ab55c2fa7ce209563f99691a7201122fb619fdd9
[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 /*
71  * When to reinvoke queueing after a resource shortage. It's 3 msecs to
72  * not change behaviour from the previous unplug mechanism, experimentation
73  * may prove this needs changing.
74  */
75 #define SCSI_QUEUE_DELAY        3
76
77 static void scsi_run_queue(struct request_queue *q);
78
79 /*
80  * Function:    scsi_unprep_request()
81  *
82  * Purpose:     Remove all preparation done for a request, including its
83  *              associated scsi_cmnd, so that it can be requeued.
84  *
85  * Arguments:   req     - request to unprepare
86  *
87  * Lock status: Assumed that no locks are held upon entry.
88  *
89  * Returns:     Nothing.
90  */
91 static void scsi_unprep_request(struct request *req)
92 {
93         struct scsi_cmnd *cmd = req->special;
94
95         blk_unprep_request(req);
96         req->special = NULL;
97
98         scsi_put_command(cmd);
99 }
100
101 /**
102  * __scsi_queue_insert - private queue insertion
103  * @cmd: The SCSI command being requeued
104  * @reason:  The reason for the requeue
105  * @unbusy: Whether the queue should be unbusied
106  *
107  * This is a private queue insertion.  The public interface
108  * scsi_queue_insert() always assumes the queue should be unbusied
109  * because it's always called before the completion.  This function is
110  * for a requeue after completion, which should only occur in this
111  * file.
112  */
113 static int __scsi_queue_insert(struct scsi_cmnd *cmd, int reason, int unbusy)
114 {
115         struct Scsi_Host *host = cmd->device->host;
116         struct scsi_device *device = cmd->device;
117         struct scsi_target *starget = scsi_target(device);
118         struct request_queue *q = device->request_queue;
119         unsigned long flags;
120
121         SCSI_LOG_MLQUEUE(1,
122                  printk("Inserting command %p into mlqueue\n", cmd));
123
124         /*
125          * Set the appropriate busy bit for the device/host.
126          *
127          * If the host/device isn't busy, assume that something actually
128          * completed, and that we should be able to queue a command now.
129          *
130          * Note that the prior mid-layer assumption that any host could
131          * always queue at least one command is now broken.  The mid-layer
132          * will implement a user specifiable stall (see
133          * scsi_host.max_host_blocked and scsi_device.max_device_blocked)
134          * if a command is requeued with no other commands outstanding
135          * either for the device or for the host.
136          */
137         switch (reason) {
138         case SCSI_MLQUEUE_HOST_BUSY:
139                 host->host_blocked = host->max_host_blocked;
140                 break;
141         case SCSI_MLQUEUE_DEVICE_BUSY:
142                 device->device_blocked = device->max_device_blocked;
143                 break;
144         case SCSI_MLQUEUE_TARGET_BUSY:
145                 starget->target_blocked = starget->max_target_blocked;
146                 break;
147         }
148
149         /*
150          * Decrement the counters, since these commands are no longer
151          * active on the host/device.
152          */
153         if (unbusy)
154                 scsi_device_unbusy(device);
155
156         /*
157          * Requeue this command.  It will go before all other commands
158          * that are already in the queue.
159          */
160         spin_lock_irqsave(q->queue_lock, flags);
161         blk_requeue_request(q, cmd->request);
162         spin_unlock_irqrestore(q->queue_lock, flags);
163
164         scsi_run_queue(q);
165
166         return 0;
167 }
168
169 /*
170  * Function:    scsi_queue_insert()
171  *
172  * Purpose:     Insert a command in the midlevel queue.
173  *
174  * Arguments:   cmd    - command that we are adding to queue.
175  *              reason - why we are inserting command to queue.
176  *
177  * Lock status: Assumed that lock is not held upon entry.
178  *
179  * Returns:     Nothing.
180  *
181  * Notes:       We do this for one of two cases.  Either the host is busy
182  *              and it cannot accept any more commands for the time being,
183  *              or the device returned QUEUE_FULL and can accept no more
184  *              commands.
185  * Notes:       This could be called either from an interrupt context or a
186  *              normal process context.
187  */
188 int scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
189 {
190         return __scsi_queue_insert(cmd, reason, 1);
191 }
192 /**
193  * scsi_execute - insert request and wait for the result
194  * @sdev:       scsi device
195  * @cmd:        scsi command
196  * @data_direction: data direction
197  * @buffer:     data buffer
198  * @bufflen:    len of buffer
199  * @sense:      optional sense buffer
200  * @timeout:    request timeout in seconds
201  * @retries:    number of times to retry request
202  * @flags:      or into request flags;
203  * @resid:      optional residual length
204  *
205  * returns the req->errors value which is the scsi_cmnd result
206  * field.
207  */
208 int scsi_execute(struct scsi_device *sdev, const unsigned char *cmd,
209                  int data_direction, void *buffer, unsigned bufflen,
210                  unsigned char *sense, int timeout, int retries, int flags,
211                  int *resid)
212 {
213         struct request *req;
214         int write = (data_direction == DMA_TO_DEVICE);
215         int ret = DRIVER_ERROR << 24;
216
217         req = blk_get_request(sdev->request_queue, write, __GFP_WAIT);
218
219         if (bufflen &&  blk_rq_map_kern(sdev->request_queue, req,
220                                         buffer, bufflen, __GFP_WAIT))
221                 goto out;
222
223         req->cmd_len = COMMAND_SIZE(cmd[0]);
224         memcpy(req->cmd, cmd, req->cmd_len);
225         req->sense = sense;
226         req->sense_len = 0;
227         req->retries = retries;
228         req->timeout = timeout;
229         req->cmd_type = REQ_TYPE_BLOCK_PC;
230         req->cmd_flags |= flags | REQ_QUIET | REQ_PREEMPT;
231
232         /*
233          * head injection *required* here otherwise quiesce won't work
234          */
235         blk_execute_rq(req->q, NULL, req, 1);
236
237         /*
238          * Some devices (USB mass-storage in particular) may transfer
239          * garbage data together with a residue indicating that the data
240          * is invalid.  Prevent the garbage from being misinterpreted
241          * and prevent security leaks by zeroing out the excess data.
242          */
243         if (unlikely(req->resid_len > 0 && req->resid_len <= bufflen))
244                 memset(buffer + (bufflen - req->resid_len), 0, req->resid_len);
245
246         if (resid)
247                 *resid = req->resid_len;
248         ret = req->errors;
249  out:
250         blk_put_request(req);
251
252         return ret;
253 }
254 EXPORT_SYMBOL(scsi_execute);
255
256
257 int scsi_execute_req(struct scsi_device *sdev, const unsigned char *cmd,
258                      int data_direction, void *buffer, unsigned bufflen,
259                      struct scsi_sense_hdr *sshdr, int timeout, int retries,
260                      int *resid)
261 {
262         char *sense = NULL;
263         int result;
264         
265         if (sshdr) {
266                 sense = kzalloc(SCSI_SENSE_BUFFERSIZE, GFP_NOIO);
267                 if (!sense)
268                         return DRIVER_ERROR << 24;
269         }
270         result = scsi_execute(sdev, cmd, data_direction, buffer, bufflen,
271                               sense, timeout, retries, 0, resid);
272         if (sshdr)
273                 scsi_normalize_sense(sense, SCSI_SENSE_BUFFERSIZE, sshdr);
274
275         kfree(sense);
276         return result;
277 }
278 EXPORT_SYMBOL(scsi_execute_req);
279
280 /*
281  * Function:    scsi_init_cmd_errh()
282  *
283  * Purpose:     Initialize cmd fields related to error handling.
284  *
285  * Arguments:   cmd     - command that is ready to be queued.
286  *
287  * Notes:       This function has the job of initializing a number of
288  *              fields related to error handling.   Typically this will
289  *              be called once for each command, as required.
290  */
291 static void scsi_init_cmd_errh(struct scsi_cmnd *cmd)
292 {
293         cmd->serial_number = 0;
294         scsi_set_resid(cmd, 0);
295         memset(cmd->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
296         if (cmd->cmd_len == 0)
297                 cmd->cmd_len = scsi_command_size(cmd->cmnd);
298 }
299
300 void scsi_device_unbusy(struct scsi_device *sdev)
301 {
302         struct Scsi_Host *shost = sdev->host;
303         struct scsi_target *starget = scsi_target(sdev);
304         unsigned long flags;
305
306         spin_lock_irqsave(shost->host_lock, flags);
307         shost->host_busy--;
308         starget->target_busy--;
309         if (unlikely(scsi_host_in_recovery(shost) &&
310                      (shost->host_failed || shost->host_eh_scheduled)))
311                 scsi_eh_wakeup(shost);
312         spin_unlock(shost->host_lock);
313         spin_lock(sdev->request_queue->queue_lock);
314         sdev->device_busy--;
315         spin_unlock_irqrestore(sdev->request_queue->queue_lock, flags);
316 }
317
318 /*
319  * Called for single_lun devices on IO completion. Clear starget_sdev_user,
320  * and call blk_run_queue for all the scsi_devices on the target -
321  * including current_sdev first.
322  *
323  * Called with *no* scsi locks held.
324  */
325 static void scsi_single_lun_run(struct scsi_device *current_sdev)
326 {
327         struct Scsi_Host *shost = current_sdev->host;
328         struct scsi_device *sdev, *tmp;
329         struct scsi_target *starget = scsi_target(current_sdev);
330         unsigned long flags;
331
332         spin_lock_irqsave(shost->host_lock, flags);
333         starget->starget_sdev_user = NULL;
334         spin_unlock_irqrestore(shost->host_lock, flags);
335
336         /*
337          * Call blk_run_queue for all LUNs on the target, starting with
338          * current_sdev. We race with others (to set starget_sdev_user),
339          * but in most cases, we will be first. Ideally, each LU on the
340          * target would get some limited time or requests on the target.
341          */
342         blk_run_queue(current_sdev->request_queue);
343
344         spin_lock_irqsave(shost->host_lock, flags);
345         if (starget->starget_sdev_user)
346                 goto out;
347         list_for_each_entry_safe(sdev, tmp, &starget->devices,
348                         same_target_siblings) {
349                 if (sdev == current_sdev)
350                         continue;
351                 if (scsi_device_get(sdev))
352                         continue;
353
354                 spin_unlock_irqrestore(shost->host_lock, flags);
355                 blk_run_queue(sdev->request_queue);
356                 spin_lock_irqsave(shost->host_lock, flags);
357         
358                 scsi_device_put(sdev);
359         }
360  out:
361         spin_unlock_irqrestore(shost->host_lock, flags);
362 }
363
364 static inline int scsi_device_is_busy(struct scsi_device *sdev)
365 {
366         if (sdev->device_busy >= sdev->queue_depth || sdev->device_blocked)
367                 return 1;
368
369         return 0;
370 }
371
372 static inline int scsi_target_is_busy(struct scsi_target *starget)
373 {
374         return ((starget->can_queue > 0 &&
375                  starget->target_busy >= starget->can_queue) ||
376                  starget->target_blocked);
377 }
378
379 static inline int scsi_host_is_busy(struct Scsi_Host *shost)
380 {
381         if ((shost->can_queue > 0 && shost->host_busy >= shost->can_queue) ||
382             shost->host_blocked || shost->host_self_blocked)
383                 return 1;
384
385         return 0;
386 }
387
388 /*
389  * Function:    scsi_run_queue()
390  *
391  * Purpose:     Select a proper request queue to serve next
392  *
393  * Arguments:   q       - last request's queue
394  *
395  * Returns:     Nothing
396  *
397  * Notes:       The previous command was completely finished, start
398  *              a new one if possible.
