756ac7c93de0dc29a5c7b57a3e2193b7585658cb
[linux-2.6.git] / drivers / scsi / scsi_lib.c
1 /*
2  *  scsi_lib.c Copyright (C) 1999 Eric Youngdale
3  *
4  *  SCSI queueing library.
5  *      Initial versions: Eric Youngdale (eric@andante.org).
6  *                        Based upon conversations with large numbers
7  *                        of people at Linux Expo.
8  */
9
10 #include <linux/bio.h>
11 #include <linux/bitops.h>
12 #include <linux/blkdev.h>
13 #include <linux/completion.h>
14 #include <linux/kernel.h>
15 #include <linux/mempool.h>
16 #include <linux/slab.h>
17 #include <linux/init.h>
18 #include <linux/pci.h>
19 #include <linux/delay.h>
20 #include <linux/hardirq.h>
21 #include <linux/scatterlist.h>
22
23 #include <scsi/scsi.h>
24 #include <scsi/scsi_cmnd.h>
25 #include <scsi/scsi_dbg.h>
26 #include <scsi/scsi_device.h>
27 #include <scsi/scsi_driver.h>
28 #include <scsi/scsi_eh.h>
29 #include <scsi/scsi_host.h>
30
31 #include "scsi_priv.h"
32 #include "scsi_logging.h"
33
34
35 #define SG_MEMPOOL_NR           ARRAY_SIZE(scsi_sg_pools)
36 #define SG_MEMPOOL_SIZE         2
37
38 struct scsi_host_sg_pool {
39         size_t          size;
40         char            *name;
41         struct kmem_cache       *slab;
42         mempool_t       *pool;
43 };
44
45 #define SP(x) { x, "sgpool-" __stringify(x) }
46 #if (SCSI_MAX_SG_SEGMENTS < 32)
47 #error SCSI_MAX_SG_SEGMENTS is too small (must be 32 or greater)
48 #endif
49 static struct scsi_host_sg_pool scsi_sg_pools[] = {
50         SP(8),
51         SP(16),
52 #if (SCSI_MAX_SG_SEGMENTS > 32)
53         SP(32),
54 #if (SCSI_MAX_SG_SEGMENTS > 64)
55         SP(64),
56 #if (SCSI_MAX_SG_SEGMENTS > 128)
57         SP(128),
58 #if (SCSI_MAX_SG_SEGMENTS > 256)
59 #error SCSI_MAX_SG_SEGMENTS is too large (256 MAX)
60 #endif
61 #endif
62 #endif
63 #endif
64         SP(SCSI_MAX_SG_SEGMENTS)
65 };
66 #undef SP
67
68 struct kmem_cache *scsi_sdb_cache;
69
70 static void scsi_run_queue(struct request_queue *q);
71
72 /*
73  * Function:    scsi_unprep_request()
74  *
75  * Purpose:     Remove all preparation done for a request, including its
76  *              associated scsi_cmnd, so that it can be requeued.
77  *
78  * Arguments:   req     - request to unprepare
79  *
80  * Lock status: Assumed that no locks are held upon entry.
81  *
82  * Returns:     Nothing.
83  */
84 static void scsi_unprep_request(struct request *req)
85 {
86         struct scsi_cmnd *cmd = req->special;
87
88         req->cmd_flags &= ~REQ_DONTPREP;
89         req->special = NULL;
90
91         scsi_put_command(cmd);
92 }
93
94 /**
95  * __scsi_queue_insert - private queue insertion
96  * @cmd: The SCSI command being requeued
97  * @reason:  The reason for the requeue
98  * @unbusy: Whether the queue should be unbusied
99  *
100  * This is a private queue insertion.  The public interface
101  * scsi_queue_insert() always assumes the queue should be unbusied
102  * because it's always called before the completion.  This function is
103  * for a requeue after completion, which should only occur in this
104  * file.
105  */
106 static int __scsi_queue_insert(struct scsi_cmnd *cmd, int reason, int unbusy)
107 {
108         struct Scsi_Host *host = cmd->device->host;
109         struct scsi_device *device = cmd->device;
110         struct scsi_target *starget = scsi_target(device);
111         struct request_queue *q = device->request_queue;
112         unsigned long flags;
113
114         SCSI_LOG_MLQUEUE(1,
115                  printk("Inserting command %p into mlqueue\n", cmd));
116
117         /*
118          * Set the appropriate busy bit for the device/host.
119          *
120          * If the host/device isn't busy, assume that something actually
121          * completed, and that we should be able to queue a command now.
122          *
123          * Note that the prior mid-layer assumption that any host could
124          * always queue at least one command is now broken.  The mid-layer
125          * will implement a user specifiable stall (see
126          * scsi_host.max_host_blocked and scsi_device.max_device_blocked)
127          * if a command is requeued with no other commands outstanding
128          * either for the device or for the host.
129          */
130         switch (reason) {
131         case SCSI_MLQUEUE_HOST_BUSY:
132                 host->host_blocked = host->max_host_blocked;
133                 break;
134         case SCSI_MLQUEUE_DEVICE_BUSY:
135                 device->device_blocked = device->max_device_blocked;
136                 break;
137         case SCSI_MLQUEUE_TARGET_BUSY:
138                 starget->target_blocked = starget->max_target_blocked;
139                 break;
140         }
141
142         /*
143          * Decrement the counters, since these commands are no longer
144          * active on the host/device.
145          */
146         if (unbusy)
147                 scsi_device_unbusy(device);
148
149         /*
150          * Requeue this command.  It will go before all other commands
151          * that are already in the queue.
152          *
153          * NOTE: there is magic here about the way the queue is plugged if
154          * we have no outstanding commands.
155          * 
156          * Although we *don't* plug the queue, we call the request
157          * function.  The SCSI request function detects the blocked condition
158          * and plugs the queue appropriately.
159          */
160         spin_lock_irqsave(q->queue_lock, flags);
161         blk_requeue_request(q, cmd->request);
162         spin_unlock_irqrestore(q->queue_lock, flags);
163
164         scsi_run_queue(q);
165
166         return 0;
167 }
168
169 /*
170  * Function:    scsi_queue_insert()
171  *
172  * Purpose:     Insert a command in the midlevel queue.
173  *
174  * Arguments:   cmd    - command that we are adding to queue.
175  *              reason - why we are inserting command to queue.
176  *
177  * Lock status: Assumed that lock is not held upon entry.
178  *
179  * Returns:     Nothing.
180  *
181  * Notes:       We do this for one of two cases.  Either the host is busy
182  *              and it cannot accept any more commands for the time being,
183  *              or the device returned QUEUE_FULL and can accept no more
184  *              commands.
185  * Notes:       This could be called either from an interrupt context or a
186  *              normal process context.
187  */
188 int scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
189 {
190         return __scsi_queue_insert(cmd, reason, 1);
191 }
192 /**
193  * scsi_execute - insert request and wait for the result
194  * @sdev:       scsi device
195  * @cmd:        scsi command
196  * @data_direction: data direction
197  * @buffer:     data buffer
198  * @bufflen:    len of buffer
199  * @sense:      optional sense buffer
200  * @timeout:    request timeout in seconds
201  * @retries:    number of times to retry request
202  * @flags:      or into request flags;
203  * @resid:      optional residual length
204  *
205  * returns the req->errors value which is the scsi_cmnd result
206  * field.
207  */
208 int scsi_execute(struct scsi_device *sdev, const unsigned char *cmd,
209                  int data_direction, void *buffer, unsigned bufflen,
210                  unsigned char *sense, int timeout, int retries, int flags,
211                  int *resid)
212 {
213         struct request *req;
214         int write = (data_direction == DMA_TO_DEVICE);
215         int ret = DRIVER_ERROR << 24;
216
217         req = blk_get_request(sdev->request_queue, write, __GFP_WAIT);
218
219         if (bufflen &&  blk_rq_map_kern(sdev->request_queue, req,
220                                         buffer, bufflen, __GFP_WAIT))
221                 goto out;
222
223         req->cmd_len = COMMAND_SIZE(cmd[0]);
224         memcpy(req->cmd, cmd, req->cmd_len);
225         req->sense = sense;
226         req->sense_len = 0;
227         req->retries = retries;
228         req->timeout = timeout;
229         req->cmd_type = REQ_TYPE_BLOCK_PC;
230         req->cmd_flags |= flags | REQ_QUIET | REQ_PREEMPT;
231
232         /*
233          * head injection *required* here otherwise quiesce won't work
234          */
235         blk_execute_rq(req->q, NULL, req, 1);
236
237         /*
238          * Some devices (USB mass-storage in particular) may transfer
239          * garbage data together with a residue indicating that the data
240          * is invalid.  Prevent the garbage from being misinterpreted
241          * and prevent security leaks by zeroing out the excess data.
242          */
243         if (unlikely(req->data_len > 0 && req->data_len <= bufflen))
244                 memset(buffer + (bufflen - req->data_len), 0, req->data_len);
245
246         if (resid)
247                 *resid = req->data_len;
248         ret = req->errors;
249  out:
250         blk_put_request(req);
251
252         return ret;
253 }
254 EXPORT_SYMBOL(scsi_execute);
255
256
257 int scsi_execute_req(struct scsi_device *sdev, const unsigned char *cmd,
258                      int data_direction, void *buffer, unsigned bufflen,
259                      struct scsi_sense_hdr *sshdr, int timeout, int retries,
260                      int *resid)
261 {
262         char *sense = NULL;
263         int result;
264         
265         if (sshdr) {
266                 sense = kzalloc(SCSI_SENSE_BUFFERSIZE, GFP_NOIO);
267                 if (!sense)
268                         return DRIVER_ERROR << 24;
269         }
270         result = scsi_execute(sdev, cmd, data_direction, buffer, bufflen,
271                               sense, timeout, retries, 0, resid);
272         if (sshdr)
273                 scsi_normalize_sense(sense, SCSI_SENSE_BUFFERSIZE, sshdr);
274
275         kfree(sense);
276         return result;
277 }
278 EXPORT_SYMBOL(scsi_execute_req);
279
280 /*
281  * Function:    scsi_init_cmd_errh()
282  *
283  * Purpose:     Initialize cmd fields related to error handling.
284  *
285  * Arguments:   cmd     - command that is ready to be queued.
286  *
287  * Notes:       This function has the job of initializing a number of
288  *              fields related to error handling.   Typically this will
289  *              be called once for each command, as required.
