Correct .gbs.conf settings
[platform/adaptation/renesas_rcar/renesas_kernel.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/export.h>
16 #include <linux/mempool.h>
17 #include <linux/slab.h>
18 #include <linux/init.h>
19 #include <linux/pci.h>
20 #include <linux/delay.h>
21 #include <linux/hardirq.h>
22 #include <linux/scatterlist.h>
23
24 #include <scsi/scsi.h>
25 #include <scsi/scsi_cmnd.h>
26 #include <scsi/scsi_dbg.h>
27 #include <scsi/scsi_device.h>
28 #include <scsi/scsi_driver.h>
29 #include <scsi/scsi_eh.h>
30 #include <scsi/scsi_host.h>
31
32 #include "scsi_priv.h"
33 #include "scsi_logging.h"
34
35
36 #define SG_MEMPOOL_NR           ARRAY_SIZE(scsi_sg_pools)
37 #define SG_MEMPOOL_SIZE         2
38
39 struct scsi_host_sg_pool {
40         size_t          size;
41         char            *name;
42         struct kmem_cache       *slab;
43         mempool_t       *pool;
44 };
45
46 #define SP(x) { x, "sgpool-" __stringify(x) }
47 #if (SCSI_MAX_SG_SEGMENTS < 32)
48 #error SCSI_MAX_SG_SEGMENTS is too small (must be 32 or greater)
49 #endif
50 static struct scsi_host_sg_pool scsi_sg_pools[] = {
51         SP(8),
52         SP(16),
53 #if (SCSI_MAX_SG_SEGMENTS > 32)
54         SP(32),
55 #if (SCSI_MAX_SG_SEGMENTS > 64)
56         SP(64),
57 #if (SCSI_MAX_SG_SEGMENTS > 128)
58         SP(128),
59 #if (SCSI_MAX_SG_SEGMENTS > 256)
60 #error SCSI_MAX_SG_SEGMENTS is too large (256 MAX)
61 #endif
62 #endif
63 #endif
64 #endif
65         SP(SCSI_MAX_SG_SEGMENTS)
66 };
67 #undef SP
68
69 struct kmem_cache *scsi_sdb_cache;
70
71 /*
72  * When to reinvoke queueing after a resource shortage. It's 3 msecs to
73  * not change behaviour from the previous unplug mechanism, experimentation
74  * may prove this needs changing.
75  */
76 #define SCSI_QUEUE_DELAY        3
77
78 /*
79  * Function:    scsi_unprep_request()
80  *
81  * Purpose:     Remove all preparation done for a request, including its
82  *              associated scsi_cmnd, so that it can be requeued.
83  *
84  * Arguments:   req     - request to unprepare
85  *
86  * Lock status: Assumed that no locks are held upon entry.
87  *
88  * Returns:     Nothing.
89  */
90 static void scsi_unprep_request(struct request *req)
91 {
92         struct scsi_cmnd *cmd = req->special;
93
94         blk_unprep_request(req);
95         req->special = NULL;
96
97         scsi_put_command(cmd);
98 }
99
100 /**
101  * __scsi_queue_insert - private queue insertion
102  * @cmd: The SCSI command being requeued
103  * @reason:  The reason for the requeue
104  * @unbusy: Whether the queue should be unbusied
105  *
106  * This is a private queue insertion.  The public interface
107  * scsi_queue_insert() always assumes the queue should be unbusied
108  * because it's always called before the completion.  This function is
109  * for a requeue after completion, which should only occur in this
110  * file.
111  */
112 static void __scsi_queue_insert(struct scsi_cmnd *cmd, int reason, int unbusy)
113 {
114         struct Scsi_Host *host = cmd->device->host;
115         struct scsi_device *device = cmd->device;
116         struct scsi_target *starget = scsi_target(device);
117         struct request_queue *q = device->request_queue;
118         unsigned long flags;
119
120         SCSI_LOG_MLQUEUE(1,
121                  printk("Inserting command %p into mlqueue\n", cmd));
122
123         /*
124          * Set the appropriate busy bit for the device/host.
125          *
126          * If the host/device isn't busy, assume that something actually
127          * completed, and that we should be able to queue a command now.
128          *
129          * Note that the prior mid-layer assumption that any host could
130          * always queue at least one command is now broken.  The mid-layer
131          * will implement a user specifiable stall (see
132          * scsi_host.max_host_blocked and scsi_device.max_device_blocked)
133          * if a command is requeued with no other commands outstanding
134          * either for the device or for the host.
135          */
136         switch (reason) {
137         case SCSI_MLQUEUE_HOST_BUSY:
138                 host->host_blocked = host->max_host_blocked;
139                 break;
140         case SCSI_MLQUEUE_DEVICE_BUSY:
141         case SCSI_MLQUEUE_EH_RETRY:
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. Schedule requeue work under
159          * lock such that the kblockd_schedule_work() call happens
160          * before blk_cleanup_queue() finishes.
161          */
162         spin_lock_irqsave(q->queue_lock, flags);
163         blk_requeue_request(q, cmd->request);
164         kblockd_schedule_work(q, &device->requeue_work);
165         spin_unlock_irqrestore(q->queue_lock, flags);
166 }
167
168 /*
169  * Function:    scsi_queue_insert()
170  *
171  * Purpose:     Insert a command in the midlevel queue.
172  *
173  * Arguments:   cmd    - command that we are adding to queue.
174  *              reason - why we are inserting command to queue.
175  *
176  * Lock status: Assumed that lock is not held upon entry.
177  *
178  * Returns:     Nothing.
179  *
180  * Notes:       We do this for one of two cases.  Either the host is busy
181  *              and it cannot accept any more commands for the time being,
182  *              or the device returned QUEUE_FULL and can accept no more
183  *              commands.
184  * Notes:       This could be called either from an interrupt context or a
185  *              normal process context.
186  */
187 void scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
188 {
189         __scsi_queue_insert(cmd, reason, 1);
190 }
191 /**
192  * scsi_execute - insert request and wait for the result
193  * @sdev:       scsi device
194  * @cmd:        scsi command
195  * @data_direction: data direction
196  * @buffer:     data buffer
197  * @bufflen:    len of buffer
198  * @sense:      optional sense buffer
199  * @timeout:    request timeout in seconds
200  * @retries:    number of times to retry request
201  * @flags:      or into request flags;
202  * @resid:      optional residual length
203  *
204  * returns the req->errors value which is the scsi_cmnd result
205  * field.
206  */
207 int scsi_execute(struct scsi_device *sdev, const unsigned char *cmd,
208                  int data_direction, void *buffer, unsigned bufflen,
209                  unsigned char *sense, int timeout, int retries, int flags,
210                  int *resid)
211 {
212         struct request *req;
213         int write = (data_direction == DMA_TO_DEVICE);
214         int ret = DRIVER_ERROR << 24;
215
216         req = blk_get_request(sdev->request_queue, write, __GFP_WAIT);
217         if (!req)
218                 return ret;
219
220         if (bufflen &&  blk_rq_map_kern(sdev->request_queue, req,
221                                         buffer, bufflen, __GFP_WAIT))
222                 goto out;
223
224         req->cmd_len = COMMAND_SIZE(cmd[0]);
225         memcpy(req->cmd, cmd, req->cmd_len);
226         req->sense = sense;
227         req->sense_len = 0;
228         req->retries = retries;
229         req->timeout = timeout;
230         req->cmd_type = REQ_TYPE_BLOCK_PC;
231         req->cmd_flags |= flags | REQ_QUIET | REQ_PREEMPT;
232
233         /*
234          * head injection *required* here otherwise quiesce won't work
235          */
236         blk_execute_rq(req->q, NULL, req, 1);
237
238         /*
239          * Some devices (USB mass-storage in particular) may transfer
240          * garbage data together with a residue indicating that the data
241          * is invalid.  Prevent the garbage from being misinterpreted
242          * and prevent security leaks by zeroing out the excess data.
243          */
244         if (unlikely(req->resid_len > 0 && req->resid_len <= bufflen))
245                 memset(buffer + (bufflen - req->resid_len), 0, req->resid_len);
246
247         if (resid)
248                 *resid = req->resid_len;
249         ret = req->errors;
250  out:
251         blk_put_request(req);
252
253         return ret;
254 }
255 EXPORT_SYMBOL(scsi_execute);
256
257 int scsi_execute_req_flags(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, int flags)
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, flags, 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_flags);
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;
404         LIST_HEAD(starved_list);
405         unsigned long flags;
406
407         shost = sdev->host;
408         if (scsi_target(sdev)->single_lun)
409                 scsi_single_lun_run(sdev);
410
411         spin_lock_irqsave(shost->host_lock, flags);
412         list_splice_init(&shost->starved_list, &starved_list);
413
414         while (!list_empty(&starved_list)) {
415                 struct request_queue *slq;
416
417                 /*
418                  * As long as shost is accepting commands and we have
419                  * starved queues, call blk_run_queue. scsi_request_fn
420                  * drops the queue_lock and can add us back to the
421                  * starved_list.
422                  *
423                  * host_lock protects the starved_list and starved_entry.
424                  * scsi_request_fn must get the host_lock before checking
425                  * or modifying starved_list or starved_entry.
426                  */
427                 if (scsi_host_is_busy(shost))
428                         break;
429
430                 sdev = list_entry(starved_list.next,
431                                   struct scsi_device, starved_entry);
432                 list_del_init(&sdev->starved_entry);
433                 if (scsi_target_is_busy(scsi_target(sdev))) {
434                         list_move_tail(&sdev->starved_entry,
435                                        &shost->starved_list);
436                         continue;
437                 }
438
439                 /*
440                  * Once we drop the host lock, a racing scsi_remove_device()
441                  * call may remove the sdev from the starved list and destroy
442                  * it and the queue.  Mitigate by taking a reference to the
443                  * queue and never touching the sdev again after we drop the
444                  * host lock.  Note: if __scsi_remove_device() invokes
445                  * blk_cleanup_queue() before the queue is run from this
446                  * function then blk_run_queue() will return immediately since
447                  * blk_cleanup_queue() marks the queue with QUEUE_FLAG_DYING.
448                  */
449                 slq = sdev->request_queue;
450                 if (!blk_get_queue(slq))
451                         continue;
452                 spin_unlock_irqrestore(shost->host_lock, flags);
453
454                 blk_run_queue(slq);
455                 blk_put_queue(slq);
456
457                 spin_lock_irqsave(shost->host_lock, flags);
458         }
459         /* put any unprocessed entries back */
460         list_splice(&starved_list, &shost->starved_list);
461         spin_unlock_irqrestore(shost->host_lock, flags);
462
463         blk_run_queue(q);
464 }
465
466 void scsi_requeue_run_queue(struct work_struct *work)
467 {
468         struct scsi_device *sdev;
469         struct request_queue *q;
470
471         sdev = container_of(work, struct scsi_device, requeue_work);
472         q = sdev->request_queue;
473         scsi_run_queue(q);
474 }
475
476 /*
477  * Function:    scsi_requeue_command()
478  *
479  * Purpose:     Handle post-processing of completed commands.