399  */
400 static void scsi_run_queue(struct request_queue *q)
401 {
402         struct scsi_device *sdev = q->queuedata;
403         struct Scsi_Host *shost = sdev->host;
404         LIST_HEAD(starved_list);
405         unsigned long flags;
406
407         if (scsi_target(sdev)->single_lun)
408                 scsi_single_lun_run(sdev);
409
410         spin_lock_irqsave(shost->host_lock, flags);
411         list_splice_init(&shost->starved_list, &starved_list);
412
413         while (!list_empty(&starved_list)) {
414                 int flagset;
415
416                 /*
417                  * As long as shost is accepting commands and we have
418                  * starved queues, call blk_run_queue. scsi_request_fn
419                  * drops the queue_lock and can add us back to the
420                  * starved_list.
421                  *
422                  * host_lock protects the starved_list and starved_entry.
423                  * scsi_request_fn must get the host_lock before checking
424                  * or modifying starved_list or starved_entry.
425                  */
426                 if (scsi_host_is_busy(shost))
427                         break;
428
429                 sdev = list_entry(starved_list.next,
430                                   struct scsi_device, starved_entry);
431                 list_del_init(&sdev->starved_entry);
432                 if (scsi_target_is_busy(scsi_target(sdev))) {
433                         list_move_tail(&sdev->starved_entry,
434                                        &shost->starved_list);
435                         continue;
436                 }
437
438                 spin_unlock(shost->host_lock);
439
440                 spin_lock(sdev->request_queue->queue_lock);
441                 flagset = test_bit(QUEUE_FLAG_REENTER, &q->queue_flags) &&
442                                 !test_bit(QUEUE_FLAG_REENTER,
443                                         &sdev->request_queue->queue_flags);
444                 if (flagset)
445                         queue_flag_set(QUEUE_FLAG_REENTER, sdev->request_queue);
446                 __blk_run_queue(sdev->request_queue);
447                 if (flagset)
448                         queue_flag_clear(QUEUE_FLAG_REENTER, sdev->request_queue);
449                 spin_unlock(sdev->request_queue->queue_lock);
450
451                 spin_lock(shost->host_lock);
452         }
453         /* put any unprocessed entries back */
454         list_splice(&starved_list, &shost->starved_list);
455         spin_unlock_irqrestore(shost->host_lock, flags);
456
457         blk_run_queue(q);
458 }
459
460 /*
461  * Function:    scsi_requeue_command()
462  *
463  * Purpose:     Handle post-processing of completed commands.
464  *
465  * Arguments:   q       - queue to operate on
466  *              cmd     - command that may need to be requeued.
467  *
468  * Returns:     Nothing
469  *
470  * Notes:       After command completion, there may be blocks left
471  *              over which weren't finished by the previous command
472  *              this can be for a number of reasons - the main one is
473  *              I/O errors in the middle of the request, in which case
474  *              we need to request the blocks that come after the bad
475  *              sector.
476  * Notes:       Upon return, cmd is a stale pointer.
477  */
478 static void scsi_requeue_command(struct request_queue *q, struct scsi_cmnd *cmd)
479 {
480         struct request *req = cmd->request;
481         unsigned long flags;
482
483         spin_lock_irqsave(q->queue_lock, flags);
484         scsi_unprep_request(req);
485         blk_requeue_request(q, req);
486         spin_unlock_irqrestore(q->queue_lock, flags);
487
488         scsi_run_queue(q);
489 }
490
491 void scsi_next_command(struct scsi_cmnd *cmd)
492 {
493         struct scsi_device *sdev = cmd->device;
494         struct request_queue *q = sdev->request_queue;
495
496         /* need to hold a reference on the device before we let go of the cmd */
497         get_device(&sdev->sdev_gendev);
498
499         scsi_put_command(cmd);
500         scsi_run_queue(q);
501
502         /* ok to remove device now */
503         put_device(&sdev->sdev_gendev);
504 }
505
506 void scsi_run_host_queues(struct Scsi_Host *shost)
507 {
508         struct scsi_device *sdev;
509
510         shost_for_each_device(sdev, shost)
511                 scsi_run_queue(sdev->request_queue);
512 }
513
514 static void __scsi_release_buffers(struct scsi_cmnd *, int);
515
516 /*
517  * Function:    scsi_end_request()
518  *
519  * Purpose:     Post-processing of completed commands (usually invoked at end
520  *              of upper level post-processing and scsi_io_completion).
521  *
522  * Arguments:   cmd      - command that is complete.
523  *              error    - 0 if I/O indicates success, < 0 for I/O error.
524  *              bytes    - number of bytes of completed I/O
525  *              requeue  - indicates whether we should requeue leftovers.
526  *
527  * Lock status: Assumed that lock is not held upon entry.
528  *
529  * Returns:     cmd if requeue required, NULL otherwise.
530  *
531  * Notes:       This is called for block device requests in order to
532  *              mark some number of sectors as complete.
533  * 
534  *              We are guaranteeing that the request queue will be goosed
535  *              at some point during this call.
536  * Notes:       If cmd was requeued, upon return it will be a stale pointer.
537  */
538 static struct scsi_cmnd *scsi_end_request(struct scsi_cmnd *cmd, int error,
539                                           int bytes, int requeue)
540 {
541         struct request_queue *q = cmd->device->request_queue;
542         struct request *req = cmd->request;
543
544         /*
545          * If there are blocks left over at the end, set up the command
546          * to queue the remainder of them.
547          */
548         if (blk_end_request(req, error, bytes)) {
549                 /* kill remainder if no retrys */
550                 if (error && scsi_noretry_cmd(cmd))
551                         blk_end_request_all(req, error);
552                 else {
553                         if (requeue) {
554                                 /*
555                                  * Bleah.  Leftovers again.  Stick the
556                                  * leftovers in the front of the
557                                  * queue, and goose the queue again.
558                                  */
559                                 scsi_release_buffers(cmd);
560                                 scsi_requeue_command(q, cmd);
561                                 cmd = NULL;
562                         }
563                         return cmd;
564                 }
565         }
566
567         /*
568          * This will goose the queue request function at the end, so we don't
569          * need to worry about launching another command.
570          */
571         __scsi_release_buffers(cmd, 0);
572         scsi_next_command(cmd);
573         return NULL;
574 }
575
576 static inline unsigned int scsi_sgtable_index(unsigned short nents)
577 {
578         unsigned int index;
579
580         BUG_ON(nents > SCSI_MAX_SG_SEGMENTS);
581
582         if (nents <= 8)
583                 index = 0;
584         else
585                 index = get_count_order(nents) - 3;
586
587         return index;
588 }
589
590 static void scsi_sg_free(struct scatterlist *sgl, unsigned int nents)
591 {
592         struct scsi_host_sg_pool *sgp;
593
594         sgp = scsi_sg_pools + scsi_sgtable_index(nents);
595         mempool_free(sgl, sgp->pool);
596 }
597
598 static struct scatterlist *scsi_sg_alloc(unsigned int nents, gfp_t gfp_mask)
599 {
600         struct scsi_host_sg_pool *sgp;
601
602         sgp = scsi_sg_pools + scsi_sgtable_index(nents);
603         return mempool_alloc(sgp->pool, gfp_mask);
604 }
605
606 static int scsi_alloc_sgtable(struct scsi_data_buffer *sdb, int nents,
607                               gfp_t gfp_mask)
608 {
609         int ret;
610
611         BUG_ON(!nents);
612
613         ret = __sg_alloc_table(&sdb->table, nents, SCSI_MAX_SG_SEGMENTS,
614                                gfp_mask, scsi_sg_alloc);
615         if (unlikely(ret))
616                 __sg_free_table(&sdb->table, SCSI_MAX_SG_SEGMENTS,
617                                 scsi_sg_free);
618
619         return ret;
620 }
621
622 static void scsi_free_sgtable(struct scsi_data_buffer *sdb)
623 {
624         __sg_free_table(&sdb->table, SCSI_MAX_SG_SEGMENTS, scsi_sg_free);
625 }
626
627 static void __scsi_release_buffers(struct scsi_cmnd *cmd, int do_bidi_check)
628 {
629
630         if (cmd->sdb.table.nents)
631                 scsi_free_sgtable(&cmd->sdb);
632
633         memset(&cmd->sdb, 0, sizeof(cmd->sdb));
634
635         if (do_bidi_check && scsi_bidi_cmnd(cmd)) {
636                 struct scsi_data_buffer *bidi_sdb =
637                         cmd->request->next_rq->special;
638                 scsi_free_sgtable(bidi_sdb);
639                 kmem_cache_free(scsi_sdb_cache, bidi_sdb);
640                 cmd->request->next_rq->special = NULL;
641         }
642
643         if (scsi_prot_sg_count(cmd))
644                 scsi_free_sgtable(cmd->prot_sdb);
645 }
646
647 /*
648  * Function:    scsi_release_buffers()
649  *
650  * Purpose:     Completion processing for block device I/O requests.
651  *
652  * Arguments:   cmd     - command that we are bailing.
653  *
654  * Lock status: Assumed that no lock is held upon entry.
655  *
656  * Returns:     Nothing
657  *
658  * Notes:       In the event that an upper level driver rejects a
659  *              command, we must release resources allocated during
660  *              the __init_io() function.  Primarily this would involve
661  *              the scatter-gather table, and potentially any bounce
662  *              buffers.
663  */
664 void scsi_release_buffers(struct scsi_cmnd *cmd)
665 {
666         __scsi_release_buffers(cmd, 1);
667 }
668 EXPORT_SYMBOL(scsi_release_buffers);
669
670 static int __scsi_error_from_host_byte(struct scsi_cmnd *cmd, int result)
671 {
672         int error = 0;
673
674         switch(host_byte(result)) {
675         case DID_TRANSPORT_FAILFAST:
676                 error = -ENOLINK;
677                 break;
678         case DID_TARGET_FAILURE:
679                 cmd->result |= (DID_OK << 16);
680                 error = -EREMOTEIO;
681                 break;
682         case DID_NEXUS_FAILURE:
683                 cmd->result |= (DID_OK << 16);
684                 error = -EBADE;
685                 break;
686         default:
687                 error = -EIO;
688                 break;
689         }
690
691         return error;
692 }
693
694 /*
695  * Function:    scsi_io_completion()
696  *
697  * Purpose:     Completion processing for block device I/O requests.
