PM: Remove CONFIG_PM_OPS
[linux-2.6.git] / drivers / scsi / scsi_lib.c
1 /*
2  *  scsi_lib.c Copyright (C) 1999 Eric Youngdale
3  *
4  *  SCSI queueing library.
5  *      Initial versions: Eric Youngdale (eric@andante.org).
6  *                        Based upon conversations with large numbers
7  *                        of people at Linux Expo.
8  */
9
10 #include <linux/bio.h>
11 #include <linux/bitops.h>
12 #include <linux/blkdev.h>
13 #include <linux/completion.h>
14 #include <linux/kernel.h>
15 #include <linux/mempool.h>
16 #include <linux/slab.h>
17 #include <linux/init.h>
18 #include <linux/pci.h>
19 #include <linux/delay.h>
20 #include <linux/hardirq.h>
21 #include <linux/scatterlist.h>
22
23 #include <scsi/scsi.h>
24 #include <scsi/scsi_cmnd.h>
25 #include <scsi/scsi_dbg.h>
26 #include <scsi/scsi_device.h>
27 #include <scsi/scsi_driver.h>
28 #include <scsi/scsi_eh.h>
29 #include <scsi/scsi_host.h>
30
31 #include "scsi_priv.h"
32 #include "scsi_logging.h"
33
34
35 #define SG_MEMPOOL_NR           ARRAY_SIZE(scsi_sg_pools)
36 #define SG_MEMPOOL_SIZE         2
37
38 struct scsi_host_sg_pool {
39         size_t          size;
40         char            *name;
41         struct kmem_cache       *slab;
42         mempool_t       *pool;
43 };
44
45 #define SP(x) { x, "sgpool-" __stringify(x) }
46 #if (SCSI_MAX_SG_SEGMENTS < 32)
47 #error SCSI_MAX_SG_SEGMENTS is too small (must be 32 or greater)
48 #endif
49 static struct scsi_host_sg_pool scsi_sg_pools[] = {
50         SP(8),
51         SP(16),
52 #if (SCSI_MAX_SG_SEGMENTS > 32)
53         SP(32),
54 #if (SCSI_MAX_SG_SEGMENTS > 64)
55         SP(64),
56 #if (SCSI_MAX_SG_SEGMENTS > 128)
57         SP(128),
58 #if (SCSI_MAX_SG_SEGMENTS > 256)
59 #error SCSI_MAX_SG_SEGMENTS is too large (256 MAX)
60 #endif
61 #endif
62 #endif
63 #endif
64         SP(SCSI_MAX_SG_SEGMENTS)
65 };
66 #undef SP
67
68 struct kmem_cache *scsi_sdb_cache;
69
70 static void scsi_run_queue(struct request_queue *q);
71
72 /*
73  * Function:    scsi_unprep_request()
74  *
75  * Purpose:     Remove all preparation done for a request, including its
76  *              associated scsi_cmnd, so that it can be requeued.
77  *
78  * Arguments:   req     - request to unprepare
79  *
80  * Lock status: Assumed that no locks are held upon entry.
81  *
82  * Returns:     Nothing.
83  */
84 static void scsi_unprep_request(struct request *req)
85 {
86         struct scsi_cmnd *cmd = req->special;
87
88         blk_unprep_request(req);
89         req->special = NULL;
90
91         scsi_put_command(cmd);
92 }
93
94 /**
95  * __scsi_queue_insert - private queue insertion
96  * @cmd: The SCSI command being requeued
97  * @reason:  The reason for the requeue
98  * @unbusy: Whether the queue should be unbusied
99  *
100  * This is a private queue insertion.  The public interface
101  * scsi_queue_insert() always assumes the queue should be unbusied
102  * because it's always called before the completion.  This function is
103  * for a requeue after completion, which should only occur in this
104  * file.
105  */
106 static int __scsi_queue_insert(struct scsi_cmnd *cmd, int reason, int unbusy)
107 {
108         struct Scsi_Host *host = cmd->device->host;
109         struct scsi_device *device = cmd->device;
110         struct scsi_target *starget = scsi_target(device);
111         struct request_queue *q = device->request_queue;
112         unsigned long flags;
113
114         SCSI_LOG_MLQUEUE(1,
115                  printk("Inserting command %p into mlqueue\n", cmd));
116
117         /*
118          * Set the appropriate busy bit for the device/host.
119          *
120          * If the host/device isn't busy, assume that something actually
121          * completed, and that we should be able to queue a command now.
122          *
123          * Note that the prior mid-layer assumption that any host could
124          * always queue at least one command is now broken.  The mid-layer
125          * will implement a user specifiable stall (see
126          * scsi_host.max_host_blocked and scsi_device.max_device_blocked)
127          * if a command is requeued with no other commands outstanding
128          * either for the device or for the host.
129          */
130         switch (reason) {
131         case SCSI_MLQUEUE_HOST_BUSY:
132                 host->host_blocked = host->max_host_blocked;
133                 break;
134         case SCSI_MLQUEUE_DEVICE_BUSY:
135                 device->device_blocked = device->max_device_blocked;
136                 break;
137         case SCSI_MLQUEUE_TARGET_BUSY:
138                 starget->target_blocked = starget->max_target_blocked;
139                 break;
140         }
141
142         /*
143          * Decrement the counters, since these commands are no longer
144          * active on the host/device.
145          */
146         if (unbusy)
147                 scsi_device_unbusy(device);
148
149         /*
150          * Requeue this command.  It will go before all other commands
151          * that are already in the queue.
152          *
153          * NOTE: there is magic here about the way the queue is plugged if
154          * we have no outstanding commands.
155          * 
156          * Although we *don't* plug the queue, we call the request
157          * function.  The SCSI request function detects the blocked condition
158          * and plugs the queue appropriately.
159          */
160         spin_lock_irqsave(q->queue_lock, flags);
161         blk_requeue_request(q, cmd->request);
162         spin_unlock_irqrestore(q->queue_lock, flags);
163
164         scsi_run_queue(q);
165
166         return 0;
167 }
168
169 /*
170  * Function:    scsi_queue_insert()
171  *
172  * Purpose:     Insert a command in the midlevel queue.
173  *
174  * Arguments:   cmd    - command that we are adding to queue.
175  *              reason - why we are inserting command to queue.
176  *
177  * Lock status: Assumed that lock is not held upon entry.
178  *
179  * Returns:     Nothing.
180  *
181  * Notes:       We do this for one of two cases.  Either the host is busy
182  *              and it cannot accept any more commands for the time being,
183  *              or the device returned QUEUE_FULL and can accept no more
184  *              commands.
185  * Notes:       This could be called either from an interrupt context or a
186  *              normal process context.
187  */
188 int scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
189 {
190         return __scsi_queue_insert(cmd, reason, 1);
191 }
192 /**
193  * scsi_execute - insert request and wait for the result
194  * @sdev:       scsi device
195  * @cmd:        scsi command
196  * @data_direction: data direction
197  * @buffer:     data buffer
198  * @bufflen:    len of buffer
199  * @sense:      optional sense buffer
200  * @timeout:    request timeout in seconds
201  * @retries:    number of times to retry request
202  * @flags:      or into request flags;
203  * @resid:      optional residual length
204  *
205  * returns the req->errors value which is the scsi_cmnd result
206  * field.
207  */
208 int scsi_execute(struct scsi_device *sdev, const unsigned char *cmd,
209                  int data_direction, void *buffer, unsigned bufflen,
210                  unsigned char *sense, int timeout, int retries, int flags,
211                  int *resid)
212 {
213         struct request *req;
214         int write = (data_direction == DMA_TO_DEVICE);
215         int ret = DRIVER_ERROR << 24;
216
217         req = blk_get_request(sdev->request_queue, write, __GFP_WAIT);
218
219         if (bufflen &&  blk_rq_map_kern(sdev->request_queue, req,
220                                         buffer, bufflen, __GFP_WAIT))
221                 goto out;
222
223         req->cmd_len = COMMAND_SIZE(cmd[0]);
224         memcpy(req->cmd, cmd, req->cmd_len);
225         req->sense = sense;
226         req->sense_len = 0;
227         req->retries = retries;
228         req->timeout = timeout;
229         req->cmd_type = REQ_TYPE_BLOCK_PC;
230         req->cmd_flags |= flags | REQ_QUIET | REQ_PREEMPT;
231
232         /*
233          * head injection *required* here otherwise quiesce won't work
234          */
235         blk_execute_rq(req->q, NULL, req, 1);
236
237         /*
238          * Some devices (USB mass-storage in particular) may transfer
239          * garbage data together with a residue indicating that the data
240          * is invalid.  Prevent the garbage from being misinterpreted
241          * and prevent security leaks by zeroing out the excess data.
242          */
243         if (unlikely(req->resid_len > 0 && req->resid_len <= bufflen))
244                 memset(buffer + (bufflen - req->resid_len), 0, req->resid_len);
245
246         if (resid)
247                 *resid = req->resid_len;
248         ret = req->errors;
249  out:
250         blk_put_request(req);
251
252         return ret;
253 }
254 EXPORT_SYMBOL(scsi_execute);
255
256
257 int scsi_execute_req(struct scsi_device *sdev, const unsigned char *cmd,
258                      int data_direction, void *buffer, unsigned bufflen,
259                      struct scsi_sense_hdr *sshdr, int timeout, int retries,
260                      int *resid)
261 {
262         char *sense = NULL;
263         int result;
264         
265         if (sshdr) {
266                 sense = kzalloc(SCSI_SENSE_BUFFERSIZE, GFP_NOIO);
267                 if (!sense)
268                         return DRIVER_ERROR << 24;
269         }
270         result = scsi_execute(sdev, cmd, data_direction, buffer, bufflen,
271                               sense, timeout, retries, 0, resid);
272         if (sshdr)
273                 scsi_normalize_sense(sense, SCSI_SENSE_BUFFERSIZE, sshdr);
274
275         kfree(sense);
276         return result;
277 }
278 EXPORT_SYMBOL(scsi_execute_req);
279
280 /*
281  * Function:    scsi_init_cmd_errh()
282  *
283  * Purpose:     Initialize cmd fields related to error handling.
