108655230b59d3d55eb4593d16c9d7f22a566980
[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                         scsi_print_command(cmd);
900                 }
901                 if (blk_end_request_err(req, error))
902                         scsi_requeue_command(q, cmd);
903                 else
904                         scsi_next_command(cmd);
905                 break;
906         case ACTION_REPREP:
907                 /* Unprep the request and put it back at the head of the queue.
908                  * A new command will be prepared and issued.
909                  */
910                 scsi_release_buffers(cmd);
911                 scsi_requeue_command(q, cmd);
912                 break;
913         case ACTION_RETRY:
914                 /* Retry the same command immediately */
915                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY, 0);
916                 break;
917         case ACTION_DELAYED_RETRY:
918                 /* Retry the same command after a delay */
919                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY, 0);
920                 break;
921         }
922 }
923
924 static int scsi_init_sgtable(struct request *req, struct scsi_data_buffer *sdb,
925                              gfp_t gfp_mask)
926 {
927         int count;
928
929         /*
930          * If sg table allocation fails, requeue request later.
931          */
932         if (unlikely(scsi_alloc_sgtable(sdb, req->nr_phys_segments,
933                                         gfp_mask))) {
934                 return BLKPREP_DEFER;
935         }
936
937         req->buffer = NULL;
938
939         /* 
940          * Next, walk the list, and fill in the addresses and sizes of
941          * each segment.
942          */
943         count = blk_rq_map_sg(req->q, req, sdb->table.sgl);
944         BUG_ON(count > sdb->table.nents);
945         sdb->table.nents = count;
946         sdb->length = blk_rq_bytes(req);
947         return BLKPREP_OK;
948 }
949
950 /*
951  * Function:    scsi_init_io()
952  *
953  * Purpose:     SCSI I/O initialize function.
954  *
955  * Arguments:   cmd   - Command descriptor we wish to initialize
956  *
957  * Returns:     0 on success
958  *              BLKPREP_DEFER if the failure is retryable
959  *              BLKPREP_KILL if the failure is fatal
960  */
961 int scsi_init_io(struct scsi_cmnd *cmd, gfp_t gfp_mask)
962 {
963         int error = scsi_init_sgtable(cmd->request, &cmd->sdb, gfp_mask);
964         if (error)
965                 goto err_exit;
966
967         if (blk_bidi_rq(cmd->request)) {
968                 struct scsi_data_buffer *bidi_sdb = kmem_cache_zalloc(
969                         scsi_sdb_cache, GFP_ATOMIC);
970                 if (!bidi_sdb) {
971                         error = BLKPREP_DEFER;
972                         goto err_exit;
973                 }
974
975                 cmd->request->next_rq->special = bidi_sdb;
976                 error = scsi_init_sgtable(cmd->request->next_rq, bidi_sdb,
977                                                                     GFP_ATOMIC);
978                 if (error)
979                         goto err_exit;
980         }
981
982         if (blk_integrity_rq(cmd->request)) {
983                 struct scsi_data_buffer *prot_sdb = cmd->prot_sdb;
984                 int ivecs, count;
985
986                 BUG_ON(prot_sdb == NULL);
987                 ivecs = blk_rq_count_integrity_sg(cmd->request);
988
989                 if (scsi_alloc_sgtable(prot_sdb, ivecs, gfp_mask)) {
990                         error = BLKPREP_DEFER;
991                         goto err_exit;
992                 }
993
994                 count = blk_rq_map_integrity_sg(cmd->request,
995                                                 prot_sdb->table.sgl);
996                 BUG_ON(unlikely(count > ivecs));
997
998                 cmd->prot_sdb = prot_sdb;
999                 cmd->prot_sdb->table.nents = count;
1000         }
1001
1002         return BLKPREP_OK ;
1003
1004 err_exit:
1005         scsi_release_buffers(cmd);
1006         if (error == BLKPREP_KILL)
1007                 scsi_put_command(cmd);
1008         else /* BLKPREP_DEFER */
1009                 scsi_unprep_request(cmd->request);
1010
1011         return error;
1012 }
1013 EXPORT_SYMBOL(scsi_init_io);
1014
1015 static struct scsi_cmnd *scsi_get_cmd_from_req(struct scsi_device *sdev,
1016                 struct request *req)
1017 {
1018         struct scsi_cmnd *cmd;
1019
1020         if (!req->special) {
1021                 cmd = scsi_get_command(sdev, GFP_ATOMIC);
1022                 if (unlikely(!cmd))
1023                         return NULL;
1024                 req->special = cmd;
1025         } else {
1026                 cmd = req->special;
1027         }
1028
1029         /* pull a tag out of the request if we have one */
1030         cmd->tag = req->tag;
1031         cmd->request = req;
1032
1033         cmd->cmnd = req->cmd;
1034
1035         return cmd;
1036 }
1037
1038 int scsi_setup_blk_pc_cmnd(struct scsi_device *sdev, struct request *req)
1039 {
1040         struct scsi_cmnd *cmd;
1041         int ret = scsi_prep_state_check(sdev, req);
1042
1043         if (ret != BLKPREP_OK)
1044                 return ret;
1045
1046         cmd = scsi_get_cmd_from_req(sdev, req);
1047         if (unlikely(!cmd))
1048                 return BLKPREP_DEFER;
1049
1050         /*
1051          * BLOCK_PC requests may transfer data, in which case they must
1052          * a bio attached to them.  Or they might contain a SCSI command
1053          * that does not transfer data, in which case they may optionally
1054          * submit a request without an attached bio.
1055          */
1056         if (req->bio) {
1057                 int ret;
1058
1059                 BUG_ON(!req->nr_phys_segments);
1060
1061                 ret = scsi_init_io(cmd, GFP_ATOMIC);
1062                 if (unlikely(ret))
1063                         return ret;
1064         } else {
1065                 BUG_ON(blk_rq_bytes(req));
1066
1067                 memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1068                 req->buffer = NULL;
1069         }
1070
1071         cmd->cmd_len = req->cmd_len;
1072         if (!blk_rq_bytes(req))
1073                 cmd->sc_data_direction = DMA_NONE;
1074         else if (rq_data_dir(req) == WRITE)
1075                 cmd->sc_data_direction = DMA_TO_DEVICE;
1076         else
1077                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1078         
1079         cmd->transfersize = blk_rq_bytes(req);
1080         cmd->allowed = req->retries;
1081         return BLKPREP_OK;
1082 }
1083 EXPORT_SYMBOL(scsi_setup_blk_pc_cmnd);
1084
1085 /*
1086  * Setup a REQ_TYPE_FS command.  These are simple read/write request
1087  * from filesystems that still need to be translated to SCSI CDBs from
1088  * the ULD.
1089  */
1090 int scsi_setup_fs_cmnd(struct scsi_device *sdev, struct request *req)
1091 {
1092         struct scsi_cmnd *cmd;
1093         int ret = scsi_prep_state_check(sdev, req);
1094
1095         if (ret != BLKPREP_OK)
1096                 return ret;
1097
1098         if (unlikely(sdev->scsi_dh_data && sdev->scsi_dh_data->scsi_dh
1099                          && sdev->scsi_dh_data->scsi_dh->prep_fn)) {
1100                 ret = sdev->scsi_dh_data->scsi_dh->prep_fn(sdev, req);
1101                 if (ret != BLKPREP_OK)
1102                         return ret;
1103         }
1104
1105         /*
1106          * Filesystem requests must transfer data.