290  */
291 static void scsi_init_cmd_errh(struct scsi_cmnd *cmd)
292 {
293         cmd->serial_number = 0;
294         scsi_set_resid(cmd, 0);
295         memset(cmd->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
296         if (cmd->cmd_len == 0)
297                 cmd->cmd_len = scsi_command_size(cmd->cmnd);
298 }
299
300 void scsi_device_unbusy(struct scsi_device *sdev)
301 {
302         struct Scsi_Host *shost = sdev->host;
303         struct scsi_target *starget = scsi_target(sdev);
304         unsigned long flags;
305
306         spin_lock_irqsave(shost->host_lock, flags);
307         shost->host_busy--;
308         starget->target_busy--;
309         if (unlikely(scsi_host_in_recovery(shost) &&
310                      (shost->host_failed || shost->host_eh_scheduled)))
311                 scsi_eh_wakeup(shost);
312         spin_unlock(shost->host_lock);
313         spin_lock(sdev->request_queue->queue_lock);
314         sdev->device_busy--;
315         spin_unlock_irqrestore(sdev->request_queue->queue_lock, flags);
316 }
317
318 /*
319  * Called for single_lun devices on IO completion. Clear starget_sdev_user,
320  * and call blk_run_queue for all the scsi_devices on the target -
321  * including current_sdev first.
322  *
323  * Called with *no* scsi locks held.
324  */
325 static void scsi_single_lun_run(struct scsi_device *current_sdev)
326 {
327         struct Scsi_Host *shost = current_sdev->host;
328         struct scsi_device *sdev, *tmp;
329         struct scsi_target *starget = scsi_target(current_sdev);
330         unsigned long flags;
331
332         spin_lock_irqsave(shost->host_lock, flags);
333         starget->starget_sdev_user = NULL;
334         spin_unlock_irqrestore(shost->host_lock, flags);
335
336         /*
337          * Call blk_run_queue for all LUNs on the target, starting with
338          * current_sdev. We race with others (to set starget_sdev_user),
339          * but in most cases, we will be first. Ideally, each LU on the
340          * target would get some limited time or requests on the target.
341          */
342         blk_run_queue(current_sdev->request_queue);
343
344         spin_lock_irqsave(shost->host_lock, flags);
345         if (starget->starget_sdev_user)
346                 goto out;
347         list_for_each_entry_safe(sdev, tmp, &starget->devices,
348                         same_target_siblings) {
349                 if (sdev == current_sdev)
350                         continue;
351                 if (scsi_device_get(sdev))
352                         continue;
353
354                 spin_unlock_irqrestore(shost->host_lock, flags);
355                 blk_run_queue(sdev->request_queue);
356                 spin_lock_irqsave(shost->host_lock, flags);
357         
358                 scsi_device_put(sdev);
359         }
360  out:
361         spin_unlock_irqrestore(shost->host_lock, flags);
362 }
363
364 static inline int scsi_device_is_busy(struct scsi_device *sdev)
365 {
366         if (sdev->device_busy >= sdev->queue_depth || sdev->device_blocked)
367                 return 1;
368
369         return 0;
370 }
371
372 static inline int scsi_target_is_busy(struct scsi_target *starget)
373 {
374         return ((starget->can_queue > 0 &&
375                  starget->target_busy >= starget->can_queue) ||
376                  starget->target_blocked);
377 }
378
379 static inline int scsi_host_is_busy(struct Scsi_Host *shost)
380 {
381         if ((shost->can_queue > 0 && shost->host_busy >= shost->can_queue) ||
382             shost->host_blocked || shost->host_self_blocked)
383                 return 1;
384
385         return 0;
386 }
387
388 /*
389  * Function:    scsi_run_queue()
390  *
391  * Purpose:     Select a proper request queue to serve next
392  *
393  * Arguments:   q       - last request's queue
394  *
395  * Returns:     Nothing
396  *
397  * Notes:       The previous command was completely finished, start
398  *              a new one if possible.
399  */
400 static void scsi_run_queue(struct request_queue *q)
401 {
402         struct scsi_device *sdev = q->queuedata;
403         struct Scsi_Host *shost = sdev->host;
404         LIST_HEAD(starved_list);
405         unsigned long flags;
406
407         if (scsi_target(sdev)->single_lun)
408                 scsi_single_lun_run(sdev);
409
410         spin_lock_irqsave(shost->host_lock, flags);
411         list_splice_init(&shost->starved_list, &starved_list);
412
413         while (!list_empty(&starved_list)) {
414                 int flagset;
415
416                 /*
417                  * As long as shost is accepting commands and we have
418                  * starved queues, call blk_run_queue. scsi_request_fn
419                  * drops the queue_lock and can add us back to the
420                  * starved_list.
421                  *
422                  * host_lock protects the starved_list and starved_entry.
423                  * scsi_request_fn must get the host_lock before checking
424                  * or modifying starved_list or starved_entry.
425                  */
426                 if (scsi_host_is_busy(shost))
427                         break;
428
429                 sdev = list_entry(starved_list.next,
430                                   struct scsi_device, starved_entry);
431                 list_del_init(&sdev->starved_entry);
432                 if (scsi_target_is_busy(scsi_target(sdev))) {
433                         list_move_tail(&sdev->starved_entry,
434                                        &shost->starved_list);
435                         continue;
436                 }
437
438                 spin_unlock(shost->host_lock);
439
440                 spin_lock(sdev->request_queue->queue_lock);
441                 flagset = test_bit(QUEUE_FLAG_REENTER, &q->queue_flags) &&
442                                 !test_bit(QUEUE_FLAG_REENTER,
443                                         &sdev->request_queue->queue_flags);
444                 if (flagset)
445                         queue_flag_set(QUEUE_FLAG_REENTER, sdev->request_queue);
446                 __blk_run_queue(sdev->request_queue);
447                 if (flagset)
448                         queue_flag_clear(QUEUE_FLAG_REENTER, sdev->request_queue);
449                 spin_unlock(sdev->request_queue->queue_lock);
450
451                 spin_lock(shost->host_lock);
452         }
453         /* put any unprocessed entries back */
454         list_splice(&starved_list, &shost->starved_list);
455         spin_unlock_irqrestore(shost->host_lock, flags);
456
457         blk_run_queue(q);
458 }
459
460 /*
461  * Function:    scsi_requeue_command()
462  *
463  * Purpose:     Handle post-processing of completed commands.
464  *
465  * Arguments:   q       - queue to operate on
466  *              cmd     - command that may need to be requeued.
467  *
468  * Returns:     Nothing
469  *
470  * Notes:       After command completion, there may be blocks left
471  *              over which weren't finished by the previous command
472  *              this can be for a number of reasons - the main one is
473  *              I/O errors in the middle of the request, in which case
474  *              we need to request the blocks that come after the bad
475  *              sector.
476  * Notes:       Upon return, cmd is a stale pointer.
477  */
478 static void scsi_requeue_command(struct request_queue *q, struct scsi_cmnd *cmd)
479 {
480         struct request *req = cmd->request;
481         unsigned long flags;
482
483         spin_lock_irqsave(q->queue_lock, flags);
484         scsi_unprep_request(req);
485         blk_requeue_request(q, req);
486         spin_unlock_irqrestore(q->queue_lock, flags);
487
488         scsi_run_queue(q);
489 }
490
491 void scsi_next_command(struct scsi_cmnd *cmd)
492 {
493         struct scsi_device *sdev = cmd->device;
494         struct request_queue *q = sdev->request_queue;
495
496         /* need to hold a reference on the device before we let go of the cmd */
497         get_device(&sdev->sdev_gendev);
498
499         scsi_put_command(cmd);
500         scsi_run_queue(q);
501
502         /* ok to remove device now */
503         put_device(&sdev->sdev_gendev);
504 }
505
506 void scsi_run_host_queues(struct Scsi_Host *shost)
507 {
508         struct scsi_device *sdev;
509
510         shost_for_each_device(sdev, shost)
511                 scsi_run_queue(sdev->request_queue);
512 }
513
514 static void __scsi_release_buffers(struct scsi_cmnd *, int);
515
516 /*
517  * Function:    scsi_end_request()
518  *
519  * Purpose:     Post-processing of completed commands (usually invoked at end
520  *              of upper level post-processing and scsi_io_completion).
521  *
522  * Arguments:   cmd      - command that is complete.
523  *              error    - 0 if I/O indicates success, < 0 for I/O error.
524  *              bytes    - number of bytes of completed I/O
525  *              requeue  - indicates whether we should requeue leftovers.
526  *
527  * Lock status: Assumed that lock is not held upon entry.
528  *
529  * Returns:     cmd if requeue required, NULL otherwise.
530  *
531  * Notes:       This is called for block device requests in order to
532  *              mark some number of sectors as complete.
533  * 
534  *              We are guaranteeing that the request queue will be goosed
535  *              at some point during this call.
536  * Notes:       If cmd was requeued, upon return it will be a stale pointer.
537  */
538 static struct scsi_cmnd *scsi_end_request(struct scsi_cmnd *cmd, int error,
539                                           int bytes, int requeue)
540 {
541         struct request_queue *q = cmd->device->request_queue;
542         struct request *req = cmd->request;
543
544         /*
545          * If there are blocks left over at the end, set up the command
546          * to queue the remainder of them.
547          */
548         if (blk_end_request(req, error, bytes)) {
549                 int leftover = (req->hard_nr_sectors << 9);
550
551                 if (blk_pc_request(req))
552                         leftover = req->data_len;
553
554                 /* kill remainder if no retrys */
555                 if (error && scsi_noretry_cmd(cmd))
556                         blk_end_request(req, error, leftover);
557                 else {
558                         if (requeue) {
559                                 /*
560                                  * Bleah.  Leftovers again.  Stick the
561                                  * leftovers in the front of the
562                                  * queue, and goose the queue again.
563                                  */
564                                 scsi_release_buffers(cmd);
565                                 scsi_requeue_command(q, cmd);
566                                 cmd = NULL;
567                         }
568                         return cmd;
569                 }
570         }
571
572         /*
573          * This will goose the queue request function at the end, so we don't
574          * need to worry about launching another command.