480  *
481  * Arguments:   q       - queue to operate on
482  *              cmd     - command that may need to be requeued.
483  *
484  * Returns:     Nothing
485  *
486  * Notes:       After command completion, there may be blocks left
487  *              over which weren't finished by the previous command
488  *              this can be for a number of reasons - the main one is
489  *              I/O errors in the middle of the request, in which case
490  *              we need to request the blocks that come after the bad
491  *              sector.
492  * Notes:       Upon return, cmd is a stale pointer.
493  */
494 static void scsi_requeue_command(struct request_queue *q, struct scsi_cmnd *cmd)
495 {
496         struct scsi_device *sdev = cmd->device;
497         struct request *req = cmd->request;
498         unsigned long flags;
499
500         /*
501          * We need to hold a reference on the device to avoid the queue being
502          * killed after the unlock and before scsi_run_queue is invoked which
503          * may happen because scsi_unprep_request() puts the command which
504          * releases its reference on the device.
505          */
506         get_device(&sdev->sdev_gendev);
507
508         spin_lock_irqsave(q->queue_lock, flags);
509         scsi_unprep_request(req);
510         blk_requeue_request(q, req);
511         spin_unlock_irqrestore(q->queue_lock, flags);
512
513         scsi_run_queue(q);
514
515         put_device(&sdev->sdev_gendev);
516 }
517
518 void scsi_next_command(struct scsi_cmnd *cmd)
519 {
520         struct scsi_device *sdev = cmd->device;
521         struct request_queue *q = sdev->request_queue;
522
523         /* need to hold a reference on the device before we let go of the cmd */
524         get_device(&sdev->sdev_gendev);
525
526         scsi_put_command(cmd);
527         scsi_run_queue(q);
528
529         /* ok to remove device now */
530         put_device(&sdev->sdev_gendev);
531 }
532
533 void scsi_run_host_queues(struct Scsi_Host *shost)
534 {
535         struct scsi_device *sdev;
536
537         shost_for_each_device(sdev, shost)
538                 scsi_run_queue(sdev->request_queue);
539 }
540
541 static void __scsi_release_buffers(struct scsi_cmnd *, int);
542
543 /*
544  * Function:    scsi_end_request()
545  *
546  * Purpose:     Post-processing of completed commands (usually invoked at end
547  *              of upper level post-processing and scsi_io_completion).
548  *
549  * Arguments:   cmd      - command that is complete.
550  *              error    - 0 if I/O indicates success, < 0 for I/O error.
551  *              bytes    - number of bytes of completed I/O
552  *              requeue  - indicates whether we should requeue leftovers.
553  *
554  * Lock status: Assumed that lock is not held upon entry.
555  *
556  * Returns:     cmd if requeue required, NULL otherwise.
557  *
558  * Notes:       This is called for block device requests in order to
559  *              mark some number of sectors as complete.
560  * 
561  *              We are guaranteeing that the request queue will be goosed
562  *              at some point during this call.
563  * Notes:       If cmd was requeued, upon return it will be a stale pointer.
564  */
565 static struct scsi_cmnd *scsi_end_request(struct scsi_cmnd *cmd, int error,
566                                           int bytes, int requeue)
567 {
568         struct request_queue *q = cmd->device->request_queue;
569         struct request *req = cmd->request;
570
571         /*
572          * If there are blocks left over at the end, set up the command
573          * to queue the remainder of them.
574          */
575         if (blk_end_request(req, error, bytes)) {
576                 /* kill remainder if no retrys */
577                 if (error && scsi_noretry_cmd(cmd))
578                         blk_end_request_all(req, error);
579                 else {
580                         if (requeue) {
581                                 /*
582                                  * Bleah.  Leftovers again.  Stick the
583                                  * leftovers in the front of the
584                                  * queue, and goose the queue again.
585                                  */
586                                 scsi_release_buffers(cmd);
587                                 scsi_requeue_command(q, cmd);
588                                 cmd = NULL;
589                         }
590                         return cmd;
591                 }
592         }
593
594         /*
595          * This will goose the queue request function at the end, so we don't
596          * need to worry about launching another command.
597          */
598         __scsi_release_buffers(cmd, 0);
599         scsi_next_command(cmd);
600         return NULL;
601 }
602
603 static inline unsigned int scsi_sgtable_index(unsigned short nents)
604 {
605         unsigned int index;
606
607         BUG_ON(nents > SCSI_MAX_SG_SEGMENTS);
608
609         if (nents <= 8)
610                 index = 0;
611         else
612                 index = get_count_order(nents) - 3;
613
614         return index;
615 }
616
617 static void scsi_sg_free(struct scatterlist *sgl, unsigned int nents)
618 {
619         struct scsi_host_sg_pool *sgp;
620
621         sgp = scsi_sg_pools + scsi_sgtable_index(nents);
622         mempool_free(sgl, sgp->pool);
623 }
624
625 static struct scatterlist *scsi_sg_alloc(unsigned int nents, gfp_t gfp_mask)
626 {
627         struct scsi_host_sg_pool *sgp;
628
629         sgp = scsi_sg_pools + scsi_sgtable_index(nents);
630         return mempool_alloc(sgp->pool, gfp_mask);
631 }
632
633 static int scsi_alloc_sgtable(struct scsi_data_buffer *sdb, int nents,
634                               gfp_t gfp_mask)
635 {
636         int ret;
637
638         BUG_ON(!nents);
639
640         ret = __sg_alloc_table(&sdb->table, nents, SCSI_MAX_SG_SEGMENTS,
641                                gfp_mask, scsi_sg_alloc);
642         if (unlikely(ret))
643                 __sg_free_table(&sdb->table, SCSI_MAX_SG_SEGMENTS,
644                                 scsi_sg_free);
645
646         return ret;
647 }
648
649 static void scsi_free_sgtable(struct scsi_data_buffer *sdb)
650 {
651         __sg_free_table(&sdb->table, SCSI_MAX_SG_SEGMENTS, scsi_sg_free);
652 }
653
654 static void __scsi_release_buffers(struct scsi_cmnd *cmd, int do_bidi_check)
655 {
656
657         if (cmd->sdb.table.nents)
658                 scsi_free_sgtable(&cmd->sdb);
659
660         memset(&cmd->sdb, 0, sizeof(cmd->sdb));
661
662         if (do_bidi_check && scsi_bidi_cmnd(cmd)) {
663                 struct scsi_data_buffer *bidi_sdb =
664                         cmd->request->next_rq->special;
665                 scsi_free_sgtable(bidi_sdb);
666                 kmem_cache_free(scsi_sdb_cache, bidi_sdb);
667                 cmd->request->next_rq->special = NULL;
668         }
669
670         if (scsi_prot_sg_count(cmd))
671                 scsi_free_sgtable(cmd->prot_sdb);
672 }
673
674 /*
675  * Function:    scsi_release_buffers()
676  *
677  * Purpose:     Completion processing for block device I/O requests.
678  *
679  * Arguments:   cmd     - command that we are bailing.
680  *
681  * Lock status: Assumed that no lock is held upon entry.
682  *
683  * Returns:     Nothing
684  *
685  * Notes:       In the event that an upper level driver rejects a
686  *              command, we must release resources allocated during
687  *              the __init_io() function.  Primarily this would involve
688  *              the scatter-gather table, and potentially any bounce
689  *              buffers.
690  */
691 void scsi_release_buffers(struct scsi_cmnd *cmd)
692 {
693         __scsi_release_buffers(cmd, 1);
694 }
695 EXPORT_SYMBOL(scsi_release_buffers);
696
697 /**
698  * __scsi_error_from_host_byte - translate SCSI error code into errno
699  * @cmd:        SCSI command (unused)
700  * @result:     scsi error code
701  *
702  * Translate SCSI error code into standard UNIX errno.
703  * Return values:
704  * -ENOLINK     temporary transport failure
705  * -EREMOTEIO   permanent target failure, do not retry
706  * -EBADE       permanent nexus failure, retry on other path
707  * -ENOSPC      No write space available
708  * -ENODATA     Medium error
709  * -EIO         unspecified I/O error
710  */
711 static int __scsi_error_from_host_byte(struct scsi_cmnd *cmd, int result)
712 {
713         int error = 0;
714
715         switch(host_byte(result)) {
716         case DID_TRANSPORT_FAILFAST:
717                 error = -ENOLINK;
718                 break;
719         case DID_TARGET_FAILURE:
720                 set_host_byte(cmd, DID_OK);
721                 error = -EREMOTEIO;
722                 break;
723         case DID_NEXUS_FAILURE:
724                 set_host_byte(cmd, DID_OK);
725                 error = -EBADE;
726                 break;
727         case DID_ALLOC_FAILURE:
728                 set_host_byte(cmd, DID_OK);
729                 error = -ENOSPC;
730                 break;
731         case DID_MEDIUM_ERROR:
732                 set_host_byte(cmd, DID_OK);
733                 error = -ENODATA;
734                 break;
735         default:
736                 error = -EIO;
737                 break;
738         }
739
740         return error;
741 }
742
743 /*
744  * Function:    scsi_io_completion()
745  *
746  * Purpose:     Completion processing for block device I/O requests.
747  *
748  * Arguments:   cmd   - command that is finished.
749  *
750  * Lock status: Assumed that no lock is held upon entry.
751  *
752  * Returns:     Nothing
753  *
754  * Notes:       This function is matched in terms of capabilities to
755  *              the function that created the scatter-gather list.
756  *              In other words, if there are no bounce buffers
757  *              (the normal case for most drivers), we don't need
758  *              the logic to deal with cleaning up afterwards.
759  *
760  *              We must call scsi_end_request().  This will finish off
761  *              the specified number of sectors.  If we are done, the
762  *              command block will be released and the queue function
763  *              will be goosed.  If we are not done then we have to
764  *              figure out what to do next:
765  *
766  *              a) We can call scsi_requeue_command().  The request
767  *                 will be unprepared and put back on the queue.  Then
768  *                 a new command will be created for it.  This should
769  *                 be used if we made forward progress, or if we want
770  *                 to switch from READ(10) to READ(6) for example.
771  *
772  *              b) We can call scsi_queue_insert().  The request will
773  *                 be put back on the queue and retried using the same
774  *                 command as before, possibly after a delay.
775  *
776  *              c) We can call blk_end_request() with -EIO to fail
777  *                 the remainder of the request.