1107          */
1108         BUG_ON(!req->nr_phys_segments);
1109
1110         cmd = scsi_get_cmd_from_req(sdev, req);
1111         if (unlikely(!cmd))
1112                 return BLKPREP_DEFER;
1113
1114         memset(cmd->cmnd, 0, BLK_MAX_CDB);
1115         return scsi_init_io(cmd, GFP_ATOMIC);
1116 }
1117 EXPORT_SYMBOL(scsi_setup_fs_cmnd);
1118
1119 int scsi_prep_state_check(struct scsi_device *sdev, struct request *req)
1120 {
1121         int ret = BLKPREP_OK;
1122
1123         /*
1124          * If the device is not in running state we will reject some
1125          * or all commands.
1126          */
1127         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1128                 switch (sdev->sdev_state) {
1129                 case SDEV_OFFLINE:
1130                         /*
1131                          * If the device is offline we refuse to process any
1132                          * commands.  The device must be brought online
1133                          * before trying any recovery commands.
1134                          */
1135                         sdev_printk(KERN_ERR, sdev,
1136                                     "rejecting I/O to offline device\n");
1137                         ret = BLKPREP_KILL;
1138                         break;
1139                 case SDEV_DEL:
1140                         /*
1141                          * If the device is fully deleted, we refuse to
1142                          * process any commands as well.
1143                          */
1144                         sdev_printk(KERN_ERR, sdev,
1145                                     "rejecting I/O to dead device\n");
1146                         ret = BLKPREP_KILL;
1147                         break;
1148                 case SDEV_QUIESCE:
1149                 case SDEV_BLOCK:
1150                 case SDEV_CREATED_BLOCK:
1151                         /*
1152                          * If the devices is blocked we defer normal commands.
1153                          */
1154                         if (!(req->cmd_flags & REQ_PREEMPT))
1155                                 ret = BLKPREP_DEFER;
1156                         break;
1157                 default:
1158                         /*
1159                          * For any other not fully online state we only allow
1160                          * special commands.  In particular any user initiated
1161                          * command is not allowed.
1162                          */
1163                         if (!(req->cmd_flags & REQ_PREEMPT))
1164                                 ret = BLKPREP_KILL;
1165                         break;
1166                 }
1167         }
1168         return ret;
1169 }
1170 EXPORT_SYMBOL(scsi_prep_state_check);
1171
1172 int scsi_prep_return(struct request_queue *q, struct request *req, int ret)
1173 {
1174         struct scsi_device *sdev = q->queuedata;
1175
1176         switch (ret) {
1177         case BLKPREP_KILL:
1178                 req->errors = DID_NO_CONNECT << 16;
1179                 /* release the command and kill it */
1180                 if (req->special) {
1181                         struct scsi_cmnd *cmd = req->special;
1182                         scsi_release_buffers(cmd);
1183                         scsi_put_command(cmd);
1184                         req->special = NULL;
1185                 }
1186                 break;
1187         case BLKPREP_DEFER:
1188                 /*
1189                  * If we defer, the blk_peek_request() returns NULL, but the
1190                  * queue must be restarted, so we plug here if no returning
1191                  * command will automatically do that.
1192                  */
1193                 if (sdev->device_busy == 0)
1194                         blk_plug_device(q);
1195                 break;
1196         default:
1197                 req->cmd_flags |= REQ_DONTPREP;
1198         }
1199
1200         return ret;
1201 }
1202 EXPORT_SYMBOL(scsi_prep_return);
1203
1204 int scsi_prep_fn(struct request_queue *q, struct request *req)
1205 {
1206         struct scsi_device *sdev = q->queuedata;
1207         int ret = BLKPREP_KILL;
1208
1209         if (req->cmd_type == REQ_TYPE_BLOCK_PC)
1210                 ret = scsi_setup_blk_pc_cmnd(sdev, req);
1211         return scsi_prep_return(q, req, ret);
1212 }
1213 EXPORT_SYMBOL(scsi_prep_fn);
1214
1215 /*
1216  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1217  * return 0.
1218  *
1219  * Called with the queue_lock held.
1220  */
1221 static inline int scsi_dev_queue_ready(struct request_queue *q,
1222                                   struct scsi_device *sdev)
1223 {
1224         if (sdev->device_busy == 0 && sdev->device_blocked) {
1225                 /*
1226                  * unblock after device_blocked iterates to zero
1227                  */
1228                 if (--sdev->device_blocked == 0) {
1229                         SCSI_LOG_MLQUEUE(3,
1230                                    sdev_printk(KERN_INFO, sdev,
1231                                    "unblocking device at zero depth\n"));
1232                 } else {
1233                         blk_plug_device(q);
1234                         return 0;
1235                 }
1236         }
1237         if (scsi_device_is_busy(sdev))
1238                 return 0;
1239
1240         return 1;
1241 }
1242
1243
1244 /*
1245  * scsi_target_queue_ready: checks if there we can send commands to target
1246  * @sdev: scsi device on starget to check.
1247  *
1248  * Called with the host lock held.
1249  */
1250 static inline int scsi_target_queue_ready(struct Scsi_Host *shost,
1251                                            struct scsi_device *sdev)
1252 {
1253         struct scsi_target *starget = scsi_target(sdev);
1254
1255         if (starget->single_lun) {
1256                 if (starget->starget_sdev_user &&
1257                     starget->starget_sdev_user != sdev)
1258                         return 0;
1259                 starget->starget_sdev_user = sdev;
1260         }
1261
1262         if (starget->target_busy == 0 && starget->target_blocked) {
1263                 /*
1264                  * unblock after target_blocked iterates to zero
1265                  */
1266                 if (--starget->target_blocked == 0) {
1267                         SCSI_LOG_MLQUEUE(3, starget_printk(KERN_INFO, starget,
1268                                          "unblocking target at zero depth\n"));
1269                 } else
1270                         return 0;
1271         }
1272
1273         if (scsi_target_is_busy(starget)) {
1274                 if (list_empty(&sdev->starved_entry)) {
1275                         list_add_tail(&sdev->starved_entry,
1276                                       &shost->starved_list);
1277                         return 0;
1278                 }
1279         }
1280
1281         /* We're OK to process the command, so we can't be starved */
1282         if (!list_empty(&sdev->starved_entry))
1283                 list_del_init(&sdev->starved_entry);
1284         return 1;
1285 }
1286
1287 /*
1288  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1289  * return 0. We must end up running the queue again whenever 0 is
1290  * returned, else IO can hang.
1291  *
1292  * Called with host_lock held.