698  *
699  * Arguments:   cmd   - command that is finished.
700  *
701  * Lock status: Assumed that no lock is held upon entry.
702  *
703  * Returns:     Nothing
704  *
705  * Notes:       This function is matched in terms of capabilities to
706  *              the function that created the scatter-gather list.
707  *              In other words, if there are no bounce buffers
708  *              (the normal case for most drivers), we don't need
709  *              the logic to deal with cleaning up afterwards.
710  *
711  *              We must call scsi_end_request().  This will finish off
712  *              the specified number of sectors.  If we are done, the
713  *              command block will be released and the queue function
714  *              will be goosed.  If we are not done then we have to
715  *              figure out what to do next:
716  *
717  *              a) We can call scsi_requeue_command().  The request
718  *                 will be unprepared and put back on the queue.  Then
719  *                 a new command will be created for it.  This should
720  *                 be used if we made forward progress, or if we want
721  *                 to switch from READ(10) to READ(6) for example.
722  *
723  *              b) We can call scsi_queue_insert().  The request will
724  *                 be put back on the queue and retried using the same
725  *                 command as before, possibly after a delay.
726  *
727  *              c) We can call blk_end_request() with -EIO to fail
728  *                 the remainder of the request.
729  */
730 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
731 {
732         int result = cmd->result;
733         struct request_queue *q = cmd->device->request_queue;
734         struct request *req = cmd->request;
735         int error = 0;
736         struct scsi_sense_hdr sshdr;
737         int sense_valid = 0;
738         int sense_deferred = 0;
739         enum {ACTION_FAIL, ACTION_REPREP, ACTION_RETRY,
740               ACTION_DELAYED_RETRY} action;
741         char *description = NULL;
742
743         if (result) {
744                 sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
745                 if (sense_valid)
746                         sense_deferred = scsi_sense_is_deferred(&sshdr);
747         }
748
749         if (req->cmd_type == REQ_TYPE_BLOCK_PC) { /* SG_IO ioctl from block level */
750                 req->errors = result;
751                 if (result) {
752                         if (sense_valid && req->sense) {
753                                 /*
754                                  * SG_IO wants current and deferred errors
755                                  */
756                                 int len = 8 + cmd->sense_buffer[7];
757
758                                 if (len > SCSI_SENSE_BUFFERSIZE)
759                                         len = SCSI_SENSE_BUFFERSIZE;
760                                 memcpy(req->sense, cmd->sense_buffer,  len);
761                                 req->sense_len = len;
762                         }
763                         if (!sense_deferred)
764                                 error = __scsi_error_from_host_byte(cmd, result);
765                 }
766
767                 req->resid_len = scsi_get_resid(cmd);
768
769                 if (scsi_bidi_cmnd(cmd)) {
770                         /*
771                          * Bidi commands Must be complete as a whole,
772                          * both sides at once.
773                          */
774                         req->next_rq->resid_len = scsi_in(cmd)->resid;
775
776                         scsi_release_buffers(cmd);
777                         blk_end_request_all(req, 0);
778
779                         scsi_next_command(cmd);
780                         return;
781                 }
782         }
783
784         /* no bidi support for !REQ_TYPE_BLOCK_PC yet */
785         BUG_ON(blk_bidi_rq(req));
786
787         /*
788          * Next deal with any sectors which we were able to correctly
789          * handle.
790          */
791         SCSI_LOG_HLCOMPLETE(1, printk("%u sectors total, "
792                                       "%d bytes done.\n",
793                                       blk_rq_sectors(req), good_bytes));
794
795         /*
796          * Recovered errors need reporting, but they're always treated
797          * as success, so fiddle the result code here.  For BLOCK_PC
798          * we already took a copy of the original into rq->errors which
799          * is what gets returned to the user
800          */
801         if (sense_valid && (sshdr.sense_key == RECOVERED_ERROR)) {
802                 /* if ATA PASS-THROUGH INFORMATION AVAILABLE skip
803                  * print since caller wants ATA registers. Only occurs on
804                  * SCSI ATA PASS_THROUGH commands when CK_COND=1
805                  */
806                 if ((sshdr.asc == 0x0) && (sshdr.ascq == 0x1d))
807                         ;
808                 else if (!(req->cmd_flags & REQ_QUIET))
809                         scsi_print_sense("", cmd);
810                 result = 0;
811                 /* BLOCK_PC may have set error */
812                 error = 0;
813         }
814
815         /*
816          * A number of bytes were successfully read.  If there
817          * are leftovers and there is some kind of error
818          * (result != 0), retry the rest.
819          */
820         if (scsi_end_request(cmd, error, good_bytes, result == 0) == NULL)
821                 return;
822
823         error = __scsi_error_from_host_byte(cmd, result);
824
825         if (host_byte(result) == DID_RESET) {
826                 /* Third party bus reset or reset for error recovery
827                  * reasons.  Just retry the command and see what
828                  * happens.
829                  */
830                 action = ACTION_RETRY;
831         } else if (sense_valid && !sense_deferred) {
832                 switch (sshdr.sense_key) {
833                 case UNIT_ATTENTION:
834                         if (cmd->device->removable) {
835                                 /* Detected disc change.  Set a bit
836                                  * and quietly refuse further access.
837                                  */
838                                 cmd->device->changed = 1;
839                                 description = "Media Changed";
840                                 action = ACTION_FAIL;
841                         } else {
842                                 /* Must have been a power glitch, or a
843                                  * bus reset.  Could not have been a
844                                  * media change, so we just retry the
845                                  * command and see what happens.
846                                  */
847                                 action = ACTION_RETRY;
848                         }
849                         break;
850                 case ILLEGAL_REQUEST:
851                         /* If we had an ILLEGAL REQUEST returned, then
852                          * we may have performed an unsupported
853                          * command.  The only thing this should be
854                          * would be a ten byte read where only a six
855                          * byte read was supported.  Also, on a system
856                          * where READ CAPACITY failed, we may have
857                          * read past the end of the disk.
858                          */
859                         if ((cmd->device->use_10_for_rw &&
860                             sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
861                             (cmd->cmnd[0] == READ_10 ||
862                              cmd->cmnd[0] == WRITE_10)) {
863                                 /* This will issue a new 6-byte command. */
864                                 cmd->device->use_10_for_rw = 0;
865                                 action = ACTION_REPREP;
866                         } else if (sshdr.asc == 0x10) /* DIX */ {
867                                 description = "Host Data Integrity Failure";
868                                 action = ACTION_FAIL;
869                                 error = -EILSEQ;
870                         /* INVALID COMMAND OPCODE or INVALID FIELD IN CDB */
871                         } else if ((sshdr.asc == 0x20 || sshdr.asc == 0x24) &&
872                                    (cmd->cmnd[0] == UNMAP ||
873                                     cmd->cmnd[0] == WRITE_SAME_16 ||
874                                     cmd->cmnd[0] == WRITE_SAME)) {
875                                 description = "Discard failure";
876                                 action = ACTION_FAIL;
877                         } else
878                                 action = ACTION_FAIL;
879                         break;
880                 case ABORTED_COMMAND:
881                         action = ACTION_FAIL;
882                         if (sshdr.asc == 0x10) { /* DIF */
883                                 description = "Target Data Integrity Failure";
884                                 error = -EILSEQ;
885                         }
886                         break;
887                 case NOT_READY:
888                         /* If the device is in the process of becoming
889                          * ready, or has a temporary blockage, retry.
890                          */
891                         if (sshdr.asc == 0x04) {
892                                 switch (sshdr.ascq) {
893                                 case 0x01: /* becoming ready */
894                                 case 0x04: /* format in progress */
895                                 case 0x05: /* rebuild in progress */
896                                 case 0x06: /* recalculation in progress */
897                                 case 0x07: /* operation in progress */
898                                 case 0x08: /* Long write in progress */
899                                 case 0x09: /* self test in progress */
900                                 case 0x14: /* space allocation in progress */
901                                         action = ACTION_DELAYED_RETRY;
902                                         break;
903                                 default:
904                                         description = "Device not ready";
905                                         action = ACTION_FAIL;
906                                         break;
907                                 }
908                         } else {
909                                 description = "Device not ready";
910                                 action = ACTION_FAIL;
911                         }
912                         break;
913                 case VOLUME_OVERFLOW:
914                         /* See SSC3rXX or current. */
915                         action = ACTION_FAIL;
916                         break;
917                 default:
918                         description = "Unhandled sense code";
919                         action = ACTION_FAIL;
920                         break;
921                 }
922         } else {
923                 description = "Unhandled error code";
924                 action = ACTION_FAIL;
925         }
926
927         switch (action) {
928         case ACTION_FAIL:
929                 /* Give up and fail the remainder of the request */
930                 scsi_release_buffers(cmd);
931                 if (!(req->cmd_flags & REQ_QUIET)) {
932                         if (description)
933                                 scmd_printk(KERN_INFO, cmd, "%s\n",
934                                             description);
935                         scsi_print_result(cmd);
936                         if (driver_byte(result) & DRIVER_SENSE)
937                                 scsi_print_sense("", cmd);
938                         scsi_print_command(cmd);
939                 }
940                 if (blk_end_request_err(req, error))
941                         scsi_requeue_command(q, cmd);
942                 else
943                         scsi_next_command(cmd);
944                 break;
945         case ACTION_REPREP:
946                 /* Unprep the request and put it back at the head of the queue.
947                  * A new command will be prepared and issued.
948                  */
949                 scsi_release_buffers(cmd);
950                 scsi_requeue_command(q, cmd);
951                 break;
952         case ACTION_RETRY:
953                 /* Retry the same command immediately */
954                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY, 0);
955                 break;
956         case ACTION_DELAYED_RETRY:
957                 /* Retry the same command after a delay */
958                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY, 0);
959                 break;
960         }
961 }
962
963 static int scsi_init_sgtable(struct request *req, struct scsi_data_buffer *sdb,
964                              gfp_t gfp_mask)
965 {
966         int count;
967
968         /*
969          * If sg table allocation fails, requeue request later.
970          */
971         if (unlikely(scsi_alloc_sgtable(sdb, req->nr_phys_segments,
972                                         gfp_mask))) {
973                 return BLKPREP_DEFER;
974         }
975
976         req->buffer = NULL;
977
978         /* 
979          * Next, walk the list, and fill in the addresses and sizes of
980          * each segment.
981          */
982         count = blk_rq_map_sg(req->q, req, sdb->table.sgl);
983         BUG_ON(count > sdb->table.nents);
984         sdb->table.nents = count;
985         sdb->length = blk_rq_bytes(req);
986         return BLKPREP_OK;
987 }
988
989 /*
990  * Function:    scsi_init_io()
991  *
992  * Purpose:     SCSI I/O initialize function.