778  */
779 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
780 {
781         int result = cmd->result;
782         struct request_queue *q = cmd->device->request_queue;
783         struct request *req = cmd->request;
784         int error = 0;
785         struct scsi_sense_hdr sshdr;
786         int sense_valid = 0;
787         int sense_deferred = 0;
788         enum {ACTION_FAIL, ACTION_REPREP, ACTION_RETRY,
789               ACTION_DELAYED_RETRY} action;
790         char *description = NULL;
791
792         if (result) {
793                 sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
794                 if (sense_valid)
795                         sense_deferred = scsi_sense_is_deferred(&sshdr);
796         }
797
798         if (req->cmd_type == REQ_TYPE_BLOCK_PC) { /* SG_IO ioctl from block level */
799                 if (result) {
800                         if (sense_valid && req->sense) {
801                                 /*
802                                  * SG_IO wants current and deferred errors
803                                  */
804                                 int len = 8 + cmd->sense_buffer[7];
805
806                                 if (len > SCSI_SENSE_BUFFERSIZE)
807                                         len = SCSI_SENSE_BUFFERSIZE;
808                                 memcpy(req->sense, cmd->sense_buffer,  len);
809                                 req->sense_len = len;
810                         }
811                         if (!sense_deferred)
812                                 error = __scsi_error_from_host_byte(cmd, result);
813                 }
814                 /*
815                  * __scsi_error_from_host_byte may have reset the host_byte
816                  */
817                 req->errors = cmd->result;
818
819                 req->resid_len = scsi_get_resid(cmd);
820
821                 if (scsi_bidi_cmnd(cmd)) {
822                         /*
823                          * Bidi commands Must be complete as a whole,
824                          * both sides at once.
825                          */
826                         req->next_rq->resid_len = scsi_in(cmd)->resid;
827
828                         scsi_release_buffers(cmd);
829                         blk_end_request_all(req, 0);
830
831                         scsi_next_command(cmd);
832                         return;
833                 }
834         } else if (blk_rq_bytes(req) == 0 && result && !sense_deferred) {
835                 /*
836                  * Certain non BLOCK_PC requests are commands that don't
837                  * actually transfer anything (FLUSH), so cannot use
838                  * good_bytes != blk_rq_bytes(req) as the signal for an error.
839                  * This sets the error explicitly for the problem case.
840                  */
841                 error = __scsi_error_from_host_byte(cmd, result);
842         }
843
844         /* no bidi support for !REQ_TYPE_BLOCK_PC yet */
845         BUG_ON(blk_bidi_rq(req));
846
847         /*
848          * Next deal with any sectors which we were able to correctly
849          * handle.
850          */
851         SCSI_LOG_HLCOMPLETE(1, printk("%u sectors total, "
852                                       "%d bytes done.\n",
853                                       blk_rq_sectors(req), good_bytes));
854
855         /*
856          * Recovered errors need reporting, but they're always treated
857          * as success, so fiddle the result code here.  For BLOCK_PC
858          * we already took a copy of the original into rq->errors which
859          * is what gets returned to the user
860          */
861         if (sense_valid && (sshdr.sense_key == RECOVERED_ERROR)) {
862                 /* if ATA PASS-THROUGH INFORMATION AVAILABLE skip
863                  * print since caller wants ATA registers. Only occurs on
864                  * SCSI ATA PASS_THROUGH commands when CK_COND=1
865                  */
866                 if ((sshdr.asc == 0x0) && (sshdr.ascq == 0x1d))
867                         ;
868                 else if (!(req->cmd_flags & REQ_QUIET))
869                         scsi_print_sense("", cmd);
870                 result = 0;
871                 /* BLOCK_PC may have set error */
872                 error = 0;
873         }
874
875         /*
876          * A number of bytes were successfully read.  If there
877          * are leftovers and there is some kind of error
878          * (result != 0), retry the rest.
879          */
880         if (scsi_end_request(cmd, error, good_bytes, result == 0) == NULL)
881                 return;
882
883         error = __scsi_error_from_host_byte(cmd, result);
884
885         if (host_byte(result) == DID_RESET) {
886                 /* Third party bus reset or reset for error recovery
887                  * reasons.  Just retry the command and see what
888                  * happens.
889                  */
890                 action = ACTION_RETRY;
891         } else if (sense_valid && !sense_deferred) {
892                 switch (sshdr.sense_key) {
893                 case UNIT_ATTENTION:
894                         if (cmd->device->removable) {
895                                 /* Detected disc change.  Set a bit
896                                  * and quietly refuse further access.
897                                  */
898                                 cmd->device->changed = 1;
899                                 description = "Media Changed";
900                                 action = ACTION_FAIL;
901                         } else {
902                                 /* Must have been a power glitch, or a
903                                  * bus reset.  Could not have been a
904                                  * media change, so we just retry the
905                                  * command and see what happens.
906                                  */
907                                 action = ACTION_RETRY;
908                         }
909                         break;
910                 case ILLEGAL_REQUEST:
911                         /* If we had an ILLEGAL REQUEST returned, then
912                          * we may have performed an unsupported
913                          * command.  The only thing this should be
914                          * would be a ten byte read where only a six
915                          * byte read was supported.  Also, on a system
916                          * where READ CAPACITY failed, we may have
917                          * read past the end of the disk.
918                          */
919                         if ((cmd->device->use_10_for_rw &&
920                             sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
921                             (cmd->cmnd[0] == READ_10 ||
922                              cmd->cmnd[0] == WRITE_10)) {
923                                 /* This will issue a new 6-byte command. */
924                                 cmd->device->use_10_for_rw = 0;
925                                 action = ACTION_REPREP;
926                         } else if (sshdr.asc == 0x10) /* DIX */ {
927                                 description = "Host Data Integrity Failure";
928                                 action = ACTION_FAIL;
929                                 error = -EILSEQ;
930                         /* INVALID COMMAND OPCODE or INVALID FIELD IN CDB */
931                         } else if (sshdr.asc == 0x20 || sshdr.asc == 0x24) {
932                                 switch (cmd->cmnd[0]) {
933                                 case UNMAP:
934                                         description = "Discard failure";
935                                         break;
936                                 case WRITE_SAME:
937                                 case WRITE_SAME_16:
938                                         if (cmd->cmnd[1] & 0x8)
939                                                 description = "Discard failure";
940                                         else
941                                                 description =
942                                                         "Write same failure";
943                                         break;
944                                 default:
945                                         description = "Invalid command failure";
946                                         break;
947                                 }
948                                 action = ACTION_FAIL;
949                                 error = -EREMOTEIO;
950                         } else
951                                 action = ACTION_FAIL;
952                         break;
953                 case ABORTED_COMMAND:
954                         action = ACTION_FAIL;
955                         if (sshdr.asc == 0x10) { /* DIF */
956                                 description = "Target Data Integrity Failure";
957                                 error = -EILSEQ;
958                         }
959                         break;
960                 case NOT_READY:
961                         /* If the device is in the process of becoming
962                          * ready, or has a temporary blockage, retry.
963                          */
964                         if (sshdr.asc == 0x04) {
965                                 switch (sshdr.ascq) {
966                                 case 0x01: /* becoming ready */
967                                 case 0x04: /* format in progress */
968                                 case 0x05: /* rebuild in progress */
969                                 case 0x06: /* recalculation in progress */
970                                 case 0x07: /* operation in progress */
971                                 case 0x08: /* Long write in progress */
972                                 case 0x09: /* self test in progress */
973                                 case 0x14: /* space allocation in progress */
974                                         action = ACTION_DELAYED_RETRY;
975                                         break;
976                                 default:
977                                         description = "Device not ready";
978                                         action = ACTION_FAIL;
979                                         break;
980                                 }
981                         } else {
982                                 description = "Device not ready";
983                                 action = ACTION_FAIL;
984                         }
985                         break;
986                 case VOLUME_OVERFLOW:
987                         /* See SSC3rXX or current. */
988                         action = ACTION_FAIL;
989                         break;
990                 default:
991                         description = "Unhandled sense code";
992                         action = ACTION_FAIL;
993                         break;
994                 }
995         } else {
996                 description = "Unhandled error code";
997                 action = ACTION_FAIL;
998         }
999
1000         switch (action) {
1001         case ACTION_FAIL:
1002                 /* Give up and fail the remainder of the request */
1003                 scsi_release_buffers(cmd);
1004                 if (!(req->cmd_flags & REQ_QUIET)) {
1005                         if (description)
1006                                 scmd_printk(KERN_INFO, cmd, "%s\n",
1007                                             description);
1008                         scsi_print_result(cmd);
1009                         if (driver_byte(result) & DRIVER_SENSE)
1010                                 scsi_print_sense("", cmd);
1011                         scsi_print_command(cmd);
1012                 }
1013                 if (blk_end_request_err(req, error))
1014                         scsi_requeue_command(q, cmd);
1015                 else
1016                         scsi_next_command(cmd);
1017                 break;
1018         case ACTION_REPREP:
1019                 /* Unprep the request and put it back at the head of the queue.
1020                  * A new command will be prepared and issued.
1021                  */
1022                 scsi_release_buffers(cmd);
1023                 scsi_requeue_command(q, cmd);
1024                 break;
1025         case ACTION_RETRY:
1026                 /* Retry the same command immediately */
1027                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY, 0);
1028                 break;
1029         case ACTION_DELAYED_RETRY:
1030                 /* Retry the same command after a delay */
1031                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY, 0);
1032                 break;
1033         }
1034 }
1035
1036 static int scsi_init_sgtable(struct request *req, struct scsi_data_buffer *sdb,
1037                              gfp_t gfp_mask)
1038 {
1039         int count;
1040
1041         /*
1042          * If sg table allocation fails, requeue request later.
1043          */
1044         if (unlikely(scsi_alloc_sgtable(sdb, req->nr_phys_segments,
1045                                         gfp_mask))) {
1046                 return BLKPREP_DEFER;
1047         }
1048
1049         req->buffer = NULL;
1050
1051         /* 
1052          * Next, walk the list, and fill in the addresses and sizes of
1053          * each segment.
1054          */
1055         count = blk_rq_map_sg(req->q, req, sdb->table.sgl);
1056         BUG_ON(count > sdb->table.nents);
1057         sdb->table.nents = count;
1058         sdb->length = blk_rq_bytes(req);
1059         return BLKPREP_OK;
1060 }
1061
1062 /*
1063  * Function:    scsi_init_io()
1064  *
1065  * Purpose:     SCSI I/O initialize function.