1293  */
1294 static inline int scsi_host_queue_ready(struct request_queue *q,
1295                                    struct Scsi_Host *shost,
1296                                    struct scsi_device *sdev)
1297 {
1298         if (scsi_host_in_recovery(shost))
1299                 return 0;
1300         if (shost->host_busy == 0 && shost->host_blocked) {
1301                 /*
1302                  * unblock after host_blocked iterates to zero
1303                  */
1304                 if (--shost->host_blocked == 0) {
1305                         SCSI_LOG_MLQUEUE(3,
1306                                 printk("scsi%d unblocking host at zero depth\n",
1307                                         shost->host_no));
1308                 } else {
1309                         return 0;
1310                 }
1311         }
1312         if (scsi_host_is_busy(shost)) {
1313                 if (list_empty(&sdev->starved_entry))
1314                         list_add_tail(&sdev->starved_entry, &shost->starved_list);
1315                 return 0;
1316         }
1317
1318         /* We're OK to process the command, so we can't be starved */
1319         if (!list_empty(&sdev->starved_entry))
1320                 list_del_init(&sdev->starved_entry);
1321
1322         return 1;
1323 }
1324
1325 /*
1326  * Busy state exporting function for request stacking drivers.
1327  *
1328  * For efficiency, no lock is taken to check the busy state of
1329  * shost/starget/sdev, since the returned value is not guaranteed and
1330  * may be changed after request stacking drivers call the function,
1331  * regardless of taking lock or not.
1332  *
1333  * When scsi can't dispatch I/Os anymore and needs to kill I/Os
1334  * (e.g. !sdev), scsi needs to return 'not busy'.
1335  * Otherwise, request stacking drivers may hold requests forever.
1336  */
1337 static int scsi_lld_busy(struct request_queue *q)
1338 {
1339         struct scsi_device *sdev = q->queuedata;
1340         struct Scsi_Host *shost;
1341         struct scsi_target *starget;
1342
1343         if (!sdev)
1344                 return 0;
1345
1346         shost = sdev->host;
1347         starget = scsi_target(sdev);
1348
1349         if (scsi_host_in_recovery(shost) || scsi_host_is_busy(shost) ||
1350             scsi_target_is_busy(starget) || scsi_device_is_busy(sdev))
1351                 return 1;
1352
1353         return 0;
1354 }
1355
1356 /*
1357  * Kill a request for a dead device
1358  */
1359 static void scsi_kill_request(struct request *req, struct request_queue *q)
1360 {
1361         struct scsi_cmnd *cmd = req->special;
1362         struct scsi_device *sdev = cmd->device;
1363         struct scsi_target *starget = scsi_target(sdev);
1364         struct Scsi_Host *shost = sdev->host;
1365
1366         blk_start_request(req);
1367
1368         if (unlikely(cmd == NULL)) {
1369                 printk(KERN_CRIT "impossible request in %s.\n",
1370                                  __func__);
1371                 BUG();
1372         }
1373
1374         scsi_init_cmd_errh(cmd);
1375         cmd->result = DID_NO_CONNECT << 16;
1376         atomic_inc(&cmd->device->iorequest_cnt);
1377
1378         /*
1379          * SCSI request completion path will do scsi_device_unbusy(),
1380          * bump busy counts.  To bump the counters, we need to dance
1381          * with the locks as normal issue path does.
1382          */
1383         sdev->device_busy++;
1384         spin_unlock(sdev->request_queue->queue_lock);
1385         spin_lock(shost->host_lock);
1386         shost->host_busy++;
1387         starget->target_busy++;
1388         spin_unlock(shost->host_lock);
1389         spin_lock(sdev->request_queue->queue_lock);
1390
1391         blk_complete_request(req);
1392 }
1393
1394 static void scsi_softirq_done(struct request *rq)
1395 {
1396         struct scsi_cmnd *cmd = rq->special;
1397         unsigned long wait_for = (cmd->allowed + 1) * rq->timeout;
1398         int disposition;
1399
1400         INIT_LIST_HEAD(&cmd->eh_entry);
1401
1402         /*
1403          * Set the serial numbers back to zero
1404          */
1405         cmd->serial_number = 0;
1406
1407         atomic_inc(&cmd->device->iodone_cnt);
1408         if (cmd->result)
1409                 atomic_inc(&cmd->device->ioerr_cnt);
1410
1411         disposition = scsi_decide_disposition(cmd);
1412         if (disposition != SUCCESS &&
1413             time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1414                 sdev_printk(KERN_ERR, cmd->device,
1415                             "timing out command, waited %lus\n",
1416                             wait_for/HZ);
1417                 disposition = SUCCESS;
1418         }
1419                         
1420         scsi_log_completion(cmd, disposition);
1421
1422         switch (disposition) {
1423                 case SUCCESS:
1424                         scsi_finish_command(cmd);
1425                         break;
1426                 case NEEDS_RETRY:
1427                         scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1428                         break;
1429                 case ADD_TO_MLQUEUE:
1430                         scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1431                         break;
1432                 default:
1433                         if (!scsi_eh_scmd_add(cmd, 0))
1434                                 scsi_finish_command(cmd);
1435         }
1436 }
1437
1438 /*
1439  * Function:    scsi_request_fn()
1440  *
1441  * Purpose:     Main strategy routine for SCSI.
1442  *
1443  * Arguments:   q       - Pointer to actual queue.
1444  *
1445  * Returns:     Nothing
1446  *
1447  * Lock status: IO request lock assumed to be held when called.
1448  */
1449 static void scsi_request_fn(struct request_queue *q)
1450 {
1451         struct scsi_device *sdev = q->queuedata;
1452         struct Scsi_Host *shost;
1453         struct scsi_cmnd *cmd;
1454         struct request *req;
1455
1456         if (!sdev) {
1457                 printk("scsi: killing requests for dead queue\n");
1458                 while ((req = blk_peek_request(q)) != NULL)
1459                         scsi_kill_request(req, q);
1460                 return;
1461         }
1462
1463         if(!get_device(&sdev->sdev_gendev))
1464                 /* We must be tearing the block queue down already */
1465                 return;
1466
1467         /*
1468          * To start with, we keep looping until the queue is empty, or until
1469          * the host is no longer able to accept any more requests.
1470          */
1471         shost = sdev->host;
1472         while (!blk_queue_plugged(q)) {
1473                 int rtn;
1474                 /*
1475                  * get next queueable request.  We do this early to make sure
1476                  * that the request is fully prepared even if we cannot 
1477                  * accept it.
1478                  */
1479                 req = blk_peek_request(q);
1480                 if (!req || !scsi_dev_queue_ready(q, sdev))
1481                         break;
1482
1483                 if (unlikely(!scsi_device_online(sdev))) {
1484                         sdev_printk(KERN_ERR, sdev,
1485                                     "rejecting I/O to offline device\n");
1486                         scsi_kill_request(req, q);
1487                         continue;
1488                 }
1489
1490
1491                 /*
1492                  * Remove the request from the request list.
