_for_ms;
1891
1892         if (use_10_for_ms) {
1893                 if (len < 8)
1894                         len = 8;
1895
1896                 cmd[0] = MODE_SENSE_10;
1897                 cmd[8] = len;
1898                 header_length = 8;
1899         } else {
1900                 if (len < 4)
1901                         len = 4;
1902
1903                 cmd[0] = MODE_SENSE;
1904                 cmd[4] = len;
1905                 header_length = 4;
1906         }
1907
1908         memset(buffer, 0, len);
1909
1910         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
1911                                   sshdr, timeout, retries, NULL);
1912
1913         /* This code looks awful: what it's doing is making sure an
1914          * ILLEGAL REQUEST sense return identifies the actual command
1915          * byte as the problem.  MODE_SENSE commands can return
1916          * ILLEGAL REQUEST if the code page isn't supported */
1917
1918         if (use_10_for_ms && !scsi_status_is_good(result) &&
1919             (driver_byte(result) & DRIVER_SENSE)) {
1920                 if (scsi_sense_valid(sshdr)) {
1921                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
1922                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
1923                                 /* 
1924                                  * Invalid command operation code
1925                                  */
1926                                 sdev->use_10_for_ms = 0;
1927                                 goto retry;
1928                         }
1929                 }
1930         }
1931
1932         if(scsi_status_is_good(result)) {
1933                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
1934                              (modepage == 6 || modepage == 8))) {
1935                         /* Initio breakage? */
1936                         header_length = 0;
1937                         data->length = 13;
1938                         data->medium_type = 0;
1939                         data->device_specific = 0;
1940                         data->longlba = 0;
1941                         data->block_descriptor_length = 0;
1942                 } else if(use_10_for_ms) {
1943                         data->length = buffer[0]*256 + buffer[1] + 2;
1944                         data->medium_type = buffer[2];
1945                         data->device_specific = buffer[3];
1946                         data->longlba = buffer[4] & 0x01;
1947                         data->block_descriptor_length = buffer[6]*256
1948                                 + buffer[7];
1949                 } else {
1950                         data->length = buffer[0] + 1;
1951                         data->medium_type = buffer[1];
1952                         data->device_specific = buffer[2];
1953                         data->block_descriptor_length = buffer[3];
1954                 }
1955                 data->header_length = header_length;
1956         }
1957
1958         return result;
1959 }
1960 EXPORT_SYMBOL(scsi_mode_sense);
1961
1962 /**
1963  *      scsi_test_unit_ready - test if unit is ready
1964  *      @sdev:  scsi device to change the state of.
1965  *      @timeout: command timeout
1966  *      @retries: number of retries before failing
1967  *      @sshdr_external: Optional pointer to struct scsi_sense_hdr for
1968  *              returning sense. Make sure that this is cleared before passing
1969  *              in.
1970  *
1971  *      Returns zero if unsuccessful or an error if TUR failed.  For
1972  *      removable media, a return of NOT_READY or UNIT_ATTENTION is
1973  *      translated to success, with the ->changed flag updated.
1974  **/
1975 int
1976 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
1977                      struct scsi_sense_hdr *sshdr_external)
1978 {
1979         char cmd[] = {
1980                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
1981         };
1982         struct scsi_sense_hdr *sshdr;
1983         int result;
1984
1985         if (!sshdr_external)
1986                 sshdr = kzalloc(sizeof(*sshdr), GFP_KERNEL);
1987         else
1988                 sshdr = sshdr_external;
1989
1990         /* try to eat the UNIT_ATTENTION if there are enough retries */
1991         do {
1992                 result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
1993                                           timeout, retries, NULL);
1994                 if (sdev->removable && scsi_sense_valid(sshdr) &&
1995                     sshdr->sense_key == UNIT_ATTENTION)
1996                         sdev->changed = 1;
1997         } while (scsi_sense_valid(sshdr) &&
1998                  sshdr->sense_key == UNIT_ATTENTION && --retries);
1999
2000         if (!sshdr)
2001                 /* could not allocate sense buffer, so can't process it */
2002                 return result;
2003
2004         if (sdev->removable && scsi_sense_valid(sshdr) &&
2005             (sshdr->sense_key == UNIT_ATTENTION ||
2006              sshdr->sense_key == NOT_READY)) {
2007                 sdev->changed = 1;
2008                 result = 0;
2009         }
2010         if (!sshdr_external)
2011                 kfree(sshdr);
2012         return result;
2013 }
2014 EXPORT_SYMBOL(scsi_test_unit_ready);
2015
2016 /**
2017  *      scsi_device_set_state - Take the given device through the device state model.