284  *
285  * Arguments:   cmd     - command that is ready to be queued.
286  *
287  * Notes:       This function has the job of initializing a number of
288  *              fields related to error handling.   Typically this will
289  *              be called once for each command, as required.
290  */
291 static void scsi_init_cmd_errh(struct scsi_cmnd *cmd)
292 {
293         cmd->serial_number = 0;
294         scsi_set_resid(cmd, 0);
295         memset(cmd->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
296         if (cmd->cmd_len == 0)
297                 cmd->cmd_len = scsi_command_size(cmd->cmnd);
298 }
299
300 void scsi_device_unbusy(struct scsi_device *sdev)
301 {
302         struct Scsi_Host *shost = sdev->host;
303         struct scsi_target *starget = scsi_target(sdev);
304         unsigned long flags;
305
306         spin_lock_irqsave(shost->host_lock, flags);
307         shost->host_busy--;
308         starget->target_busy--;
309         if (unlikely(scsi_host_in_recovery(shost) &&
310                      (shost->host_failed || shost->host_eh_scheduled)))
311                 scsi_eh_wakeup(shost);
312         spin_unlock(shost->host_lock);
313         spin_lock(sdev->request_queue->queue_lock);
314         sdev->device_busy--;
315         spin_unlock_irqrestore(sdev->request_queue->queue_lock, flags);
316 }
317
318 /*
319  * Called for single_lun devices on IO completion. Clear starget_sdev_user,
320  * and call blk_run_queue for all the scsi_devices on the target -
321  * including current_sdev first.
322  *
323  * Called with *no* scsi locks held.
324  */
325 static void scsi_single_lun_run(struct scsi_device *current_sdev)
326 {
327         struct Scsi_Host *shost = current_sdev->host;
328         struct scsi_device *sdev, *tmp;
329         struct scsi_target *starget = scsi_target(current_sdev);
330         unsigned long flags;
331
332         spin_lock_irqsave(shost->host_lock, flags);
333         starget->starget_sdev_user = NULL;
334         spin_unlock_irqrestore(shost->host_lock, flags);
335
336         /*
337          * Call blk_run_queue for all LUNs on the target, starting with
338          * current_sdev. We race with others (to set starget_sdev_user),
339          * but in most cases, we will be first. Ideally, each LU on the
340          * target would get some limited time or requests on the target.
341          */
342         blk_run_queue(current_sdev->request_queue);
343
344         spin_lock_irqsave(shost->host_lock, flags);
345         if (starget->starget_sdev_user)
346                 goto out;
347         list_for_each_entry_safe(sdev, tmp, &starget->devices,
348                         same_target_siblings) {
349                 if (sdev == current_sdev)
350                         continue;
351                 if (scsi_device_get(sdev))
352                         continue;
353
354                 spin_unlock_irqrestore(shost->host_lock, flags);
355                 blk_run_queue(sdev->request_queue);
356                 spin_lock_irqsave(shost->host_lock, flags);
357         
358                 scsi_device_put(sdev);
359         }
360  out:
361         spin_unlock_irqrestore(shost->host_lock, flags);
362 }
363
364 static inline int scsi_device_is_busy(struct scsi_device *sdev)
365 {
366         if (sdev->device_busy >= sdev->queue_depth || sdev->device_blocked)
367                 return 1;
368
369         return 0;
370 }
371
372 static inline int scsi_target_is_busy(struct scsi_target *starget)
373 {
374         return ((starget->can_queue > 0 &&
375                  starget->target_busy >= starget->can_queue) ||
376                  starget->target_blocked);
377 }
378
379 static inline int scsi_host_is_busy(struct Scsi_Host *shost)
380 {
381         if ((shost->can_queue > 0 && shost->host_busy >= shost->can_queue) ||
382             shost->host_blocked || shost->host_self_blocked)
383                 return 1;
384
385         return 0;
386 }
387
388 /*
389  * Function:    scsi_run_queue()
390  *
391  * Purpose:     Select a proper request queue to serve next
392  *
393  * Arguments:   q       - last request's queue
394  *
395  * Returns:     Nothing
396  *
397  * Notes:       The previous command was completely finished, start
398  *              a new one if possible.
399  */
400 static void scsi_run_queue(struct request_queue *q)
401 {
402         struct scsi_device *sdev = q->queuedata;
403         struct Scsi_Host *shost = sdev->host;
404         LIST_HEAD(starved_list);
405         unsigned long flags;
406
407         if (scsi_target(sdev)->single_lun)
408                 scsi_single_lun_run(sdev);
409
410         spin_lock_irqsave(shost->host_lock, flags);
411         list_splice_init(&shost->starved_list, &starved_list);
412
413         while (!list_empty(&starved_list)) {
414                 int flagset;
415
416                 /*
417                  * As long as shost is accepting commands and we have
418                  * starved queues, call blk_run_queue. scsi_request_fn
419                  * drops the queue_lock and can add us back to the
420                  * starved_list.
421                  *
422                  * host_lock protects the starved_list and starved_entry.
423                  * scsi_request_fn must get the host_lock before checking
424                  * or modifying starved_list or starved_entry.
425                  */
426                 if (scsi_host_is_busy(shost))
427                         break;
428
429                 sdev = list_entry(starved_list.next,
430                                   struct scsi_device, starved_entry);
431                 list_del_init(&sdev->starved_entry);
432                 if (scsi_target_is_busy(scsi_target(sdev))) {
433                         list_move_tail(&sdev->starved_entry,
434                                        &shost->starved_list);
435                         continue;
436                 }
437
438                 spin_unlock(shost->host_lock);
439
440                 spin_lock(sdev->request_queue->queue_lock);
441                 flagset = test_bit(QUEUE_FLAG_REENTER, &q->queue_flags) &&
442                                 !test_bit(QUEUE_FLAG_REENTER,
443                                         &sdev->request_queue->queue_flags);
444                 if (flagset)
445                         queue_flag_set(QUEUE_FLAG_REENTER, sdev->request_queue);
446                 __blk_run_queue(sdev->request_queue, false);
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 (req->cmd_type == REQ_TYPE_BLOCK_PC) { /* 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                         scsi_release_buffers(cmd);
753                         blk_end_request_all(req, 0);
754
755                         scsi_next_command(cmd);
756                         return;
757                 }
758         }
759
760         /* no bidi support for !REQ_TYPE_BLOCK_PC yet */
761         BUG_ON(blk_bidi_rq(req));
762
763         /*
764          * Next deal with any sectors which we were able to correctly
765          * handle.
766          */
767         SCSI_LOG_HLCOMPLETE(1, printk("%u sectors total, "
768                                       "%d bytes done.\n",
769                                       blk_rq_sectors(req), good_bytes));
770
771         /*
772          * Recovered errors need reporting, but they're always treated
773          * as success, so fiddle the result code here.  For BLOCK_PC
774          * we already took a copy of the original into rq->errors which
775          * is what gets returned to the user
776          */
777         if (sense_valid && (sshdr.sense_key == RECOVERED_ERROR)) {
778                 /* if ATA PASS-THROUGH INFORMATION AVAILABLE skip
779                  * print since caller wants ATA registers. Only occurs on
780                  * SCSI ATA PASS_THROUGH commands when CK_COND=1
781                  */
782                 if ((sshdr.asc == 0x0) && (sshdr.ascq == 0x1d))
783                         ;
784                 else if (!(req->cmd_flags & REQ_QUIET))
785                         scsi_print_sense("", cmd);
786                 result = 0;
787                 /* BLOCK_PC may have set error */
788                 error = 0;
789         }
790
791         /*
792          * A number of bytes were successfully read.  If there
793          * are leftovers and there is some kind of error
794          * (result != 0), retry the rest.
795          */
796         if (scsi_end_request(cmd, error, good_bytes, result == 0) == NULL)
797                 return;
798
799         error = -EIO;
800
801         if (host_byte(result) == DID_RESET) {
802                 /* Third party bus reset or reset for error recovery
803                  * reasons.  Just retry the command and see what
804                  * happens.
805                  */
806                 action = ACTION_RETRY;
807         } else if (sense_valid && !sense_deferred) {
808                 switch (sshdr.sense_key) {
809                 case UNIT_ATTENTION:
810                         if (cmd->device->removable) {
811                                 /* Detected disc change.  Set a bit
812                                  * and quietly refuse further access.
813                                  */
814                                 cmd->device->changed = 1;
815                                 description = "Media Changed";
816                                 action = ACTION_FAIL;
817                         } else {
818                                 /* Must have been a power glitch, or a
819                                  * bus reset.  Could not have been a
820                                  * media change, so we just retry the
821                                  * command and see what happens.
822                                  */
823                                 action = ACTION_RETRY;
824                         }
825                         break;
826                 case ILLEGAL_REQUEST:
827                         /* If we had an ILLEGAL REQUEST returned, then
828                          * we may have performed an unsupported
829                          * command.  The only thing this should be
830                          * would be a ten byte read where only a six
831                          * byte read was supported.  Also, on a system
832                          * where READ CAPACITY failed, we may have
833                          * read past the end of the disk.