1066  *
1067  * Arguments:   cmd   - Command descriptor we wish to initialize
1068  *
1069  * Returns:     0 on success
1070  *              BLKPREP_DEFER if the failure is retryable
1071  *              BLKPREP_KILL if the failure is fatal
1072  */
1073 int scsi_init_io(struct scsi_cmnd *cmd, gfp_t gfp_mask)
1074 {
1075         struct request *rq = cmd->request;
1076
1077         int error = scsi_init_sgtable(rq, &cmd->sdb, gfp_mask);
1078         if (error)
1079                 goto err_exit;
1080
1081         if (blk_bidi_rq(rq)) {
1082                 struct scsi_data_buffer *bidi_sdb = kmem_cache_zalloc(
1083                         scsi_sdb_cache, GFP_ATOMIC);
1084                 if (!bidi_sdb) {
1085                         error = BLKPREP_DEFER;
1086                         goto err_exit;
1087                 }
1088
1089                 rq->next_rq->special = bidi_sdb;
1090                 error = scsi_init_sgtable(rq->next_rq, bidi_sdb, GFP_ATOMIC);
1091                 if (error)
1092                         goto err_exit;
1093         }
1094
1095         if (blk_integrity_rq(rq)) {
1096                 struct scsi_data_buffer *prot_sdb = cmd->prot_sdb;
1097                 int ivecs, count;
1098
1099                 BUG_ON(prot_sdb == NULL);
1100                 ivecs = blk_rq_count_integrity_sg(rq->q, rq->bio);
1101
1102                 if (scsi_alloc_sgtable(prot_sdb, ivecs, gfp_mask)) {
1103                         error = BLKPREP_DEFER;
1104                         goto err_exit;
1105                 }
1106
1107                 count = blk_rq_map_integrity_sg(rq->q, rq->bio,
1108                                                 prot_sdb->table.sgl);
1109                 BUG_ON(unlikely(count > ivecs));
1110                 BUG_ON(unlikely(count > queue_max_integrity_segments(rq->q)));
1111
1112                 cmd->prot_sdb = prot_sdb;
1113                 cmd->prot_sdb->table.nents = count;
1114         }
1115
1116         return BLKPREP_OK ;
1117
1118 err_exit:
1119         scsi_release_buffers(cmd);
1120         cmd->request->special = NULL;
1121         scsi_put_command(cmd);
1122         return error;
1123 }
1124 EXPORT_SYMBOL(scsi_init_io);
1125
1126 static struct scsi_cmnd *scsi_get_cmd_from_req(struct scsi_device *sdev,
1127                 struct request *req)
1128 {
1129         struct scsi_cmnd *cmd;
1130
1131         if (!req->special) {
1132                 cmd = scsi_get_command(sdev, GFP_ATOMIC);
1133                 if (unlikely(!cmd))
1134                         return NULL;
1135                 req->special = cmd;
1136         } else {
1137                 cmd = req->special;
1138         }
1139
1140         /* pull a tag out of the request if we have one */
1141         cmd->tag = req->tag;
1142         cmd->request = req;
1143
1144         cmd->cmnd = req->cmd;
1145         cmd->prot_op = SCSI_PROT_NORMAL;
1146
1147         return cmd;
1148 }
1149
1150 int scsi_setup_blk_pc_cmnd(struct scsi_device *sdev, struct request *req)
1151 {
1152         struct scsi_cmnd *cmd;
1153         int ret = scsi_prep_state_check(sdev, req);
1154
1155         if (ret != BLKPREP_OK)
1156                 return ret;
1157
1158         cmd = scsi_get_cmd_from_req(sdev, req);
1159         if (unlikely(!cmd))
1160                 return BLKPREP_DEFER;
1161
1162         /*
1163          * BLOCK_PC requests may transfer data, in which case they must
1164          * a bio attached to them.  Or they might contain a SCSI command
1165          * that does not transfer data, in which case they may optionally
1166          * submit a request without an attached bio.
1167          */
1168         if (req->bio) {
1169                 int ret;
1170
1171                 BUG_ON(!req->nr_phys_segments);
1172
1173                 ret = scsi_init_io(cmd, GFP_ATOMIC);
1174                 if (unlikely(ret))
1175                         return ret;
1176         } else {
1177                 BUG_ON(blk_rq_bytes(req));
1178
1179                 memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1180                 req->buffer = NULL;
1181         }
1182
1183         cmd->cmd_len = req->cmd_len;
1184         if (!blk_rq_bytes(req))
1185                 cmd->sc_data_direction = DMA_NONE;
1186         else if (rq_data_dir(req) == WRITE)
1187                 cmd->sc_data_direction = DMA_TO_DEVICE;
1188         else
1189                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1190         
1191         cmd->transfersize = blk_rq_bytes(req);
1192         cmd->allowed = req->retries;
1193         return BLKPREP_OK;
1194 }
1195 EXPORT_SYMBOL(scsi_setup_blk_pc_cmnd);
1196
1197 /*
1198  * Setup a REQ_TYPE_FS command.  These are simple read/write request
1199  * from filesystems that still need to be translated to SCSI CDBs from
1200  * the ULD.
1201  */
1202 int scsi_setup_fs_cmnd(struct scsi_device *sdev, struct request *req)
1203 {
1204         struct scsi_cmnd *cmd;
1205         int ret = scsi_prep_state_check(sdev, req);
1206
1207         if (ret != BLKPREP_OK)
1208                 return ret;
1209
1210         if (unlikely(sdev->scsi_dh_data && sdev->scsi_dh_data->scsi_dh
1211                          && sdev->scsi_dh_data->scsi_dh->prep_fn)) {
1212                 ret = sdev->scsi_dh_data->scsi_dh->prep_fn(sdev, req);
1213                 if (ret != BLKPREP_OK)
1214                         return ret;
1215         }
1216
1217         /*
1218          * Filesystem requests must transfer data.
1219          */
1220         BUG_ON(!req->nr_phys_segments);
1221
1222         cmd = scsi_get_cmd_from_req(sdev, req);
1223         if (unlikely(!cmd))
1224                 return BLKPREP_DEFER;
1225
1226         memset(cmd->cmnd, 0, BLK_MAX_CDB);
1227         return scsi_init_io(cmd, GFP_ATOMIC);
1228 }
1229 EXPORT_SYMBOL(scsi_setup_fs_cmnd);
1230
1231 int scsi_prep_state_check(struct scsi_device *sdev, struct request *req)
1232 {
1233         int ret = BLKPREP_OK;
1234
1235         /*
1236          * If the device is not in running state we will reject some
1237          * or all commands.
1238          */
1239         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1240                 switch (sdev->sdev_state) {
1241                 case SDEV_OFFLINE:
1242                 case SDEV_TRANSPORT_OFFLINE:
1243                         /*
1244                          * If the device is offline we refuse to process any
1245                          * commands.  The device must be brought online
1246                          * before trying any recovery commands.
1247                          */
1248                         sdev_printk(KERN_ERR, sdev,
1249                                     "rejecting I/O to offline device\n");
1250                         ret = BLKPREP_KILL;
1251                         break;
1252                 case SDEV_DEL:
1253                         /*
1254                          * If the device is fully deleted, we refuse to
1255                          * process any commands as well.
1256                          */
1257                         sdev_printk(KERN_ERR, sdev,
1258                                     "rejecting I/O to dead device\n");
1259                         ret = BLKPREP_KILL;
1260                         break;
1261                 case SDEV_QUIESCE:
1262                 case SDEV_BLOCK:
1263                 case SDEV_CREATED_BLOCK:
1264                         /*
1265                          * If the devices is blocked we defer normal commands.
1266                          */
1267                         if (!(req->cmd_flags & REQ_PREEMPT))
1268                                 ret = BLKPREP_DEFER;
1269                         break;
1270                 default:
1271                         /*
1272                          * For any other not fully online state we only allow
1273                          * special commands.  In particular any user initiated
1274                          * command is not allowed.
1275                          */
1276                         if (!(req->cmd_flags & REQ_PREEMPT))
1277                                 ret = BLKPREP_KILL;
1278                         break;
1279                 }
1280         }
1281         return ret;
1282 }
1283 EXPORT_SYMBOL(scsi_prep_state_check);
1284
1285 int scsi_prep_return(struct request_queue *q, struct request *req, int ret)
1286 {
1287         struct scsi_device *sdev = q->queuedata;
1288
1289         switch (ret) {
1290         case BLKPREP_KILL:
1291                 req->errors = DID_NO_CONNECT << 16;
1292                 /* release the command and kill it */
1293                 if (req->special) {
1294                         struct scsi_cmnd *cmd = req->special;
1295                         scsi_release_buffers(cmd);
1296                         scsi_put_command(cmd);
1297                         req->special = NULL;
1298                 }
1299                 break;
1300         case BLKPREP_DEFER:
1301                 /*
1302                  * If we defer, the blk_peek_request() returns NULL, but the
1303                  * queue must be restarted, so we schedule a callback to happen
1304                  * shortly.
1305                  */
1306                 if (sdev->device_busy == 0)
1307                         blk_delay_queue(q, SCSI_QUEUE_DELAY);
1308                 break;
1309         default:
1310                 req->cmd_flags |= REQ_DONTPREP;
1311         }
1312
1313         return ret;
1314 }
1315 EXPORT_SYMBOL(scsi_prep_return);
1316
1317 int scsi_prep_fn(struct request_queue *q, struct request *req)
1318 {
1319         struct scsi_device *sdev = q->queuedata;
1320         int ret = BLKPREP_KILL;
1321
1322         if (req->cmd_type == REQ_TYPE_BLOCK_PC)
1323                 ret = scsi_setup_blk_pc_cmnd(sdev, req);
1324         return scsi_prep_return(q, req, ret);
1325 }
1326 EXPORT_SYMBOL(scsi_prep_fn);
1327
1328 /*
1329  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1330  * return 0.
1331  *
1332  * Called with the queue_lock held.
1333  */
1334 static inline int scsi_dev_queue_ready(struct request_queue *q,
1335                                   struct scsi_device *sdev)
1336 {
1337         if (sdev->device_busy == 0 && sdev->device_blocked) {
1338                 /*
1339                  * unblock after device_blocked iterates to zero
1340                  */
1341                 if (--sdev->device_blocked == 0) {
1342                         SCSI_LOG_MLQUEUE(3,
1343                                    sdev_printk(KERN_INFO, sdev,
1344                                    "unblocking device at zero depth\n"));
1345                 } else {
1346                         blk_delay_queue(q, SCSI_QUEUE_DELAY);
1347                         return 0;
1348                 }
1349         }
1350         if (scsi_device_is_busy(sdev))
1351                 return 0;
1352
1353         return 1;
1354 }
1355
1356
1357 /*
1358  * scsi_target_queue_ready: checks if there we can send commands to target
1359  * @sdev: scsi device on starget to check.
1360  *
1361  * Called with the host lock held.
1362  */
1363 static inline int scsi_target_queue_ready(struct Scsi_Host *shost,
1364                                            struct scsi_device *sdev)
1365 {
1366         struct scsi_target *starget = scsi_target(sdev);
1367
1368         if (starget->single_lun) {
1369                 if (starget->starget_sdev_user &&
1370                     starget->starget_sdev_user != sdev)
1371                         return 0;
1372                 starget->starget_sdev_user = sdev;
1373         }
1374
1375         if (starget->target_busy == 0 && starget->target_blocked) {
1376                 /*
1377                  * unblock after target_blocked iterates to zero
1378                  */
1379                 if (--starget->target_blocked == 0) {
1380                         SCSI_LOG_MLQUEUE(3, starget_printk(KERN_INFO, starget,
1381                                          "unblocking target at zero depth\n"));
1382                 } else
1383                         return 0;
1384         }
1385
1386         if (scsi_target_is_busy(starget)) {
1387                 list_move_tail(&sdev->starved_entry, &shost->starved_list);
1388                 return 0;
1389         }
1390
1391         return 1;
1392 }
1393
1394 /*
1395  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1396  * return 0. We must end up running the queue again whenever 0 is
1397  * returned, else IO can hang.
1398  *
1399  * Called with host_lock held.