1493                  */
1494                 if (!(blk_queue_tagged(q) && !blk_queue_start_tag(q, req)))
1495                         blk_start_request(req);
1496                 sdev->device_busy++;
1497
1498                 spin_unlock(q->queue_lock);
1499                 cmd = req->special;
1500                 if (unlikely(cmd == NULL)) {
1501                         printk(KERN_CRIT "impossible request in %s.\n"
1502                                          "please mail a stack trace to "
1503                                          "linux-scsi@vger.kernel.org\n",
1504                                          __func__);
1505                         blk_dump_rq_flags(req, "foo");
1506                         BUG();
1507                 }
1508                 spin_lock(shost->host_lock);
1509
1510                 /*
1511                  * We hit this when the driver is using a host wide
1512                  * tag map. For device level tag maps the queue_depth check
1513                  * in the device ready fn would prevent us from trying
1514                  * to allocate a tag. Since the map is a shared host resource
1515                  * we add the dev to the starved list so it eventually gets
1516                  * a run when a tag is freed.
1517                  */
1518                 if (blk_queue_tagged(q) && !blk_rq_tagged(req)) {
1519                         if (list_empty(&sdev->starved_entry))
1520                                 list_add_tail(&sdev->starved_entry,
1521                                               &shost->starved_list);
1522                         goto not_ready;
1523                 }
1524
1525                 if (!scsi_target_queue_ready(shost, sdev))
1526                         goto not_ready;
1527
1528                 if (!scsi_host_queue_ready(q, shost, sdev))
1529                         goto not_ready;
1530
1531                 scsi_target(sdev)->target_busy++;
1532                 shost->host_busy++;
1533
1534                 /*
1535                  * XXX(hch): This is rather suboptimal, scsi_dispatch_cmd will
1536                  *              take the lock again.
1537                  */
1538                 spin_unlock_irq(shost->host_lock);
1539
1540                 /*
1541                  * Finally, initialize any error handling parameters, and set up
1542                  * the timers for timeouts.
1543                  */
1544                 scsi_init_cmd_errh(cmd);
1545
1546                 /*
1547                  * Dispatch the command to the low-level driver.
1548                  */
1549                 rtn = scsi_dispatch_cmd(cmd);
1550                 spin_lock_irq(q->queue_lock);
1551                 if(rtn) {
1552                         /* we're refusing the command; because of
1553                          * the way locks get dropped, we need to 
1554                          * check here if plugging is required */
1555                         if(sdev->device_busy == 0)
1556                                 blk_plug_device(q);
1557
1558                         break;
1559                 }
1560         }
1561
1562         goto out;
1563
1564  not_ready:
1565         spin_unlock_irq(shost->host_lock);
1566
1567         /*
1568          * lock q, handle tag, requeue req, and decrement device_busy. We
1569          * must return with queue_lock held.
1570          *
1571          * Decrementing device_busy without checking it is OK, as all such
1572          * cases (host limits or settings) should run the queue at some
1573          * later time.
1574          */
1575         spin_lock_irq(q->queue_lock);
1576         blk_requeue_request(q, req);
1577         sdev->device_busy--;
1578         if(sdev->device_busy == 0)
1579                 blk_plug_device(q);
1580  out:
1581         /* must be careful here...if we trigger the ->remove() function
1582          * we cannot be holding the q lock */
1583         spin_unlock_irq(q->queue_lock);
1584         put_device(&sdev->sdev_gendev);
1585         spin_lock_irq(q->queue_lock);
1586 }
1587
1588 u64 scsi_calculate_bounce_limit(struct Scsi_Host *shost)
1589 {
1590         struct device *host_dev;
1591         u64 bounce_limit = 0xffffffff;
1592
1593         if (shost->unchecked_isa_dma)
1594                 return BLK_BOUNCE_ISA;
1595         /*
1596          * Platforms with virtual-DMA translation
1597          * hardware have no practical limit.
1598          */
1599         if (!PCI_DMA_BUS_IS_PHYS)
1600                 return BLK_BOUNCE_ANY;
1601
1602         host_dev = scsi_get_device(shost);
1603         if (host_dev && host_dev->dma_mask)
1604                 bounce_limit = *host_dev->dma_mask;
1605
1606         return bounce_limit;
1607 }
1608 EXPORT_SYMBOL(scsi_calculate_bounce_limit);
1609
1610 struct request_queue *__scsi_alloc_queue(struct Scsi_Host *shost,
1611                                          request_fn_proc *request_fn)
1612 {
1613         struct request_queue *q;
1614         struct device *dev = shost->shost_gendev.parent;
1615
1616         q = blk_init_queue(request_fn, NULL);
1617         if (!q)
1618                 return NULL;
1619
1620         /*
1621          * this limit is imposed by hardware restrictions
1622          */
1623         blk_queue_max_hw_segments(q, shost->sg_tablesize);
1624         blk_queue_max_phys_segments(q, SCSI_MAX_SG_CHAIN_SEGMENTS);
1625
1626         blk_queue_max_sectors(q, shost->max_sectors);
1627         blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
1628         blk_queue_segment_boundary(q, shost->dma_boundary);
1629         dma_set_seg_boundary(dev, shost->dma_boundary);
1630
1631         blk_queue_max_segment_size(q, dma_get_max_seg_size(dev));
1632
1633         /* New queue, no concurrency on queue_flags */
1634         if (!shost->use_clustering)
1635                 queue_flag_clear_unlocked(QUEUE_FLAG_CLUSTER, q);
1636
1637         /*
1638          * set a reasonable default alignment on word boundaries: the
1639          * host and device may alter it using
1640          * blk_queue_update_dma_alignment() later.
1641          */
1642         blk_queue_dma_alignment(q, 0x03);
1643
1644         return q;
1645 }
1646 EXPORT_SYMBOL(__scsi_alloc_queue);
1647
1648 struct request_queue *scsi_alloc_queue(struct scsi_device *sdev)
1649 {
1650         struct request_queue *q;
1651
1652         q = __scsi_alloc_queue(sdev->host, scsi_request_fn);
1653         if (!q)
1654                 return NULL;
1655
1656         blk_queue_prep_rq(q, scsi_prep_fn);
1657         blk_queue_softirq_done(q, scsi_softirq_done);
1658         blk_queue_rq_timed_out(q, scsi_times_out);
1659         blk_queue_lld_busy(q, scsi_lld_busy);
1660         return q;
1661 }
1662
1663 void scsi_free_queue(struct request_queue *q)
1664 {
1665         blk_cleanup_queue(q);
1666 }
1667
1668 /*
1669  * Function:    scsi_block_requests()
1670  *
1671  * Purpose:     Utility function used by low-level drivers to prevent further
1672  *              commands from being queued to the device.
1673  *
1674  * Arguments:   shost       - Host in question
1675  *
1676  * Returns:     Nothing
1677  *
1678  * Lock status: No locks are assumed held.
1679  *
1680  * Notes:       There is no timer nor any other means by which the requests
1681  *              get unblocked other than the low-level driver calling
1682  *              scsi_unblock_requests().
1683  */
1684 void scsi_block_requests(struct Scsi_Host *shost)
1685 {
1686         shost->host_self_blocked = 1;
1687 }
1688 EXPORT_SYMBOL(scsi_block_requests);
1689
1690 /*
1691  * Function:    scsi_unblock_requests()
1692  *
1693  * Purpose:     Utility function used by low-level drivers to allow further
1694  *              commands from being queued to the device.