834                          */
835                         if ((cmd->device->use_10_for_rw &&
836                             sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
837                             (cmd->cmnd[0] == READ_10 ||
838                              cmd->cmnd[0] == WRITE_10)) {
839                                 /* This will issue a new 6-byte command. */
840                                 cmd->device->use_10_for_rw = 0;
841                                 action = ACTION_REPREP;
842                         } else if (sshdr.asc == 0x10) /* DIX */ {
843                                 description = "Host Data Integrity Failure";
844                                 action = ACTION_FAIL;
845                                 error = -EILSEQ;
846                         } else
847                                 action = ACTION_FAIL;
848                         break;
849                 case ABORTED_COMMAND:
850                         action = ACTION_FAIL;
851                         if (sshdr.asc == 0x10) { /* DIF */
852                                 description = "Target Data Integrity Failure";
853                                 error = -EILSEQ;
854                         }
855                         break;
856                 case NOT_READY:
857                         /* If the device is in the process of becoming
858                          * ready, or has a temporary blockage, retry.
859                          */
860                         if (sshdr.asc == 0x04) {
861                                 switch (sshdr.ascq) {
862                                 case 0x01: /* becoming ready */
863                                 case 0x04: /* format in progress */
864                                 case 0x05: /* rebuild in progress */
865                                 case 0x06: /* recalculation in progress */
866                                 case 0x07: /* operation in progress */
867                                 case 0x08: /* Long write in progress */
868                                 case 0x09: /* self test in progress */
869                                 case 0x14: /* space allocation in progress */
870                                         action = ACTION_DELAYED_RETRY;
871                                         break;
872                                 default:
873                                         description = "Device not ready";
874                                         action = ACTION_FAIL;
875                                         break;
876                                 }
877                         } else {
878                                 description = "Device not ready";
879                                 action = ACTION_FAIL;
880                         }
881                         break;
882                 case VOLUME_OVERFLOW:
883                         /* See SSC3rXX or current. */
884                         action = ACTION_FAIL;
885                         break;
886                 default:
887                         description = "Unhandled sense code";
888                         action = ACTION_FAIL;
889                         break;
890                 }
891         } else {
892                 description = "Unhandled error code";
893                 action = ACTION_FAIL;
894         }
895
896         switch (action) {
897         case ACTION_FAIL:
898                 /* Give up and fail the remainder of the request */
899                 scsi_release_buffers(cmd);
900                 if (!(req->cmd_flags & REQ_QUIET)) {
901                         if (description)
902                                 scmd_printk(KERN_INFO, cmd, "%s\n",
903                                             description);
904                         scsi_print_result(cmd);
905                         if (driver_byte(result) & DRIVER_SENSE)
906                                 scsi_print_sense("", cmd);
907                         scsi_print_command(cmd);
908                 }
909                 if (blk_end_request_err(req, error))
910                         scsi_requeue_command(q, cmd);
911                 else
912                         scsi_next_command(cmd);
913                 break;
914         case ACTION_REPREP:
915                 /* Unprep the request and put it back at the head of the queue.
916                  * A new command will be prepared and issued.
917                  */
918                 scsi_release_buffers(cmd);
919                 scsi_requeue_command(q, cmd);
920                 break;
921         case ACTION_RETRY:
922                 /* Retry the same command immediately */
923                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY, 0);
924                 break;
925         case ACTION_DELAYED_RETRY:
926                 /* Retry the same command after a delay */
927                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY, 0);
928                 break;
929         }
930 }
931
932 static int scsi_init_sgtable(struct request *req, struct scsi_data_buffer *sdb,
933                              gfp_t gfp_mask)
934 {
935         int count;
936
937         /*
938          * If sg table allocation fails, requeue request later.
939          */
940         if (unlikely(scsi_alloc_sgtable(sdb, req->nr_phys_segments,
941                                         gfp_mask))) {
942                 return BLKPREP_DEFER;
943         }
944
945         req->buffer = NULL;
946
947         /* 
948          * Next, walk the list, and fill in the addresses and sizes of
949          * each segment.
950          */
951         count = blk_rq_map_sg(req->q, req, sdb->table.sgl);
952         BUG_ON(count > sdb->table.nents);
953         sdb->table.nents = count;
954         sdb->length = blk_rq_bytes(req);
955         return BLKPREP_OK;
956 }
957
958 /*
959  * Function:    scsi_init_io()
960  *
961  * Purpose:     SCSI I/O initialize function.
962  *
963  * Arguments:   cmd   - Command descriptor we wish to initialize
964  *
965  * Returns:     0 on success
966  *              BLKPREP_DEFER if the failure is retryable
967  *              BLKPREP_KILL if the failure is fatal
968  */
969 int scsi_init_io(struct scsi_cmnd *cmd, gfp_t gfp_mask)
970 {
971         struct request *rq = cmd->request;
972
973         int error = scsi_init_sgtable(rq, &cmd->sdb, gfp_mask);
974         if (error)
975                 goto err_exit;
976
977         if (blk_bidi_rq(rq)) {
978                 struct scsi_data_buffer *bidi_sdb = kmem_cache_zalloc(
979                         scsi_sdb_cache, GFP_ATOMIC);
980                 if (!bidi_sdb) {
981                         error = BLKPREP_DEFER;
982                         goto err_exit;
983                 }
984
985                 rq->next_rq->special = bidi_sdb;
986                 error = scsi_init_sgtable(rq->next_rq, bidi_sdb, GFP_ATOMIC);
987                 if (error)
988                         goto err_exit;
989         }
990
991         if (blk_integrity_rq(rq)) {
992                 struct scsi_data_buffer *prot_sdb = cmd->prot_sdb;
993                 int ivecs, count;
994
995                 BUG_ON(prot_sdb == NULL);
996                 ivecs = blk_rq_count_integrity_sg(rq->q, rq->bio);
997
998                 if (scsi_alloc_sgtable(prot_sdb, ivecs, gfp_mask)) {
999                         error = BLKPREP_DEFER;
1000                         goto err_exit;
1001                 }
1002
1003                 count = blk_rq_map_integrity_sg(rq->q, rq->bio,
1004                                                 prot_sdb->table.sgl);
1005                 BUG_ON(unlikely(count > ivecs));
1006                 BUG_ON(unlikely(count > queue_max_integrity_segments(rq->q)));
1007
1008                 cmd->prot_sdb = prot_sdb;
1009                 cmd->prot_sdb->table.nents = count;
1010         }
1011
1012         return BLKPREP_OK ;
1013
1014 err_exit:
1015         scsi_release_buffers(cmd);
1016         cmd->request->special = NULL;
1017         scsi_put_command(cmd);
1018         return error;
1019 }
1020 EXPORT_SYMBOL(scsi_init_io);
1021
1022 static struct scsi_cmnd *scsi_get_cmd_from_req(struct scsi_device *sdev,
1023                 struct request *req)
1024 {
1025         struct scsi_cmnd *cmd;
1026
1027         if (!req->special) {
1028                 cmd = scsi_get_command(sdev, GFP_ATOMIC);
1029                 if (unlikely(!cmd))
1030                         return NULL;
1031                 req->special = cmd;
1032         } else {
1033                 cmd = req->special;
1034         }
1035
1036         /* pull a tag out of the request if we have one */
1037         cmd->tag = req->tag;
1038         cmd->request = req;
1039
1040         cmd->cmnd = req->cmd;
1041
1042         return cmd;
1043 }
1044
1045 int scsi_setup_blk_pc_cmnd(struct scsi_device *sdev, struct request *req)
1046 {
1047         struct scsi_cmnd *cmd;
1048         int ret = scsi_prep_state_check(sdev, req);
1049
1050         if (ret != BLKPREP_OK)
1051                 return ret;
1052
1053         cmd = scsi_get_cmd_from_req(sdev, req);
1054         if (unlikely(!cmd))
1055                 return BLKPREP_DEFER;
1056
1057         /*
1058          * BLOCK_PC requests may transfer data, in which case they must
1059          * a bio attached to them.  Or they might contain a SCSI command
1060          * that does not transfer data, in which case they may optionally
1061          * submit a request without an attached bio.
1062          */
1063         if (req->bio) {
1064                 int ret;
1065
1066                 BUG_ON(!req->nr_phys_segments);
1067
1068                 ret = scsi_init_io(cmd, GFP_ATOMIC);
1069                 if (unlikely(ret))
1070                         return ret;
1071         } else {
1072                 BUG_ON(blk_rq_bytes(req));
1073
1074                 memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1075                 req->buffer = NULL;
1076         }
1077
1078         cmd->cmd_len = req->cmd_len;
1079         if (!blk_rq_bytes(req))
1080                 cmd->sc_data_direction = DMA_NONE;
1081         else if (rq_data_dir(req) == WRITE)
1082                 cmd->sc_data_direction = DMA_TO_DEVICE;
1083         else
1084                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1085         
1086         cmd->transfersize = blk_rq_bytes(req);
1087         cmd->allowed = req->retries;
1088         return BLKPREP_OK;
1089 }
1090 EXPORT_SYMBOL(scsi_setup_blk_pc_cmnd);
1091
1092 /*
1093  * Setup a REQ_TYPE_FS command.  These are simple read/write request
1094  * from filesystems that still need to be translated to SCSI CDBs from
1095  * the ULD.
1096  */
1097 int scsi_setup_fs_cmnd(struct scsi_device *sdev, struct request *req)
1098 {
1099         struct scsi_cmnd *cmd;
1100         int ret = scsi_prep_state_check(sdev, req);
1101
1102         if (ret != BLKPREP_OK)
1103                 return ret;
1104
1105         if (unlikely(sdev->scsi_dh_data && sdev->scsi_dh_data->scsi_dh
1106                          && sdev->scsi_dh_data->scsi_dh->prep_fn)) {
1107                 ret = sdev->scsi_dh_data->scsi_dh->prep_fn(sdev, req);
1108                 if (ret != BLKPREP_OK)
1109                         return ret;
1110         }
1111
1112         /*
1113          * Filesystem requests must transfer data.