575          */
576         __scsi_release_buffers(cmd, 0);
577         scsi_next_command(cmd);
578         return NULL;
579 }
580
581 static inline unsigned int scsi_sgtable_index(unsigned short nents)
582 {
583         unsigned int index;
584
585         BUG_ON(nents > SCSI_MAX_SG_SEGMENTS);
586
587         if (nents <= 8)
588                 index = 0;
589         else
590                 index = get_count_order(nents) - 3;
591
592         return index;
593 }
594
595 static void scsi_sg_free(struct scatterlist *sgl, unsigned int nents)
596 {
597         struct scsi_host_sg_pool *sgp;
598
599         sgp = scsi_sg_pools + scsi_sgtable_index(nents);
600         mempool_free(sgl, sgp->pool);
601 }
602
603 static struct scatterlist *scsi_sg_alloc(unsigned int nents, gfp_t gfp_mask)
604 {
605         struct scsi_host_sg_pool *sgp;
606
607         sgp = scsi_sg_pools + scsi_sgtable_index(nents);
608         return mempool_alloc(sgp->pool, gfp_mask);
609 }
610
611 static int scsi_alloc_sgtable(struct scsi_data_buffer *sdb, int nents,
612                               gfp_t gfp_mask)
613 {
614         int ret;
615
616         BUG_ON(!nents);
617
618         ret = __sg_alloc_table(&sdb->table, nents, SCSI_MAX_SG_SEGMENTS,
619                                gfp_mask, scsi_sg_alloc);
620         if (unlikely(ret))
621                 __sg_free_table(&sdb->table, SCSI_MAX_SG_SEGMENTS,
622                                 scsi_sg_free);
623
624         return ret;
625 }
626
627 static void scsi_free_sgtable(struct scsi_data_buffer *sdb)
628 {
629         __sg_free_table(&sdb->table, SCSI_MAX_SG_SEGMENTS, scsi_sg_free);
630 }
631
632 static void __scsi_release_buffers(struct scsi_cmnd *cmd, int do_bidi_check)
633 {
634
635         if (cmd->sdb.table.nents)
636                 scsi_free_sgtable(&cmd->sdb);
637
638         memset(&cmd->sdb, 0, sizeof(cmd->sdb));
639
640         if (do_bidi_check && scsi_bidi_cmnd(cmd)) {
641                 struct scsi_data_buffer *bidi_sdb =
642                         cmd->request->next_rq->special;
643                 scsi_free_sgtable(bidi_sdb);
644                 kmem_cache_free(scsi_sdb_cache, bidi_sdb);
645                 cmd->request->next_rq->special = NULL;
646         }
647
648         if (scsi_prot_sg_count(cmd))
649                 scsi_free_sgtable(cmd->prot_sdb);
650 }
651
652 /*
653  * Function:    scsi_release_buffers()
654  *
655  * Purpose:     Completion processing for block device I/O requests.
656  *
657  * Arguments:   cmd     - command that we are bailing.
658  *
659  * Lock status: Assumed that no lock is held upon entry.
660  *
661  * Returns:     Nothing
662  *
663  * Notes:       In the event that an upper level driver rejects a
664  *              command, we must release resources allocated during
665  *              the __init_io() function.  Primarily this would involve
666  *              the scatter-gather table, and potentially any bounce
667  *              buffers.
668  */
669 void scsi_release_buffers(struct scsi_cmnd *cmd)
670 {
671         __scsi_release_buffers(cmd, 1);
672 }
673 EXPORT_SYMBOL(scsi_release_buffers);
674
675 /*
676  * Bidi commands Must be complete as a whole, both sides at once.
677  * If part of the bytes were written and lld returned
678  * scsi_in()->resid and/or scsi_out()->resid this information will be left
679  * in req->data_len and req->next_rq->data_len. The upper-layer driver can
680  * decide what to do with this information.
681  */
682 static void scsi_end_bidi_request(struct scsi_cmnd *cmd)
683 {
684         struct request *req = cmd->request;
685         unsigned int dlen = req->data_len;
686         unsigned int next_dlen = req->next_rq->data_len;
687
688         req->data_len = scsi_out(cmd)->resid;
689         req->next_rq->data_len = scsi_in(cmd)->resid;
690
691         /* The req and req->next_rq have not been completed */
692         BUG_ON(blk_end_bidi_request(req, 0, dlen, next_dlen));
693
694         scsi_release_buffers(cmd);
695
696         /*
697          * This will goose the queue request function at the end, so we don't
698          * need to worry about launching another command.
699          */
700         scsi_next_command(cmd);
701 }
702
703 /*
704  * Function:    scsi_io_completion()
705  *
706  * Purpose:     Completion processing for block device I/O requests.
707  *
708  * Arguments:   cmd   - command that is finished.
709  *
710  * Lock status: Assumed that no lock is held upon entry.
711  *
712  * Returns:     Nothing
713  *
714  * Notes:       This function is matched in terms of capabilities to
715  *              the function that created the scatter-gather list.
716  *              In other words, if there are no bounce buffers
717  *              (the normal case for most drivers), we don't need
718  *              the logic to deal with cleaning up afterwards.
719  *
720  *              We must call scsi_end_request().  This will finish off
721  *              the specified number of sectors.  If we are done, the
722  *              command block will be released and the queue function
723  *              will be goosed.  If we are not done then we have to
724  *              figure out what to do next:
725  *
726  *              a) We can call scsi_requeue_command().  The request
727  *                 will be unprepared and put back on the queue.  Then
728  *                 a new command will be created for it.  This should
729  *                 be used if we made forward progress, or if we want
730  *                 to switch from READ(10) to READ(6) for example.
731  *
732  *              b) We can call scsi_queue_insert().  The request will
733  *                 be put back on the queue and retried using the same
734  *                 command as before, possibly after a delay.
735  *
736  *              c) We can call blk_end_request() with -EIO to fail
737  *                 the remainder of the request.
738  */
739 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
740 {
741         int result = cmd->result;
742         int this_count;
743         struct request_queue *q = cmd->device->request_queue;
744         struct request *req = cmd->request;
745         int error = 0;
746         struct scsi_sense_hdr sshdr;
747         int sense_valid = 0;
748         int sense_deferred = 0;
749         enum {ACTION_FAIL, ACTION_REPREP, ACTION_RETRY,
750               ACTION_DELAYED_RETRY} action;
751         char *description = NULL;
752
753         if (result) {
754                 sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
755                 if (sense_valid)
756                         sense_deferred = scsi_sense_is_deferred(&sshdr);
757         }
758
759         if (blk_pc_request(req)) { /* SG_IO ioctl from block level */
760                 req->errors = result;
761                 if (result) {
762                         if (sense_valid && req->sense) {
763                                 /*
764                                  * SG_IO wants current and deferred errors
765                                  */
766                                 int len = 8 + cmd->sense_buffer[7];
767
768                                 if (len > SCSI_SENSE_BUFFERSIZE)
769                                         len = SCSI_SENSE_BUFFERSIZE;
770                                 memcpy(req->sense, cmd->sense_buffer,  len);
771                                 req->sense_len = len;
772                         }
773                         if (!sense_deferred)
774                                 error = -EIO;
775                 }
776                 if (scsi_bidi_cmnd(cmd)) {
777                         /* will also release_buffers */
778                         scsi_end_bidi_request(cmd);
779                         return;
780                 }
781                 req->data_len = scsi_get_resid(cmd);
782         }
783
784         BUG_ON(blk_bidi_rq(req)); /* bidi not support for !blk_pc_request yet */
785
786         /*
787          * Next deal with any sectors which we were able to correctly
788          * handle.
789          */
790         SCSI_LOG_HLCOMPLETE(1, printk("%ld sectors total, "
791                                       "%d bytes done.\n",
792                                       req->nr_sectors, good_bytes));
793
794         /*
795          * Recovered errors need reporting, but they're always treated
796          * as success, so fiddle the result code here.  For BLOCK_PC
797          * we already took a copy of the original into rq->errors which
798          * is what gets returned to the user
799          */
800         if (sense_valid && sshdr.sense_key == RECOVERED_ERROR) {
801                 if (!(req->cmd_flags & REQ_QUIET))
802                         scsi_print_sense("", cmd);
803                 result = 0;
804                 /* BLOCK_PC may have set error */
805                 error = 0;
806         }
807
808         /*
809          * A number of bytes were successfully read.  If there
810          * are leftovers and there is some kind of error
811          * (result != 0), retry the rest.
812          */
813         if (scsi_end_request(cmd, error, good_bytes, result == 0) == NULL)
814                 return;
815         this_count = blk_rq_bytes(req);
816
817         error = -EIO;
818
819         if (host_byte(result) == DID_RESET) {
820                 /* Third party bus reset or reset for error recovery
821                  * reasons.  Just retry the command and see what
822                  * happens.
823                  */
824                 action = ACTION_RETRY;
825         } else if (sense_valid && !sense_deferred) {
826                 switch (sshdr.sense_key) {
827                 case UNIT_ATTENTION:
828                         if (cmd->device->removable) {
829                                 /* Detected disc change.  Set a bit
830                                  * and quietly refuse further access.
831                                  */
832                                 cmd->device->changed = 1;
833                                 description = "Media Changed";
834                                 action = ACTION_FAIL;
835                         } else {
836                                 /* Must have been a power glitch, or a
837                                  * bus reset.  Could not have been a
838                                  * media change, so we just retry the
839                                  * command and see what happens.
840                                  */
841                                 action = ACTION_RETRY;
842                         }
843                         break;
844                 case ILLEGAL_REQUEST:
845                         /* If we had an ILLEGAL REQUEST returned, then
846                          * we may have performed an unsupported
847                          * command.  The only thing this should be
848                          * would be a ten byte read where only a six
849                          * byte read was supported.  Also, on a system
850                          * where READ CAPACITY failed, we may have
851                          * read past the end of the disk.
852                          */
853                         if ((cmd->device->use_10_for_rw &&
854                             sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
855                             (cmd->cmnd[0] == READ_10 ||
856                              cmd->cmnd[0] == WRITE_10)) {
857                                 /* This will issue a new 6-byte command. */
858                                 cmd->device->use_10_for_rw = 0;
859                                 action = ACTION_REPREP;
860                         } else if (sshdr.asc == 0x10) /* DIX */ {
861                                 description = "Host Data Integrity Failure";
862                                 action = ACTION_FAIL;
863                                 error = -EILSEQ;
864                         } else
865                                 action = ACTION_FAIL;
866                         break;
867                 case ABORTED_COMMAND:
868                         action = ACTION_FAIL;
869                         if (sshdr.asc == 0x10) { /* DIF */
870                                 description = "Target Data Integrity Failure";
871                                 error = -EILSEQ;
872                         }
873                         break;
874                 case NOT_READY:
875                         /* If the device is in the process of becoming
876                          * ready, or has a temporary blockage, retry.