1400  */
1401 static inline int scsi_host_queue_ready(struct request_queue *q,
1402                                    struct Scsi_Host *shost,
1403                                    struct scsi_device *sdev)
1404 {
1405         if (scsi_host_in_recovery(shost))
1406                 return 0;
1407         if (shost->host_busy == 0 && shost->host_blocked) {
1408                 /*
1409                  * unblock after host_blocked iterates to zero
1410                  */
1411                 if (--shost->host_blocked == 0) {
1412                         SCSI_LOG_MLQUEUE(3,
1413                                 printk("scsi%d unblocking host at zero depth\n",
1414                                         shost->host_no));
1415                 } else {
1416                         return 0;
1417                 }
1418         }
1419         if (scsi_host_is_busy(shost)) {
1420                 if (list_empty(&sdev->starved_entry))
1421                         list_add_tail(&sdev->starved_entry, &shost->starved_list);
1422                 return 0;
1423         }
1424
1425         /* We're OK to process the command, so we can't be starved */
1426         if (!list_empty(&sdev->starved_entry))
1427                 list_del_init(&sdev->starved_entry);
1428
1429         return 1;
1430 }
1431
1432 /*
1433  * Busy state exporting function for request stacking drivers.
1434  *
1435  * For efficiency, no lock is taken to check the busy state of
1436  * shost/starget/sdev, since the returned value is not guaranteed and
1437  * may be changed after request stacking drivers call the function,
1438  * regardless of taking lock or not.
1439  *
1440  * When scsi can't dispatch I/Os anymore and needs to kill I/Os scsi
1441  * needs to return 'not busy'. Otherwise, request stacking drivers
1442  * may hold requests forever.
1443  */
1444 static int scsi_lld_busy(struct request_queue *q)
1445 {
1446         struct scsi_device *sdev = q->queuedata;
1447         struct Scsi_Host *shost;
1448
1449         if (blk_queue_dying(q))
1450                 return 0;
1451
1452         shost = sdev->host;
1453
1454         /*
1455          * Ignore host/starget busy state.
1456          * Since block layer does not have a concept of fairness across
1457          * multiple queues, congestion of host/starget needs to be handled
1458          * in SCSI layer.
1459          */
1460         if (scsi_host_in_recovery(shost) || scsi_device_is_busy(sdev))
1461                 return 1;
1462
1463         return 0;
1464 }
1465
1466 /*
1467  * Kill a request for a dead device
1468  */
1469 static void scsi_kill_request(struct request *req, struct request_queue *q)
1470 {
1471         struct scsi_cmnd *cmd = req->special;
1472         struct scsi_device *sdev;
1473         struct scsi_target *starget;
1474         struct Scsi_Host *shost;
1475
1476         blk_start_request(req);
1477
1478         scmd_printk(KERN_INFO, cmd, "killing request\n");
1479
1480         sdev = cmd->device;
1481         starget = scsi_target(sdev);
1482         shost = sdev->host;
1483         scsi_init_cmd_errh(cmd);
1484         cmd->result = DID_NO_CONNECT << 16;
1485         atomic_inc(&cmd->device->iorequest_cnt);
1486
1487         /*
1488          * SCSI request completion path will do scsi_device_unbusy(),
1489          * bump busy counts.  To bump the counters, we need to dance
1490          * with the locks as normal issue path does.
1491          */
1492         sdev->device_busy++;
1493         spin_unlock(sdev->request_queue->queue_lock);
1494         spin_lock(shost->host_lock);
1495         shost->host_busy++;
1496         starget->target_busy++;
1497         spin_unlock(shost->host_lock);
1498         spin_lock(sdev->request_queue->queue_lock);
1499
1500         blk_complete_request(req);
1501 }
1502
1503 static void scsi_softirq_done(struct request *rq)
1504 {
1505         struct scsi_cmnd *cmd = rq->special;
1506         unsigned long wait_for = (cmd->allowed + 1) * rq->timeout;
1507         int disposition;
1508
1509         INIT_LIST_HEAD(&cmd->eh_entry);
1510
1511         atomic_inc(&cmd->device->iodone_cnt);
1512         if (cmd->result)
1513                 atomic_inc(&cmd->device->ioerr_cnt);
1514
1515         disposition = scsi_decide_disposition(cmd);
1516         if (disposition != SUCCESS &&
1517             time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1518                 sdev_printk(KERN_ERR, cmd->device,
1519                             "timing out command, waited %lus\n",
1520                             wait_for/HZ);
1521                 disposition = SUCCESS;
1522         }
1523                         
1524         scsi_log_completion(cmd, disposition);
1525
1526         switch (disposition) {
1527                 case SUCCESS:
1528                         scsi_finish_command(cmd);
1529                         break;
1530                 case NEEDS_RETRY:
1531                         scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1532                         break;
1533                 case ADD_TO_MLQUEUE:
1534                         scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1535                         break;
1536                 default:
1537                         if (!scsi_eh_scmd_add(cmd, 0))
1538                                 scsi_finish_command(cmd);
1539         }
1540 }
1541
1542 /*
1543  * Function:    scsi_request_fn()
1544  *
1545  * Purpose:     Main strategy routine for SCSI.
1546  *
1547  * Arguments:   q       - Pointer to actual queue.
1548  *
1549  * Returns:     Nothing
1550  *
1551  * Lock status: IO request lock assumed to be held when called.
1552  */
1553 static void scsi_request_fn(struct request_queue *q)
1554 {
1555         struct scsi_device *sdev = q->queuedata;
1556         struct Scsi_Host *shost;
1557         struct scsi_cmnd *cmd;
1558         struct request *req;
1559
1560         if(!get_device(&sdev->sdev_gendev))
1561                 /* We must be tearing the block queue down already */
1562                 return;
1563
1564         /*
1565          * To start with, we keep looping until the queue is empty, or until
1566          * the host is no longer able to accept any more requests.
1567          */
1568         shost = sdev->host;
1569         for (;;) {
1570                 int rtn;
1571                 /*
1572                  * get next queueable request.  We do this early to make sure
1573                  * that the request is fully prepared even if we cannot 
1574                  * accept it.
1575                  */
1576                 req = blk_peek_request(q);
1577                 if (!req || !scsi_dev_queue_ready(q, sdev))
1578                         break;
1579
1580                 if (unlikely(!scsi_device_online(sdev))) {
1581                         sdev_printk(KERN_ERR, sdev,
1582                                     "rejecting I/O to offline device\n");
1583                         scsi_kill_request(req, q);
1584                         continue;
1585                 }
1586
1587
1588                 /*
1589                  * Remove the request from the request list.
1590                  */
1591                 if (!(blk_queue_tagged(q) && !blk_queue_start_tag(q, req)))
1592                         blk_start_request(req);
1593                 sdev->device_busy++;
1594
1595                 spin_unlock(q->queue_lock);
1596                 cmd = req->special;
1597                 if (unlikely(cmd == NULL)) {
1598                         printk(KERN_CRIT "impossible request in %s.\n"
1599                                          "please mail a stack trace to "
1600                                          "linux-scsi@vger.kernel.org\n",
1601                                          __func__);
1602                         blk_dump_rq_flags(req, "foo");
1603                         BUG();
1604                 }
1605                 spin_lock(shost->host_lock);
1606
1607                 /*
1608                  * We hit this when the driver is using a host wide
1609                  * tag map. For device level tag maps the queue_depth check
1610                  * in the device ready fn would prevent us from trying
1611                  * to allocate a tag. Since the map is a shared host resource
1612                  * we add the dev to the starved list so it eventually gets
1613                  * a run when a tag is freed.
1614                  */
1615                 if (blk_queue_tagged(q) && !blk_rq_tagged(req)) {
1616                         if (list_empty(&sdev->starved_entry))
1617                                 list_add_tail(&sdev->starved_entry,
1618                                               &shost->starved_list);
1619                         goto not_ready;
1620                 }
1621
1622                 if (!scsi_target_queue_ready(shost, sdev))
1623                         goto not_ready;
1624
1625                 if (!scsi_host_queue_ready(q, shost, sdev))
1626                         goto not_ready;
1627
1628                 scsi_target(sdev)->target_busy++;
1629                 shost->host_busy++;
1630
1631                 /*
1632                  * XXX(hch): This is rather suboptimal, scsi_dispatch_cmd will
1633                  *              take the lock again.
1634                  */
1635                 spin_unlock_irq(shost->host_lock);
1636
1637                 /*
1638                  * Finally, initialize any error handling parameters, and set up
1639                  * the timers for timeouts.
1640                  */
1641                 scsi_init_cmd_errh(cmd);
1642
1643                 /*
1644                  * Dispatch the command to the low-level driver.
1645                  */
1646                 rtn = scsi_dispatch_cmd(cmd);
1647                 spin_lock_irq(q->queue_lock);
1648                 if (rtn)
1649                         goto out_delay;
1650         }
1651
1652         goto out;
1653
1654  not_ready:
1655         spin_unlock_irq(shost->host_lock);
1656
1657         /*
1658          * lock q, handle tag, requeue req, and decrement device_busy. We
1659          * must return with queue_lock held.
1660          *
1661          * Decrementing device_busy without checking it is OK, as all such
1662          * cases (host limits or settings) should run the queue at some
1663          * later time.
1664          */
1665         spin_lock_irq(q->queue_lock);
1666         blk_requeue_request(q, req);
1667         sdev->device_busy--;
1668 out_delay:
1669         if (sdev->device_busy == 0)
1670                 blk_delay_queue(q, SCSI_QUEUE_DELAY);
1671 out:
1672         /* must be careful here...if we trigger the ->remove() function
1673          * we cannot be holding the q lock */
1674         spin_unlock_irq(q->queue_lock);
1675         put_device(&sdev->sdev_gendev);
1676         spin_lock_irq(q->queue_lock);
1677 }
1678
1679 u64 scsi_calculate_bounce_limit(struct Scsi_Host *shost)
1680 {
1681         struct device *host_dev;
1682         u64 bounce_limit = 0xffffffff;
1683
1684         if (shost->unchecked_isa_dma)
1685                 return BLK_BOUNCE_ISA;
1686         /*
1687          * Platforms with virtual-DMA translation
1688          * hardware have no practical limit.
1689          */
1690         if (!PCI_DMA_BUS_IS_PHYS)
1691                 return BLK_BOUNCE_ANY;
1692
1693         host_dev = scsi_get_device(shost);
1694         if (host_dev && host_dev->dma_mask)
1695                 bounce_limit = (u64)dma_max_pfn(host_dev) << PAGE_SHIFT;
1696
1697         return bounce_limit;
1698 }
1699 EXPORT_SYMBOL(scsi_calculate_bounce_limit);
1700
1701 struct request_queue *__scsi_alloc_queue(struct Scsi_Host *shost,
1702                                          request_fn_proc *request_fn)
1703 {
1704         struct request_queue *q;
1705         struct device *dev = shost->dma_dev;
1706
1707         q = blk_init_queue(request_fn, NULL);
1708         if (!q)
1709                 return NULL;
1710
1711         /*
1712          * this limit is imposed by hardware restrictions
1713          */
1714         blk_queue_max_segments(q, min_t(unsigned short, shost->sg_tablesize,
1715                                         SCSI_MAX_SG_CHAIN_SEGMENTS));
1716
1717         if (scsi_host_prot_dma(shost)) {
1718                 shost->sg_prot_tablesize =
1719                         min_not_zero(shost->sg_prot_tablesize,
1720                                      (unsigned short)SCSI_MAX_PROT_SG_SEGMENTS);
1721                 BUG_ON(shost->sg_prot_tablesize < shost->sg_tablesize);
1722                 blk_queue_max_integrity_segments(q, shost->sg_prot_tablesize);
1723         }
1724
1725         blk_queue_max_hw_sectors(q, shost->max_sectors);
1726         blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
1727         blk_queue_segment_boundary(q, shost->dma_boundary);
1728         dma_set_seg_boundary(dev, shost->dma_boundary);
1729
1730         blk_queue_max_segment_size(q, dma_get_max_seg_size(dev));
1731
1732         if (!shost->use_clustering)
1733                 q->limits.cluster = 0;
1734
1735         /*
1736          * set a reasonable default alignment on word boundaries: the
1737          * host and device may alter it using
1738          * blk_queue_update_dma_alignment() later.