1114          */
1115         BUG_ON(!req->nr_phys_segments);
1116
1117         cmd = scsi_get_cmd_from_req(sdev, req);
1118         if (unlikely(!cmd))
1119                 return BLKPREP_DEFER;
1120
1121         memset(cmd->cmnd, 0, BLK_MAX_CDB);
1122         return scsi_init_io(cmd, GFP_ATOMIC);
1123 }
1124 EXPORT_SYMBOL(scsi_setup_fs_cmnd);
1125
1126 int scsi_prep_state_check(struct scsi_device *sdev, struct request *req)
1127 {
1128         int ret = BLKPREP_OK;
1129
1130         /*
1131          * If the device is not in running state we will reject some
1132          * or all commands.
1133          */
1134         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1135                 switch (sdev->sdev_state) {
1136                 case SDEV_OFFLINE:
1137                         /*
1138                          * If the device is offline we refuse to process any
1139                          * commands.  The device must be brought online
1140                          * before trying any recovery commands.
1141                          */
1142                         sdev_printk(KERN_ERR, sdev,
1143                                     "rejecting I/O to offline device\n");
1144                         ret = BLKPREP_KILL;
1145                         break;
1146                 case SDEV_DEL:
1147                         /*
1148                          * If the device is fully deleted, we refuse to
1149                          * process any commands as well.
1150                          */
1151                         sdev_printk(KERN_ERR, sdev,
1152                                     "rejecting I/O to dead device\n");
1153                         ret = BLKPREP_KILL;
1154                         break;
1155                 case SDEV_QUIESCE:
1156                 case SDEV_BLOCK:
1157                 case SDEV_CREATED_BLOCK:
1158                         /*
1159                          * If the devices is blocked we defer normal commands.
1160                          */
1161                         if (!(req->cmd_flags & REQ_PREEMPT))
1162                                 ret = BLKPREP_DEFER;
1163                         break;
1164                 default:
1165                         /*
1166                          * For any other not fully online state we only allow
1167                          * special commands.  In particular any user initiated
1168                          * command is not allowed.
1169                          */
1170                         if (!(req->cmd_flags & REQ_PREEMPT))
1171                                 ret = BLKPREP_KILL;
1172                         break;
1173                 }
1174         }
1175         return ret;
1176 }
1177 EXPORT_SYMBOL(scsi_prep_state_check);
1178
1179 int scsi_prep_return(struct request_queue *q, struct request *req, int ret)
1180 {
1181         struct scsi_device *sdev = q->queuedata;
1182
1183         switch (ret) {
1184         case BLKPREP_KILL:
1185                 req->errors = DID_NO_CONNECT << 16;
1186                 /* release the command and kill it */
1187                 if (req->special) {
1188                         struct scsi_cmnd *cmd = req->special;
1189                         scsi_release_buffers(cmd);
1190                         scsi_put_command(cmd);
1191                         req->special = NULL;
1192                 }
1193                 break;
1194         case BLKPREP_DEFER:
1195                 /*
1196                  * If we defer, the blk_peek_request() returns NULL, but the
1197                  * queue must be restarted, so we plug here if no returning
1198                  * command will automatically do that.
1199                  */
1200                 if (sdev->device_busy == 0)
1201                         blk_plug_device(q);
1202                 break;
1203         default:
1204                 req->cmd_flags |= REQ_DONTPREP;
1205         }
1206
1207         return ret;
1208 }
1209 EXPORT_SYMBOL(scsi_prep_return);
1210
1211 int scsi_prep_fn(struct request_queue *q, struct request *req)
1212 {
1213         struct scsi_device *sdev = q->queuedata;
1214         int ret = BLKPREP_KILL;
1215
1216         if (req->cmd_type == REQ_TYPE_BLOCK_PC)
1217                 ret = scsi_setup_blk_pc_cmnd(sdev, req);
1218         return scsi_prep_return(q, req, ret);
1219 }
1220 EXPORT_SYMBOL(scsi_prep_fn);
1221
1222 /*
1223  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1224  * return 0.
1225  *
1226  * Called with the queue_lock held.
1227  */
1228 static inline int scsi_dev_queue_ready(struct request_queue *q,
1229                                   struct scsi_device *sdev)
1230 {
1231         if (sdev->device_busy == 0 && sdev->device_blocked) {
1232                 /*
1233                  * unblock after device_blocked iterates to zero
1234                  */
1235                 if (--sdev->device_blocked == 0) {
1236                         SCSI_LOG_MLQUEUE(3,
1237                                    sdev_printk(KERN_INFO, sdev,
1238                                    "unblocking device at zero depth\n"));
1239                 } else {
1240                         blk_plug_device(q);
1241                         return 0;
1242                 }
1243         }
1244         if (scsi_device_is_busy(sdev))
1245                 return 0;
1246
1247         return 1;
1248 }
1249
1250
1251 /*
1252  * scsi_target_queue_ready: checks if there we can send commands to target
1253  * @sdev: scsi device on starget to check.
1254  *
1255  * Called with the host lock held.
1256  */
1257 static inline int scsi_target_queue_ready(struct Scsi_Host *shost,
1258                                            struct scsi_device *sdev)
1259 {
1260         struct scsi_target *starget = scsi_target(sdev);
1261
1262         if (starget->single_lun) {
1263                 if (starget->starget_sdev_user &&
1264                     starget->starget_sdev_user != sdev)
1265                         return 0;
1266                 starget->starget_sdev_user = sdev;
1267         }
1268
1269         if (starget->target_busy == 0 && starget->target_blocked) {
1270                 /*
1271                  * unblock after target_blocked iterates to zero
1272                  */
1273                 if (--starget->target_blocked == 0) {
1274                         SCSI_LOG_MLQUEUE(3, starget_printk(KERN_INFO, starget,
1275                                          "unblocking target at zero depth\n"));
1276                 } else
1277                         return 0;
1278         }
1279
1280         if (scsi_target_is_busy(starget)) {
1281                 if (list_empty(&sdev->starved_entry))
1282                         list_add_tail(&sdev->starved_entry,
1283                                       &shost->starved_list);
1284                 return 0;
1285         }
1286
1287         /* We're OK to process the command, so we can't be starved */
1288         if (!list_empty(&sdev->starved_entry))
1289                 list_del_init(&sdev->starved_entry);
1290         return 1;
1291 }
1292
1293 /*
1294  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1295  * return 0. We must end up running the queue again whenever 0 is
1296  * returned, else IO can hang.
1297  *
1298  * Called with host_lock held.
1299  */
1300 static inline int scsi_host_queue_ready(struct request_queue *q,
1301                                    struct Scsi_Host *shost,
1302                                    struct scsi_device *sdev)
1303 {
1304         if (scsi_host_in_recovery(shost))
1305                 return 0;
1306         if (shost->host_busy == 0 && shost->host_blocked) {
1307                 /*
1308                  * unblock after host_blocked iterates to zero
1309                  */
1310                 if (--shost->host_blocked == 0) {
1311                         SCSI_LOG_MLQUEUE(3,
1312                                 printk("scsi%d unblocking host at zero depth\n",
1313                                         shost->host_no));
1314                 } else {
1315                         return 0;
1316                 }
1317         }
1318         if (scsi_host_is_busy(shost)) {
1319                 if (list_empty(&sdev->starved_entry))
1320                         list_add_tail(&sdev->starved_entry, &shost->starved_list);
1321                 return 0;
1322         }
1323
1324         /* We're OK to process the command, so we can't be starved */
1325         if (!list_empty(&sdev->starved_entry))
1326                 list_del_init(&sdev->starved_entry);
1327
1328         return 1;
1329 }
1330
1331 /*
1332  * Busy state exporting function for request stacking drivers.
1333  *
1334  * For efficiency, no lock is taken to check the busy state of
1335  * shost/starget/sdev, since the returned value is not guaranteed and
1336  * may be changed after request stacking drivers call the function,
1337  * regardless of taking lock or not.
1338  *
1339  * When scsi can't dispatch I/Os anymore and needs to kill I/Os
1340  * (e.g. !sdev), scsi needs to return 'not busy'.
1341  * Otherwise, request stacking drivers may hold requests forever.
1342  */
1343 static int scsi_lld_busy(struct request_queue *q)
1344 {
1345         struct scsi_device *sdev = q->queuedata;
1346         struct Scsi_Host *shost;
1347         struct scsi_target *starget;
1348
1349         if (!sdev)
1350                 return 0;
1351
1352         shost = sdev->host;
1353         starget = scsi_target(sdev);
1354
1355         if (scsi_host_in_recovery(shost) || scsi_host_is_busy(shost) ||
1356             scsi_target_is_busy(starget) || scsi_device_is_busy(sdev))
1357                 return 1;
1358
1359         return 0;
1360 }
1361
1362 /*
1363  * Kill a request for a dead device
1364  */
1365 static void scsi_kill_request(struct request *req, struct request_queue *q)
1366 {
1367         struct scsi_cmnd *cmd = req->special;
1368         struct scsi_device *sdev;
1369         struct scsi_target *starget;
1370         struct Scsi_Host *shost;
1371
1372         blk_start_request(req);
1373
1374         sdev = cmd->device;
1375         starget = scsi_target(sdev);
1376         shost = sdev->host;
1377         scsi_init_cmd_errh(cmd);
1378         cmd->result = DID_NO_CONNECT << 16;
1379         atomic_inc(&cmd->device->iorequest_cnt);
1380
1381         /*
1382          * SCSI request completion path will do scsi_device_unbusy(),
1383          * bump busy counts.  To bump the counters, we need to dance
1384          * with the locks as normal issue path does.