2018  *      @sdev:  scsi device to change the state of.
2019  *      @state: state to change to.
2020  *
2021  *      Returns zero if unsuccessful or an error if the requested 
2022  *      transition is illegal.
2023  */
2024 int
2025 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2026 {
2027         enum scsi_device_state oldstate = sdev->sdev_state;
2028
2029         if (state == oldstate)
2030                 return 0;
2031
2032         switch (state) {
2033         case SDEV_CREATED:
2034                 switch (oldstate) {
2035                 case SDEV_CREATED_BLOCK:
2036                         break;
2037                 default:
2038                         goto illegal;
2039                 }
2040                 break;
2041                         
2042         case SDEV_RUNNING:
2043                 switch (oldstate) {
2044                 case SDEV_CREATED:
2045                 case SDEV_OFFLINE:
2046                 case SDEV_QUIESCE:
2047                 case SDEV_BLOCK:
2048                         break;
2049                 default:
2050                         goto illegal;
2051                 }
2052                 break;
2053
2054         case SDEV_QUIESCE:
2055                 switch (oldstate) {
2056                 case SDEV_RUNNING:
2057                 case SDEV_OFFLINE:
2058                         break;
2059                 default:
2060                         goto illegal;
2061                 }
2062                 break;
2063
2064         case SDEV_OFFLINE:
2065                 switch (oldstate) {
2066                 case SDEV_CREATED:
2067                 case SDEV_RUNNING:
2068                 case SDEV_QUIESCE:
2069                 case SDEV_BLOCK:
2070                         break;
2071                 default:
2072                         goto illegal;
2073                 }
2074                 break;
2075
2076         case SDEV_BLOCK:
2077                 switch (oldstate) {
2078                 case SDEV_RUNNING:
2079                 case SDEV_CREATED_BLOCK:
2080                         break;
2081                 default:
2082                         goto illegal;
2083                 }
2084                 break;
2085
2086         case SDEV_CREATED_BLOCK:
2087                 switch (oldstate) {
2088                 case SDEV_CREATED:
2089                         break;
2090                 default:
2091                         goto illegal;
2092                 }
2093                 break;
2094
2095         case SDEV_CANCEL:
2096                 switch (oldstate) {
2097                 case SDEV_CREATED:
2098                 case SDEV_RUNNING:
2099                 case SDEV_QUIESCE:
2100                 case SDEV_OFFLINE:
2101                 case SDEV_BLOCK:
2102                         break;
2103                 default:
2104                         goto illegal;
2105                 }
2106                 break;
2107
2108         case SDEV_DEL:
2109                 switch (oldstate) {
2110                 case SDEV_CREATED:
2111                 case SDEV_RUNNING:
2112                 case SDEV_OFFLINE:
2113                 case SDEV_CANCEL:
2114                         break;
2115                 default:
2116                         goto illegal;
2117                 }
2118                 break;
2119
2120         }
2121         sdev->sdev_state = state;
2122         return 0;
2123
2124  illegal:
2125         SCSI_LOG_ERROR_RECOVERY(1, 
2126                                 sdev_printk(KERN_ERR, sdev,
2127                                             "Illegal state transition %s->%s\n",
2128                                             scsi_device_state_name(oldstate),
2129                                             scsi_device_state_name(state))
2130                                 );
2131         return -EINVAL;
2132 }
2133 EXPORT_SYMBOL(scsi_device_set_state);
2134
2135 /**
2136  *      sdev_evt_emit - emit a single SCSI device uevent
2137  *      @sdev: associated SCSI device
2138  *      @evt: event to emit
2139  *
2140  *      Send a single uevent (scsi_event) to the associated scsi_device.
2141  */
2142 static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2143 {
2144         int idx = 0;
2145         char *envp[3];
2146
2147         switch (evt->evt_type) {
2148         case SDEV_EVT_MEDIA_CHANGE:
2149                 envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2150                 break;
2151
2152         default:
2153                 /* do nothing */
2154                 break;
2155         }
2156
2157         envp[idx++] = NULL;
2158
2159         kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2160 }
2161
2162 /**
2163  *      sdev_evt_thread - send a uevent for each scsi event
2164  *      @work: work struct for scsi_device
2165  *
2166  *      Dispatch queued events to their associated scsi_device kobjects
2167  *      as uevents.