877                          */
878                         if (sshdr.asc == 0x04) {
879                                 switch (sshdr.ascq) {
880                                 case 0x01: /* becoming ready */
881                                 case 0x04: /* format in progress */
882                                 case 0x05: /* rebuild in progress */
883                                 case 0x06: /* recalculation in progress */
884                                 case 0x07: /* operation in progress */
885                                 case 0x08: /* Long write in progress */
886                                 case 0x09: /* self test in progress */
887                                         action = ACTION_DELAYED_RETRY;
888                                         break;
889                                 default:
890                                         description = "Device not ready";
891                                         action = ACTION_FAIL;
892                                         break;
893                                 }
894                         } else {
895                                 description = "Device not ready";
896                                 action = ACTION_FAIL;
897                         }
898                         break;
899                 case VOLUME_OVERFLOW:
900                         /* See SSC3rXX or current. */
901                         action = ACTION_FAIL;
902                         break;
903                 default:
904                         description = "Unhandled sense code";
905                         action = ACTION_FAIL;
906                         break;
907                 }
908         } else {
909                 description = "Unhandled error code";
910                 action = ACTION_FAIL;
911         }
912
913         switch (action) {
914         case ACTION_FAIL:
915                 /* Give up and fail the remainder of the request */
916                 scsi_release_buffers(cmd);
917                 if (!(req->cmd_flags & REQ_QUIET)) {
918                         if (description)
919                                 scmd_printk(KERN_INFO, cmd, "%s\n",
920                                             description);
921                         scsi_print_result(cmd);
922                         if (driver_byte(result) & DRIVER_SENSE)
923                                 scsi_print_sense("", cmd);
924                 }
925                 blk_end_request_all(req, -EIO);
926                 scsi_next_command(cmd);
927                 break;
928         case ACTION_REPREP:
929                 /* Unprep the request and put it back at the head of the queue.
930                  * A new command will be prepared and issued.
931                  */
932                 scsi_release_buffers(cmd);
933                 scsi_requeue_command(q, cmd);
934                 break;
935         case ACTION_RETRY:
936                 /* Retry the same command immediately */
937                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY, 0);
938                 break;
939         case ACTION_DELAYED_RETRY:
940                 /* Retry the same command after a delay */
941                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY, 0);
942                 break;
943         }
944 }
945
946 static int scsi_init_sgtable(struct request *req, struct scsi_data_buffer *sdb,
947                              gfp_t gfp_mask)
948 {
949         int count;
950
951         /*
952          * If sg table allocation fails, requeue request later.
953          */
954         if (unlikely(scsi_alloc_sgtable(sdb, req->nr_phys_segments,
955                                         gfp_mask))) {
956                 return BLKPREP_DEFER;
957         }
958
959         req->buffer = NULL;
960
961         /* 
962          * Next, walk the list, and fill in the addresses and sizes of
963          * each segment.
964          */
965         count = blk_rq_map_sg(req->q, req, sdb->table.sgl);
966         BUG_ON(count > sdb->table.nents);
967         sdb->table.nents = count;
968         if (blk_pc_request(req))
969                 sdb->length = req->data_len;
970         else
971                 sdb->length = req->nr_sectors << 9;
972         return BLKPREP_OK;
973 }
974
975 /*
976  * Function:    scsi_init_io()
977  *
978  * Purpose:     SCSI I/O initialize function.
979  *
980  * Arguments:   cmd   - Command descriptor we wish to initialize
981  *
982  * Returns:     0 on success
983  *              BLKPREP_DEFER if the failure is retryable
984  *              BLKPREP_KILL if the failure is fatal
985  */
986 int scsi_init_io(struct scsi_cmnd *cmd, gfp_t gfp_mask)
987 {
988         int error = scsi_init_sgtable(cmd->request, &cmd->sdb, gfp_mask);
989         if (error)
990                 goto err_exit;
991
992         if (blk_bidi_rq(cmd->request)) {
993                 struct scsi_data_buffer *bidi_sdb = kmem_cache_zalloc(
994                         scsi_sdb_cache, GFP_ATOMIC);
995                 if (!bidi_sdb) {
996                         error = BLKPREP_DEFER;
997                         goto err_exit;
998                 }
999
1000                 cmd->request->next_rq->special = bidi_sdb;
1001                 error = scsi_init_sgtable(cmd->request->next_rq, bidi_sdb,
1002                                                                     GFP_ATOMIC);
1003                 if (error)
1004                         goto err_exit;
1005         }
1006
1007         if (blk_integrity_rq(cmd->request)) {
1008                 struct scsi_data_buffer *prot_sdb = cmd->prot_sdb;
1009                 int ivecs, count;
1010
1011                 BUG_ON(prot_sdb == NULL);
1012                 ivecs = blk_rq_count_integrity_sg(cmd->request);
1013
1014                 if (scsi_alloc_sgtable(prot_sdb, ivecs, gfp_mask)) {
1015                         error = BLKPREP_DEFER;
1016                         goto err_exit;
1017                 }
1018
1019                 count = blk_rq_map_integrity_sg(cmd->request,
1020                                                 prot_sdb->table.sgl);
1021                 BUG_ON(unlikely(count > ivecs));
1022
1023                 cmd->prot_sdb = prot_sdb;
1024                 cmd->prot_sdb->table.nents = count;
1025         }
1026
1027         return BLKPREP_OK ;
1028
1029 err_exit:
1030         scsi_release_buffers(cmd);
1031         if (error == BLKPREP_KILL)
1032                 scsi_put_command(cmd);
1033         else /* BLKPREP_DEFER */
1034                 scsi_unprep_request(cmd->request);
1035
1036         return error;
1037 }
1038 EXPORT_SYMBOL(scsi_init_io);
1039
1040 static struct scsi_cmnd *scsi_get_cmd_from_req(struct scsi_device *sdev,
1041                 struct request *req)
1042 {
1043         struct scsi_cmnd *cmd;
1044
1045         if (!req->special) {
1046                 cmd = scsi_get_command(sdev, GFP_ATOMIC);
1047                 if (unlikely(!cmd))
1048                         return NULL;
1049                 req->special = cmd;
1050         } else {
1051                 cmd = req->special;
1052         }
1053
1054         /* pull a tag out of the request if we have one */
1055         cmd->tag = req->tag;
1056         cmd->request = req;
1057
1058         cmd->cmnd = req->cmd;
1059
1060         return cmd;
1061 }
1062
1063 int scsi_setup_blk_pc_cmnd(struct scsi_device *sdev, struct request *req)
1064 {
1065         struct scsi_cmnd *cmd;
1066         int ret = scsi_prep_state_check(sdev, req);
1067
1068         if (ret != BLKPREP_OK)
1069                 return ret;
1070
1071         cmd = scsi_get_cmd_from_req(sdev, req);
1072         if (unlikely(!cmd))
1073                 return BLKPREP_DEFER;
1074
1075         /*
1076          * BLOCK_PC requests may transfer data, in which case they must
1077          * a bio attached to them.  Or they might contain a SCSI command
1078          * that does not transfer data, in which case they may optionally
1079          * submit a request without an attached bio.
1080          */
1081         if (req->bio) {
1082                 int ret;
1083
1084                 BUG_ON(!req->nr_phys_segments);
1085
1086                 ret = scsi_init_io(cmd, GFP_ATOMIC);
1087                 if (unlikely(ret))
1088                         return ret;
1089         } else {
1090                 BUG_ON(req->data_len);
1091                 BUG_ON(req->data);
1092
1093                 memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1094                 req->buffer = NULL;
1095         }
1096
1097         cmd->cmd_len = req->cmd_len;
1098         if (!req->data_len)
1099                 cmd->sc_data_direction = DMA_NONE;
1100         else if (rq_data_dir(req) == WRITE)
1101                 cmd->sc_data_direction = DMA_TO_DEVICE;
1102         else
1103                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1104         
1105         cmd->transfersize = req->data_len;
1106         cmd->allowed = req->retries;
1107         return BLKPREP_OK;
1108 }
1109 EXPORT_SYMBOL(scsi_setup_blk_pc_cmnd);
1110
1111 /*
1112  * Setup a REQ_TYPE_FS command.  These are simple read/write request
1113  * from filesystems that still need to be translated to SCSI CDBs from
1114  * the ULD.
1115  */
1116 int scsi_setup_fs_cmnd(struct scsi_device *sdev, struct request *req)
1117 {
1118         struct scsi_cmnd *cmd;
1119         int ret = scsi_prep_state_check(sdev, req);
1120
1121         if (ret != BLKPREP_OK)
1122                 return ret;
1123
1124         if (unlikely(sdev->scsi_dh_data && sdev->scsi_dh_data->scsi_dh
1125                          && sdev->scsi_dh_data->scsi_dh->prep_fn)) {
1126                 ret = sdev->scsi_dh_data->scsi_dh->prep_fn(sdev, req);
1127                 if (ret != BLKPREP_OK)
1128                         return ret;
1129         }
1130
1131         /*
1132          * Filesystem requests must transfer data.
1133          */
1134         BUG_ON(!req->nr_phys_segments);
1135
1136         cmd = scsi_get_cmd_from_req(sdev, req);
1137         if (unlikely(!cmd))
1138                 return BLKPREP_DEFER;
1139
1140         memset(cmd->cmnd, 0, BLK_MAX_CDB);
1141         return scsi_init_io(cmd, GFP_ATOMIC);
1142 }
1143 EXPORT_SYMBOL(scsi_setup_fs_cmnd);
1144
1145 int scsi_prep_state_check(struct scsi_device *sdev, struct request *req)
1146 {
1147         int ret = BLKPREP_OK;
1148
1149         /*
1150          * If the device is not in running state we will reject some
1151          * or all commands.
1152          */
1153         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1154                 switch (sdev->sdev_state) {
1155                 case SDEV_OFFLINE:
1156                         /*
1157                          * If the device is offline we refuse to process any
1158                          * commands.  The device must be brought online
1159                          * before trying any recovery commands.
1160                          */
1161                         sdev_printk(KERN_ERR, sdev,
1162                                     "rejecting I/O to offline device\n");
1163                         ret = BLKPREP_KILL;
1164                         break;
1165                 case SDEV_DEL:
1166                         /*
1167                          * If the device is fully deleted, we refuse to
1168                          * process any commands as well.