1695  *
1696  * Arguments:   shost       - Host in question
1697  *
1698  * Returns:     Nothing
1699  *
1700  * Lock status: No locks are assumed held.
1701  *
1702  * Notes:       There is no timer nor any other means by which the requests
1703  *              get unblocked other than the low-level driver calling
1704  *              scsi_unblock_requests().
1705  *
1706  *              This is done as an API function so that changes to the
1707  *              internals of the scsi mid-layer won't require wholesale
1708  *              changes to drivers that use this feature.
1709  */
1710 void scsi_unblock_requests(struct Scsi_Host *shost)
1711 {
1712         shost->host_self_blocked = 0;
1713         scsi_run_host_queues(shost);
1714 }
1715 EXPORT_SYMBOL(scsi_unblock_requests);
1716
1717 int __init scsi_init_queue(void)
1718 {
1719         int i;
1720
1721         scsi_sdb_cache = kmem_cache_create("scsi_data_buffer",
1722                                            sizeof(struct scsi_data_buffer),
1723                                            0, 0, NULL);
1724         if (!scsi_sdb_cache) {
1725                 printk(KERN_ERR "SCSI: can't init scsi sdb cache\n");
1726                 return -ENOMEM;
1727         }
1728
1729         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1730                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1731                 int size = sgp->size * sizeof(struct scatterlist);
1732
1733                 sgp->slab = kmem_cache_create(sgp->name, size, 0,
1734                                 SLAB_HWCACHE_ALIGN, NULL);
1735                 if (!sgp->slab) {
1736                         printk(KERN_ERR "SCSI: can't init sg slab %s\n",
1737                                         sgp->name);
1738                         goto cleanup_sdb;
1739                 }
1740
1741                 sgp->pool = mempool_create_slab_pool(SG_MEMPOOL_SIZE,
1742                                                      sgp->slab);
1743                 if (!sgp->pool) {
1744                         printk(KERN_ERR "SCSI: can't init sg mempool %s\n",
1745                                         sgp->name);
1746                         goto cleanup_sdb;
1747                 }
1748         }
1749
1750         return 0;
1751
1752 cleanup_sdb:
1753         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1754                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1755                 if (sgp->pool)
1756                         mempool_destroy(sgp->pool);
1757                 if (sgp->slab)
1758                         kmem_cache_destroy(sgp->slab);
1759         }
1760         kmem_cache_destroy(scsi_sdb_cache);
1761
1762         return -ENOMEM;
1763 }
1764
1765 void scsi_exit_queue(void)
1766 {
1767         int i;
1768
1769         kmem_cache_destroy(scsi_sdb_cache);
1770
1771         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1772                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1773                 mempool_destroy(sgp->pool);
1774                 kmem_cache_destroy(sgp->slab);
1775         }
1776 }
1777
1778 /**
1779  *      scsi_mode_select - issue a mode select
1780  *      @sdev:  SCSI device to be queried
1781  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
1782  *      @sp:    Save page bit (0 == don't save, 1 == save)
1783  *      @modepage: mode page being requested
1784  *      @buffer: request buffer (may not be smaller than eight bytes)
1785  *      @len:   length of request buffer.
1786  *      @timeout: command timeout
1787  *      @retries: number of retries before failing
1788  *      @data: returns a structure abstracting the mode header data
1789  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
1790  *              must be SCSI_SENSE_BUFFERSIZE big.
1791  *
1792  *      Returns zero if successful; negative error number or scsi
1793  *      status on error
1794  *
1795  */
1796 int
1797 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
1798                  unsigned char *buffer, int len, int timeout, int retries,
1799                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1800 {
1801         unsigned char cmd[10];
1802         unsigned char *real_buffer;
1803         int ret;
1804
1805         memset(cmd, 0, sizeof(cmd));
1806         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
1807
1808         if (sdev->use_10_for_ms) {
1809                 if (len > 65535)
1810                         return -EINVAL;
1811                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
1812                 if (!real_buffer)
1813                         return -ENOMEM;
1814                 memcpy(real_buffer + 8, buffer, len);
1815                 len += 8;
1816                 real_buffer[0] = 0;
1817                 real_buffer[1] = 0;
1818                 real_buffer[2] = data->medium_type;
1819                 real_buffer[3] = data->device_specific;
1820                 real_buffer[4] = data->longlba ? 0x01 : 0;
1821                 real_buffer[5] = 0;
1822                 real_buffer[6] = data->block_descriptor_length >> 8;
1823                 real_buffer[7] = data->block_descriptor_length;
1824
1825                 cmd[0] = MODE_SELECT_10;
1826                 cmd[7] = len >> 8;
1827                 cmd[8] = len;
1828         } else {
1829                 if (len > 255 || data->block_descriptor_length > 255 ||
1830                     data->longlba)
1831                         return -EINVAL;
1832
1833                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
1834                 if (!real_buffer)
1835                         return -ENOMEM;
1836                 memcpy(real_buffer + 4, buffer, len);
1837                 len += 4;
1838                 real_buffer[0] = 0;
1839                 real_buffer[1] = data->medium_type;
1840                 real_buffer[2] = data->device_specific;
1841                 real_buffer[3] = data->block_descriptor_length;
1842                 
1843
1844                 cmd[0] = MODE_SELECT;
1845                 cmd[4] = len;
1846         }
1847
1848         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
1849                                sshdr, timeout, retries, NULL);
1850         kfree(real_buffer);
1851         return ret;
1852 }
1853 EXPORT_SYMBOL_GPL(scsi_mode_select);
1854
1855 /**
1856  *      scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
1857  *      @sdev:  SCSI device to be queried
1858  *      @dbd:   set if mode sense will allow block descriptors to be returned
1859  *      @modepage: mode page being requested
1860  *      @buffer: request buffer (may not be smaller than eight bytes)
1861  *      @len:   length of request buffer.
1862  *      @timeout: command timeout
1863  *      @retries: number of retries before failing
1864  *      @data: returns a structure abstracting the mode header data
1865  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
1866  *              must be SCSI_SENSE_BUFFERSIZE big.
1867  *
1868  *      Returns zero if unsuccessful, or the header offset (either 4
1869  *      or 8 depending on whether a six or ten byte command was
1870  *      issued) if successful.