1739          */
1740         blk_queue_dma_alignment(q, 0x03);
1741
1742         return q;
1743 }
1744 EXPORT_SYMBOL(__scsi_alloc_queue);
1745
1746 struct request_queue *scsi_alloc_queue(struct scsi_device *sdev)
1747 {
1748         struct request_queue *q;
1749
1750         q = __scsi_alloc_queue(sdev->host, scsi_request_fn);
1751         if (!q)
1752                 return NULL;
1753
1754         blk_queue_prep_rq(q, scsi_prep_fn);
1755         blk_queue_softirq_done(q, scsi_softirq_done);
1756         blk_queue_rq_timed_out(q, scsi_times_out);
1757         blk_queue_lld_busy(q, scsi_lld_busy);
1758         return q;
1759 }
1760
1761 /*
1762  * Function:    scsi_block_requests()
1763  *
1764  * Purpose:     Utility function used by low-level drivers to prevent further
1765  *              commands from being queued to the device.
1766  *
1767  * Arguments:   shost       - Host in question
1768  *
1769  * Returns:     Nothing
1770  *
1771  * Lock status: No locks are assumed held.
1772  *
1773  * Notes:       There is no timer nor any other means by which the requests
1774  *              get unblocked other than the low-level driver calling
1775  *              scsi_unblock_requests().
1776  */
1777 void scsi_block_requests(struct Scsi_Host *shost)
1778 {
1779         shost->host_self_blocked = 1;
1780 }
1781 EXPORT_SYMBOL(scsi_block_requests);
1782
1783 /*
1784  * Function:    scsi_unblock_requests()
1785  *
1786  * Purpose:     Utility function used by low-level drivers to allow further
1787  *              commands from being queued to the device.
1788  *
1789  * Arguments:   shost       - Host in question
1790  *
1791  * Returns:     Nothing
1792  *
1793  * Lock status: No locks are assumed held.
1794  *
1795  * Notes:       There is no timer nor any other means by which the requests
1796  *              get unblocked other than the low-level driver calling
1797  *              scsi_unblock_requests().
1798  *
1799  *              This is done as an API function so that changes to the
1800  *              internals of the scsi mid-layer won't require wholesale
1801  *              changes to drivers that use this feature.
1802  */
1803 void scsi_unblock_requests(struct Scsi_Host *shost)
1804 {
1805         shost->host_self_blocked = 0;
1806         scsi_run_host_queues(shost);
1807 }
1808 EXPORT_SYMBOL(scsi_unblock_requests);
1809
1810 int __init scsi_init_queue(void)
1811 {
1812         int i;
1813
1814         scsi_sdb_cache = kmem_cache_create("scsi_data_buffer",
1815                                            sizeof(struct scsi_data_buffer),
1816                                            0, 0, NULL);
1817         if (!scsi_sdb_cache) {
1818                 printk(KERN_ERR "SCSI: can't init scsi sdb cache\n");
1819                 return -ENOMEM;
1820         }
1821
1822         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1823                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1824                 int size = sgp->size * sizeof(struct scatterlist);
1825
1826                 sgp->slab = kmem_cache_create(sgp->name, size, 0,
1827                                 SLAB_HWCACHE_ALIGN, NULL);
1828                 if (!sgp->slab) {
1829                         printk(KERN_ERR "SCSI: can't init sg slab %s\n",
1830                                         sgp->name);
1831                         goto cleanup_sdb;
1832                 }
1833
1834                 sgp->pool = mempool_create_slab_pool(SG_MEMPOOL_SIZE,
1835                                                      sgp->slab);
1836                 if (!sgp->pool) {
1837                         printk(KERN_ERR "SCSI: can't init sg mempool %s\n",
1838                                         sgp->name);
1839                         goto cleanup_sdb;
1840                 }
1841         }
1842
1843         return 0;
1844
1845 cleanup_sdb:
1846         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1847                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1848                 if (sgp->pool)
1849                         mempool_destroy(sgp->pool);
1850                 if (sgp->slab)
1851                         kmem_cache_destroy(sgp->slab);
1852         }
1853         kmem_cache_destroy(scsi_sdb_cache);
1854
1855         return -ENOMEM;
1856 }
1857
1858 void scsi_exit_queue(void)
1859 {
1860         int i;
1861
1862         kmem_cache_destroy(scsi_sdb_cache);
1863
1864         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1865                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1866                 mempool_destroy(sgp->pool);
1867                 kmem_cache_destroy(sgp->slab);
1868         }
1869 }
1870
1871 /**
1872  *      scsi_mode_select - issue a mode select
1873  *      @sdev:  SCSI device to be queried
1874  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
1875  *      @sp:    Save page bit (0 == don't save, 1 == save)
1876  *      @modepage: mode page being requested
1877  *      @buffer: request buffer (may not be smaller than eight bytes)
1878  *      @len:   length of request buffer.
1879  *      @timeout: command timeout
1880  *      @retries: number of retries before failing
1881  *      @data: returns a structure abstracting the mode header data
1882  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
1883  *              must be SCSI_SENSE_BUFFERSIZE big.
1884  *
1885  *      Returns zero if successful; negative error number or scsi
1886  *      status on error
1887  *
1888  */
1889 int
1890 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
1891                  unsigned char *buffer, int len, int timeout, int retries,
1892                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1893 {
1894         unsigned char cmd[10];
1895         unsigned char *real_buffer;
1896         int ret;
1897
1898         memset(cmd, 0, sizeof(cmd));
1899         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
1900
1901         if (sdev->use_10_for_ms) {
1902                 if (len > 65535)
1903                         return -EINVAL;
1904                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
1905                 if (!real_buffer)
1906                         return -ENOMEM;
1907                 memcpy(real_buffer + 8, buffer, len);
1908                 len += 8;
1909                 real_buffer[0] = 0;
1910                 real_buffer[1] = 0;
1911                 real_buffer[2] = data->medium_type;
1912                 real_buffer[3] = data->device_specific;
1913                 real_buffer[4] = data->longlba ? 0x01 : 0;
1914                 real_buffer[5] = 0;
1915                 real_buffer[6] = data->block_descriptor_length >> 8;
1916                 real_buffer[7] = data->block_descriptor_length;
1917
1918                 cmd[0] = MODE_SELECT_10;
1919                 cmd[7] = len >> 8;
1920                 cmd[8] = len;
1921         } else {
1922                 if (len > 255 || data->block_descriptor_length > 255 ||
1923                     data->longlba)
1924                         return -EINVAL;
1925
1926                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
1927                 if (!real_buffer)
1928                         return -ENOMEM;
1929                 memcpy(real_buffer + 4, buffer, len);
1930                 len += 4;
1931                 real_buffer[0] = 0;
1932                 real_buffer[1] = data->medium_type;
1933                 real_buffer[2] = data->device_specific;
1934                 real_buffer[3] = data->block_descriptor_length;
1935                 
1936
1937                 cmd[0] = MODE_SELECT;
1938                 cmd[4] = len;
1939         }
1940
1941         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
1942                                sshdr, timeout, retries, NULL);
1943         kfree(real_buffer);
1944         return ret;
1945 }
1946 EXPORT_SYMBOL_GPL(scsi_mode_select);
1947
1948 /**
1949  *      scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
1950  *      @sdev:  SCSI device to be queried
1951  *      @dbd:   set if mode sense will allow block descriptors to be returned
1952  *      @modepage: mode page being requested
1953  *      @buffer: request buffer (may not be smaller than eight bytes)
1954  *      @len:   length of request buffer.
1955  *      @timeout: command timeout
1956  *      @retries: number of retries before failing
1957  *      @data: returns a structure abstracting the mode header data
1958  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
1959  *              must be SCSI_SENSE_BUFFERSIZE big.
1960  *
1961  *      Returns zero if unsuccessful, or the header offset (either 4
1962  *      or 8 depending on whether a six or ten byte command was
1963  *      issued) if successful.
1964  */
1965 int
1966 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
1967                   unsigned char *buffer, int len, int timeout, int retries,
1968                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1969 {
1970         unsigned char cmd[12];
1971         int use_10_for_ms;
1972         int header_length;
1973         int result;
1974         struct scsi_sense_hdr my_sshdr;
1975
1976         memset(data, 0, sizeof(*data));
1977         memset(&cmd[0], 0, 12);
1978         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
1979         cmd[2] = modepage;
1980
1981         /* caller might not be interested in sense, but we need it */
1982         if (!sshdr)
1983                 sshdr = &my_sshdr;
1984
1985  retry:
1986         use_10_for_ms = sdev->use_10_for_ms;
1987
1988         if (use_10_for_ms) {
1989                 if (len < 8)
1990                         len = 8;
1991
1992                 cmd[0] = MODE_SENSE_10;
1993                 cmd[8] = len;
1994                 header_length = 8;
1995         } else {
1996                 if (len < 4)
1997                         len = 4;
1998
1999                 cmd[0] = MODE_SENSE;
2000                 cmd[4] = len;
2001                 header_length = 4;
2002         }
2003
2004         memset(buffer, 0, len);
2005
2006         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
2007                                   sshdr, timeout, retries, NULL);
2008
2009         /* This code looks awful: what it's doing is making sure an
2010          * ILLEGAL REQUEST sense return identifies the actual command
2011          * byte as the problem.  MODE_SENSE commands can return
2012          * ILLEGAL REQUEST if the code page isn't supported */
2013
2014         if (use_10_for_ms && !scsi_status_is_good(result) &&
2015             (driver_byte(result) & DRIVER_SENSE)) {
2016                 if (scsi_sense_valid(sshdr)) {
2017                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
2018                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
2019                                 /* 
2020                                  * Invalid command operation code
2021                                  */
2022                                 sdev->use_10_for_ms = 0;
2023                                 goto retry;
2024                         }
2025                 }
2026         }
2027
2028         if(scsi_status_is_good(result)) {
2029                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
2030                              (modepage == 6 || modepage == 8))) {
2031                         /* Initio breakage? */
2032                         header_length = 0;
2033                         data->length = 13;
2034                         data->medium_type = 0;
2035                         data->device_specific = 0;
2036                         data->longlba = 0;
2037                         data->block_descriptor_length = 0;
2038                 } else if(use_10_for_ms) {
2039                         data->length = buffer[0]*256 + buffer[1] + 2;
2040                         data->medium_type = buffer[2];
2041                         data->device_specific = buffer[3];
2042                         data->longlba = buffer[4] & 0x01;
2043                         data->block_descriptor_length = buffer[6]*256
2044                                 + buffer[7];
2045                 } else {
2046                         data->length = buffer[0] + 1;
2047                         data->medium_type = buffer[1];
2048                         data->device_specific = buffer[2];
2049                         data->block_descriptor_length = buffer[3];
2050                 }
2051                 data->header_length = header_length;
2052         }
2053
2054         return result;
2055 }
2056 EXPORT_SYMBOL(scsi_mode_sense);
2057
2058 /**
2059  *      scsi_test_unit_ready - test if unit is ready
2060  *      @sdev:  scsi device to change the state of.