993  *
994  * Arguments:   cmd   - Command descriptor we wish to initialize
995  *
996  * Returns:     0 on success
997  *              BLKPREP_DEFER if the failure is retryable
998  *              BLKPREP_KILL if the failure is fatal
999  */
1000 int scsi_init_io(struct scsi_cmnd *cmd, gfp_t gfp_mask)
1001 {
1002         struct request *rq = cmd->request;
1003
1004         int error = scsi_init_sgtable(rq, &cmd->sdb, gfp_mask);
1005         if (error)
1006                 goto err_exit;
1007
1008         if (blk_bidi_rq(rq)) {
1009                 struct scsi_data_buffer *bidi_sdb = kmem_cache_zalloc(
1010                         scsi_sdb_cache, GFP_ATOMIC);
1011                 if (!bidi_sdb) {
1012                         error = BLKPREP_DEFER;
1013                         goto err_exit;
1014                 }
1015
1016                 rq->next_rq->special = bidi_sdb;
1017                 error = scsi_init_sgtable(rq->next_rq, bidi_sdb, GFP_ATOMIC);
1018                 if (error)
1019                         goto err_exit;
1020         }
1021
1022         if (blk_integrity_rq(rq)) {
1023                 struct scsi_data_buffer *prot_sdb = cmd->prot_sdb;
1024                 int ivecs, count;
1025
1026                 BUG_ON(prot_sdb == NULL);
1027                 ivecs = blk_rq_count_integrity_sg(rq->q, rq->bio);
1028
1029                 if (scsi_alloc_sgtable(prot_sdb, ivecs, gfp_mask)) {
1030                         error = BLKPREP_DEFER;
1031                         goto err_exit;
1032                 }
1033
1034                 count = blk_rq_map_integrity_sg(rq->q, rq->bio,
1035                                                 prot_sdb->table.sgl);
1036                 BUG_ON(unlikely(count > ivecs));
1037                 BUG_ON(unlikely(count > queue_max_integrity_segments(rq->q)));
1038
1039                 cmd->prot_sdb = prot_sdb;
1040                 cmd->prot_sdb->table.nents = count;
1041         }
1042
1043         return BLKPREP_OK ;
1044
1045 err_exit:
1046         scsi_release_buffers(cmd);
1047         cmd->request->special = NULL;
1048         scsi_put_command(cmd);
1049         return error;
1050 }
1051 EXPORT_SYMBOL(scsi_init_io);
1052
1053 static struct scsi_cmnd *scsi_get_cmd_from_req(struct scsi_device *sdev,
1054                 struct request *req)
1055 {
1056         struct scsi_cmnd *cmd;
1057
1058         if (!req->special) {
1059                 cmd = scsi_get_command(sdev, GFP_ATOMIC);
1060                 if (unlikely(!cmd))
1061                         return NULL;
1062                 req->special = cmd;
1063         } else {
1064                 cmd = req->special;
1065         }
1066
1067         /* pull a tag out of the request if we have one */
1068         cmd->tag = req->tag;
1069         cmd->request = req;
1070
1071         cmd->cmnd = req->cmd;
1072         cmd->prot_op = SCSI_PROT_NORMAL;
1073
1074         return cmd;
1075 }
1076
1077 int scsi_setup_blk_pc_cmnd(struct scsi_device *sdev, struct request *req)
1078 {
1079         struct scsi_cmnd *cmd;
1080         int ret = scsi_prep_state_check(sdev, req);
1081
1082         if (ret != BLKPREP_OK)
1083                 return ret;
1084
1085         cmd = scsi_get_cmd_from_req(sdev, req);
1086         if (unlikely(!cmd))
1087                 return BLKPREP_DEFER;
1088
1089         /*
1090          * BLOCK_PC requests may transfer data, in which case they must
1091          * a bio attached to them.  Or they might contain a SCSI command
1092          * that does not transfer data, in which case they may optionally
1093          * submit a request without an attached bio.
1094          */
1095         if (req->bio) {
1096                 int ret;
1097
1098                 BUG_ON(!req->nr_phys_segments);
1099
1100                 ret = scsi_init_io(cmd, GFP_ATOMIC);
1101                 if (unlikely(ret))
1102                         return ret;
1103         } else {
1104                 BUG_ON(blk_rq_bytes(req));
1105
1106                 memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1107                 req->buffer = NULL;
1108         }
1109
1110         cmd->cmd_len = req->cmd_len;
1111         if (!blk_rq_bytes(req))
1112                 cmd->sc_data_direction = DMA_NONE;
1113         else if (rq_data_dir(req) == WRITE)
1114                 cmd->sc_data_direction = DMA_TO_DEVICE;
1115         else
1116                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1117         
1118         cmd->transfersize = blk_rq_bytes(req);
1119         cmd->allowed = req->retries;
1120         return BLKPREP_OK;
1121 }
1122 EXPORT_SYMBOL(scsi_setup_blk_pc_cmnd);
1123
1124 /*
1125  * Setup a REQ_TYPE_FS command.  These are simple read/write request
1126  * from filesystems that still need to be translated to SCSI CDBs from
1127  * the ULD.
1128  */
1129 int scsi_setup_fs_cmnd(struct scsi_device *sdev, struct request *req)
1130 {
1131         struct scsi_cmnd *cmd;
1132         int ret = scsi_prep_state_check(sdev, req);
1133
1134         if (ret != BLKPREP_OK)
1135                 return ret;
1136
1137         if (unlikely(sdev->scsi_dh_data && sdev->scsi_dh_data->scsi_dh
1138                          && sdev->scsi_dh_data->scsi_dh->prep_fn)) {
1139                 ret = sdev->scsi_dh_data->scsi_dh->prep_fn(sdev, req);
1140                 if (ret != BLKPREP_OK)
1141                         return ret;
1142         }
1143
1144         /*
1145          * Filesystem requests must transfer data.
1146          */
1147         BUG_ON(!req->nr_phys_segments);
1148
1149         cmd = scsi_get_cmd_from_req(sdev, req);
1150         if (unlikely(!cmd))
1151                 return BLKPREP_DEFER;
1152
1153         memset(cmd->cmnd, 0, BLK_MAX_CDB);
1154         return scsi_init_io(cmd, GFP_ATOMIC);
1155 }
1156 EXPORT_SYMBOL(scsi_setup_fs_cmnd);
1157
1158 int scsi_prep_state_check(struct scsi_device *sdev, struct request *req)
1159 {
1160         int ret = BLKPREP_OK;
1161
1162         /*
1163          * If the device is not in running state we will reject some
1164          * or all commands.
1165          */
1166         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1167                 switch (sdev->sdev_state) {
1168                 case SDEV_OFFLINE:
1169                         /*
1170                          * If the device is offline we refuse to process any
1171                          * commands.  The device must be brought online
1172                          * before trying any recovery commands.
1173                          */
1174                         sdev_printk(KERN_ERR, sdev,
1175                                     "rejecting I/O to offline device\n");
1176                         ret = BLKPREP_KILL;
1177                         break;
1178                 case SDEV_DEL:
1179                         /*
1180                          * If the device is fully deleted, we refuse to
1181                          * process any commands as well.
1182                          */
1183                         sdev_printk(KERN_ERR, sdev,
1184                                     "rejecting I/O to dead device\n");
1185                         ret = BLKPREP_KILL;
1186                         break;
1187                 case SDEV_QUIESCE:
1188                 case SDEV_BLOCK:
1189                 case SDEV_CREATED_BLOCK:
1190                         /*
1191                          * If the devices is blocked we defer normal commands.
1192                          */
1193                         if (!(req->cmd_flags & REQ_PREEMPT))
1194                                 ret = BLKPREP_DEFER;
1195                         break;
1196                 default:
1197                         /*
1198                          * For any other not fully online state we only allow
1199                          * special commands.  In particular any user initiated
1200                          * command is not allowed.
1201                          */
1202                         if (!(req->cmd_flags & REQ_PREEMPT))
1203                                 ret = BLKPREP_KILL;
1204                         break;
1205                 }
1206         }
1207         return ret;
1208 }
1209 EXPORT_SYMBOL(scsi_prep_state_check);
1210
1211 int scsi_prep_return(struct request_queue *q, struct request *req, int ret)
1212 {
1213         struct scsi_device *sdev = q->queuedata;
1214
1215         switch (ret) {
1216         case BLKPREP_KILL:
1217                 req->errors = DID_NO_CONNECT << 16;
1218                 /* release the command and kill it */
1219                 if (req->special) {
1220                         struct scsi_cmnd *cmd = req->special;
1221                         scsi_release_buffers(cmd);
1222                         scsi_put_command(cmd);
1223                         req->special = NULL;
1224                 }
1225                 break;
1226         case BLKPREP_DEFER:
1227                 /*
1228                  * If we defer, the blk_peek_request() returns NULL, but the
1229                  * queue must be restarted, so we schedule a callback to happen
1230                  * shortly.
1231                  */
1232                 if (sdev->device_busy == 0)
1233                         blk_delay_queue(q, SCSI_QUEUE_DELAY);
1234                 break;
1235         default:
1236                 req->cmd_flags |= REQ_DONTPREP;
1237         }
1238
1239         return ret;
1240 }
1241 EXPORT_SYMBOL(scsi_prep_return);
1242
1243 int scsi_prep_fn(struct request_queue *q, struct request *req)
1244 {
1245         struct scsi_device *sdev = q->queuedata;
1246         int ret = BLKPREP_KILL;
1247
1248         if (req->cmd_type == REQ_TYPE_BLOCK_PC)
1249                 ret = scsi_setup_blk_pc_cmnd(sdev, req);
1250         return scsi_prep_return(q, req, ret);
1251 }
1252 EXPORT_SYMBOL(scsi_prep_fn);
1253
1254 /*
1255  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1256  * return 0.
1257  *
1258  * Called with the queue_lock held.
1259  */
1260 static inline int scsi_dev_queue_ready(struct request_queue *q,
1261                                   struct scsi_device *sdev)
1262 {
1263         if (sdev->device_busy == 0 && sdev->device_blocked) {
1264                 /*
1265                  * unblock after device_blocked iterates to zero
1266                  */
1267                 if (--sdev->device_blocked == 0) {
1268                         SCSI_LOG_MLQUEUE(3,
1269                                    sdev_printk(KERN_INFO, sdev,
1270                                    "unblocking device at zero depth\n"));
1271                 } else {
1272                         blk_delay_queue(q, SCSI_QUEUE_DELAY);
1273                         return 0;
1274                 }
1275         }
1276         if (scsi_device_is_busy(sdev))
1277                 return 0;
1278
1279         return 1;
1280 }
1281
1282
1283 /*
1284  * scsi_target_queue_ready: checks if there we can send commands to target
1285  * @sdev: scsi device on starget to check.