1169                          */
1170                         sdev_printk(KERN_ERR, sdev,
1171                                     "rejecting I/O to dead device\n");
1172                         ret = BLKPREP_KILL;
1173                         break;
1174                 case SDEV_QUIESCE:
1175                 case SDEV_BLOCK:
1176                 case SDEV_CREATED_BLOCK:
1177                         /*
1178                          * If the devices is blocked we defer normal commands.
1179                          */
1180                         if (!(req->cmd_flags & REQ_PREEMPT))
1181                                 ret = BLKPREP_DEFER;
1182                         break;
1183                 default:
1184                         /*
1185                          * For any other not fully online state we only allow
1186                          * special commands.  In particular any user initiated
1187                          * command is not allowed.
1188                          */
1189                         if (!(req->cmd_flags & REQ_PREEMPT))
1190                                 ret = BLKPREP_KILL;
1191                         break;
1192                 }
1193         }
1194         return ret;
1195 }
1196 EXPORT_SYMBOL(scsi_prep_state_check);
1197
1198 int scsi_prep_return(struct request_queue *q, struct request *req, int ret)
1199 {
1200         struct scsi_device *sdev = q->queuedata;
1201
1202         switch (ret) {
1203         case BLKPREP_KILL:
1204                 req->errors = DID_NO_CONNECT << 16;
1205                 /* release the command and kill it */
1206                 if (req->special) {
1207                         struct scsi_cmnd *cmd = req->special;
1208                         scsi_release_buffers(cmd);
1209                         scsi_put_command(cmd);
1210                         req->special = NULL;
1211                 }
1212                 break;
1213         case BLKPREP_DEFER:
1214                 /*
1215                  * If we defer, the elv_next_request() returns NULL, but the
1216                  * queue must be restarted, so we plug here if no returning
1217                  * command will automatically do that.
1218                  */
1219                 if (sdev->device_busy == 0)
1220                         blk_plug_device(q);
1221                 break;
1222         default:
1223                 req->cmd_flags |= REQ_DONTPREP;
1224         }
1225
1226         return ret;
1227 }
1228 EXPORT_SYMBOL(scsi_prep_return);
1229
1230 int scsi_prep_fn(struct request_queue *q, struct request *req)
1231 {
1232         struct scsi_device *sdev = q->queuedata;
1233         int ret = BLKPREP_KILL;
1234
1235         if (req->cmd_type == REQ_TYPE_BLOCK_PC)
1236                 ret = scsi_setup_blk_pc_cmnd(sdev, req);
1237         return scsi_prep_return(q, req, ret);
1238 }
1239
1240 /*
1241  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1242  * return 0.
1243  *
1244  * Called with the queue_lock held.
1245  */
1246 static inline int scsi_dev_queue_ready(struct request_queue *q,
1247                                   struct scsi_device *sdev)
1248 {
1249         if (sdev->device_busy == 0 && sdev->device_blocked) {
1250                 /*
1251                  * unblock after device_blocked iterates to zero
1252                  */
1253                 if (--sdev->device_blocked == 0) {
1254                         SCSI_LOG_MLQUEUE(3,
1255                                    sdev_printk(KERN_INFO, sdev,
1256                                    "unblocking device at zero depth\n"));
1257                 } else {
1258                         blk_plug_device(q);
1259                         return 0;
1260                 }
1261         }
1262         if (scsi_device_is_busy(sdev))
1263                 return 0;
1264
1265         return 1;
1266 }
1267
1268
1269 /*
1270  * scsi_target_queue_ready: checks if there we can send commands to target
1271  * @sdev: scsi device on starget to check.
1272  *
1273  * Called with the host lock held.
1274  */
1275 static inline int scsi_target_queue_ready(struct Scsi_Host *shost,
1276                                            struct scsi_device *sdev)
1277 {
1278         struct scsi_target *starget = scsi_target(sdev);
1279
1280         if (starget->single_lun) {
1281                 if (starget->starget_sdev_user &&
1282                     starget->starget_sdev_user != sdev)
1283                         return 0;
1284                 starget->starget_sdev_user = sdev;
1285         }
1286
1287         if (starget->target_busy == 0 && starget->target_blocked) {
1288                 /*
1289                  * unblock after target_blocked iterates to zero
1290                  */
1291                 if (--starget->target_blocked == 0) {
1292                         SCSI_LOG_MLQUEUE(3, starget_printk(KERN_INFO, starget,
1293                                          "unblocking target at zero depth\n"));
1294                 } else {
1295                         blk_plug_device(sdev->request_queue);
1296                         return 0;
1297                 }
1298         }
1299
1300         if (scsi_target_is_busy(starget)) {
1301                 if (list_empty(&sdev->starved_entry)) {
1302                         list_add_tail(&sdev->starved_entry,
1303                                       &shost->starved_list);
1304                         return 0;
1305                 }
1306         }
1307
1308         /* We're OK to process the command, so we can't be starved */
1309         if (!list_empty(&sdev->starved_entry))
1310                 list_del_init(&sdev->starved_entry);
1311         return 1;
1312 }
1313
1314 /*
1315  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1316  * return 0. We must end up running the queue again whenever 0 is
1317  * returned, else IO can hang.
1318  *
1319  * Called with host_lock held.
1320  */
1321 static inline int scsi_host_queue_ready(struct request_queue *q,
1322                                    struct Scsi_Host *shost,
1323                                    struct scsi_device *sdev)
1324 {
1325         if (scsi_host_in_recovery(shost))
1326                 return 0;
1327         if (shost->host_busy == 0 && shost->host_blocked) {
1328                 /*
1329                  * unblock after host_blocked iterates to zero
1330                  */
1331                 if (--shost->host_blocked == 0) {
1332                         SCSI_LOG_MLQUEUE(3,
1333                                 printk("scsi%d unblocking host at zero depth\n",
1334                                         shost->host_no));
1335                 } else {
1336                         return 0;
1337                 }
1338         }
1339         if (scsi_host_is_busy(shost)) {
1340                 if (list_empty(&sdev->starved_entry))
1341                         list_add_tail(&sdev->starved_entry, &shost->starved_list);
1342                 return 0;
1343         }
1344
1345         /* We're OK to process the command, so we can't be starved */
1346         if (!list_empty(&sdev->starved_entry))
1347                 list_del_init(&sdev->starved_entry);
1348
1349         return 1;
1350 }
1351
1352 /*
1353  * Busy state exporting function for request stacking drivers.
1354  *
1355  * For efficiency, no lock is taken to check the busy state of
1356  * shost/starget/sdev, since the returned value is not guaranteed and
1357  * may be changed after request stacking drivers call the function,
1358  * regardless of taking lock or not.
1359  *
1360  * When scsi can't dispatch I/Os anymore and needs to kill I/Os
1361  * (e.g. !sdev), scsi needs to return 'not busy'.
1362  * Otherwise, request stacking drivers may hold requests forever.
1363  */
1364 static int scsi_lld_busy(struct request_queue *q)
1365 {
1366         struct scsi_device *sdev = q->queuedata;
1367         struct Scsi_Host *shost;
1368         struct scsi_target *starget;
1369
1370         if (!sdev)
1371                 return 0;
1372
1373         shost = sdev->host;
1374         starget = scsi_target(sdev);
1375
1376         if (scsi_host_in_recovery(shost) || scsi_host_is_busy(shost) ||
1377             scsi_target_is_busy(starget) || scsi_device_is_busy(sdev))
1378                 return 1;
1379
1380         return 0;
1381 }
1382
1383 /*
1384  * Kill a request for a dead device
1385  */
1386 static void scsi_kill_request(struct request *req, struct request_queue *q)
1387 {
1388         struct scsi_cmnd *cmd = req->special;
1389         struct scsi_device *sdev = cmd->device;
1390         struct scsi_target *starget = scsi_target(sdev);
1391         struct Scsi_Host *shost = sdev->host;
1392
1393         blkdev_dequeue_request(req);
1394
1395         if (unlikely(cmd == NULL)) {
1396                 printk(KERN_CRIT "impossible request in %s.\n",
1397                                  __func__);
1398                 BUG();
1399         }
1400
1401         scsi_init_cmd_errh(cmd);
1402         cmd->result = DID_NO_CONNECT << 16;
1403         atomic_inc(&cmd->device->iorequest_cnt);
1404
1405         /*
1406          * SCSI request completion path will do scsi_device_unbusy(),
1407          * bump busy counts.  To bump the counters, we need to dance
1408          * with the locks as normal issue path does.
1409          */
1410         sdev->device_busy++;
1411         spin_unlock(sdev->request_queue->queue_lock);
1412         spin_lock(shost->host_lock);
1413         shost->host_busy++;
1414         starget->target_busy++;
1415         spin_unlock(shost->host_lock);
1416         spin_lock(sdev->request_queue->queue_lock);
1417
1418         blk_complete_request(req);
1419 }
1420
1421 static void scsi_softirq_done(struct request *rq)
1422 {
1423         struct scsi_cmnd *cmd = rq->special;
1424         unsigned long wait_for = (cmd->allowed + 1) * rq->timeout;
1425         int disposition;
1426
1427         INIT_LIST_HEAD(&cmd->eh_entry);
1428
1429         /*
1430          * Set the serial numbers back to zero
1431          */
1432         cmd->serial_number = 0;
1433
1434         atomic_inc(&cmd->device->iodone_cnt);
1435         if (cmd->result)
1436                 atomic_inc(&cmd->device->ioerr_cnt);
1437
1438         disposition = scsi_decide_disposition(cmd);
1439         if (disposition != SUCCESS &&
1440             time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1441                 sdev_printk(KERN_ERR, cmd->device,
1442                             "timing out command, waited %lus\n",
1443                             wait_for/HZ);
1444                 disposition = SUCCESS;
1445         }
1446                         
1447         scsi_log_completion(cmd, disposition);
1448
1449         switch (disposition) {
1450                 case SUCCESS:
1451                         scsi_finish_command(cmd);
1452                         break;
1453                 case NEEDS_RETRY:
1454                         scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1455                         break;
1456                 case ADD_TO_MLQUEUE:
1457                         scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1458                         break;
1459                 default:
1460                         if (!scsi_eh_scmd_add(cmd, 0))
1461                                 scsi_finish_command(cmd);
1462         }
1463 }
1464
1465 /*
1466  * Function:    scsi_request_fn()
1467  *
1468  * Purpose:     Main strategy routine for SCSI.