2061  *      @timeout: command timeout
2062  *      @retries: number of retries before failing
2063  *      @sshdr_external: Optional pointer to struct scsi_sense_hdr for
2064  *              returning sense. Make sure that this is cleared before passing
2065  *              in.
2066  *
2067  *      Returns zero if unsuccessful or an error if TUR failed.  For
2068  *      removable media, UNIT_ATTENTION sets ->changed flag.
2069  **/
2070 int
2071 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
2072                      struct scsi_sense_hdr *sshdr_external)
2073 {
2074         char cmd[] = {
2075                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
2076         };
2077         struct scsi_sense_hdr *sshdr;
2078         int result;
2079
2080         if (!sshdr_external)
2081                 sshdr = kzalloc(sizeof(*sshdr), GFP_KERNEL);
2082         else
2083                 sshdr = sshdr_external;
2084
2085         /* try to eat the UNIT_ATTENTION if there are enough retries */
2086         do {
2087                 result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
2088                                           timeout, retries, NULL);
2089                 if (sdev->removable && scsi_sense_valid(sshdr) &&
2090                     sshdr->sense_key == UNIT_ATTENTION)
2091                         sdev->changed = 1;
2092         } while (scsi_sense_valid(sshdr) &&
2093                  sshdr->sense_key == UNIT_ATTENTION && --retries);
2094
2095         if (!sshdr_external)
2096                 kfree(sshdr);
2097         return result;
2098 }
2099 EXPORT_SYMBOL(scsi_test_unit_ready);
2100
2101 /**
2102  *      scsi_device_set_state - Take the given device through the device state model.
2103  *      @sdev:  scsi device to change the state of.
2104  *      @state: state to change to.
2105  *
2106  *      Returns zero if unsuccessful or an error if the requested 
2107  *      transition is illegal.
2108  */
2109 int
2110 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2111 {
2112         enum scsi_device_state oldstate = sdev->sdev_state;
2113
2114         if (state == oldstate)
2115                 return 0;
2116
2117         switch (state) {
2118         case SDEV_CREATED:
2119                 switch (oldstate) {
2120                 case SDEV_CREATED_BLOCK:
2121                         break;
2122                 default:
2123                         goto illegal;
2124                 }
2125                 break;
2126                         
2127         case SDEV_RUNNING:
2128                 switch (oldstate) {
2129                 case SDEV_CREATED:
2130                 case SDEV_OFFLINE:
2131                 case SDEV_TRANSPORT_OFFLINE:
2132                 case SDEV_QUIESCE:
2133                 case SDEV_BLOCK:
2134                         break;
2135                 default:
2136                         goto illegal;
2137                 }
2138                 break;
2139
2140         case SDEV_QUIESCE:
2141                 switch (oldstate) {
2142                 case SDEV_RUNNING:
2143                 case SDEV_OFFLINE:
2144                 case SDEV_TRANSPORT_OFFLINE:
2145                         break;
2146                 default:
2147                         goto illegal;
2148                 }
2149                 break;
2150
2151         case SDEV_OFFLINE:
2152         case SDEV_TRANSPORT_OFFLINE:
2153                 switch (oldstate) {
2154                 case SDEV_CREATED:
2155                 case SDEV_RUNNING:
2156                 case SDEV_QUIESCE:
2157                 case SDEV_BLOCK:
2158                         break;
2159                 default:
2160                         goto illegal;
2161                 }
2162                 break;
2163
2164         case SDEV_BLOCK:
2165                 switch (oldstate) {
2166                 case SDEV_RUNNING:
2167                 case SDEV_CREATED_BLOCK:
2168                         break;
2169                 default:
2170                         goto illegal;
2171                 }
2172                 break;
2173
2174         case SDEV_CREATED_BLOCK:
2175                 switch (oldstate) {
2176                 case SDEV_CREATED:
2177                         break;
2178                 default:
2179                         goto illegal;
2180                 }
2181                 break;
2182
2183         case SDEV_CANCEL:
2184                 switch (oldstate) {
2185                 case SDEV_CREATED:
2186                 case SDEV_RUNNING:
2187                 case SDEV_QUIESCE:
2188                 case SDEV_OFFLINE:
2189                 case SDEV_TRANSPORT_OFFLINE:
2190                 case SDEV_BLOCK:
2191                         break;
2192                 default:
2193                         goto illegal;
2194                 }
2195                 break;
2196
2197         case SDEV_DEL:
2198                 switch (oldstate) {
2199                 case SDEV_CREATED:
2200                 case SDEV_RUNNING:
2201                 case SDEV_OFFLINE:
2202                 case SDEV_TRANSPORT_OFFLINE:
2203                 case SDEV_CANCEL:
2204                 case SDEV_CREATED_BLOCK:
2205                         break;
2206                 default:
2207                         goto illegal;
2208                 }
2209                 break;
2210
2211         }
2212         sdev->sdev_state = state;
2213         return 0;
2214
2215  illegal:
2216         SCSI_LOG_ERROR_RECOVERY(1, 
2217                                 sdev_printk(KERN_ERR, sdev,
2218                                             "Illegal state transition %s->%s\n",
2219                                             scsi_device_state_name(oldstate),
2220                                             scsi_device_state_name(state))
2221                                 );
2222         return -EINVAL;
2223 }
2224 EXPORT_SYMBOL(scsi_device_set_state);
2225
2226 /**
2227  *      sdev_evt_emit - emit a single SCSI device uevent
2228  *      @sdev: associated SCSI device
2229  *      @evt: event to emit
2230  *
2231  *      Send a single uevent (scsi_event) to the associated scsi_device.
2232  */
2233 static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2234 {
2235         int idx = 0;
2236         char *envp[3];
2237
2238         switch (evt->evt_type) {
2239         case SDEV_EVT_MEDIA_CHANGE:
2240                 envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2241                 break;
2242         case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2243                 envp[idx++] = "SDEV_UA=INQUIRY_DATA_HAS_CHANGED";
2244                 break;
2245         case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2246                 envp[idx++] = "SDEV_UA=CAPACITY_DATA_HAS_CHANGED";
2247                 break;
2248         case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2249                envp[idx++] = "SDEV_UA=THIN_PROVISIONING_SOFT_THRESHOLD_REACHED";
2250                 break;
2251         case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2252                 envp[idx++] = "SDEV_UA=MODE_PARAMETERS_CHANGED";
2253                 break;
2254         case SDEV_EVT_LUN_CHANGE_REPORTED:
2255                 envp[idx++] = "SDEV_UA=REPORTED_LUNS_DATA_HAS_CHANGED";
2256                 break;
2257         default:
2258                 /* do nothing */
2259                 break;
2260         }
2261
2262         envp[idx++] = NULL;
2263
2264         kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2265 }
2266
2267 /**
2268  *      sdev_evt_thread - send a uevent for each scsi event
2269  *      @work: work struct for scsi_device
2270  *
2271  *      Dispatch queued events to their associated scsi_device kobjects
2272  *      as uevents.
2273  */
2274 void scsi_evt_thread(struct work_struct *work)
2275 {
2276         struct scsi_device *sdev;
2277         enum scsi_device_event evt_type;
2278         LIST_HEAD(event_list);
2279
2280         sdev = container_of(work, struct scsi_device, event_work);
2281
2282         for (evt_type = SDEV_EVT_FIRST; evt_type <= SDEV_EVT_LAST; evt_type++)
2283                 if (test_and_clear_bit(evt_type, sdev->pending_events))
2284                         sdev_evt_send_simple(sdev, evt_type, GFP_KERNEL);
2285
2286         while (1) {
2287                 struct scsi_event *evt;
2288                 struct list_head *this, *tmp;
2289                 unsigned long flags;
2290
2291                 spin_lock_irqsave(&sdev->list_lock, flags);
2292                 list_splice_init(&sdev->event_list, &event_list);
2293                 spin_unlock_irqrestore(&sdev->list_lock, flags);
2294
2295                 if (list_empty(&event_list))
2296                         break;
2297
2298                 list_for_each_safe(this, tmp, &event_list) {
2299                         evt = list_entry(this, struct scsi_event, node);
2300                         list_del(&evt->node);
2301                         scsi_evt_emit(sdev, evt);
2302                         kfree(evt);
2303                 }
2304         }
2305 }
2306
2307 /**
2308  *      sdev_evt_send - send asserted event to uevent thread
2309  *      @sdev: scsi_device event occurred on
2310  *      @evt: event to send
2311  *
2312  *      Assert scsi device event asynchronously.
2313  */
2314 void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2315 {
2316         unsigned long flags;
2317
2318 #if 0
2319         /* FIXME: currently this check eliminates all media change events
2320          * for polled devices.  Need to update to discriminate between AN
2321          * and polled events */
2322         if (!test_bit(evt->evt_type, sdev->supported_events)) {
2323                 kfree(evt);
2324                 return;
2325         }
2326 #endif
2327
2328         spin_lock_irqsave(&sdev->list_lock, flags);
2329         list_add_tail(&evt->node, &sdev->event_list);
2330         schedule_work(&sdev->event_work);
2331         spin_unlock_irqrestore(&sdev->list_lock, flags);
2332 }
2333 EXPORT_SYMBOL_GPL(sdev_evt_send);
2334
2335 /**
2336  *      sdev_evt_alloc - allocate a new scsi event
2337  *      @evt_type: type of event to allocate
2338  *      @gfpflags: GFP flags for allocation
2339  *
2340  *      Allocates and returns a new scsi_event.
2341  */
2342 struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2343                                   gfp_t gfpflags)
2344 {
2345         struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2346         if (!evt)
2347                 return NULL;
2348
2349         evt->evt_type = evt_type;
2350         INIT_LIST_HEAD(&evt->node);
2351
2352         /* evt_type-specific initialization, if any */
2353         switch (evt_type) {
2354         case SDEV_EVT_MEDIA_CHANGE:
2355         case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2356         case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2357         case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2358         case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2359         case SDEV_EVT_LUN_CHANGE_REPORTED:
2360         default:
2361                 /* do nothing */
2362                 break;
2363         }
2364
2365         return evt;
2366 }
2367 EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2368
2369 /**
2370  *      sdev_evt_send_simple - send asserted event to uevent thread
2371  *      @sdev: scsi_device event occurred on
2372  *      @evt_type: type of event to send
2373  *      @gfpflags: GFP flags for allocation
2374  *
2375  *      Assert scsi device event asynchronously, given an event type.