1286  *
1287  * Called with the host lock held.
1288  */
1289 static inline int scsi_target_queue_ready(struct Scsi_Host *shost,
1290                                            struct scsi_device *sdev)
1291 {
1292         struct scsi_target *starget = scsi_target(sdev);
1293
1294         if (starget->single_lun) {
1295                 if (starget->starget_sdev_user &&
1296                     starget->starget_sdev_user != sdev)
1297                         return 0;
1298                 starget->starget_sdev_user = sdev;
1299         }
1300
1301         if (starget->target_busy == 0 && starget->target_blocked) {
1302                 /*
1303                  * unblock after target_blocked iterates to zero
1304                  */
1305                 if (--starget->target_blocked == 0) {
1306                         SCSI_LOG_MLQUEUE(3, starget_printk(KERN_INFO, starget,
1307                                          "unblocking target at zero depth\n"));
1308                 } else
1309                         return 0;
1310         }
1311
1312         if (scsi_target_is_busy(starget)) {
1313                 if (list_empty(&sdev->starved_entry))
1314                         list_add_tail(&sdev->starved_entry,
1315                                       &shost->starved_list);
1316                 return 0;
1317         }
1318
1319         /* We're OK to process the command, so we can't be starved */
1320         if (!list_empty(&sdev->starved_entry))
1321                 list_del_init(&sdev->starved_entry);
1322         return 1;
1323 }
1324
1325 /*
1326  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1327  * return 0. We must end up running the queue again whenever 0 is
1328  * returned, else IO can hang.
1329  *
1330  * Called with host_lock held.
1331  */
1332 static inline int scsi_host_queue_ready(struct request_queue *q,
1333                                    struct Scsi_Host *shost,
1334                                    struct scsi_device *sdev)
1335 {
1336         if (scsi_host_in_recovery(shost))
1337                 return 0;
1338         if (shost->host_busy == 0 && shost->host_blocked) {
1339                 /*
1340                  * unblock after host_blocked iterates to zero
1341                  */
1342                 if (--shost->host_blocked == 0) {
1343                         SCSI_LOG_MLQUEUE(3,
1344                                 printk("scsi%d unblocking host at zero depth\n",
1345                                         shost->host_no));
1346                 } else {
1347                         return 0;
1348                 }
1349         }
1350         if (scsi_host_is_busy(shost)) {
1351                 if (list_empty(&sdev->starved_entry))
1352                         list_add_tail(&sdev->starved_entry, &shost->starved_list);
1353                 return 0;
1354         }
1355
1356         /* We're OK to process the command, so we can't be starved */
1357         if (!list_empty(&sdev->starved_entry))
1358                 list_del_init(&sdev->starved_entry);
1359
1360         return 1;
1361 }
1362
1363 /*
1364  * Busy state exporting function for request stacking drivers.
1365  *
1366  * For efficiency, no lock is taken to check the busy state of
1367  * shost/starget/sdev, since the returned value is not guaranteed and
1368  * may be changed after request stacking drivers call the function,
1369  * regardless of taking lock or not.
1370  *
1371  * When scsi can't dispatch I/Os anymore and needs to kill I/Os
1372  * (e.g. !sdev), scsi needs to return 'not busy'.
1373  * Otherwise, request stacking drivers may hold requests forever.
1374  */
1375 static int scsi_lld_busy(struct request_queue *q)
1376 {
1377         struct scsi_device *sdev = q->queuedata;
1378         struct Scsi_Host *shost;
1379         struct scsi_target *starget;
1380
1381         if (!sdev)
1382                 return 0;
1383
1384         shost = sdev->host;
1385         starget = scsi_target(sdev);
1386
1387         if (scsi_host_in_recovery(shost) || scsi_host_is_busy(shost) ||
1388             scsi_target_is_busy(starget) || scsi_device_is_busy(sdev))
1389                 return 1;
1390
1391         return 0;
1392 }
1393
1394 /*
1395  * Kill a request for a dead device
1396  */
1397 static void scsi_kill_request(struct request *req, struct request_queue *q)
1398 {
1399         struct scsi_cmnd *cmd = req->special;
1400         struct scsi_device *sdev;
1401         struct scsi_target *starget;
1402         struct Scsi_Host *shost;
1403
1404         blk_start_request(req);
1405
1406         sdev = cmd->device;
1407         starget = scsi_target(sdev);
1408         shost = sdev->host;
1409         scsi_init_cmd_errh(cmd);
1410         cmd->result = DID_NO_CONNECT << 16;
1411         atomic_inc(&cmd->device->iorequest_cnt);
1412
1413         /*
1414          * SCSI request completion path will do scsi_device_unbusy(),
1415          * bump busy counts.  To bump the counters, we need to dance
1416          * with the locks as normal issue path does.
1417          */
1418         sdev->device_busy++;
1419         spin_unlock(sdev->request_queue->queue_lock);
1420         spin_lock(shost->host_lock);
1421         shost->host_busy++;
1422         starget->target_busy++;
1423         spin_unlock(shost->host_lock);
1424         spin_lock(sdev->request_queue->queue_lock);
1425
1426         blk_complete_request(req);
1427 }
1428
1429 static void scsi_softirq_done(struct request *rq)
1430 {
1431         struct scsi_cmnd *cmd = rq->special;
1432         unsigned long wait_for = (cmd->allowed + 1) * rq->timeout;
1433         int disposition;
1434
1435         INIT_LIST_HEAD(&cmd->eh_entry);
1436
1437         atomic_inc(&cmd->device->iodone_cnt);
1438         if (cmd->result)
1439                 atomic_inc(&cmd->device->ioerr_cnt);
1440
1441         disposition = scsi_decide_disposition(cmd);
1442         if (disposition != SUCCESS &&
1443             time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1444                 sdev_printk(KERN_ERR, cmd->device,
1445                             "timing out command, waited %lus\n",
1446                             wait_for/HZ);
1447                 disposition = SUCCESS;
1448         }
1449                         
1450         scsi_log_completion(cmd, disposition);
1451
1452         switch (disposition) {
1453                 case SUCCESS:
1454                         scsi_finish_command(cmd);
1455                         break;
1456                 case NEEDS_RETRY:
1457                         scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1458                         break;
1459                 case ADD_TO_MLQUEUE:
1460                         scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1461                         break;
1462                 default:
1463                         if (!scsi_eh_scmd_add(cmd, 0))
1464                                 scsi_finish_command(cmd);
1465         }
1466 }
1467
1468 /*
1469  * Function:    scsi_request_fn()
1470  *
1471  * Purpose:     Main strategy routine for SCSI.
1472  *
1473  * Arguments:   q       - Pointer to actual queue.
1474  *
1475  * Returns:     Nothing
1476  *
1477  * Lock status: IO request lock assumed to be held when called.
1478  */
1479 static void scsi_request_fn(struct request_queue *q)
1480 {
1481         struct scsi_device *sdev = q->queuedata;
1482         struct Scsi_Host *shost;
1483         struct scsi_cmnd *cmd;
1484         struct request *req;
1485
1486         if (!sdev) {
1487                 printk("scsi: killing requests for dead queue\n");
1488                 while ((req = blk_peek_request(q)) != NULL)
1489                         scsi_kill_request(req, q);
1490                 return;
1491         }
1492
1493         if(!get_device(&sdev->sdev_gendev))
1494                 /* We must be tearing the block queue down already */
1495                 return;
1496
1497         /*
1498          * To start with, we keep looping until the queue is empty, or until
1499          * the host is no longer able to accept any more requests.
1500          */
1501         shost = sdev->host;
1502         for (;;) {
1503                 int rtn;
1504                 /*
1505                  * get next queueable request.  We do this early to make sure
1506                  * that the request is fully prepared even if we cannot 
1507                  * accept it.
1508                  */
1509                 req = blk_peek_request(q);
1510                 if (!req || !scsi_dev_queue_ready(q, sdev))
1511                         break;
1512
1513                 if (unlikely(!scsi_device_online(sdev))) {
1514                         sdev_printk(KERN_ERR, sdev,
1515                                     "rejecting I/O to offline device\n");
1516                         scsi_kill_request(req, q);
1517                         continue;
1518                 }
1519
1520
1521                 /*
1522                  * Remove the request from the request list.
1523                  */
1524                 if (!(blk_queue_tagged(q) && !blk_queue_start_tag(q, req)))
1525                         blk_start_request(req);
1526                 sdev->device_busy++;
1527
1528                 spin_unlock(q->queue_lock);
1529                 cmd = req->special;
1530                 if (unlikely(cmd == NULL)) {
1531                         printk(KERN_CRIT "impossible request in %s.\n"
1532                                          "please mail a stack trace to "
1533                                          "linux-scsi@vger.kernel.org\n",
1534                                          __func__);
1535                         blk_dump_rq_flags(req, "foo");
1536                         BUG();
1537                 }
1538                 spin_lock(shost->host_lock);
1539
1540                 /*
1541                  * We hit this when the driver is using a host wide
1542                  * tag map. For device level tag maps the queue_depth check
1543                  * in the device ready fn would prevent us from trying
1544                  * to allocate a tag. Since the map is a shared host resource
1545                  * we add the dev to the starved list so it eventually gets
1546                  * a run when a tag is freed.
1547                  */
1548                 if (blk_queue_tagged(q) && !blk_rq_tagged(req)) {
1549                         if (list_empty(&sdev->starved_entry))
1550                                 list_add_tail(&sdev->starved_entry,
1551                                               &shost->starved_list);
1552                         goto not_ready;
1553                 }
1554
1555                 if (!scsi_target_queue_ready(shost, sdev))
1556                         goto not_ready;
1557
1558                 if (!scsi_host_queue_ready(q, shost, sdev))
1559                         goto not_ready;
1560
1561                 scsi_target(sdev)->target_busy++;
1562                 shost->host_busy++;
1563
1564                 /*
1565                  * XXX(hch): This is rather suboptimal, scsi_dispatch_cmd will
1566                  *              take the lock again.
1567                  */
1568                 spin_unlock_irq(shost->host_lock);
1569
1570                 /*
1571                  * Finally, initialize any error handling parameters, and set up
1572                  * the timers for timeouts.
1573                  */
1574                 scsi_init_cmd_errh(cmd);
1575
1576                 /*
1577                  * Dispatch the command to the low-level driver.
1578                  */
1579                 rtn = scsi_dispatch_cmd(cmd);
1580                 spin_lock_irq(q->queue_lock);
1581                 if (rtn)
1582                         goto out_delay;
1583         }
1584
1585         goto out;
1586
1587  not_ready:
1588         spin_unlock_irq(shost->host_lock);
1589
1590         /*
1591          * lock q, handle tag, requeue req, and decrement device_busy. We
1592          * must return with queue_lock held.
1593          *
1594          * Decrementing device_busy without checking it is OK, as all such
1595          * cases (host limits or settings) should run the queue at some
1596          * later time.