1385          */
1386         sdev->device_busy++;
1387         spin_unlock(sdev->request_queue->queue_lock);
1388         spin_lock(shost->host_lock);
1389         shost->host_busy++;
1390         starget->target_busy++;
1391         spin_unlock(shost->host_lock);
1392         spin_lock(sdev->request_queue->queue_lock);
1393
1394         blk_complete_request(req);
1395 }
1396
1397 static void scsi_softirq_done(struct request *rq)
1398 {
1399         struct scsi_cmnd *cmd = rq->special;
1400         unsigned long wait_for = (cmd->allowed + 1) * rq->timeout;
1401         int disposition;
1402
1403         INIT_LIST_HEAD(&cmd->eh_entry);
1404
1405         atomic_inc(&cmd->device->iodone_cnt);
1406         if (cmd->result)
1407                 atomic_inc(&cmd->device->ioerr_cnt);
1408
1409         disposition = scsi_decide_disposition(cmd);
1410         if (disposition != SUCCESS &&
1411             time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1412                 sdev_printk(KERN_ERR, cmd->device,
1413                             "timing out command, waited %lus\n",
1414                             wait_for/HZ);
1415                 disposition = SUCCESS;
1416         }
1417                         
1418         scsi_log_completion(cmd, disposition);
1419
1420         switch (disposition) {
1421                 case SUCCESS:
1422                         scsi_finish_command(cmd);
1423                         break;
1424                 case NEEDS_RETRY:
1425                         scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1426                         break;
1427                 case ADD_TO_MLQUEUE:
1428                         scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1429                         break;
1430                 default:
1431                         if (!scsi_eh_scmd_add(cmd, 0))
1432                                 scsi_finish_command(cmd);
1433         }
1434 }
1435
1436 /*
1437  * Function:    scsi_request_fn()
1438  *
1439  * Purpose:     Main strategy routine for SCSI.
1440  *
1441  * Arguments:   q       - Pointer to actual queue.
1442  *
1443  * Returns:     Nothing
1444  *
1445  * Lock status: IO request lock assumed to be held when called.
1446  */
1447 static void scsi_request_fn(struct request_queue *q)
1448 {
1449         struct scsi_device *sdev = q->queuedata;
1450         struct Scsi_Host *shost;
1451         struct scsi_cmnd *cmd;
1452         struct request *req;
1453
1454         if (!sdev) {
1455                 printk("scsi: killing requests for dead queue\n");
1456                 while ((req = blk_peek_request(q)) != NULL)
1457                         scsi_kill_request(req, q);
1458                 return;
1459         }
1460
1461         if(!get_device(&sdev->sdev_gendev))
1462                 /* We must be tearing the block queue down already */
1463                 return;
1464
1465         /*
1466          * To start with, we keep looping until the queue is empty, or until
1467          * the host is no longer able to accept any more requests.
1468          */
1469         shost = sdev->host;
1470         while (!blk_queue_plugged(q)) {
1471                 int rtn;
1472                 /*
1473                  * get next queueable request.  We do this early to make sure
1474                  * that the request is fully prepared even if we cannot 
1475                  * accept it.
1476                  */
1477                 req = blk_peek_request(q);
1478                 if (!req || !scsi_dev_queue_ready(q, sdev))
1479                         break;
1480
1481                 if (unlikely(!scsi_device_online(sdev))) {
1482                         sdev_printk(KERN_ERR, sdev,
1483                                     "rejecting I/O to offline device\n");
1484                         scsi_kill_request(req, q);
1485                         continue;
1486                 }
1487
1488
1489                 /*
1490                  * Remove the request from the request list.
1491                  */
1492                 if (!(blk_queue_tagged(q) && !blk_queue_start_tag(q, req)))
1493                         blk_start_request(req);
1494                 sdev->device_busy++;
1495
1496                 spin_unlock(q->queue_lock);
1497                 cmd = req->special;
1498                 if (unlikely(cmd == NULL)) {
1499                         printk(KERN_CRIT "impossible request in %s.\n"
1500                                          "please mail a stack trace to "
1501                                          "linux-scsi@vger.kernel.org\n",
1502                                          __func__);
1503                         blk_dump_rq_flags(req, "foo");
1504                         BUG();
1505                 }
1506                 spin_lock(shost->host_lock);
1507
1508                 /*
1509                  * We hit this when the driver is using a host wide
1510                  * tag map. For device level tag maps the queue_depth check
1511                  * in the device ready fn would prevent us from trying
1512                  * to allocate a tag. Since the map is a shared host resource
1513                  * we add the dev to the starved list so it eventually gets
1514                  * a run when a tag is freed.
1515                  */
1516                 if (blk_queue_tagged(q) && !blk_rq_tagged(req)) {
1517                         if (list_empty(&sdev->starved_entry))
1518                                 list_add_tail(&sdev->starved_entry,
1519                                               &shost->starved_list);
1520                         goto not_ready;
1521                 }
1522
1523                 if (!scsi_target_queue_ready(shost, sdev))
1524                         goto not_ready;
1525
1526                 if (!scsi_host_queue_ready(q, shost, sdev))
1527                         goto not_ready;
1528
1529                 scsi_target(sdev)->target_busy++;
1530                 shost->host_busy++;
1531
1532                 /*
1533                  * XXX(hch): This is rather suboptimal, scsi_dispatch_cmd will
1534                  *              take the lock again.
1535                  */
1536                 spin_unlock_irq(shost->host_lock);
1537
1538                 /*
1539                  * Finally, initialize any error handling parameters, and set up
1540                  * the timers for timeouts.
1541                  */
1542                 scsi_init_cmd_errh(cmd);
1543
1544                 /*
1545                  * Dispatch the command to the low-level driver.
1546                  */
1547                 rtn = scsi_dispatch_cmd(cmd);
1548                 spin_lock_irq(q->queue_lock);
1549                 if(rtn) {
1550                         /* we're refusing the command; because of
1551                          * the way locks get dropped, we need to 
1552                          * check here if plugging is required */
1553                         if(sdev->device_busy == 0)
1554                                 blk_plug_device(q);
1555
1556                         break;
1557                 }
1558         }
1559
1560         goto out;
1561
1562  not_ready:
1563         spin_unlock_irq(shost->host_lock);
1564
1565         /*
1566          * lock q, handle tag, requeue req, and decrement device_busy. We
1567          * must return with queue_lock held.
1568          *
1569          * Decrementing device_busy without checking it is OK, as all such
1570          * cases (host limits or settings) should run the queue at some
1571          * later time.
1572          */
1573         spin_lock_irq(q->queue_lock);
1574         blk_requeue_request(q, req);
1575         sdev->device_busy--;
1576         if(sdev->device_busy == 0)
1577                 blk_plug_device(q);
1578  out:
1579         /* must be careful here...if we trigger the ->remove() function
1580          * we cannot be holding the q lock */
1581         spin_unlock_irq(q->queue_lock);
1582         put_device(&sdev->sdev_gendev);
1583         spin_lock_irq(q->queue_lock);
1584 }
1585
1586 u64 scsi_calculate_bounce_limit(struct Scsi_Host *shost)
1587 {
1588         struct device *host_dev;
1589         u64 bounce_limit = 0xffffffff;
1590
1591         if (shost->unchecked_isa_dma)
1592                 return BLK_BOUNCE_ISA;
1593         /*
1594          * Platforms with virtual-DMA translation
1595          * hardware have no practical limit.
1596          */
1597         if (!PCI_DMA_BUS_IS_PHYS)
1598                 return BLK_BOUNCE_ANY;
1599
1600         host_dev = scsi_get_device(shost);
1601         if (host_dev && host_dev->dma_mask)
1602                 bounce_limit = *host_dev->dma_mask;
1603
1604         return bounce_limit;
1605 }
1606 EXPORT_SYMBOL(scsi_calculate_bounce_limit);
1607
1608 struct request_queue *__scsi_alloc_queue(struct Scsi_Host *shost,
1609                                          request_fn_proc *request_fn)
1610 {
1611         struct request_queue *q;
1612         struct device *dev = shost->shost_gendev.parent;
1613
1614         q = blk_init_queue(request_fn, NULL);
1615         if (!q)
1616                 return NULL;
1617
1618         /*
1619          * this limit is imposed by hardware restrictions
1620          */
1621         blk_queue_max_segments(q, min_t(unsigned short, shost->sg_tablesize,
1622                                         SCSI_MAX_SG_CHAIN_SEGMENTS));
1623
1624         if (scsi_host_prot_dma(shost)) {
1625                 shost->sg_prot_tablesize =
1626                         min_not_zero(shost->sg_prot_tablesize,
1627                                      (unsigned short)SCSI_MAX_PROT_SG_SEGMENTS);
1628                 BUG_ON(shost->sg_prot_tablesize < shost->sg_tablesize);
1629                 blk_queue_max_integrity_segments(q, shost->sg_prot_tablesize);
1630         }
1631
1632         blk_queue_max_hw_sectors(q, shost->max_sectors);
1633         blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
1634         blk_queue_segment_boundary(q, shost->dma_boundary);
1635         dma_set_seg_boundary(dev, shost->dma_boundary);
1636
1637         blk_queue_max_segment_size(q, dma_get_max_seg_size(dev));
1638
1639         if (!shost->use_clustering)
1640                 q->limits.cluster = 0;
1641
1642         /*
1643          * set a reasonable default alignment on word boundaries: the
1644          * host and device may alter it using
1645          * blk_queue_update_dma_alignment() later.