2376  */
2377 void sdev_evt_send_simple(struct scsi_device *sdev,
2378                           enum scsi_device_event evt_type, gfp_t gfpflags)
2379 {
2380         struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2381         if (!evt) {
2382                 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2383                             evt_type);
2384                 return;
2385         }
2386
2387         sdev_evt_send(sdev, evt);
2388 }
2389 EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2390
2391 /**
2392  *      scsi_device_quiesce - Block user issued commands.
2393  *      @sdev:  scsi device to quiesce.
2394  *
2395  *      This works by trying to transition to the SDEV_QUIESCE state
2396  *      (which must be a legal transition).  When the device is in this
2397  *      state, only special requests will be accepted, all others will
2398  *      be deferred.  Since special requests may also be requeued requests,
2399  *      a successful return doesn't guarantee the device will be 
2400  *      totally quiescent.
2401  *
2402  *      Must be called with user context, may sleep.
2403  *
2404  *      Returns zero if unsuccessful or an error if not.
2405  */
2406 int
2407 scsi_device_quiesce(struct scsi_device *sdev)
2408 {
2409         int err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2410         if (err)
2411                 return err;
2412
2413         scsi_run_queue(sdev->request_queue);
2414         while (sdev->device_busy) {
2415                 msleep_interruptible(200);
2416                 scsi_run_queue(sdev->request_queue);
2417         }
2418         return 0;
2419 }
2420 EXPORT_SYMBOL(scsi_device_quiesce);
2421
2422 /**
2423  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2424  *      @sdev:  scsi device to resume.
2425  *
2426  *      Moves the device from quiesced back to running and restarts the
2427  *      queues.
2428  *
2429  *      Must be called with user context, may sleep.
2430  */
2431 void scsi_device_resume(struct scsi_device *sdev)
2432 {
2433         /* check if the device state was mutated prior to resume, and if
2434          * so assume the state is being managed elsewhere (for example
2435          * device deleted during suspend)
2436          */
2437         if (sdev->sdev_state != SDEV_QUIESCE ||
2438             scsi_device_set_state(sdev, SDEV_RUNNING))
2439                 return;
2440         scsi_run_queue(sdev->request_queue);
2441 }
2442 EXPORT_SYMBOL(scsi_device_resume);
2443
2444 static void
2445 device_quiesce_fn(struct scsi_device *sdev, void *data)
2446 {
2447         scsi_device_quiesce(sdev);
2448 }
2449
2450 void
2451 scsi_target_quiesce(struct scsi_target *starget)
2452 {
2453         starget_for_each_device(starget, NULL, device_quiesce_fn);
2454 }
2455 EXPORT_SYMBOL(scsi_target_quiesce);
2456
2457 static void
2458 device_resume_fn(struct scsi_device *sdev, void *data)
2459 {
2460         scsi_device_resume(sdev);
2461 }
2462
2463 void
2464 scsi_target_resume(struct scsi_target *starget)
2465 {
2466         starget_for_each_device(starget, NULL, device_resume_fn);
2467 }
2468 EXPORT_SYMBOL(scsi_target_resume);
2469
2470 /**
2471  * scsi_internal_device_block - internal function to put a device temporarily into the SDEV_BLOCK state
2472  * @sdev:       device to block
2473  *
2474  * Block request made by scsi lld's to temporarily stop all
2475  * scsi commands on the specified device.  Called from interrupt
2476  * or normal process context.
2477  *
2478  * Returns zero if successful or error if not
2479  *
2480  * Notes:       
2481  *      This routine transitions the device to the SDEV_BLOCK state
2482  *      (which must be a legal transition).  When the device is in this
2483  *      state, all commands are deferred until the scsi lld reenables
2484  *      the device with scsi_device_unblock or device_block_tmo fires.
2485  */
2486 int
2487 scsi_internal_device_block(struct scsi_device *sdev)
2488 {
2489         struct request_queue *q = sdev->request_queue;
2490         unsigned long flags;
2491         int err = 0;
2492
2493         err = scsi_device_set_state(sdev, SDEV_BLOCK);
2494         if (err) {
2495                 err = scsi_device_set_state(sdev, SDEV_CREATED_BLOCK);
2496
2497                 if (err)
2498                         return err;
2499         }
2500
2501         /* 
2502          * The device has transitioned to SDEV_BLOCK.  Stop the
2503          * block layer from calling the midlayer with this device's
2504          * request queue. 
2505          */
2506         spin_lock_irqsave(q->queue_lock, flags);
2507         blk_stop_queue(q);
2508         spin_unlock_irqrestore(q->queue_lock, flags);
2509
2510         return 0;
2511 }
2512 EXPORT_SYMBOL_GPL(scsi_internal_device_block);
2513  
2514 /**
2515  * scsi_internal_device_unblock - resume a device after a block request
2516  * @sdev:       device to resume
2517  * @new_state:  state to set devices to after unblocking
2518  *
2519  * Called by scsi lld's or the midlayer to restart the device queue
2520  * for the previously suspended scsi device.  Called from interrupt or
2521  * normal process context.
2522  *
2523  * Returns zero if successful or error if not.
2524  *
2525  * Notes:       
2526  *      This routine transitions the device to the SDEV_RUNNING state
2527  *      or to one of the offline states (which must be a legal transition)
2528  *      allowing the midlayer to goose the queue for this device.
2529  */
2530 int
2531 scsi_internal_device_unblock(struct scsi_device *sdev,
2532                              enum scsi_device_state new_state)
2533 {
2534         struct request_queue *q = sdev->request_queue; 
2535         unsigned long flags;
2536
2537         /*
2538          * Try to transition the scsi device to SDEV_RUNNING or one of the
2539          * offlined states and goose the device queue if successful.
2540          */
2541         if ((sdev->sdev_state == SDEV_BLOCK) ||
2542             (sdev->sdev_state == SDEV_TRANSPORT_OFFLINE))
2543                 sdev->sdev_state = new_state;
2544         else if (sdev->sdev_state == SDEV_CREATED_BLOCK) {
2545                 if (new_state == SDEV_TRANSPORT_OFFLINE ||
2546                     new_state == SDEV_OFFLINE)
2547                         sdev->sdev_state = new_state;
2548                 else
2549                         sdev->sdev_state = SDEV_CREATED;
2550         } else if (sdev->sdev_state != SDEV_CANCEL &&
2551                  sdev->sdev_state != SDEV_OFFLINE)
2552                 return -EINVAL;
2553
2554         spin_lock_irqsave(q->queue_lock, flags);
2555         blk_start_queue(q);
2556         spin_unlock_irqrestore(q->queue_lock, flags);
2557
2558         return 0;
2559 }
2560 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock);
2561
2562 static void
2563 device_block(struct scsi_device *sdev, void *data)
2564 {
2565         scsi_internal_device_block(sdev);
2566 }
2567
2568 static int
2569 target_block(struct device *dev, void *data)
2570 {
2571         if (scsi_is_target_device(dev))
2572                 starget_for_each_device(to_scsi_target(dev), NULL,
2573                                         device_block);
2574         return 0;
2575 }
2576
2577 void
2578 scsi_target_block(struct device *dev)
2579 {
2580         if (scsi_is_target_device(dev))
2581                 starget_for_each_device(to_scsi_target(dev), NULL,
2582                                         device_block);
2583         else
2584                 device_for_each_child(dev, NULL, target_block);
2585 }
2586 EXPORT_SYMBOL_GPL(scsi_target_block);
2587
2588 static void
2589 device_unblock(struct scsi_device *sdev, void *data)
2590 {
2591         scsi_internal_device_unblock(sdev, *(enum scsi_device_state *)data);
2592 }
2593
2594 static int
2595 target_unblock(struct device *dev, void *data)
2596 {
2597         if (scsi_is_target_device(dev))
2598                 starget_for_each_device(to_scsi_target(dev), data,
2599                                         device_unblock);
2600         return 0;
2601 }
2602
2603 void
2604 scsi_target_unblock(struct device *dev, enum scsi_device_state new_state)
2605 {
2606         if (scsi_is_target_device(dev))
2607                 starget_for_each_device(to_scsi_target(dev), &new_state,
2608                                         device_unblock);
2609         else
2610                 device_for_each_child(dev, &new_state, target_unblock);
2611 }
2612 EXPORT_SYMBOL_GPL(scsi_target_unblock);
2613
2614 /**
2615  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2616  * @sgl:        scatter-gather list
2617  * @sg_count:   number of segments in sg
2618  * @offset:     offset in bytes into sg, on return offset into the mapped area
2619  * @len:        bytes to map, on return number of bytes mapped
2620  *
2621  * Returns virtual address of the start of the mapped page
2622  */
2623 void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
2624                           size_t *offset, size_t *len)
2625 {
2626         int i;
2627         size_t sg_len = 0, len_complete = 0;
2628         struct scatterlist *sg;
2629         struct page *page;
2630
2631         WARN_ON(!irqs_disabled());
2632
2633         for_each_sg(sgl, sg, sg_count, i) {
2634                 len_complete = sg_len; /* Complete sg-entries */
2635                 sg_len += sg->length;
2636                 if (sg_len > *offset)
2637                         break;
2638         }
2639
2640         if (unlikely(i == sg_count)) {
2641                 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2642                         "elements %d\n",
2643                        __func__, sg_len, *offset, sg_count);
2644                 WARN_ON(1);
2645                 return NULL;
2646         }
2647
2648         /* Offset starting from the beginning of first page in this sg-entry */
2649         *offset = *offset - len_complete + sg->offset;
2650
2651         /* Assumption: contiguous pages can be accessed as "page + i" */
2652         page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
2653         *offset &= ~PAGE_MASK;
2654
2655         /* Bytes in this sg-entry from *offset to the end of the page */
2656         sg_len = PAGE_SIZE - *offset;
2657         if (*len > sg_len)
2658                 *len = sg_len;
2659
2660         return kmap_atomic(page);
2661 }
2662 EXPORT_SYMBOL(scsi_kmap_atomic_sg);
2663
2664 /**
2665  * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
2666  * @virt:       virtual address to be unmapped
2667  */
2668 void scsi_kunmap_atomic_sg(void *virt)
2669 {
2670         kunmap_atomic(virt);
2671 }
2672 EXPORT_SYMBOL(scsi_kunmap_atomic_sg);
2673
2674 void sdev_disable_disk_events(struct scsi_device *sdev)
2675 {
2676         atomic_inc(&sdev->disk_events_disable_depth);
2677 }
2678 EXPORT_SYMBOL(sdev_disable_disk_events);
2679
2680 void sdev_enable_disk_events(struct scsi_device *sdev)
2681 {
2682         if (WARN_ON_ONCE(atomic_read(&sdev->disk_events_disable_depth) <= 0))
2683                 return;
2684         atomic_dec(&sdev->disk_events_disable_depth);
2685 }
2686 EXPORT_SYMBOL(sdev_enable_disk_events);