1597          */
1598         spin_lock_irq(q->queue_lock);
1599         blk_requeue_request(q, req);
1600         sdev->device_busy--;
1601 out_delay:
1602         if (sdev->device_busy == 0)
1603                 blk_delay_queue(q, SCSI_QUEUE_DELAY);
1604 out:
1605         /* must be careful here...if we trigger the ->remove() function
1606          * we cannot be holding the q lock */
1607         spin_unlock_irq(q->queue_lock);
1608         put_device(&sdev->sdev_gendev);
1609         spin_lock_irq(q->queue_lock);
1610 }
1611
1612 u64 scsi_calculate_bounce_limit(struct Scsi_Host *shost)
1613 {
1614         struct device *host_dev;
1615         u64 bounce_limit = 0xffffffff;
1616
1617         if (shost->unchecked_isa_dma)
1618                 return BLK_BOUNCE_ISA;
1619         /*
1620          * Platforms with virtual-DMA translation
1621          * hardware have no practical limit.
1622          */
1623         if (!PCI_DMA_BUS_IS_PHYS)
1624                 return BLK_BOUNCE_ANY;
1625
1626         host_dev = scsi_get_device(shost);
1627         if (host_dev && host_dev->dma_mask)
1628                 bounce_limit = *host_dev->dma_mask;
1629
1630         return bounce_limit;
1631 }
1632 EXPORT_SYMBOL(scsi_calculate_bounce_limit);
1633
1634 struct request_queue *__scsi_alloc_queue(struct Scsi_Host *shost,
1635                                          request_fn_proc *request_fn)
1636 {
1637         struct request_queue *q;
1638         struct device *dev = shost->shost_gendev.parent;
1639
1640         q = blk_init_queue(request_fn, NULL);
1641         if (!q)
1642                 return NULL;
1643
1644         /*
1645          * this limit is imposed by hardware restrictions
1646          */
1647         blk_queue_max_segments(q, min_t(unsigned short, shost->sg_tablesize,
1648                                         SCSI_MAX_SG_CHAIN_SEGMENTS));
1649
1650         if (scsi_host_prot_dma(shost)) {
1651                 shost->sg_prot_tablesize =
1652                         min_not_zero(shost->sg_prot_tablesize,
1653                                      (unsigned short)SCSI_MAX_PROT_SG_SEGMENTS);
1654                 BUG_ON(shost->sg_prot_tablesize < shost->sg_tablesize);
1655                 blk_queue_max_integrity_segments(q, shost->sg_prot_tablesize);
1656         }
1657
1658         blk_queue_max_hw_sectors(q, shost->max_sectors);
1659         blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
1660         blk_queue_segment_boundary(q, shost->dma_boundary);
1661         dma_set_seg_boundary(dev, shost->dma_boundary);
1662
1663         blk_queue_max_segment_size(q, dma_get_max_seg_size(dev));
1664
1665         if (!shost->use_clustering)
1666                 q->limits.cluster = 0;
1667
1668         /*
1669          * set a reasonable default alignment on word boundaries: the
1670          * host and device may alter it using
1671          * blk_queue_update_dma_alignment() later.
1672          */
1673         blk_queue_dma_alignment(q, 0x03);
1674
1675         return q;
1676 }
1677 EXPORT_SYMBOL(__scsi_alloc_queue);
1678
1679 struct request_queue *scsi_alloc_queue(struct scsi_device *sdev)
1680 {
1681         struct request_queue *q;
1682
1683         q = __scsi_alloc_queue(sdev->host, scsi_request_fn);
1684         if (!q)
1685                 return NULL;
1686
1687         blk_queue_prep_rq(q, scsi_prep_fn);
1688         blk_queue_softirq_done(q, scsi_softirq_done);
1689         blk_queue_rq_timed_out(q, scsi_times_out);
1690         blk_queue_lld_busy(q, scsi_lld_busy);
1691         return q;
1692 }
1693
1694 void scsi_free_queue(struct request_queue *q)
1695 {
1696         blk_cleanup_queue(q);
1697 }
1698
1699 /*
1700  * Function:    scsi_block_requests()
1701  *
1702  * Purpose:     Utility function used by low-level drivers to prevent further
1703  *              commands from being queued to the device.
1704  *
1705  * Arguments:   shost       - Host in question
1706  *
1707  * Returns:     Nothing
1708  *
1709  * Lock status: No locks are assumed held.
1710  *
1711  * Notes:       There is no timer nor any other means by which the requests
1712  *              get unblocked other than the low-level driver calling
1713  *              scsi_unblock_requests().
1714  */
1715 void scsi_block_requests(struct Scsi_Host *shost)
1716 {
1717         shost->host_self_blocked = 1;
1718 }
1719 EXPORT_SYMBOL(scsi_block_requests);
1720
1721 /*
1722  * Function:    scsi_unblock_requests()
1723  *
1724  * Purpose:     Utility function used by low-level drivers to allow further
1725  *              commands from being queued to the device.
1726  *
1727  * Arguments:   shost       - Host in question
1728  *
1729  * Returns:     Nothing
1730  *
1731  * Lock status: No locks are assumed held.
1732  *
1733  * Notes:       There is no timer nor any other means by which the requests
1734  *              get unblocked other than the low-level driver calling
1735  *              scsi_unblock_requests().
1736  *
1737  *              This is done as an API function so that changes to the
1738  *              internals of the scsi mid-layer won't require wholesale
1739  *              changes to drivers that use this feature.
1740  */
1741 void scsi_unblock_requests(struct Scsi_Host *shost)
1742 {
1743         shost->host_self_blocked = 0;
1744         scsi_run_host_queues(shost);
1745 }
1746 EXPORT_SYMBOL(scsi_unblock_requests);
1747
1748 int __init scsi_init_queue(void)
1749 {
1750         int i;
1751
1752         scsi_sdb_cache = kmem_cache_create("scsi_data_buffer",
1753                                            sizeof(struct scsi_data_buffer),
1754                                            0, 0, NULL);
1755         if (!scsi_sdb_cache) {
1756                 printk(KERN_ERR "SCSI: can't init scsi sdb cache\n");
1757                 return -ENOMEM;
1758         }
1759
1760         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1761                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1762                 int size = sgp->size * sizeof(struct scatterlist);
1763
1764                 sgp->slab = kmem_cache_create(sgp->name, size, 0,
1765                                 SLAB_HWCACHE_ALIGN, NULL);
1766                 if (!sgp->slab) {
1767                         printk(KERN_ERR "SCSI: can't init sg slab %s\n",
1768                                         sgp->name);
1769                         goto cleanup_sdb;
1770                 }
1771
1772                 sgp->pool = mempool_create_slab_pool(SG_MEMPOOL_SIZE,
1773                                                      sgp->slab);
1774                 if (!sgp->pool) {
1775                         printk(KERN_ERR "SCSI: can't init sg mempool %s\n",
1776                                         sgp->name);
1777                         goto cleanup_sdb;
1778                 }
1779         }
1780
1781         return 0;
1782
1783 cleanup_sdb:
1784         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1785                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1786                 if (sgp->pool)
1787                         mempool_destroy(sgp->pool);
1788                 if (sgp->slab)
1789                         kmem_cache_destroy(sgp->slab);
1790         }
1791         kmem_cache_destroy(scsi_sdb_cache);
1792
1793         return -ENOMEM;
1794 }
1795
1796 void scsi_exit_queue(void)
1797 {
1798         int i;
1799
1800         kmem_cache_destroy(scsi_sdb_cache);
1801
1802         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1803                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1804                 mempool_destroy(sgp->pool);
1805                 kmem_cache_destroy(sgp->slab);
1806         }
1807 }
1808
1809 /**
1810  *      scsi_mode_select - issue a mode select
1811  *      @sdev:  SCSI device to be queried
1812  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
1813  *      @sp:    Save page bit (0 == don't save, 1 == save)
1814  *      @modepage: mode page being requested
1815  *      @buffer: request buffer (may not be smaller than eight bytes)
1816  *      @len:   length of request buffer.
1817  *      @timeout: command timeout
1818  *      @retries: number of retries before failing
1819  *      @data: returns a structure abstracting the mode header data
1820  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
1821  *              must be SCSI_SENSE_BUFFERSIZE big.
1822  *
1823  *      Returns zero if successful; negative error number or scsi
1824  *      status on error
1825  *
1826  */
1827 int
1828 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
1829                  unsigned char *buffer, int len, int timeout, int retries,
1830                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1831 {
1832         unsigned char cmd[10];
1833         unsigned char *real_buffer;
1834         int ret;
1835
1836         memset(cmd, 0, sizeof(cmd));
1837         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
1838
1839         if (sdev->use_10_for_ms) {
1840                 if (len > 65535)
1841                         return -EINVAL;
1842                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
1843                 if (!real_buffer)
1844                         return -ENOMEM;
1845                 memcpy(real_buffer + 8, buffer, len);
1846                 len += 8;
1847                 real_buffer[0] = 0;
1848                 real_buffer[1] = 0;
1849                 real_buffer[2] = data->medium_type;
1850                 real_buffer[3] = data->device_specific;
1851                 real_buffer[4] = data->longlba ? 0x01 : 0;
1852                 real_buffer[5] = 0;
1853                 real_buffer[6] = data->block_descriptor_length >> 8;
1854                 real_buffer[7] = data->block_descriptor_length;
1855
1856                 cmd[0] = MODE_SELECT_10;
1857                 cmd[7] = len >> 8;
1858                 cmd[8] = len;
1859         } else {
1860                 if (len > 255 || data->block_descriptor_length > 255 ||
1861                     data->longlba)
1862                         return -EINVAL;
1863
1864                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
1865                 if (!real_buffer)
1866                         return -ENOMEM;
1867                 memcpy(real_buffer + 4, buffer, len);
1868                 len += 4;
1869                 real_buffer[0] = 0;
1870                 real_buffer[1] = data->medium_type;
1871                 real_buffer[2] = data->device_specific;
1872                 real_buffer[3] = data->block_descriptor_length;
1873                 
1874
1875                 cmd[0] = MODE_SELECT;
1876                 cmd[4] = len;
1877         }
1878
1879         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
1880                                sshdr, timeout, retries, NULL);
1881         kfree(real_buffer);
1882         return ret;
1883 }
1884 EXPORT_SYMBOL_GPL(scsi_mode_select);
1885
1886 /**
1887  *      scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
1888  *      @sdev:  SCSI device to be queried
1889  *      @dbd:   set if mode sense will allow block descriptors to be returned
1890  *      @modepage: mode page being requested
1891  *      @buffer: request buffer (may not be smaller than eight bytes)
1892  *      @len:   length of request buffer.
1893  *      @timeout: command timeout
1894  *      @retries: number of retries before failing
1895  *      @data: returns a structure abstracting the mode header data
1896  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
1897  *              must be SCSI_SENSE_BUFFERSIZE big.