1871  */
1872 int
1873 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
1874                   unsigned char *buffer, int len, int timeout, int retries,
1875                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1876 {
1877         unsigned char cmd[12];
1878         int use_10_for_ms;
1879         int header_length;
1880         int result;
1881         struct scsi_sense_hdr my_sshdr;
1882
1883         memset(data, 0, sizeof(*data));
1884         memset(&cmd[0], 0, 12);
1885         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
1886         cmd[2] = modepage;
1887
1888         /* caller might not be interested in sense, but we need it */
1889         if (!sshdr)
1890                 sshdr = &my_sshdr;
1891
1892  retry:
1893         use_10_for_ms = sdev->use_10_for_ms;
1894
1895         if (use_10_for_ms) {
1896                 if (len < 8)
1897                         len = 8;
1898
1899                 cmd[0] = MODE_SENSE_10;
1900                 cmd[8] = len;
1901                 header_length = 8;
1902         } else {
1903                 if (len < 4)
1904                         len = 4;
1905
1906                 cmd[0] = MODE_SENSE;
1907                 cmd[4] = len;
1908                 header_length = 4;
1909         }
1910
1911         memset(buffer, 0, len);
1912
1913         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
1914                                   sshdr, timeout, retries, NULL);
1915
1916         /* This code looks awful: what it's doing is making sure an
1917          * ILLEGAL REQUEST sense return identifies the actual command
1918          * byte as the problem.  MODE_SENSE commands can return
1919          * ILLEGAL REQUEST if the code page isn't supported */
1920
1921         if (use_10_for_ms && !scsi_status_is_good(result) &&
1922             (driver_byte(result) & DRIVER_SENSE)) {
1923                 if (scsi_sense_valid(sshdr)) {
1924                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
1925                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
1926                                 /* 
1927                                  * Invalid command operation code
1928                                  */
1929                                 sdev->use_10_for_ms = 0;
1930                                 goto retry;
1931                         }
1932                 }
1933         }
1934
1935         if(scsi_status_is_good(result)) {
1936                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
1937                              (modepage == 6 || modepage == 8))) {
1938                         /* Initio breakage? */
1939                         header_length = 0;
1940                         data->length = 13;
1941                         data->medium_type = 0;
1942                         data->device_specific = 0;
1943                         data->longlba = 0;
1944                         data->block_descriptor_length = 0;
1945                 } else if(use_10_for_ms) {
1946                         data->length = buffer[0]*256 + buffer[1] + 2;
1947                         data->medium_type = buffer[2];
1948                         data->device_specific = buffer[3];
1949                         data->longlba = buffer[4] & 0x01;
1950                         data->block_descriptor_length = buffer[6]*256
1951                                 + buffer[7];
1952                 } else {
1953                         data->length = buffer[0] + 1;
1954                         data->medium_type = buffer[1];
1955                         data->device_specific = buffer[2];
1956                         data->block_descriptor_length = buffer[3];
1957                 }
1958                 data->header_length = header_length;
1959         }
1960
1961         return result;
1962 }
1963 EXPORT_SYMBOL(scsi_mode_sense);
1964
1965 /**
1966  *      scsi_test_unit_ready - test if unit is ready
1967  *      @sdev:  scsi device to change the state of.
1968  *      @timeout: command timeout
1969  *      @retries: number of retries before failing
1970  *      @sshdr_external: Optional pointer to struct scsi_sense_hdr for
1971  *              returning sense. Make sure that this is cleared before passing
1972  *              in.
1973  *
1974  *      Returns zero if unsuccessful or an error if TUR failed.  For
1975  *      removable media, a return of NOT_READY or UNIT_ATTENTION is
1976  *      translated to success, with the ->changed flag updated.
1977  **/
1978 int
1979 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
1980                      struct scsi_sense_hdr *sshdr_external)
1981 {
1982         char cmd[] = {
1983                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
1984         };
1985         struct scsi_sense_hdr *sshdr;
1986         int result;
1987
1988         if (!sshdr_external)
1989                 sshdr = kzalloc(sizeof(*sshdr), GFP_KERNEL);
1990         else
1991                 sshdr = sshdr_external;
1992
1993         /* try to eat the UNIT_ATTENTION if there are enough retries */
1994         do {
1995                 result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
1996                                           timeout, retries, NULL);
1997                 if (sdev->removable && scsi_sense_valid(sshdr) &&
1998                     sshdr->sense_key == UNIT_ATTENTION)
1999                         sdev->changed = 1;
2000         } while (scsi_sense_valid(sshdr) &&
2001                  sshdr->sense_key == UNIT_ATTENTION && --retries);
2002
2003         if (!sshdr)
2004                 /* could not allocate sense buffer, so can't process it */
2005                 return result;
2006
2007         if (sdev->removable && scsi_sense_valid(sshdr) &&
2008             (sshdr->sense_key == UNIT_ATTENTION ||
2009              sshdr->sense_key == NOT_READY)) {
2010                 sdev->changed = 1;
2011                 result = 0;
2012         }
2013         if (!sshdr_external)
2014                 kfree(sshdr);
2015         return result;
2016 }
2017 EXPORT_SYMBOL(scsi_test_unit_ready);
2018
2019 /**
2020  *      scsi_device_set_state - Take the given device through the device state model.
2021  *      @sdev:  scsi device to change the state of.
2022  *      @state: state to change to.
2023  *
2024  *      Returns zero if unsuccessful or an error if the requested 
2025  *      transition is illegal.
2026  */
2027 int
2028 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2029 {
2030         enum scsi_device_state oldstate = sdev->sdev_state;
2031
2032         if (state == oldstate)
2033                 return 0;
2034
2035         switch (state) {
2036         case SDEV_CREATED:
2037                 switch (oldstate) {
2038                 case SDEV_CREATED_BLOCK:
2039                         break;
2040                 default:
2041                         goto illegal;
2042                 }
2043                 break;
2044                         
2045         case SDEV_RUNNING:
2046                 switch (oldstate) {
2047                 case SDEV_CREATED:
2048                 case SDEV_OFFLINE:
2049                 case SDEV_QUIESCE:
2050                 case SDEV_BLOCK:
2051                         break;
2052                 default:
2053                         goto illegal;
2054                 }
2055                 break;
2056
2057         case SDEV_QUIESCE:
2058                 switch (oldstate) {
2059                 case SDEV_RUNNING:
2060                 case SDEV_OFFLINE:
2061                         break;
2062                 default:
2063                         goto illegal;
2064                 }
2065                 break;
2066
2067         case SDEV_OFFLINE:
2068                 switch (oldstate) {
2069                 case SDEV_CREATED:
2070                 case SDEV_RUNNING:
2071                 case SDEV_QUIESCE:
2072                 case SDEV_BLOCK:
2073                         break;
2074                 default:
2075                         goto illegal;
2076                 }
2077                 break;
2078
2079         case SDEV_BLOCK:
2080                 switch (oldstate) {
2081                 case SDEV_RUNNING:
2082                 case SDEV_CREATED_BLOCK:
2083                         break;
2084                 default:
2085                         goto illegal;
2086                 }
2087                 break;
2088
2089         case SDEV_CREATED_BLOCK:
2090                 switch (oldstate) {
2091                 case SDEV_CREATED:
2092                         break;
2093                 default:
2094                         goto illegal;
2095                 }
2096                 break;
2097
2098         case SDEV_CANCEL:
2099                 switch (oldstate) {
2100                 case SDEV_CREATED:
2101                 case SDEV_RUNNING:
2102                 case SDEV_QUIESCE:
2103                 case SDEV_OFFLINE:
2104                 case SDEV_BLOCK:
2105                         break;
2106                 default:
2107                         goto illegal;
2108                 }
2109                 break;
2110
2111         case SDEV_DEL:
2112                 switch (oldstate) {
2113                 case SDEV_CREATED:
2114                 case SDEV_RUNNING:
2115                 case SDEV_OFFLINE:
2116                 case SDEV_CANCEL:
2117                         break;
2118                 default:
2119                         goto illegal;
2120                 }
2121                 break;
2122
2123         }
2124         sdev->sdev_state = state;
2125         return 0;
2126
2127  illegal:
2128         SCSI_LOG_ERROR_RECOVERY(1, 
2129                                 sdev_printk(KERN_ERR, sdev,
2130                                             "Illegal state transition %s->%s\n",
2131                                             scsi_device_state_name(oldstate),
2132                                             scsi_device_state_name(state))
2133                                 );
2134         return -EINVAL;
2135 }
2136 EXPORT_SYMBOL(scsi_device_set_state);
2137
2138 /**
2139  *      sdev_evt_emit - emit a single SCSI device uevent
2140  *      @sdev: associated SCSI device
2141  *      @evt: event to emit
2142  *
2143  *      Send a single uevent (scsi_event) to the associated scsi_device.