1646          */
1647         blk_queue_dma_alignment(q, 0x03);
1648
1649         return q;
1650 }
1651 EXPORT_SYMBOL(__scsi_alloc_queue);
1652
1653 struct request_queue *scsi_alloc_queue(struct scsi_device *sdev)
1654 {
1655         struct request_queue *q;
1656
1657         q = __scsi_alloc_queue(sdev->host, scsi_request_fn);
1658         if (!q)
1659                 return NULL;
1660
1661         blk_queue_prep_rq(q, scsi_prep_fn);
1662         blk_queue_softirq_done(q, scsi_softirq_done);
1663         blk_queue_rq_timed_out(q, scsi_times_out);
1664         blk_queue_lld_busy(q, scsi_lld_busy);
1665         return q;
1666 }
1667
1668 void scsi_free_queue(struct request_queue *q)
1669 {
1670         blk_cleanup_queue(q);
1671 }
1672
1673 /*
1674  * Function:    scsi_block_requests()
1675  *
1676  * Purpose:     Utility function used by low-level drivers to prevent further
1677  *              commands from being queued to the device.
1678  *
1679  * Arguments:   shost       - Host in question
1680  *
1681  * Returns:     Nothing
1682  *
1683  * Lock status: No locks are assumed held.
1684  *
1685  * Notes:       There is no timer nor any other means by which the requests
1686  *              get unblocked other than the low-level driver calling
1687  *              scsi_unblock_requests().
1688  */
1689 void scsi_block_requests(struct Scsi_Host *shost)
1690 {
1691         shost->host_self_blocked = 1;
1692 }
1693 EXPORT_SYMBOL(scsi_block_requests);
1694
1695 /*
1696  * Function:    scsi_unblock_requests()
1697  *
1698  * Purpose:     Utility function used by low-level drivers to allow further
1699  *              commands from being queued to the device.
1700  *
1701  * Arguments:   shost       - Host in question
1702  *
1703  * Returns:     Nothing
1704  *
1705  * Lock status: No locks are assumed held.
1706  *
1707  * Notes:       There is no timer nor any other means by which the requests
1708  *              get unblocked other than the low-level driver calling
1709  *              scsi_unblock_requests().
1710  *
1711  *              This is done as an API function so that changes to the
1712  *              internals of the scsi mid-layer won't require wholesale
1713  *              changes to drivers that use this feature.
1714  */
1715 void scsi_unblock_requests(struct Scsi_Host *shost)
1716 {
1717         shost->host_self_blocked = 0;
1718         scsi_run_host_queues(shost);
1719 }
1720 EXPORT_SYMBOL(scsi_unblock_requests);
1721
1722 int __init scsi_init_queue(void)
1723 {
1724         int i;
1725
1726         scsi_sdb_cache = kmem_cache_create("scsi_data_buffer",
1727                                            sizeof(struct scsi_data_buffer),
1728                                            0, 0, NULL);
1729         if (!scsi_sdb_cache) {
1730                 printk(KERN_ERR "SCSI: can't init scsi sdb cache\n");
1731                 return -ENOMEM;
1732         }
1733
1734         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1735                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1736                 int size = sgp->size * sizeof(struct scatterlist);
1737
1738                 sgp->slab = kmem_cache_create(sgp->name, size, 0,
1739                                 SLAB_HWCACHE_ALIGN, NULL);
1740                 if (!sgp->slab) {
1741                         printk(KERN_ERR "SCSI: can't init sg slab %s\n",
1742                                         sgp->name);
1743                         goto cleanup_sdb;
1744                 }
1745
1746                 sgp->pool = mempool_create_slab_pool(SG_MEMPOOL_SIZE,
1747                                                      sgp->slab);
1748                 if (!sgp->pool) {
1749                         printk(KERN_ERR "SCSI: can't init sg mempool %s\n",
1750                                         sgp->name);
1751                         goto cleanup_sdb;
1752                 }
1753         }
1754
1755         return 0;
1756
1757 cleanup_sdb:
1758         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1759                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1760                 if (sgp->pool)
1761                         mempool_destroy(sgp->pool);
1762                 if (sgp->slab)
1763                         kmem_cache_destroy(sgp->slab);
1764         }
1765         kmem_cache_destroy(scsi_sdb_cache);
1766
1767         return -ENOMEM;
1768 }
1769
1770 void scsi_exit_queue(void)
1771 {
1772         int i;
1773
1774         kmem_cache_destroy(scsi_sdb_cache);
1775
1776         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1777                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1778                 mempool_destroy(sgp->pool);
1779                 kmem_cache_destroy(sgp->slab);
1780         }
1781 }
1782
1783 /**
1784  *      scsi_mode_select - issue a mode select
1785  *      @sdev:  SCSI device to be queried
1786  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
1787  *      @sp:    Save page bit (0 == don't save, 1 == save)
1788  *      @modepage: mode page being requested
1789  *      @buffer: request buffer (may not be smaller than eight bytes)
1790  *      @len:   length of request buffer.
1791  *      @timeout: command timeout
1792  *      @retries: number of retries before failing
1793  *      @data: returns a structure abstracting the mode header data
1794  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
1795  *              must be SCSI_SENSE_BUFFERSIZE big.
1796  *
1797  *      Returns zero if successful; negative error number or scsi
1798  *      status on error
1799  *
1800  */
1801 int
1802 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
1803                  unsigned char *buffer, int len, int timeout, int retries,
1804                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1805 {
1806         unsigned char cmd[10];
1807         unsigned char *real_buffer;
1808         int ret;
1809
1810         memset(cmd, 0, sizeof(cmd));
1811         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
1812
1813         if (sdev->use_10_for_ms) {
1814                 if (len > 65535)
1815                         return -EINVAL;
1816                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
1817                 if (!real_buffer)
1818                         return -ENOMEM;
1819                 memcpy(real_buffer + 8, buffer, len);
1820                 len += 8;
1821                 real_buffer[0] = 0;
1822                 real_buffer[1] = 0;
1823                 real_buffer[2] = data->medium_type;
1824                 real_buffer[3] = data->device_specific;
1825                 real_buffer[4] = data->longlba ? 0x01 : 0;
1826                 real_buffer[5] = 0;
1827                 real_buffer[6] = data->block_descriptor_length >> 8;
1828                 real_buffer[7] = data->block_descriptor_length;
1829
1830                 cmd[0] = MODE_SELECT_10;
1831                 cmd[7] = len >> 8;
1832                 cmd[8] = len;
1833         } else {
1834                 if (len > 255 || data->block_descriptor_length > 255 ||
1835                     data->longlba)
1836                         return -EINVAL;
1837
1838                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
1839                 if (!real_buffer)
1840                         return -ENOMEM;
1841                 memcpy(real_buffer + 4, buffer, len);
1842                 len += 4;
1843                 real_buffer[0] = 0;
1844                 real_buffer[1] = data->medium_type;
1845                 real_buffer[2] = data->device_specific;
1846                 real_buffer[3] = data->block_descriptor_length;
1847                 
1848
1849                 cmd[0] = MODE_SELECT;
1850                 cmd[4] = len;
1851         }
1852
1853         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
1854                                sshdr, timeout, retries, NULL);
1855         kfree(real_buffer);
1856         return ret;
1857 }
1858 EXPORT_SYMBOL_GPL(scsi_mode_select);
1859
1860 /**
1861  *      scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
1862  *      @sdev:  SCSI device to be queried
1863  *      @dbd:   set if mode sense will allow block descriptors to be returned
1864  *      @modepage: mode page being requested
1865  *      @buffer: request buffer (may not be smaller than eight bytes)
1866  *      @len:   length of request buffer.
1867  *      @timeout: command timeout
1868  *      @retries: number of retries before failing
1869  *      @data: returns a structure abstracting the mode header data
1870  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
1871  *              must be SCSI_SENSE_BUFFERSIZE big.
1872  *
1873  *      Returns zero if unsuccessful, or the header offset (either 4
1874  *      or 8 depending on whether a six or ten byte command was
1875  *      issued) if successful.
1876  */
1877 int
1878 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
1879                   unsigned char *buffer, int len, int timeout, int retries,
1880                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1881 {
1882         unsigned char cmd[12];
1883         int use_10_for_ms;
1884         int header_length;
1885         int result;
1886         struct scsi_sense_hdr my_sshdr;
1887
1888         memset(data, 0, sizeof(*data));
1889         memset(&cmd[0], 0, 12);
1890         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
1891         cmd[2] = modepage;
1892
1893         /* caller might not be interested in sense, but we need it */
1894         if (!sshdr)
1895                 sshdr = &my_sshdr;
1896
1897  retry:
1898         use_10_for_ms = sdev->use_10_for_ms;
1899
1900         if (use_10_for_ms) {
1901                 if (len < 8)
1902                         len = 8;
1903
1904                 cmd[0] = MODE_SENSE_10;
1905                 cmd[8] = len;
1906                 header_length = 8;
1907         } else {
1908                 if (len < 4)
1909                         len = 4;
1910
1911                 cmd[0] = MODE_SENSE;
1912                 cmd[4] = len;
1913                 header_length = 4;
1914         }
1915
1916         memset(buffer, 0, len);
1917
1918         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
1919                                   sshdr, timeout, retries, NULL);
1920
1921         /* This code looks awful: what it's doing is making sure an
1922          * ILLEGAL REQUEST sense return identifies the actual command
1923          * byte as the problem.  MODE_SENSE commands can return
1924          * ILLEGAL REQUEST if the code page isn't supported */
1925
1926         if (use_10_for_ms && !scsi_status_is_good(result) &&
1927             (driver_byte(result) & DRIVER_SENSE)) {
1928                 if (scsi_sense_valid(sshdr)) {
1929                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
1930                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
1931                                 /* 
1932                                  * Invalid command operation code
1933                                  */
1934                                 sdev->use_10_for_ms = 0;
1935                                 goto retry;
1936                         }
1937                 }
1938         }
1939
1940         if(scsi_status_is_good(result)) {
1941                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
1942                              (modepage == 6 || modepage == 8))) {
1943                         /* Initio breakage? */
1944                         header_length = 0;
1945                         data->length = 13;
1946                         data->medium_type = 0;
1947                         data->device_specific = 0;
1948                         data->longlba = 0;
1949                         data->block_descriptor_length = 0;
1950                 } else if(use_10_for_ms) {
1951                         data->length = buffer[0]*256 + buffer[1] + 2;
1952                         data->medium_type = buffer[2];
1953                         data->device_specific = buffer[3];
1954                         data->longlba = buffer[4] & 0x01;
1955                         data->block_descriptor_length = buffer[6]*256
1956                                 + buffer[7];
1957                 } else {
1958                         data->length = buffer[0] + 1;
1959                         data->medium_type = buffer[1];
1960                         data->device_specific = buffer[2];
1961                         data->block_descriptor_length = buffer[3];
1962                 }
1963                 data->header_length = header_length;
1964         }
1965
1966         return result;
1967 }
1968 EXPORT_SYMBOL(scsi_mode_sense);
1969
1970 /**
1971  *      scsi_test_unit_ready - test if unit is ready
1972  *      @sdev:  scsi device to change the state of.