1898  *
1899  *      Returns zero if unsuccessful, or the header offset (either 4
1900  *      or 8 depending on whether a six or ten byte command was
1901  *      issued) if successful.
1902  */
1903 int
1904 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
1905                   unsigned char *buffer, int len, int timeout, int retries,
1906                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1907 {
1908         unsigned char cmd[12];
1909         int use_10_for_ms;
1910         int header_length;
1911         int result;
1912         struct scsi_sense_hdr my_sshdr;
1913
1914         memset(data, 0, sizeof(*data));
1915         memset(&cmd[0], 0, 12);
1916         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
1917         cmd[2] = modepage;
1918
1919         /* caller might not be interested in sense, but we need it */
1920         if (!sshdr)
1921                 sshdr = &my_sshdr;
1922
1923  retry:
1924         use_10_for_ms = sdev->use_10_for_ms;
1925
1926         if (use_10_for_ms) {
1927                 if (len < 8)
1928                         len = 8;
1929
1930                 cmd[0] = MODE_SENSE_10;
1931                 cmd[8] = len;
1932                 header_length = 8;
1933         } else {
1934                 if (len < 4)
1935                         len = 4;
1936
1937                 cmd[0] = MODE_SENSE;
1938                 cmd[4] = len;
1939                 header_length = 4;
1940         }
1941
1942         memset(buffer, 0, len);
1943
1944         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
1945                                   sshdr, timeout, retries, NULL);
1946
1947         /* This code looks awful: what it's doing is making sure an
1948          * ILLEGAL REQUEST sense return identifies the actual command
1949          * byte as the problem.  MODE_SENSE commands can return
1950          * ILLEGAL REQUEST if the code page isn't supported */
1951
1952         if (use_10_for_ms && !scsi_status_is_good(result) &&
1953             (driver_byte(result) & DRIVER_SENSE)) {
1954                 if (scsi_sense_valid(sshdr)) {
1955                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
1956                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
1957                                 /* 
1958                                  * Invalid command operation code
1959                                  */
1960                                 sdev->use_10_for_ms = 0;
1961                                 goto retry;
1962                         }
1963                 }
1964         }
1965
1966         if(scsi_status_is_good(result)) {
1967                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
1968                              (modepage == 6 || modepage == 8))) {
1969                         /* Initio breakage? */
1970                         header_length = 0;
1971                         data->length = 13;
1972                         data->medium_type = 0;
1973                         data->device_specific = 0;
1974                         data->longlba = 0;
1975                         data->block_descriptor_length = 0;
1976                 } else if(use_10_for_ms) {
1977                         data->length = buffer[0]*256 + buffer[1] + 2;
1978                         data->medium_type = buffer[2];
1979                         data->device_specific = buffer[3];
1980                         data->longlba = buffer[4] & 0x01;
1981                         data->block_descriptor_length = buffer[6]*256
1982                                 + buffer[7];
1983                 } else {
1984                         data->length = buffer[0] + 1;
1985                         data->medium_type = buffer[1];
1986                         data->device_specific = buffer[2];
1987                         data->block_descriptor_length = buffer[3];
1988                 }
1989                 data->header_length = header_length;
1990         }
1991
1992         return result;
1993 }
1994 EXPORT_SYMBOL(scsi_mode_sense);
1995
1996 /**
1997  *      scsi_test_unit_ready - test if unit is ready
1998  *      @sdev:  scsi device to change the state of.
1999  *      @timeout: command timeout
2000  *      @retries: number of retries before failing
2001  *      @sshdr_external: Optional pointer to struct scsi_sense_hdr for
2002  *              returning sense. Make sure that this is cleared before passing
2003  *              in.
2004  *
2005  *      Returns zero if unsuccessful or an error if TUR failed.  For
2006  *      removable media, UNIT_ATTENTION sets ->changed flag.
2007  **/
2008 int
2009 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
2010                      struct scsi_sense_hdr *sshdr_external)
2011 {
2012         char cmd[] = {
2013                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
2014         };
2015         struct scsi_sense_hdr *sshdr;
2016         int result;
2017
2018         if (!sshdr_external)
2019                 sshdr = kzalloc(sizeof(*sshdr), GFP_KERNEL);
2020         else
2021                 sshdr = sshdr_external;
2022
2023         /* try to eat the UNIT_ATTENTION if there are enough retries */
2024         do {
2025                 result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
2026                                           timeout, retries, NULL);
2027                 if (sdev->removable && scsi_sense_valid(sshdr) &&
2028                     sshdr->sense_key == UNIT_ATTENTION)
2029                         sdev->changed = 1;
2030         } while (scsi_sense_valid(sshdr) &&
2031                  sshdr->sense_key == UNIT_ATTENTION && --retries);
2032
2033         if (!sshdr_external)
2034                 kfree(sshdr);
2035         return result;
2036 }
2037 EXPORT_SYMBOL(scsi_test_unit_ready);
2038
2039 /**
2040  *      scsi_device_set_state - Take the given device through the device state model.
2041  *      @sdev:  scsi device to change the state of.
2042  *      @state: state to change to.
2043  *
2044  *      Returns zero if unsuccessful or an error if the requested 
2045  *      transition is illegal.
2046  */
2047 int
2048 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2049 {
2050         enum scsi_device_state oldstate = sdev->sdev_state;
2051
2052         if (state == oldstate)
2053                 return 0;
2054
2055         switch (state) {
2056         case SDEV_CREATED:
2057                 switch (oldstate) {
2058                 case SDEV_CREATED_BLOCK:
2059                         break;
2060                 default:
2061                         goto illegal;
2062                 }
2063                 break;
2064                         
2065         case SDEV_RUNNING:
2066                 switch (oldstate) {
2067                 case SDEV_CREATED:
2068                 case SDEV_OFFLINE:
2069                 case SDEV_QUIESCE:
2070                 case SDEV_BLOCK:
2071                         break;
2072                 default:
2073                         goto illegal;
2074                 }
2075                 break;
2076
2077         case SDEV_QUIESCE:
2078                 switch (oldstate) {
2079                 case SDEV_RUNNING:
2080                 case SDEV_OFFLINE:
2081                         break;
2082                 default:
2083                         goto illegal;
2084                 }
2085                 break;
2086
2087         case SDEV_OFFLINE:
2088                 switch (oldstate) {
2089                 case SDEV_CREATED:
2090                 case SDEV_RUNNING:
2091                 case SDEV_QUIESCE:
2092                 case SDEV_BLOCK:
2093                         break;
2094                 default:
2095                         goto illegal;
2096                 }
2097                 break;
2098
2099         case SDEV_BLOCK:
2100                 switch (oldstate) {
2101                 case SDEV_RUNNING:
2102                 case SDEV_CREATED_BLOCK:
2103                         break;
2104                 default:
2105                         goto illegal;
2106                 }
2107                 break;
2108
2109         case SDEV_CREATED_BLOCK:
2110                 switch (oldstate) {
2111                 case SDEV_CREATED:
2112                         break;
2113                 default:
2114                         goto illegal;
2115                 }
2116                 break;
2117
2118         case SDEV_CANCEL:
2119                 switch (oldstate) {
2120                 case SDEV_CREATED:
2121                 case SDEV_RUNNING:
2122                 case SDEV_QUIESCE:
2123                 case SDEV_OFFLINE:
2124                 case SDEV_BLOCK:
2125                         break;
2126                 default:
2127                         goto illegal;
2128                 }
2129                 break;
2130
2131         case SDEV_DEL:
2132                 switch (oldstate) {
2133                 case SDEV_CREATED:
2134                 case SDEV_RUNNING:
2135                 case SDEV_OFFLINE:
2136                 case SDEV_CANCEL:
2137                         break;
2138                 default:
2139                         goto illegal;
2140                 }
2141                 break;
2142
2143         }
2144         sdev->sdev_state = state;
2145         return 0;
2146
2147  illegal:
2148         SCSI_LOG_ERROR_RECOVERY(1, 
2149                                 sdev_printk(KERN_ERR, sdev,
2150                                             "Illegal state transition %s->%s\n",
2151                                             scsi_device_state_name(oldstate),
2152                                             scsi_device_state_name(state))
2153                                 );
2154         return -EINVAL;
2155 }
2156 EXPORT_SYMBOL(scsi_device_set_state);
2157
2158 /**
2159  *      sdev_evt_emit - emit a single SCSI device uevent
2160  *      @sdev: associated SCSI device
2161  *      @evt: event to emit
2162  *
2163  *      Send a single uevent (scsi_event) to the associated scsi_device.
2164  */
2165 static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2166 {
2167         int idx = 0;
2168         char *envp[3];
2169
2170         switch (evt->evt_type) {
2171         case SDEV_EVT_MEDIA_CHANGE:
2172                 envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2173                 break;
2174
2175         default:
2176                 /* do nothing */
2177                 break;
2178         }
2179
2180         envp[idx++] = NULL;
2181
2182         kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2183 }
2184
2185 /**
2186  *      sdev_evt_thread - send a uevent for each scsi event
2187  *      @work: work struct for scsi_device
2188  *
2189  *      Dispatch queued events to their associated scsi_device kobjects
2190  *      as uevents.
2191  */
2192 void scsi_evt_thread(struct work_struct *work)
2193 {
2194         struct scsi_device *sdev;
2195         LIST_HEAD(event_list);
2196
2197         sdev = container_of(work, struct scsi_device, event_work);
2198
2199         while (1) {
2200                 struct scsi_event *evt;
2201                 struct list_head *this, *tmp;
2202                 unsigned long flags;
2203
2204                 spin_lock_irqsave(&sdev->list_lock, flags);
2205                 list_splice_init(&sdev->event_list, &event_list);
2206                 spin_unlock_irqrestore(&sdev->list_lock, flags);
2207
2208                 if (list_empty(&event_list))
2209                         break;
2210
2211                 list_for_each_safe(this, tmp, &event_list) {
2212                         evt = list_entry(this, struct scsi_event, node);
2213                         list_del(&evt->node);
2214                         scsi_evt_emit(sdev, evt);
2215                         kfree(evt);
2216                 }
2217         }
2218 }
2219
2220 /**
2221  *      sdev_evt_send - send asserted event to uevent thread
2222  *      @sdev: scsi_device event occurred on
2223  *      @evt: event to send
2224  *
2225  *      Assert scsi device event asynchronously.
2226  */
2227 void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2228 {
2229         unsigned long flags;
2230
2231 #if 0
2232         /* FIXME: currently this check eliminates all media change events
2233          * for polled devices.  Need to update to discriminate between AN
2234          * and polled events */
2235         if (!test_bit(evt->evt_type, sdev->supported_events)) {
2236                 kfree(evt);
2237                 return;
2238         }
2239 #endif
2240
2241         spin_lock_irqsave(&sdev->list_lock, flags);
2242         list_add_tail(&evt->node, &sdev->event_list);
2243         schedule_work(&sdev->event_work);
2244         spin_unlock_irqrestore(&sdev->list_lock, flags);
2245 }
2246 EXPORT_SYMBOL_GPL(sdev_evt_send);
2247
2248 /**
2249  *      sdev_evt_alloc - allocate a new scsi event
2250  *      @evt_type: type of event to allocate
2251  *      @gfpflags: GFP flags for allocation
2252  *
2253  *      Allocates and returns a new scsi_event.