2144  */
2145 static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2146 {
2147         int idx = 0;
2148         char *envp[3];
2149
2150         switch (evt->evt_type) {
2151         case SDEV_EVT_MEDIA_CHANGE:
2152                 envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2153                 break;
2154
2155         default:
2156                 /* do nothing */
2157                 break;
2158         }
2159
2160         envp[idx++] = NULL;
2161
2162         kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2163 }
2164
2165 /**
2166  *      sdev_evt_thread - send a uevent for each scsi event
2167  *      @work: work struct for scsi_device
2168  *
2169  *      Dispatch queued events to their associated scsi_device kobjects
2170  *      as uevents.
2171  */
2172 void scsi_evt_thread(struct work_struct *work)
2173 {
2174         struct scsi_device *sdev;
2175         LIST_HEAD(event_list);
2176
2177         sdev = container_of(work, struct scsi_device, event_work);
2178
2179         while (1) {
2180                 struct scsi_event *evt;
2181                 struct list_head *this, *tmp;
2182                 unsigned long flags;
2183
2184                 spin_lock_irqsave(&sdev->list_lock, flags);
2185                 list_splice_init(&sdev->event_list, &event_list);
2186                 spin_unlock_irqrestore(&sdev->list_lock, flags);
2187
2188                 if (list_empty(&event_list))
2189                         break;
2190
2191                 list_for_each_safe(this, tmp, &event_list) {
2192                         evt = list_entry(this, struct scsi_event, node);
2193                         list_del(&evt->node);
2194                         scsi_evt_emit(sdev, evt);
2195                         kfree(evt);
2196                 }
2197         }
2198 }
2199
2200 /**
2201  *      sdev_evt_send - send asserted event to uevent thread
2202  *      @sdev: scsi_device event occurred on
2203  *      @evt: event to send
2204  *
2205  *      Assert scsi device event asynchronously.
2206  */
2207 void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2208 {
2209         unsigned long flags;
2210
2211 #if 0
2212         /* FIXME: currently this check eliminates all media change events
2213          * for polled devices.  Need to update to discriminate between AN
2214          * and polled events */
2215         if (!test_bit(evt->evt_type, sdev->supported_events)) {
2216                 kfree(evt);
2217                 return;
2218         }
2219 #endif
2220
2221         spin_lock_irqsave(&sdev->list_lock, flags);
2222         list_add_tail(&evt->node, &sdev->event_list);
2223         schedule_work(&sdev->event_work);
2224         spin_unlock_irqrestore(&sdev->list_lock, flags);
2225 }
2226 EXPORT_SYMBOL_GPL(sdev_evt_send);
2227
2228 /**
2229  *      sdev_evt_alloc - allocate a new scsi event
2230  *      @evt_type: type of event to allocate
2231  *      @gfpflags: GFP flags for allocation
2232  *
2233  *      Allocates and returns a new scsi_event.
2234  */
2235 struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2236                                   gfp_t gfpflags)
2237 {
2238         struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2239         if (!evt)
2240                 return NULL;
2241
2242         evt->evt_type = evt_type;
2243         INIT_LIST_HEAD(&evt->node);
2244
2245         /* evt_type-specific initialization, if any */
2246         switch (evt_type) {
2247         case SDEV_EVT_MEDIA_CHANGE:
2248         default:
2249                 /* do nothing */
2250                 break;
2251         }
2252
2253         return evt;
2254 }
2255 EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2256
2257 /**
2258  *      sdev_evt_send_simple - send asserted event to uevent thread
2259  *      @sdev: scsi_device event occurred on
2260  *      @evt_type: type of event to send
2261  *      @gfpflags: GFP flags for allocation
2262  *
2263  *      Assert scsi device event asynchronously, given an event type.
2264  */
2265 void sdev_evt_send_simple(struct scsi_device *sdev,
2266                           enum scsi_device_event evt_type, gfp_t gfpflags)
2267 {
2268         struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2269         if (!evt) {
2270                 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2271                             evt_type);
2272                 return;
2273         }
2274
2275         sdev_evt_send(sdev, evt);
2276 }
2277 EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2278
2279 /**
2280  *      scsi_device_quiesce - Block user issued commands.
2281  *      @sdev:  scsi device to quiesce.
2282  *
2283  *      This works by trying to transition to the SDEV_QUIESCE state
2284  *      (which must be a legal transition).  When the device is in this
2285  *      state, only special requests will be accepted, all others will
2286  *      be deferred.  Since special requests may also be requeued requests,
2287  *      a successful return doesn't guarantee the device will be 
2288  *      totally quiescent.
2289  *
2290  *      Must be called with user context, may sleep.
2291  *
2292  *      Returns zero if unsuccessful or an error if not.
2293  */
2294 int
2295 scsi_device_quiesce(struct scsi_device *sdev)
2296 {
2297         int err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2298         if (err)
2299                 return err;
2300
2301         scsi_run_queue(sdev->request_queue);
2302         while (sdev->device_busy) {
2303                 msleep_interruptible(200);
2304                 scsi_run_queue(sdev->request_queue);
2305         }
2306         return 0;
2307 }
2308 EXPORT_SYMBOL(scsi_device_quiesce);
2309
2310 /**
2311  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2312  *      @sdev:  scsi device to resume.
2313  *
2314  *      Moves the device from quiesced back to running and restarts the
2315  *      queues.
2316  *
2317  *      Must be called with user context, may sleep.
2318  */
2319 void
2320 scsi_device_resume(struct scsi_device *sdev)
2321 {
2322         if(scsi_device_set_state(sdev, SDEV_RUNNING))
2323                 return;
2324         scsi_run_queue(sdev->request_queue);
2325 }
2326 EXPORT_SYMBOL(scsi_device_resume);
2327
2328 static void
2329 device_quiesce_fn(struct scsi_device *sdev, void *data)
2330 {
2331         scsi_device_quiesce(sdev);
2332 }
2333
2334 void
2335 scsi_target_quiesce(struct scsi_target *starget)
2336 {
2337         starget_for_each_device(starget, NULL, device_quiesce_fn);
2338 }
2339 EXPORT_SYMBOL(scsi_target_quiesce);
2340
2341 static void
2342 device_resume_fn(struct scsi_device *sdev, void *data)
2343 {
2344         scsi_device_resume(sdev);
2345 }
2346
2347 void
2348 scsi_target_resume(struct scsi_target *starget)
2349 {
2350         starget_for_each_device(starget, NULL, device_resume_fn);
2351 }
2352 EXPORT_SYMBOL(scsi_target_resume);
2353
2354 /**
2355  * scsi_internal_device_block - internal function to put a device temporarily into the SDEV_BLOCK state
2356  * @sdev:       device to block
2357  *
2358  * Block request made by scsi lld's to temporarily stop all
2359  * scsi commands on the specified device.  Called from interrupt
2360  * or normal process context.