2168  */
2169 void scsi_evt_thread(struct work_struct *work)
2170 {
2171         struct scsi_device *sdev;
2172         LIST_HEAD(event_list);
2173
2174         sdev = container_of(work, struct scsi_device, event_work);
2175
2176         while (1) {
2177                 struct scsi_event *evt;
2178                 struct list_head *this, *tmp;
2179                 unsigned long flags;
2180
2181                 spin_lock_irqsave(&sdev->list_lock, flags);
2182                 list_splice_init(&sdev->event_list, &event_list);
2183                 spin_unlock_irqrestore(&sdev->list_lock, flags);
2184
2185                 if (list_empty(&event_list))
2186                         break;
2187
2188                 list_for_each_safe(this, tmp, &event_list) {
2189                         evt = list_entry(this, struct scsi_event, node);
2190                         list_del(&evt->node);
2191                         scsi_evt_emit(sdev, evt);
2192                         kfree(evt);
2193                 }
2194         }
2195 }
2196
2197 /**
2198  *      sdev_evt_send - send asserted event to uevent thread
2199  *      @sdev: scsi_device event occurred on
2200  *      @evt: event to send
2201  *
2202  *      Assert scsi device event asynchronously.
2203  */
2204 void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2205 {
2206         unsigned long flags;
2207
2208 #if 0
2209         /* FIXME: currently this check eliminates all media change events
2210          * for polled devices.  Need to update to discriminate between AN
2211          * and polled events */
2212         if (!test_bit(evt->evt_type, sdev->supported_events)) {
2213                 kfree(evt);
2214                 return;
2215         }
2216 #endif
2217
2218         spin_lock_irqsave(&sdev->list_lock, flags);
2219         list_add_tail(&evt->node, &sdev->event_list);
2220         schedule_work(&sdev->event_work);
2221         spin_unlock_irqrestore(&sdev->list_lock, flags);
2222 }
2223 EXPORT_SYMBOL_GPL(sdev_evt_send);
2224
2225 /**
2226  *      sdev_evt_alloc - allocate a new scsi event
2227  *      @evt_type: type of event to allocate
2228  *      @gfpflags: GFP flags for allocation
2229  *
2230  *      Allocates and returns a new scsi_event.
2231  */
2232 struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2233                                   gfp_t gfpflags)
2234 {
2235         struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2236         if (!evt)
2237                 return NULL;
2238
2239         evt->evt_type = evt_type;
2240         INIT_LIST_HEAD(&evt->node);
2241
2242         /* evt_type-specific initialization, if any */
2243         switch (evt_type) {
2244         case SDEV_EVT_MEDIA_CHANGE:
2245         default:
2246                 /* do nothing */
2247                 break;
2248         }
2249
2250         return evt;
2251 }
2252 EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2253
2254 /**
2255  *      sdev_evt_send_simple - send asserted event to uevent thread
2256  *      @sdev: scsi_device event occurred on
2257  *      @evt_type: type of event to send
2258  *      @gfpflags: GFP flags for allocation
2259  *
2260  *      Assert scsi device event asynchronously, given an event type.
2261  */
2262 void sdev_evt_send_simple(struct scsi_device *sdev,
2263                           enum scsi_device_event evt_type, gfp_t gfpflags)
2264 {
2265         struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2266         if (!evt) {
2267                 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2268                             evt_type);
2269                 return;
2270         }
2271
2272         sdev_evt_send(sdev, evt);
2273 }
2274 EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2275
2276 /**
2277  *      scsi_device_quiesce - Block user issued commands.
2278  *      @sdev:  scsi device to quiesce.
2279  *
2280  *      This works by trying to transition to the SDEV_QUIESCE state
2281  *      (which must be a legal transition).  When the device is in this
2282  *      state, only special requests will be accepted, all others will
2283  *      be deferred.  Since special requests may also be requeued requests,
2284  *      a successful return doesn't guarantee the device will be 
2285  *      totally quiescent.
2286  *
2287  *      Must be called with user context, may sleep.
2288  *
2289  *      Returns zero if unsuccessful or an error if not.