2254  */
2255 struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2256                                   gfp_t gfpflags)
2257 {
2258         struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2259         if (!evt)
2260                 return NULL;
2261
2262         evt->evt_type = evt_type;
2263         INIT_LIST_HEAD(&evt->node);
2264
2265         /* evt_type-specific initialization, if any */
2266         switch (evt_type) {
2267         case SDEV_EVT_MEDIA_CHANGE:
2268         default:
2269                 /* do nothing */
2270                 break;
2271         }
2272
2273         return evt;
2274 }
2275 EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2276
2277 /**
2278  *      sdev_evt_send_simple - send asserted event to uevent thread
2279  *      @sdev: scsi_device event occurred on
2280  *      @evt_type: type of event to send
2281  *      @gfpflags: GFP flags for allocation
2282  *
2283  *      Assert scsi device event asynchronously, given an event type.
2284  */
2285 void sdev_evt_send_simple(struct scsi_device *sdev,
2286                           enum scsi_device_event evt_type, gfp_t gfpflags)
2287 {
2288         struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2289         if (!evt) {
2290                 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2291                             evt_type);
2292                 return;
2293         }
2294
2295         sdev_evt_send(sdev, evt);
2296 }
2297 EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2298
2299 /**
2300  *      scsi_device_quiesce - Block user issued commands.
2301  *      @sdev:  scsi device to quiesce.
2302  *
2303  *      This works by trying to transition to the SDEV_QUIESCE state
2304  *      (which must be a legal transition).  When the device is in this
2305  *      state, only special requests will be accepted, all others will
2306  *      be deferred.  Since special requests may also be requeued requests,
2307  *      a successful return doesn't guarantee the device will be 
2308  *      totally quiescent.
2309  *
2310  *      Must be called with user context, may sleep.
2311  *
2312  *      Returns zero if unsuccessful or an error if not.
2313  */
2314 int
2315 scsi_device_quiesce(struct scsi_device *sdev)
2316 {
2317         int err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2318         if (err)
2319                 return err;
2320
2321         scsi_run_queue(sdev->request_queue);
2322         while (sdev->device_busy) {
2323                 msleep_interruptible(200);
2324                 scsi_run_queue(sdev->request_queue);
2325         }
2326         return 0;
2327 }
2328 EXPORT_SYMBOL(scsi_device_quiesce);
2329
2330 /**
2331  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2332  *      @sdev:  scsi device to resume.
2333  *
2334  *      Moves the device from quiesced back to running and restarts the
2335  *      queues.
2336  *
2337  *      Must be called with user context, may sleep.
2338  */
2339 void
2340 scsi_device_resume(struct scsi_device *sdev)
2341 {
2342         if(scsi_device_set_state(sdev, SDEV_RUNNING))
2343                 return;
2344         scsi_run_queue(sdev->request_queue);
2345 }
2346 EXPORT_SYMBOL(scsi_device_resume);
2347
2348 static void
2349 device_quiesce_fn(struct scsi_device *sdev, void *data)
2350 {
2351         scsi_device_quiesce(sdev);
2352 }
2353
2354 void
2355 scsi_target_quiesce(struct scsi_target *starget)
2356 {
2357         starget_for_each_device(starget, NULL, device_quiesce_fn);
2358 }
2359 EXPORT_SYMBOL(scsi_target_quiesce);
2360
2361 static void
2362 device_resume_fn(struct scsi_device *sdev, void *data)
2363 {
2364         scsi_device_resume(sdev);
2365 }
2366
2367 void
2368 scsi_target_resume(struct scsi_target *starget)
2369 {
2370         starget_for_each_device(starget, NULL, device_resume_fn);
2371 }
2372 EXPORT_SYMBOL(scsi_target_resume);
2373
2374 /**
2375  * scsi_internal_device_block - internal function to put a device temporarily into the SDEV_BLOCK state
2376  * @sdev:       device to block
2377  *
2378  * Block request made by scsi lld's to temporarily stop all
2379  * scsi commands on the specified device.  Called from interrupt
2380  * or normal process context.
2381  *
2382  * Returns zero if successful or error if not
2383  *
2384  * Notes:       
2385  *      This routine transitions the device to the SDEV_BLOCK state
2386  *      (which must be a legal transition).  When the device is in this
2387  *      state, all commands are deferred until the scsi lld reenables
2388  *      the device with scsi_device_unblock or device_block_tmo fires.
2389  *      This routine assumes the host_lock is held on entry.
2390  */
2391 int
2392 scsi_internal_device_block(struct scsi_device *sdev)
2393 {
2394         struct request_queue *q = sdev->request_queue;
2395         unsigned long flags;
2396         int err = 0;
2397
2398         err = scsi_device_set_state(sdev, SDEV_BLOCK);
2399         if (err) {
2400                 err = scsi_device_set_state(sdev, SDEV_CREATED_BLOCK);
2401
2402                 if (err)
2403                         return err;
2404         }
2405
2406         /* 
2407          * The device has transitioned to SDEV_BLOCK.  Stop the
2408          * block layer from calling the midlayer with this device's
2409          * request queue. 
2410          */
2411         spin_lock_irqsave(q->queue_lock, flags);
2412         blk_stop_queue(q);
2413         spin_unlock_irqrestore(q->queue_lock, flags);
2414
2415         return 0;
2416 }
2417 EXPORT_SYMBOL_GPL(scsi_internal_device_block);
2418  
2419 /**
2420  * scsi_internal_device_unblock - resume a device after a block request
2421  * @sdev:       device to resume
2422  *
2423  * Called by scsi lld's or the midlayer to restart the device queue
2424  * for the previously suspended scsi device.  Called from interrupt or
2425  * normal process context.
2426  *
2427  * Returns zero if successful or error if not.
2428  *
2429  * Notes:       
2430  *      This routine transitions the device to the SDEV_RUNNING state
2431  *      (which must be a legal transition) allowing the midlayer to
2432  *      goose the queue for this device.  This routine assumes the 
2433  *      host_lock is held upon entry.
2434  */
2435 int
2436 scsi_internal_device_unblock(struct scsi_device *sdev)
2437 {
2438         struct request_queue *q = sdev->request_queue; 
2439         unsigned long flags;
2440         
2441         /* 
2442          * Try to transition the scsi device to SDEV_RUNNING
2443          * and goose the device queue if successful.  
2444          */
2445         if (sdev->sdev_state == SDEV_BLOCK)
2446                 sdev->sdev_state = SDEV_RUNNING;
2447         else if (sdev->sdev_state == SDEV_CREATED_BLOCK)
2448                 sdev->sdev_state = SDEV_CREATED;
2449         else if (sdev->sdev_state != SDEV_CANCEL &&
2450                  sdev->sdev_state != SDEV_OFFLINE)
2451                 return -EINVAL;
2452
2453         spin_lock_irqsave(q->queue_lock, flags);
2454         blk_start_queue(q);
2455         spin_unlock_irqrestore(q->queue_lock, flags);
2456
2457         return 0;
2458 }
2459 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock);
2460
2461 static void
2462 device_block(struct scsi_device *sdev, void *data)
2463 {
2464         scsi_internal_device_block(sdev);
2465 }
2466
2467 static int
2468 target_block(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_block);
2473         return 0;
2474 }
2475
2476 void
2477 scsi_target_block(struct device *dev)
2478 {
2479         if (scsi_is_target_device(dev))
2480                 starget_for_each_device(to_scsi_target(dev), NULL,
2481                                         device_block);
2482         else
2483                 device_for_each_child(dev, NULL, target_block);
2484 }
2485 EXPORT_SYMBOL_GPL(scsi_target_block);
2486
2487 static void
2488 device_unblock(struct scsi_device *sdev, void *data)
2489 {
2490         scsi_internal_device_unblock(sdev);
2491 }
2492
2493 static int
2494 target_unblock(struct device *dev, void *data)
2495 {
2496         if (scsi_is_target_device(dev))
2497                 starget_for_each_device(to_scsi_target(dev), NULL,
2498                                         device_unblock);
2499         return 0;
2500 }
2501
2502 void
2503 scsi_target_unblock(struct device *dev)
2504 {
2505         if (scsi_is_target_device(dev))
2506                 starget_for_each_device(to_scsi_target(dev), NULL,
2507                                         device_unblock);
2508         else
2509                 device_for_each_child(dev, NULL, target_unblock);
2510 }
2511 EXPORT_SYMBOL_GPL(scsi_target_unblock);
2512
2513 /**
2514  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2515  * @sgl:        scatter-gather list
2516  * @sg_count:   number of segments in sg
2517  * @offset:     offset in bytes into sg, on return offset into the mapped area
2518  * @len:        bytes to map, on return number of bytes mapped
2519  *
2520  * Returns virtual address of the start of the mapped page
2521  */
2522 void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
2523                           size_t *offset, size_t *len)
2524 {
2525         int i;
2526         size_t sg_len = 0, len_complete = 0;
2527         struct scatterlist *sg;
2528         struct page *page;
2529
2530         WARN_ON(!irqs_disabled());
2531
2532         for_each_sg(sgl, sg, sg_count, i) {
2533                 len_complete = sg_len; /* Complete sg-entries */
2534                 sg_len += sg->length;
2535                 if (sg_len > *offset)
2536                         break;
2537         }
2538
2539         if (unlikely(i == sg_count)) {
2540                 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2541                         "elements %d\n",
2542                        __func__, sg_len, *offset, sg_count);
2543                 WARN_ON(1);
2544                 return NULL;
2545         }
2546
2547         /* Offset starting from the beginning of first page in this sg-entry */
2548         *offset = *offset - len_complete + sg->offset;
2549
2550         /* Assumption: contiguous pages can be accessed as "page + i" */
2551         page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
2552         *offset &= ~PAGE_MASK;
2553
2554         /* Bytes in this sg-entry from *offset to the end of the page */
2555         sg_len = PAGE_SIZE - *offset;
2556         if (*len > sg_len)
2557                 *len = sg_len;
2558
2559         return kmap_atomic(page, KM_BIO_SRC_IRQ);
2560 }
2561 EXPORT_SYMBOL(scsi_kmap_atomic_sg);
2562
2563 /**
2564  * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
2565  * @virt:       virtual address to be unmapped
2566  */
2567 void scsi_kunmap_atomic_sg(void *virt)
2568 {
2569         kunmap_atomic(virt, KM_BIO_SRC_IRQ);
2570 }
2571 EXPORT_SYMBOL(scsi_kunmap_atomic_sg);