1469  *
1470  * Arguments:   q       - Pointer to actual queue.
1471  *
1472  * Returns:     Nothing
1473  *
1474  * Lock status: IO request lock assumed to be held when called.
1475  */
1476 static void scsi_request_fn(struct request_queue *q)
1477 {
1478         struct scsi_device *sdev = q->queuedata;
1479         struct Scsi_Host *shost;
1480         struct scsi_cmnd *cmd;
1481         struct request *req;
1482
1483         if (!sdev) {
1484                 printk("scsi: killing requests for dead queue\n");
1485                 while ((req = elv_next_request(q)) != NULL)
1486                         scsi_kill_request(req, q);
1487                 return;
1488         }
1489
1490         if(!get_device(&sdev->sdev_gendev))
1491                 /* We must be tearing the block queue down already */
1492                 return;
1493
1494         /*
1495          * To start with, we keep looping until the queue is empty, or until
1496          * the host is no longer able to accept any more requests.
1497          */
1498         shost = sdev->host;
1499         while (!blk_queue_plugged(q)) {
1500                 int rtn;
1501                 /*
1502                  * get next queueable request.  We do this early to make sure
1503                  * that the request is fully prepared even if we cannot 
1504                  * accept it.
1505                  */
1506                 req = elv_next_request(q);
1507                 if (!req || !scsi_dev_queue_ready(q, sdev))
1508                         break;
1509
1510                 if (unlikely(!scsi_device_online(sdev))) {
1511                         sdev_printk(KERN_ERR, sdev,
1512                                     "rejecting I/O to offline device\n");
1513                         scsi_kill_request(req, q);
1514                         continue;
1515                 }
1516
1517
1518                 /*
1519                  * Remove the request from the request list.
1520                  */
1521                 if (!(blk_queue_tagged(q) && !blk_queue_start_tag(q, req)))
1522                         blkdev_dequeue_request(req);
1523                 sdev->device_busy++;
1524
1525                 spin_unlock(q->queue_lock);
1526                 cmd = req->special;
1527                 if (unlikely(cmd == NULL)) {
1528                         printk(KERN_CRIT "impossible request in %s.\n"
1529                                          "please mail a stack trace to "
1530                                          "linux-scsi@vger.kernel.org\n",
1531                                          __func__);
1532                         blk_dump_rq_flags(req, "foo");
1533                         BUG();
1534                 }
1535                 spin_lock(shost->host_lock);
1536
1537                 /*
1538                  * We hit this when the driver is using a host wide
1539                  * tag map. For device level tag maps the queue_depth check
1540                  * in the device ready fn would prevent us from trying
1541                  * to allocate a tag. Since the map is a shared host resource
1542                  * we add the dev to the starved list so it eventually gets
1543                  * a run when a tag is freed.
1544                  */
1545                 if (blk_queue_tagged(q) && !blk_rq_tagged(req)) {
1546                         if (list_empty(&sdev->starved_entry))
1547                                 list_add_tail(&sdev->starved_entry,
1548                                               &shost->starved_list);
1549                         goto not_ready;
1550                 }
1551
1552                 if (!scsi_target_queue_ready(shost, sdev))
1553                         goto not_ready;
1554
1555                 if (!scsi_host_queue_ready(q, shost, sdev))
1556                         goto not_ready;
1557
1558                 scsi_target(sdev)->target_busy++;
1559                 shost->host_busy++;
1560
1561                 /*
1562                  * XXX(hch): This is rather suboptimal, scsi_dispatch_cmd will
1563                  *              take the lock again.
1564                  */
1565                 spin_unlock_irq(shost->host_lock);
1566
1567                 /*
1568                  * Finally, initialize any error handling parameters, and set up
1569                  * the timers for timeouts.
1570                  */
1571                 scsi_init_cmd_errh(cmd);
1572
1573                 /*
1574                  * Dispatch the command to the low-level driver.
1575                  */
1576                 rtn = scsi_dispatch_cmd(cmd);
1577                 spin_lock_irq(q->queue_lock);
1578                 if(rtn) {
1579                         /* we're refusing the command; because of
1580                          * the way locks get dropped, we need to 
1581                          * check here if plugging is required */
1582                         if(sdev->device_busy == 0)
1583                                 blk_plug_device(q);
1584
1585                         break;
1586                 }
1587         }
1588
1589         goto out;
1590
1591  not_ready:
1592         spin_unlock_irq(shost->host_lock);
1593
1594         /*
1595          * lock q, handle tag, requeue req, and decrement device_busy. We
1596          * must return with queue_lock held.
1597          *
1598          * Decrementing device_busy without checking it is OK, as all such
1599          * cases (host limits or settings) should run the queue at some
1600          * later time.
1601          */
1602         spin_lock_irq(q->queue_lock);
1603         blk_requeue_request(q, req);
1604         sdev->device_busy--;
1605         if(sdev->device_busy == 0)
1606                 blk_plug_device(q);
1607  out:
1608         /* must be careful here...if we trigger the ->remove() function
1609          * we cannot be holding the q lock */
1610         spin_unlock_irq(q->queue_lock);
1611         put_device(&sdev->sdev_gendev);
1612         spin_lock_irq(q->queue_lock);
1613 }
1614
1615 u64 scsi_calculate_bounce_limit(struct Scsi_Host *shost)
1616 {
1617         struct device *host_dev;
1618         u64 bounce_limit = 0xffffffff;
1619
1620         if (shost->unchecked_isa_dma)
1621                 return BLK_BOUNCE_ISA;
1622         /*
1623          * Platforms with virtual-DMA translation
1624          * hardware have no practical limit.
1625          */
1626         if (!PCI_DMA_BUS_IS_PHYS)
1627                 return BLK_BOUNCE_ANY;
1628
1629         host_dev = scsi_get_device(shost);
1630         if (host_dev && host_dev->dma_mask)
1631                 bounce_limit = *host_dev->dma_mask;
1632
1633         return bounce_limit;
1634 }
1635 EXPORT_SYMBOL(scsi_calculate_bounce_limit);
1636
1637 struct request_queue *__scsi_alloc_queue(struct Scsi_Host *shost,
1638                                          request_fn_proc *request_fn)
1639 {
1640         struct request_queue *q;
1641         struct device *dev = shost->shost_gendev.parent;
1642
1643         q = blk_init_queue(request_fn, NULL);
1644         if (!q)
1645                 return NULL;
1646
1647         /*
1648          * this limit is imposed by hardware restrictions
1649          */
1650         blk_queue_max_hw_segments(q, shost->sg_tablesize);
1651         blk_queue_max_phys_segments(q, SCSI_MAX_SG_CHAIN_SEGMENTS);
1652
1653         blk_queue_max_sectors(q, shost->max_sectors);
1654         blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
1655         blk_queue_segment_boundary(q, shost->dma_boundary);
1656         dma_set_seg_boundary(dev, shost->dma_boundary);
1657
1658         blk_queue_max_segment_size(q, dma_get_max_seg_size(dev));
1659
1660         /* New queue, no concurrency on queue_flags */
1661         if (!shost->use_clustering)
1662                 queue_flag_clear_unlocked(QUEUE_FLAG_CLUSTER, q);
1663
1664         /*
1665          * set a reasonable default alignment on word boundaries: the
1666          * host and device may alter it using
1667          * blk_queue_update_dma_alignment() later.
1668          */
1669         blk_queue_dma_alignment(q, 0x03);
1670
1671         return q;
1672 }
1673 EXPORT_SYMBOL(__scsi_alloc_queue);
1674
1675 struct request_queue *scsi_alloc_queue(struct scsi_device *sdev)
1676 {
1677         struct request_queue *q;
1678
1679         q = __scsi_alloc_queue(sdev->host, scsi_request_fn);
1680         if (!q)
1681                 return NULL;
1682
1683         blk_queue_prep_rq(q, scsi_prep_fn);
1684         blk_queue_softirq_done(q, scsi_softirq_done);
1685         blk_queue_rq_timed_out(q, scsi_times_out);
1686         blk_queue_lld_busy(q, scsi_lld_busy);
1687         return q;
1688 }
1689
1690 void scsi_free_queue(struct request_queue *q)
1691 {
1692         blk_cleanup_queue(q);
1693 }
1694
1695 /*
1696  * Function:    scsi_block_requests()
1697  *
1698  * Purpose:     Utility function used by low-level drivers to prevent further
1699  *              commands from being queued to the device.
1700  *
1701  * Arguments:   shost       - Host in question
1702  *
1703  * Returns:     Nothing
1704  *
1705  * Lock status: No locks are assumed held.
1706  *
1707  * Notes:       There is no timer nor any other means by which the requests
1708  *              get unblocked other than the low-level driver calling
1709  *              scsi_unblock_requests().
1710  */
1711 void scsi_block_requests(struct Scsi_Host *shost)
1712 {
1713         shost->host_self_blocked = 1;
1714 }
1715 EXPORT_SYMBOL(scsi_block_requests);
1716
1717 /*
1718  * Function:    scsi_unblock_requests()
1719  *
1720  * Purpose:     Utility function used by low-level drivers to allow further
1721  *              commands from being queued to the device.
1722  *
1723  * Arguments:   shost       - Host in question
1724  *
1725  * Returns:     Nothing
1726  *
1727  * Lock status: No locks are assumed held.
1728  *
1729  * Notes:       There is no timer nor any other means by which the requests
1730  *              get unblocked other than the low-level driver calling
1731  *              scsi_unblock_requests().
1732  *
1733  *              This is done as an API function so that changes to the
1734  *              internals of the scsi mid-layer won't require wholesale
1735  *              changes to drivers that use this feature.