1973  *      @timeout: command timeout
1974  *      @retries: number of retries before failing
1975  *      @sshdr_external: Optional pointer to struct scsi_sense_hdr for
1976  *              returning sense. Make sure that this is cleared before passing
1977  *              in.
1978  *
1979  *      Returns zero if unsuccessful or an error if TUR failed.  For
1980  *      removable media, UNIT_ATTENTION sets ->changed flag.
1981  **/
1982 int
1983 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
1984                      struct scsi_sense_hdr *sshdr_external)
1985 {
1986         char cmd[] = {
1987                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
1988         };
1989         struct scsi_sense_hdr *sshdr;
1990         int result;
1991
1992         if (!sshdr_external)
1993                 sshdr = kzalloc(sizeof(*sshdr), GFP_KERNEL);
1994         else
1995                 sshdr = sshdr_external;
1996
1997         /* try to eat the UNIT_ATTENTION if there are enough retries */
1998         do {
1999                 result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
2000                                           timeout, retries, NULL);
2001                 if (sdev->removable && scsi_sense_valid(sshdr) &&
2002                     sshdr->sense_key == UNIT_ATTENTION)
2003                         sdev->changed = 1;
2004         } while (scsi_sense_valid(sshdr) &&
2005                  sshdr->sense_key == UNIT_ATTENTION && --retries);
2006
2007         if (!sshdr_external)
2008                 kfree(sshdr);
2009         return result;
2010 }
2011 EXPORT_SYMBOL(scsi_test_unit_ready);
2012
2013 /**
2014  *      scsi_device_set_state - Take the given device through the device state model.
2015  *      @sdev:  scsi device to change the state of.
2016  *      @state: state to change to.
2017  *
2018  *      Returns zero if unsuccessful or an error if the requested 
2019  *      transition is illegal.
2020  */
2021 int
2022 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2023 {
2024         enum scsi_device_state oldstate = sdev->sdev_state;
2025
2026         if (state == oldstate)
2027                 return 0;
2028
2029         switch (state) {
2030         case SDEV_CREATED:
2031                 switch (oldstate) {
2032                 case SDEV_CREATED_BLOCK:
2033                         break;
2034                 default:
2035                         goto illegal;
2036                 }
2037                 break;
2038                         
2039         case SDEV_RUNNING:
2040                 switch (oldstate) {
2041                 case SDEV_CREATED:
2042                 case SDEV_OFFLINE:
2043                 case SDEV_QUIESCE:
2044                 case SDEV_BLOCK:
2045                         break;
2046                 default:
2047                         goto illegal;
2048                 }
2049                 break;
2050
2051         case SDEV_QUIESCE:
2052                 switch (oldstate) {
2053                 case SDEV_RUNNING:
2054                 case SDEV_OFFLINE:
2055                         break;
2056                 default:
2057                         goto illegal;
2058                 }
2059                 break;
2060
2061         case SDEV_OFFLINE:
2062                 switch (oldstate) {
2063                 case SDEV_CREATED:
2064                 case SDEV_RUNNING:
2065                 case SDEV_QUIESCE:
2066                 case SDEV_BLOCK:
2067                         break;
2068                 default:
2069                         goto illegal;
2070                 }
2071                 break;
2072
2073         case SDEV_BLOCK:
2074                 switch (oldstate) {
2075                 case SDEV_RUNNING:
2076                 case SDEV_CREATED_BLOCK:
2077                         break;
2078                 default:
2079                         goto illegal;
2080                 }
2081                 break;
2082
2083         case SDEV_CREATED_BLOCK:
2084                 switch (oldstate) {
2085                 case SDEV_CREATED:
2086                         break;
2087                 default:
2088                         goto illegal;
2089                 }
2090                 break;
2091
2092         case SDEV_CANCEL:
2093                 switch (oldstate) {
2094                 case SDEV_CREATED:
2095                 case SDEV_RUNNING:
2096                 case SDEV_QUIESCE:
2097                 case SDEV_OFFLINE:
2098                 case SDEV_BLOCK:
2099                         break;
2100                 default:
2101                         goto illegal;
2102                 }
2103                 break;
2104
2105         case SDEV_DEL:
2106                 switch (oldstate) {
2107                 case SDEV_CREATED:
2108                 case SDEV_RUNNING:
2109                 case SDEV_OFFLINE:
2110                 case SDEV_CANCEL:
2111                         break;
2112                 default:
2113                         goto illegal;
2114                 }
2115                 break;
2116
2117         }
2118         sdev->sdev_state = state;
2119         return 0;
2120
2121  illegal:
2122         SCSI_LOG_ERROR_RECOVERY(1, 
2123                                 sdev_printk(KERN_ERR, sdev,
2124                                             "Illegal state transition %s->%s\n",
2125                                             scsi_device_state_name(oldstate),
2126                                             scsi_device_state_name(state))
2127                                 );
2128         return -EINVAL;
2129 }
2130 EXPORT_SYMBOL(scsi_device_set_state);
2131
2132 /**
2133  *      sdev_evt_emit - emit a single SCSI device uevent
2134  *      @sdev: associated SCSI device
2135  *      @evt: event to emit
2136  *
2137  *      Send a single uevent (scsi_event) to the associated scsi_device.
2138  */
2139 static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2140 {
2141         int idx = 0;
2142         char *envp[3];
2143
2144         switch (evt->evt_type) {
2145         case SDEV_EVT_MEDIA_CHANGE:
2146                 envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2147                 break;
2148
2149         default:
2150                 /* do nothing */
2151                 break;
2152         }
2153
2154         envp[idx++] = NULL;
2155
2156         kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2157 }
2158
2159 /**
2160  *      sdev_evt_thread - send a uevent for each scsi event
2161  *      @work: work struct for scsi_device
2162  *
2163  *      Dispatch queued events to their associated scsi_device kobjects
2164  *      as uevents.
2165  */
2166 void scsi_evt_thread(struct work_struct *work)
2167 {
2168         struct scsi_device *sdev;
2169         LIST_HEAD(event_list);
2170
2171         sdev = container_of(work, struct scsi_device, event_work);
2172
2173         while (1) {
2174                 struct scsi_event *evt;
2175                 struct list_head *this, *tmp;
2176                 unsigned long flags;
2177
2178                 spin_lock_irqsave(&sdev->list_lock, flags);
2179                 list_splice_init(&sdev->event_list, &event_list);
2180                 spin_unlock_irqrestore(&sdev->list_lock, flags);
2181
2182                 if (list_empty(&event_list))
2183                         break;
2184
2185                 list_for_each_safe(this, tmp, &event_list) {
2186                         evt = list_entry(this, struct scsi_event, node);
2187                         list_del(&evt->node);
2188                         scsi_evt_emit(sdev, evt);
2189                         kfree(evt);
2190                 }
2191         }
2192 }
2193
2194 /**
2195  *      sdev_evt_send - send asserted event to uevent thread
2196  *      @sdev: scsi_device event occurred on
2197  *      @evt: event to send
2198  *
2199  *      Assert scsi device event asynchronously.
2200  */
2201 void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2202 {
2203         unsigned long flags;
2204
2205 #if 0
2206         /* FIXME: currently this check eliminates all media change events
2207          * for polled devices.  Need to update to discriminate between AN
2208          * and polled events */
2209         if (!test_bit(evt->evt_type, sdev->supported_events)) {
2210                 kfree(evt);
2211                 return;
2212         }
2213 #endif
2214
2215         spin_lock_irqsave(&sdev->list_lock, flags);
2216         list_add_tail(&evt->node, &sdev->event_list);
2217         schedule_work(&sdev->event_work);
2218         spin_unlock_irqrestore(&sdev->list_lock, flags);
2219 }
2220 EXPORT_SYMBOL_GPL(sdev_evt_send);
2221
2222 /**
2223  *      sdev_evt_alloc - allocate a new scsi event
2224  *      @evt_type: type of event to allocate
2225  *      @gfpflags: GFP flags for allocation
2226  *
2227  *      Allocates and returns a new scsi_event.
2228  */
2229 struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2230                                   gfp_t gfpflags)
2231 {
2232         struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2233         if (!evt)
2234                 return NULL;
2235
2236         evt->evt_type = evt_type;
2237         INIT_LIST_HEAD(&evt->node);
2238
2239         /* evt_type-specific initialization, if any */
2240         switch (evt_type) {
2241         case SDEV_EVT_MEDIA_CHANGE:
2242         default:
2243                 /* do nothing */
2244                 break;
2245         }
2246
2247         return evt;
2248 }
2249 EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2250
2251 /**
2252  *      sdev_evt_send_simple - send asserted event to uevent thread
2253  *      @sdev: scsi_device event occurred on
2254  *      @evt_type: type of event to send
2255  *      @gfpflags: GFP flags for allocation
2256  *
2257  *      Assert scsi device event asynchronously, given an event type.