2361  *
2362  * Returns zero if successful or error if not
2363  *
2364  * Notes:       
2365  *      This routine transitions the device to the SDEV_BLOCK state
2366  *      (which must be a legal transition).  When the device is in this
2367  *      state, all commands are deferred until the scsi lld reenables
2368  *      the device with scsi_device_unblock or device_block_tmo fires.
2369  *      This routine assumes the host_lock is held on entry.
2370  */
2371 int
2372 scsi_internal_device_block(struct scsi_device *sdev)
2373 {
2374         struct request_queue *q = sdev->request_queue;
2375         unsigned long flags;
2376         int err = 0;
2377
2378         err = scsi_device_set_state(sdev, SDEV_BLOCK);
2379         if (err) {
2380                 err = scsi_device_set_state(sdev, SDEV_CREATED_BLOCK);
2381
2382                 if (err)
2383                         return err;
2384         }
2385
2386         /* 
2387          * The device has transitioned to SDEV_BLOCK.  Stop the
2388          * block layer from calling the midlayer with this device's
2389          * request queue. 
2390          */
2391         spin_lock_irqsave(q->queue_lock, flags);
2392         blk_stop_queue(q);
2393         spin_unlock_irqrestore(q->queue_lock, flags);
2394
2395         return 0;
2396 }
2397 EXPORT_SYMBOL_GPL(scsi_internal_device_block);
2398  
2399 /**
2400  * scsi_internal_device_unblock - resume a device after a block request
2401  * @sdev:       device to resume
2402  *
2403  * Called by scsi lld's or the midlayer to restart the device queue
2404  * for the previously suspended scsi device.  Called from interrupt or
2405  * normal process context.
2406  *
2407  * Returns zero if successful or error if not.
2408  *
2409  * Notes:       
2410  *      This routine transitions the device to the SDEV_RUNNING state
2411  *      (which must be a legal transition) allowing the midlayer to
2412  *      goose the queue for this device.  This routine assumes the 
2413  *      host_lock is held upon entry.
2414  */
2415 int
2416 scsi_internal_device_unblock(struct scsi_device *sdev)
2417 {
2418         struct request_queue *q = sdev->request_queue; 
2419         unsigned long flags;
2420         
2421         /* 
2422          * Try to transition the scsi device to SDEV_RUNNING
2423          * and goose the device queue if successful.  
2424          */
2425         if (sdev->sdev_state == SDEV_BLOCK)
2426                 sdev->sdev_state = SDEV_RUNNING;
2427         else if (sdev->sdev_state == SDEV_CREATED_BLOCK)
2428                 sdev->sdev_state = SDEV_CREATED;
2429         else
2430                 return -EINVAL;
2431
2432         spin_lock_irqsave(q->queue_lock, flags);
2433         blk_start_queue(q);
2434         spin_unlock_irqrestore(q->queue_lock, flags);
2435
2436         return 0;
2437 }
2438 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock);
2439
2440 static void
2441 device_block(struct scsi_device *sdev, void *data)
2442 {
2443         scsi_internal_device_block(sdev);
2444 }
2445
2446 static int
2447 target_block(struct device *dev, void *data)
2448 {
2449         if (scsi_is_target_device(dev))
2450                 starget_for_each_device(to_scsi_target(dev), NULL,
2451                                         device_block);
2452         return 0;
2453 }
2454
2455 void
2456 scsi_target_block(struct device *dev)
2457 {
2458         if (scsi_is_target_device(dev))
2459                 starget_for_each_device(to_scsi_target(dev), NULL,
2460                                         device_block);
2461         else
2462                 device_for_each_child(dev, NULL, target_block);
2463 }
2464 EXPORT_SYMBOL_GPL(scsi_target_block);
2465
2466 static void
2467 device_unblock(struct scsi_device *sdev, void *data)
2468 {
2469         scsi_internal_device_unblock(sdev);
2470 }
2471
2472 static int
2473 target_unblock(struct device *dev, void *data)
2474 {
2475         if (scsi_is_target_device(dev))
2476                 starget_for_each_device(to_scsi_target(dev), NULL,
2477                                         device_unblock);
2478         return 0;
2479 }
2480
2481 void
2482 scsi_target_unblock(struct device *dev)
2483 {
2484         if (scsi_is_target_device(dev))
2485                 starget_for_each_device(to_scsi_target(dev), NULL,
2486                                         device_unblock);
2487         else
2488                 device_for_each_child(dev, NULL, target_unblock);
2489 }
2490 EXPORT_SYMBOL_GPL(scsi_target_unblock);
2491
2492 /**
2493  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2494  * @sgl:        scatter-gather list
2495  * @sg_count:   number of segments in sg
2496  * @offset:     offset in bytes into sg, on return offset into the mapped area
2497  * @len:        bytes to map, on return number of bytes mapped
2498  *
2499  * Returns virtual address of the start of the mapped page
2500  */
2501 void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
2502                           size_t *offset, size_t *len)
2503 {
2504         int i;
2505         size_t sg_len = 0, len_complete = 0;
2506         struct scatterlist *sg;
2507         struct page *page;
2508
2509         WARN_ON(!irqs_disabled());
2510
2511         for_each_sg(sgl, sg, sg_count, i) {
2512                 len_complete = sg_len; /* Complete sg-entries */
2513                 sg_len += sg->length;
2514                 if (sg_len > *offset)
2515                         break;
2516         }
2517
2518         if (unlikely(i == sg_count)) {
2519                 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2520                         "elements %d\n",
2521                        __func__, sg_len, *offset, sg_count);
2522                 WARN_ON(1);
2523                 return NULL;
2524         }
2525
2526         /* Offset starting from the beginning of first page in this sg-entry */
2527         *offset = *offset - len_complete + sg->offset;
2528
2529         /* Assumption: contiguous pages can be accessed as "page + i" */
2530         page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
2531         *offset &= ~PAGE_MASK;
2532
2533         /* Bytes in this sg-entry from *offset to the end of the page */
2534         sg_len = PAGE_SIZE - *offset;
2535         if (*len > sg_len)
2536                 *len = sg_len;
2537
2538         return kmap_atomic(page, KM_BIO_SRC_IRQ);
2539 }
2540 EXPORT_SYMBOL(scsi_kmap_atomic_sg);
2541
2542 /**
2543  * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
2544  * @virt:       virtual address to be unmapped
2545  */
2546 void scsi_kunmap_atomic_sg(void *virt)
2547 {
2548         kunmap_atomic(virt, KM_BIO_SRC_IRQ);
2549 }
2550 EXPORT_SYMBOL(scsi_kunmap_atomic_sg);