2290  */
2291 int
2292 scsi_device_quiesce(struct scsi_device *sdev)
2293 {
2294         int err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2295         if (err)
2296                 return err;
2297
2298         scsi_run_queue(sdev->request_queue);
2299         while (sdev->device_busy) {
2300                 msleep_interruptible(200);
2301                 scsi_run_queue(sdev->request_queue);
2302         }
2303         return 0;
2304 }
2305 EXPORT_SYMBOL(scsi_device_quiesce);
2306
2307 /**
2308  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2309  *      @sdev:  scsi device to resume.
2310  *
2311  *      Moves the device from quiesced back to running and restarts the
2312  *      queues.
2313  *
2314  *      Must be called with user context, may sleep.
2315  */
2316 void
2317 scsi_device_resume(struct scsi_device *sdev)
2318 {
2319         if(scsi_device_set_state(sdev, SDEV_RUNNING))
2320                 return;
2321         scsi_run_queue(sdev->request_queue);
2322 }
2323 EXPORT_SYMBOL(scsi_device_resume);
2324
2325 static void
2326 device_quiesce_fn(struct scsi_device *sdev, void *data)
2327 {
2328         scsi_device_quiesce(sdev);
2329 }
2330
2331 void
2332 scsi_target_quiesce(struct scsi_target *starget)
2333 {
2334         starget_for_each_device(starget, NULL, device_quiesce_fn);
2335 }
2336 EXPORT_SYMBOL(scsi_target_quiesce);
2337
2338 static void
2339 device_resume_fn(struct scsi_device *sdev, void *data)
2340 {
2341         scsi_device_resume(sdev);
2342 }
2343
2344 void
2345 scsi_target_resume(struct scsi_target *starget)
2346 {
2347         starget_for_each_device(starget, NULL, device_resume_fn);
2348 }
2349 EXPORT_SYMBOL(scsi_target_resume);
2350
2351 /**
2352  * scsi_internal_device_block - internal function to put a device temporarily into the SDEV_BLOCK state
2353  * @sdev:       device to block
2354  *
2355  * Block request made by scsi lld's to temporarily stop all
2356  * scsi commands on the specified device.  Called from interrupt
2357  * or normal process context.
2358  *
2359  * Returns zero if successful or error if not
2360  *
2361  * Notes:       
2362  *      This routine transitions the device to the SDEV_BLOCK state
2363  *      (which must be a legal transition).  When the device is in this
2364  *      state, all commands are deferred until the scsi lld reenables
2365  *      the device with scsi_device_unblock or device_block_tmo fires.
2366  *      This routine assumes the host_lock is held on entry.
2367  */
2368 int
2369 scsi_internal_device_block(struct scsi_device *sdev)
2370 {
2371         struct request_queue *q = sdev->request_queue;
2372         unsigned long flags;
2373         int err = 0;
2374
2375         err = scsi_device_set_state(sdev, SDEV_BLOCK);
2376         if (err) {
2377                 err = scsi_device_set_state(sdev, SDEV_CREATED_BLOCK);
2378
2379                 if (err)
2380                         return err;
2381         }
2382
2383         /* 
2384          * The device has transitioned to SDEV_BLOCK.  Stop the
2385          * block layer from calling the midlayer with this device's
2386          * request queue. 
2387          */
2388         spin_lock_irqsave(q->queue_lock, flags);
2389         blk_stop_queue(q);
2390         spin_unlock_irqrestore(q->queue_lock, flags);
2391
2392         return 0;
2393 }
2394 EXPORT_SYMBOL_GPL(scsi_internal_device_block);
2395  
2396 /**
2397  * scsi_internal_device_unblock - resume a device after a block request
2398  * @sdev:       device to resume
2399  *
2400  * Called by scsi lld's or the midlayer to restart the device queue
2401  * for the previously suspended scsi device.  Called from interrupt or
2402  * normal process context.
2403  *
2404  * Returns zero if successful or error if not.
2405  *
2406  * Notes:       
2407  *      This routine transitions the device to the SDEV_RUNNING state
2408  *      (which must be a legal transition) allowing the midlayer to
2409  *      goose the queue for this device.  This routine assumes the 
2410  *      host_lock is held upon entry.
2411  */
2412 int
2413 scsi_internal_device_unblock(struct scsi_device *sdev)
2414 {
2415         struct request_queue *q = sdev->request_queue; 
2416         unsigned long flags;
2417         
2418         /* 
2419          * Try to transition the scsi device to SDEV_RUNNING
2420          * and goose the device queue if successful.  