2258  */
2259 void sdev_evt_send_simple(struct scsi_device *sdev,
2260                           enum scsi_device_event evt_type, gfp_t gfpflags)
2261 {
2262         struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2263         if (!evt) {
2264                 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2265                             evt_type);
2266                 return;
2267         }
2268
2269         sdev_evt_send(sdev, evt);
2270 }
2271 EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2272
2273 /**
2274  *      scsi_device_quiesce - Block user issued commands.
2275  *      @sdev:  scsi device to quiesce.
2276  *
2277  *      This works by trying to transition to the SDEV_QUIESCE state
2278  *      (which must be a legal transition).  When the device is in this
2279  *      state, only special requests will be accepted, all others will
2280  *      be deferred.  Since special requests may also be requeued requests,
2281  *      a successful return doesn't guarantee the device will be 
2282  *      totally quiescent.
2283  *
2284  *      Must be called with user context, may sleep.
2285  *
2286  *      Returns zero if unsuccessful or an error if not.
2287  */
2288 int
2289 scsi_device_quiesce(struct scsi_device *sdev)
2290 {
2291         int err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2292         if (err)
2293                 return err;
2294
2295         scsi_run_queue(sdev->request_queue);
2296         while (sdev->device_busy) {
2297                 msleep_interruptible(200);
2298                 scsi_run_queue(sdev->request_queue);
2299         }
2300         return 0;
2301 }
2302 EXPORT_SYMBOL(scsi_device_quiesce);
2303
2304 /**
2305  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2306  *      @sdev:  scsi device to resume.
2307  *
2308  *      Moves the device from quiesced back to running and restarts the
2309  *      queues.
2310  *
2311  *      Must be called with user context, may sleep.
2312  */
2313 void
2314 scsi_device_resume(struct scsi_device *sdev)
2315 {
2316         if(scsi_device_set_state(sdev, SDEV_RUNNING))
2317                 return;
2318         scsi_run_queue(sdev->request_queue);
2319 }
2320 EXPORT_SYMBOL(scsi_device_resume);
2321
2322 static void
2323 device_quiesce_fn(struct scsi_device *sdev, void *data)
2324 {
2325         scsi_device_quiesce(sdev);
2326 }
2327
2328 void
2329 scsi_target_quiesce(struct scsi_target *starget)
2330 {
2331         starget_for_each_device(starget, NULL, device_quiesce_fn);
2332 }
2333 EXPORT_SYMBOL(scsi_target_quiesce);
2334
2335 static void
2336 device_resume_fn(struct scsi_device *sdev, void *data)
2337 {
2338         scsi_device_resume(sdev);
2339 }
2340
2341 void
2342 scsi_target_resume(struct scsi_target *starget)
2343 {
2344         starget_for_each_device(starget, NULL, device_resume_fn);
2345 }
2346 EXPORT_SYMBOL(scsi_target_resume);
2347
2348 /**
2349  * scsi_internal_device_block - internal function to put a device temporarily into the SDEV_BLOCK state
2350  * @sdev:       device to block
2351  *
2352  * Block request made by scsi lld's to temporarily stop all
2353  * scsi commands on the specified device.  Called from interrupt
2354  * or normal process context.
2355  *
2356  * Returns zero if successful or error if not
2357  *
2358  * Notes:       
2359  *      This routine transitions the device to the SDEV_BLOCK state
2360  *      (which must be a legal transition).  When the device is in this
2361  *      state, all commands are deferred until the scsi lld reenables
2362  *      the device with scsi_device_unblock or device_block_tmo fires.
2363  *      This routine assumes the host_lock is held on entry.
2364  */
2365 int
2366 scsi_internal_device_block(struct scsi_device *sdev)
2367 {
2368         struct request_queue *q = sdev->request_queue;
2369         unsigned long flags;
2370         int err = 0;
2371
2372         err = scsi_device_set_state(sdev, SDEV_BLOCK);
2373         if (err) {
2374                 err = scsi_device_set_state(sdev, SDEV_CREATED_BLOCK);
2375
2376                 if (err)
2377                         return err;
2378         }
2379
2380         /* 
2381          * The device has transitioned to SDEV_BLOCK.  Stop the
2382          * block layer from calling the midlayer with this device's
2383          * request queue. 
2384          */
2385         spin_lock_irqsave(q->queue_lock, flags);
2386         blk_stop_queue(q);
2387         spin_unlock_irqrestore(q->queue_lock, flags);
2388
2389         return 0;
2390 }
2391 EXPORT_SYMBOL_GPL(scsi_internal_device_block);
2392  
2393 /**
2394  * scsi_internal_device_unblock - resume a device after a block request
2395  * @sdev:       device to resume
2396  *
2397  * Called by scsi lld's or the midlayer to restart the device queue
2398  * for the previously suspended scsi device.  Called from interrupt or
2399  * normal process context.
2400  *
2401  * Returns zero if successful or error if not.
2402  *
2403  * Notes:       
2404  *      This routine transitions the device to the SDEV_RUNNING state
2405  *      (which must be a legal transition) allowing the midlayer to
2406  *      goose the queue for this device.  This routine assumes the 
2407  *      host_lock is held upon entry.
2408  */
2409 int
2410 scsi_internal_device_unblock(struct scsi_device *sdev)
2411 {
2412         struct request_queue *q = sdev->request_queue; 
2413         unsigned long flags;
2414         
2415         /* 
2416          * Try to transition the scsi device to SDEV_RUNNING
2417          * and goose the device queue if successful.  
2418          */
2419         if (sdev->sdev_state == SDEV_BLOCK)
2420                 sdev->sdev_state = SDEV_RUNNING;
2421         else if (sdev->sdev_state == SDEV_CREATED_BLOCK)
2422                 sdev->sdev_state = SDEV_CREATED;
2423         else if (sdev->sdev_state != SDEV_CANCEL &&
2424                  sdev->sdev_state != SDEV_OFFLINE)
2425                 return -EINVAL;
2426
2427         spin_lock_irqsave(q->queue_lock, flags);
2428         blk_start_queue(q);
2429         spin_unlock_irqrestore(q->queue_lock, flags);
2430
2431         return 0;
2432 }
2433 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock);
2434
2435 static void
2436 device_block(struct scsi_device *sdev, void *data)
2437 {
2438         scsi_internal_device_block(sdev);
2439 }
2440
2441 static int
2442 target_block(struct device *dev, void *data)
2443 {
2444         if (scsi_is_target_device(dev))
2445                 starget_for_each_device(to_scsi_target(dev), NULL,
2446                                         device_block);
2447         return 0;
2448 }
2449
2450 void
2451 scsi_target_block(struct device *dev)
2452 {
2453         if (scsi_is_target_device(dev))
2454                 starget_for_each_device(to_scsi_target(dev), NULL,
2455                                         device_block);
2456         else
2457                 device_for_each_child(dev, NULL, target_block);
2458 }
2459 EXPORT_SYMBOL_GPL(scsi_target_block);
2460
2461 static void
2462 device_unblock(struct scsi_device *sdev, void *data)
2463 {
2464         scsi_internal_device_unblock(sdev);
2465 }
2466
2467 static int
2468 target_unblock(struct device *dev, void *data)
2469 {
2470         if (scsi_is_target_device(dev))
2471                 starget_for_each_device(to_scsi_target(dev), NULL,
2472                                         device_unblock);
2473         return 0;
2474 }
2475
2476 void
2477 scsi_target_unblock(struct device *dev)
2478 {
2479         if (scsi_is_target_device(dev))
2480                 starget_for_each_device(to_scsi_target(dev), NULL,
2481                                         device_unblock);
2482         else
2483                 device_for_each_child(dev, NULL, target_unblock);
2484 }
2485 EXPORT_SYMBOL_GPL(scsi_target_unblock);
2486
2487 /**
2488  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2489  * @sgl:        scatter-gather list
2490  * @sg_count:   number of segments in sg
2491  * @offset:     offset in bytes into sg, on return offset into the mapped area
2492  * @len:        bytes to map, on return number of bytes mapped
2493  *
2494  * Returns virtual address of the start of the mapped page
2495  */
2496 void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
2497                           size_t *offset, size_t *len)
2498 {
2499         int i;
2500         size_t sg_len = 0, len_complete = 0;
2501         struct scatterlist *sg;
2502         struct page *page;
2503
2504         WARN_ON(!irqs_disabled());
2505
2506         for_each_sg(sgl, sg, sg_count, i) {
2507                 len_complete = sg_len; /* Complete sg-entries */
2508                 sg_len += sg->length;
2509                 if (sg_len > *offset)
2510                         break;
2511         }
2512
2513         if (unlikely(i == sg_count)) {
2514                 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2515                         "elements %d\n",
2516                        __func__, sg_len, *offset, sg_count);
2517                 WARN_ON(1);
2518                 return NULL;
2519         }
2520
2521         /* Offset starting from the beginning of first page in this sg-entry */
2522         *offset = *offset - len_complete + sg->offset;
2523
2524         /* Assumption: contiguous pages can be accessed as "page + i" */
2525         page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
2526         *offset &= ~PAGE_MASK;
2527
2528         /* Bytes in this sg-entry from *offset to the end of the page */
2529         sg_len = PAGE_SIZE - *offset;
2530         if (*len > sg_len)
2531                 *len = sg_len;
2532
2533         return kmap_atomic(page, KM_BIO_SRC_IRQ);
2534 }
2535 EXPORT_SYMBOL(scsi_kmap_atomic_sg);
2536
2537 /**
2538  * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
2539  * @virt:       virtual address to be unmapped
2540  */
2541 void scsi_kunmap_atomic_sg(void *virt)
2542 {
2543         kunmap_atomic(virt, KM_BIO_SRC_IRQ);
2544 }
2545 EXPORT_SYMBOL(scsi_kunmap_atomic_sg);