1736  */
1737 void scsi_unblock_requests(struct Scsi_Host *shost)
1738 {
1739         shost->host_self_blocked = 0;
1740         scsi_run_host_queues(shost);
1741 }
1742 EXPORT_SYMBOL(scsi_unblock_requests);
1743
1744 int __init scsi_init_queue(void)
1745 {
1746         int i;
1747
1748         scsi_sdb_cache = kmem_cache_create("scsi_data_buffer",
1749                                            sizeof(struct scsi_data_buffer),
1750                                            0, 0, NULL);
1751         if (!scsi_sdb_cache) {
1752                 printk(KERN_ERR "SCSI: can't init scsi sdb cache\n");
1753                 return -ENOMEM;
1754         }
1755
1756         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1757                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1758                 int size = sgp->size * sizeof(struct scatterlist);
1759
1760                 sgp->slab = kmem_cache_create(sgp->name, size, 0,
1761                                 SLAB_HWCACHE_ALIGN, NULL);
1762                 if (!sgp->slab) {
1763                         printk(KERN_ERR "SCSI: can't init sg slab %s\n",
1764                                         sgp->name);
1765                         goto cleanup_sdb;
1766                 }
1767
1768                 sgp->pool = mempool_create_slab_pool(SG_MEMPOOL_SIZE,
1769                                                      sgp->slab);
1770                 if (!sgp->pool) {
1771                         printk(KERN_ERR "SCSI: can't init sg mempool %s\n",
1772                                         sgp->name);
1773                         goto cleanup_sdb;
1774                 }
1775         }
1776
1777         return 0;
1778
1779 cleanup_sdb:
1780         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1781                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1782                 if (sgp->pool)
1783                         mempool_destroy(sgp->pool);
1784                 if (sgp->slab)
1785                         kmem_cache_destroy(sgp->slab);
1786         }
1787         kmem_cache_destroy(scsi_sdb_cache);
1788
1789         return -ENOMEM;
1790 }
1791
1792 void scsi_exit_queue(void)
1793 {
1794         int i;
1795
1796         kmem_cache_destroy(scsi_sdb_cache);
1797
1798         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1799                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1800                 mempool_destroy(sgp->pool);
1801                 kmem_cache_destroy(sgp->slab);
1802         }
1803 }
1804
1805 /**
1806  *      scsi_mode_select - issue a mode select
1807  *      @sdev:  SCSI device to be queried
1808  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
1809  *      @sp:    Save page bit (0 == don't save, 1 == save)
1810  *      @modepage: mode page being requested
1811  *      @buffer: request buffer (may not be smaller than eight bytes)
1812  *      @len:   length of request buffer.
1813  *      @timeout: command timeout
1814  *      @retries: number of retries before failing
1815  *      @data: returns a structure abstracting the mode header data
1816  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
1817  *              must be SCSI_SENSE_BUFFERSIZE big.
1818  *
1819  *      Returns zero if successful; negative error number or scsi
1820  *      status on error
1821  *
1822  */
1823 int
1824 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
1825                  unsigned char *buffer, int len, int timeout, int retries,
1826                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1827 {
1828         unsigned char cmd[10];
1829         unsigned char *real_buffer;
1830         int ret;
1831
1832         memset(cmd, 0, sizeof(cmd));
1833         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
1834
1835         if (sdev->use_10_for_ms) {
1836                 if (len > 65535)
1837                         return -EINVAL;
1838                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
1839                 if (!real_buffer)
1840                         return -ENOMEM;
1841                 memcpy(real_buffer + 8, buffer, len);
1842                 len += 8;
1843                 real_buffer[0] = 0;
1844                 real_buffer[1] = 0;
1845                 real_buffer[2] = data->medium_type;
1846                 real_buffer[3] = data->device_specific;
1847                 real_buffer[4] = data->longlba ? 0x01 : 0;
1848                 real_buffer[5] = 0;
1849                 real_buffer[6] = data->block_descriptor_length >> 8;
1850                 real_buffer[7] = data->block_descriptor_length;
1851
1852                 cmd[0] = MODE_SELECT_10;
1853                 cmd[7] = len >> 8;
1854                 cmd[8] = len;
1855         } else {
1856                 if (len > 255 || data->block_descriptor_length > 255 ||
1857                     data->longlba)
1858                         return -EINVAL;
1859
1860                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
1861                 if (!real_buffer)
1862                         return -ENOMEM;
1863                 memcpy(real_buffer + 4, buffer, len);
1864                 len += 4;
1865                 real_buffer[0] = 0;
1866                 real_buffer[1] = data->medium_type;
1867                 real_buffer[2] = data->device_specific;
1868                 real_buffer[3] = data->block_descriptor_length;
1869                 
1870
1871                 cmd[0] = MODE_SELECT;
1872                 cmd[4] = len;
1873         }
1874
1875         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
1876                                sshdr, timeout, retries, NULL);
1877         kfree(real_buffer);
1878         return ret;
1879 }
1880 EXPORT_SYMBOL_GPL(scsi_mode_select);
1881
1882 /**
1883  *      scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
1884  *      @sdev:  SCSI device to be queried
1885  *      @dbd:   set if mode sense will allow block descriptors to be returned
1886  *      @modepage: mode page being requested
1887  *      @buffer: request buffer (may not be smaller than eight bytes)
1888  *      @len:   length of request buffer.
1889  *      @timeout: command timeout
1890  *      @retries: number of retries before failing
1891  *      @data: returns a structure abstracting the mode header data
1892  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
1893  *              must be SCSI_SENSE_BUFFERSIZE big.
1894  *
1895  *      Returns zero if unsuccessful, or the header offset (either 4
1896  *      or 8 depending on whether a six or ten byte command was
1897  *      issued) if successful.
1898  */
1899 int
1900 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
1901                   unsigned char *buffer, int len, int timeout, int retries,
1902                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1903 {
1904         unsigned char cmd[12];
1905         int use_10_for_ms;
1906         int header_length;
1907         int result;
1908         struct scsi_sense_hdr my_sshdr;
1909
1910         memset(data, 0, sizeof(*data));
1911         memset(&cmd[0], 0, 12);
1912         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
1913         cmd[2] = modepage;
1914
1915         /* caller might not be interested in sense, but we need it */
1916         if (!sshdr)
1917                 sshdr = &my_sshdr;
1918
1919  retry:
1920         use_10_for_ms = sdev->use_10_for_ms;
1921
1922         if (use_10_for_ms) {
1923                 if (len < 8)
1924                         len = 8;
1925
1926                 cmd[0] = MODE_SENSE_10;
1927                 cmd[8] = len;
1928                 header_length = 8;
1929         } else {
1930                 if (len < 4)
1931                         len = 4;
1932
1933                 cmd[0] = MODE_SENSE;
1934                 cmd[4] = len;
1935                 header_length = 4;
1936         }
1937
1938         memset(buffer, 0, len);
1939
1940         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
1941                                   sshdr, timeout, retries, NULL);
1942
1943         /* This code looks awful: what it's doing is making sure an
1944          * ILLEGAL REQUEST sense return identifies the actual command
1945          * byte as the problem.  MODE_SENSE commands can return
1946          * ILLEGAL REQUEST if the code page isn't supported */
1947
1948         if (use_10_for_ms && !scsi_status_is_good(result) &&
1949             (driver_byte(result) & DRIVER_SENSE)) {
1950                 if (scsi_sense_valid(sshdr)) {
1951                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
1952                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
1953                                 /* 
1954                                  * Invalid command operation code
1955                                  */
1956                                 sdev->use_10_for_ms = 0;
1957                                 goto retry;
1958                         }
1959                 }
1960         }
1961
1962         if(scsi_status_is_good(result)) {
1963                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
1964                              (modepage == 6 || modepage == 8))) {
1965                         /* Initio breakage? */
1966                         header_length = 0;
1967                         data->length = 13;
1968                         data->medium_type = 0;
1969                         data->device_specific = 0;
1970                         data->longlba = 0;
1971                         data->block_descriptor_length = 0;
1972                 } else if(use_10_for_ms) {
1973                         data->length = buffer[0]*256 + buffer[1] + 2;
1974                         data->medium_type = buffer[2];
1975                         data->device_specific = buffer[3];
1976                         data->longlba = buffer[4] & 0x01;
1977                         data->block_descriptor_length = buffer[6]*256
1978                                 + buffer[7];
1979                 } else {
1980                         data->length = buffer[0] + 1;
1981                         data->medium_type = buffer[1];
1982                         data->device_specific = buffer[2];
1983                         data->block_descriptor_length = buffer[3];
1984                 }
1985                 data->header_length = header_length;
1986         }
1987
1988         return result;
1989 }
1990 EXPORT_SYMBOL(scsi_mode_sense);
1991
1992 /**
1993  *      scsi_test_unit_ready - test if unit is ready
1994  *      @sdev:  scsi device to change the state of.
1995  *      @timeout: command timeout
1996  *      @retries: number of retries before failing
1997  *      @sshdr_external: Optional pointer to struct scsi_sense_hdr for
1998  *              returning sense. Make sure that this is cleared before passing
1999  *              in.
2000  *
2001  *      Returns zero if unsuccessful or an error if TUR failed.  For
2002  *      removable media, a return of NOT_READY or UNIT_ATTENTION is
2003  *      translated to success, with the ->changed flag updated.
2004  **/
2005 int
2006 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
2007                      struct scsi_sense_hdr *sshdr_external)
2008 {
2009         char cmd[] = {
2010                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
2011         };
2012         struct scsi_sense_hdr *sshdr;
2013         int result;
2014
2015         if (!sshdr_external)
2016                 sshdr = kzalloc(sizeof(*sshdr), GFP_KERNEL);
2017         else
2018                 sshdr = sshdr_external;
2019
2020         /* try to eat the UNIT_ATTENTION if there are enough retries */
2021         do {
2022                 result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
2023                                           timeout, retries, NULL);
2024                 if (sdev->removable && scsi_sense_valid(sshdr) &&
2025                     sshdr->sense_key == UNIT_ATTENTION)
2026                         sdev->changed = 1;
2027         } while (scsi_sense_valid(sshdr) &&
2028                  sshdr->sense_key == UNIT_ATTENTION && --retries);
2029
2030         if (!sshdr)
2031                 /* could not allocate sense buffer, so can't process it */
2032                 return result;
2033
2034         if (sdev->removable && scsi_sense_valid(sshdr) &&
2035             (sshdr->sense_key == UNIT_ATTENTION ||
2036              sshdr->sense_key == NOT_READY)) {
2037                 sdev->changed = 1;
2038                 result = 0;
2039         }
2040         if (!sshdr_external)
2041                 kfree(sshdr);
2042         return result;
2043 }
2044 EXPORT_SYMBOL(scsi_test_unit_ready);
2045
2046 /**
2047  *      scsi_device_set_state - Take the given device through the device state model.
2048  *      @sdev:  scsi device to change the state of.
2049  *      @state: state to change to.
2050  *
2051  *      Returns zero if unsuccessful or an error if the requested 
2052  *      transition is illegal.