2421          */
2422         if (sdev->sdev_state == SDEV_BLOCK)
2423                 sdev->sdev_state = SDEV_RUNNING;
2424         else if (sdev->sdev_state == SDEV_CREATED_BLOCK)
2425                 sdev->sdev_state = SDEV_CREATED;
2426         else
2427                 return -EINVAL;
2428
2429         spin_lock_irqsave(q->queue_lock, flags);
2430         blk_start_queue(q);
2431         spin_unlock_irqrestore(q->queue_lock, flags);
2432
2433         return 0;
2434 }
2435 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock);
2436
2437 static void
2438 device_block(struct scsi_device *sdev, void *data)
2439 {
2440         scsi_internal_device_block(sdev);
2441 }
2442
2443 static int
2444 target_block(struct device *dev, void *data)
2445 {
2446         if (scsi_is_target_device(dev))
2447                 starget_for_each_device(to_scsi_target(dev), NULL,
2448                                         device_block);
2449         return 0;
2450 }
2451
2452 void
2453 scsi_target_block(struct device *dev)
2454 {
2455         if (scsi_is_target_device(dev))
2456                 starget_for_each_device(to_scsi_target(dev), NULL,
2457                                         device_block);
2458         else
2459                 device_for_each_child(dev, NULL, target_block);
2460 }
2461 EXPORT_SYMBOL_GPL(scsi_target_block);
2462
2463 static void
2464 device_unblock(struct scsi_device *sdev, void *data)
2465 {
2466         scsi_internal_device_unblock(sdev);
2467 }
2468
2469 static int
2470 target_unblock(struct device *dev, void *data)
2471 {
2472         if (scsi_is_target_device(dev))
2473                 starget_for_each_device(to_scsi_target(dev), NULL,
2474                                         device_unblock);
2475         return 0;
2476 }
2477
2478 void
2479 scsi_target_unblock(struct device *dev)
2480 {
2481         if (scsi_is_target_device(dev))
2482                 starget_for_each_device(to_scsi_target(dev), NULL,
2483                                         device_unblock);
2484         else
2485                 device_for_each_child(dev, NULL, target_unblock);
2486 }
2487 EXPORT_SYMBOL_GPL(scsi_target_unblock);
2488
2489 /**
2490  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2491  * @sgl:        scatter-gather list
2492  * @sg_count:   number of segments in sg
2493  * @offset:     offset in bytes into sg, on return offset into the mapped area
2494  * @len:        bytes to map, on return number of bytes mapped
2495  *
2496  * Returns virtual address of the start of the mapped page
2497  */
2498 void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
2499                           size_t *offset, size_t *len)
2500 {
2501         int i;
2502         size_t sg_len = 0, len_complete = 0;
2503         struct scatterlist *sg;
2504         struct page *page;
2505
2506         WARN_ON(!irqs_disabled());
2507
2508         for_each_sg(sgl, sg, sg_count, i) {
2509                 len_complete = sg_len; /* Complete sg-entries */
2510                 sg_len += sg->length;
2511                 if (sg_len > *offset)
2512                         break;
2513         }
2514
2515         if (unlikely(i == sg_count)) {
2516                 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2517                         "elements %d\n",
2518                        __func__, sg_len, *offset, sg_count);
2519                 WARN_ON(1);
2520                 return NULL;
2521         }
2522
2523         /* Offset starting from the beginning of first page in this sg-entry */
2524         *offset = *offset - len_complete + sg->offset;
2525
2526         /* Assumption: contiguous pages can be accessed as "page + i" */
2527         page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
2528         *offset &= ~PAGE_MASK;
2529
2530         /* Bytes in this sg-entry from *offset to the end of the page */
2531         sg_len = PAGE_SIZE - *offset;
2532         if (*len > sg_len)
2533                 *len = sg_len;
2534
2535         return kmap_atomic(page, KM_BIO_SRC_IRQ);
2536 }
2537 EXPORT_SYMBOL(scsi_kmap_atomic_sg);
2538
2539 /**
2540  * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
2541  * @virt:       virtual address to be unmapped
2542  */
2543 void scsi_kunmap_atomic_sg(void *virt)
2544 {
2545         kunmap_atomic(virt, KM_BIO_SRC_IRQ);
2546 }
2547 EXPORT_SYMBOL(scsi_kunmap_atomic_sg);