2053  */
2054 int
2055 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2056 {
2057         enum scsi_device_state oldstate = sdev->sdev_state;
2058
2059         if (state == oldstate)
2060                 return 0;
2061
2062         switch (state) {
2063         case SDEV_CREATED:
2064                 switch (oldstate) {
2065                 case SDEV_CREATED_BLOCK:
2066                         break;
2067                 default:
2068                         goto illegal;
2069                 }
2070                 break;
2071                         
2072         case SDEV_RUNNING:
2073                 switch (oldstate) {
2074                 case SDEV_CREATED:
2075                 case SDEV_OFFLINE:
2076                 case SDEV_QUIESCE:
2077                 case SDEV_BLOCK:
2078                         break;
2079                 default:
2080                         goto illegal;
2081                 }
2082                 break;
2083
2084         case SDEV_QUIESCE:
2085                 switch (oldstate) {
2086                 case SDEV_RUNNING:
2087                 case SDEV_OFFLINE:
2088                         break;
2089                 default:
2090                         goto illegal;
2091                 }
2092                 break;
2093
2094         case SDEV_OFFLINE:
2095                 switch (oldstate) {
2096                 case SDEV_CREATED:
2097                 case SDEV_RUNNING:
2098                 case SDEV_QUIESCE:
2099                 case SDEV_BLOCK:
2100                         break;
2101                 default:
2102                         goto illegal;
2103                 }
2104                 break;
2105
2106         case SDEV_BLOCK:
2107                 switch (oldstate) {
2108                 case SDEV_RUNNING:
2109                 case SDEV_CREATED_BLOCK:
2110                         break;
2111                 default:
2112                         goto illegal;
2113                 }
2114                 break;
2115
2116         case SDEV_CREATED_BLOCK:
2117                 switch (oldstate) {
2118                 case SDEV_CREATED:
2119                         break;
2120                 default:
2121                         goto illegal;
2122                 }
2123                 break;
2124
2125         case SDEV_CANCEL:
2126                 switch (oldstate) {
2127                 case SDEV_CREATED:
2128                 case SDEV_RUNNING:
2129                 case SDEV_QUIESCE:
2130                 case SDEV_OFFLINE:
2131                 case SDEV_BLOCK:
2132                         break;
2133                 default:
2134                         goto illegal;
2135                 }
2136                 break;
2137
2138         case SDEV_DEL:
2139                 switch (oldstate) {
2140                 case SDEV_CREATED:
2141                 case SDEV_RUNNING:
2142                 case SDEV_OFFLINE:
2143                 case SDEV_CANCEL:
2144                         break;
2145                 default:
2146                         goto illegal;
2147                 }
2148                 break;
2149
2150         }
2151         sdev->sdev_state = state;
2152         return 0;
2153
2154  illegal:
2155         SCSI_LOG_ERROR_RECOVERY(1, 
2156                                 sdev_printk(KERN_ERR, sdev,
2157                                             "Illegal state transition %s->%s\n",
2158                                             scsi_device_state_name(oldstate),
2159                                             scsi_device_state_name(state))
2160                                 );
2161         return -EINVAL;
2162 }
2163 EXPORT_SYMBOL(scsi_device_set_state);
2164
2165 /**
2166  *      sdev_evt_emit - emit a single SCSI device uevent
2167  *      @sdev: associated SCSI device
2168  *      @evt: event to emit
2169  *
2170  *      Send a single uevent (scsi_event) to the associated scsi_device.
2171  */
2172 static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2173 {
2174         int idx = 0;
2175         char *envp[3];
2176
2177         switch (evt->evt_type) {
2178         case SDEV_EVT_MEDIA_CHANGE:
2179                 envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2180                 break;
2181
2182         default:
2183                 /* do nothing */
2184                 break;
2185         }
2186
2187         envp[idx++] = NULL;
2188
2189         kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2190 }
2191
2192 /**
2193  *      sdev_evt_thread - send a uevent for each scsi event
2194  *      @work: work struct for scsi_device
2195  *
2196  *      Dispatch queued events to their associated scsi_device kobjects
2197  *      as uevents.
2198  */
2199 void scsi_evt_thread(struct work_struct *work)
2200 {
2201         struct scsi_device *sdev;
2202         LIST_HEAD(event_list);
2203
2204         sdev = container_of(work, struct scsi_device, event_work);
2205
2206         while (1) {
2207                 struct scsi_event *evt;
2208                 struct list_head *this, *tmp;
2209                 unsigned long flags;
2210
2211                 spin_lock_irqsave(&sdev->list_lock, flags);
2212                 list_splice_init(&sdev->event_list, &event_list);
2213                 spin_unlock_irqrestore(&sdev->list_lock, flags);
2214
2215                 if (list_empty(&event_list))
2216                         break;
2217
2218                 list_for_each_safe(this, tmp, &event_list) {
2219                         evt = list_entry(this, struct scsi_event, node);
2220                         list_del(&evt->node);
2221                         scsi_evt_emit(sdev, evt);
2222                         kfree(evt);
2223                 }
2224         }
2225 }
2226
2227 /**
2228  *      sdev_evt_send - send asserted event to uevent thread
2229  *      @sdev: scsi_device event occurred on
2230  *      @evt: event to send
2231  *
2232  *      Assert scsi device event asynchronously.
2233  */
2234 void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2235 {
2236         unsigned long flags;
2237
2238 #if 0
2239         /* FIXME: currently this check eliminates all media change events
2240          * for polled devices.  Need to update to discriminate between AN
2241          * and polled events */
2242         if (!test_bit(evt->evt_type, sdev->supported_events)) {
2243                 kfree(evt);
2244                 return;
2245         }
2246 #endif
2247
2248         spin_lock_irqsave(&sdev->list_lock, flags);
2249         list_add_tail(&evt->node, &sdev->event_list);
2250         schedule_work(&sdev->event_work);
2251         spin_unlock_irqrestore(&sdev->list_lock, flags);
2252 }
2253 EXPORT_SYMBOL_GPL(sdev_evt_send);
2254
2255 /**
2256  *      sdev_evt_alloc - allocate a new scsi event
2257  *      @evt_type: type of event to allocate
2258  *      @gfpflags: GFP flags for allocation
2259  *
2260  *      Allocates and returns a new scsi_event.
2261  */
2262 struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2263                                   gfp_t gfpflags)
2264 {
2265         struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2266         if (!evt)
2267                 return NULL;
2268
2269         evt->evt_type = evt_type;
2270         INIT_LIST_HEAD(&evt->node);
2271
2272         /* evt_type-specific initialization, if any */
2273         switch (evt_type) {
2274         case SDEV_EVT_MEDIA_CHANGE:
2275         default:
2276                 /* do nothing */
2277                 break;
2278         }
2279
2280         return evt;
2281 }
2282 EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2283
2284 /**
2285  *      sdev_evt_send_simple - send asserted event to uevent thread
2286  *      @sdev: scsi_device event occurred on
2287  *      @evt_type: type of event to send
2288  *      @gfpflags: GFP flags for allocation
2289  *
2290  *      Assert scsi device event asynchronously, given an event type.
2291  */
2292 void sdev_evt_send_simple(struct scsi_device *sdev,
2293                           enum scsi_device_event evt_type, gfp_t gfpflags)
2294 {
2295         struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2296         if (!evt) {
2297                 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2298                             evt_type);
2299                 return;
2300         }
2301
2302         sdev_evt_send(sdev, evt);
2303 }
2304 EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2305
2306 /**
2307  *      scsi_device_quiesce - Block user issued commands.
2308  *      @sdev:  scsi device to quiesce.
2309  *
2310  *      This works by trying to transition to the SDEV_QUIESCE state
2311  *      (which must be a legal transition).  When the device is in this
2312  *      state, only special requests will be accepted, all others will
2313  *      be deferred.  Since special requests may also be requeued requests,
2314  *      a successful return doesn't guarantee the device will be 
2315  *      totally quiescent.
2316  *
2317  *      Must be called with user context, may sleep.
2318  *
2319  *      Returns zero if unsuccessful or an error if not.
2320  */
2321 int
2322 scsi_device_quiesce(struct scsi_device *sdev)
2323 {
2324         int err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2325         if (err)
2326                 return err;
2327
2328         scsi_run_queue(sdev->request_queue);
2329         while (sdev->device_busy) {
2330                 msleep_interruptible(200);
2331                 scsi_run_queue(sdev->request_queue);
2332         }
2333         return 0;
2334 }
2335 EXPORT_SYMBOL(scsi_device_quiesce);
2336
2337 /**
2338  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2339  *      @sdev:  scsi device to resume.
2340  *
2341  *      Moves the device from quiesced back to running and restarts the
2342  *      queues.
2343  *
2344  *      Must be called with user context, may sleep.
2345  */
2346 void
2347 scsi_device_resume(struct scsi_device *sdev)
2348 {
2349         if(scsi_device_set_state(sdev, SDEV_RUNNING))
2350                 return;
2351         scsi_run_queue(sdev->request_queue);
2352 }
2353 EXPORT_SYMBOL(scsi_device_resume);
2354
2355 static void
2356 device_quiesce_fn(struct scsi_device *sdev, void *data)
2357 {
2358         scsi_device_quiesce(sdev);
2359 }
2360
2361 void
2362 scsi_target_quiesce(struct scsi_target *starget)
2363 {
2364         starget_for_each_device(starget, NULL, device_quiesce_fn);
2365 }
2366 EXPORT_SYMBOL(scsi_target_quiesce);
2367
2368 static void
2369 device_resume_fn(struct scsi_device *sdev, void *data)
2370 {
2371         scsi_device_resume(sdev);
2372 }
2373
2374 void
2375 scsi_target_resume(struct scsi_target *starget)
2376 {
2377         starget_for_each_device(starget, NULL, device_resume_fn);
2378 }
2379 EXPORT_SYMBOL(scsi_target_resume);
2380
2381 /**
2382  * scsi_internal_device_block - internal function to put a device temporarily into the SDEV_BLOCK state
2383  * @sdev:       device to block
2384  *
2385  * Block request made by scsi lld's to temporarily stop all
2386  * scsi commands on the specified device.  Called from interrupt
2387  * or normal process context.
2388  *
2389  * Returns zero if successful or error if not
2390  *
2391  * Notes:       
2392  *      This routine transitions the device to the SDEV_BLOCK state
2393  *      (which must be a legal transition).  When the device is in this
2394  *      state, all commands are deferred until the scsi lld reenables
2395  *      the device with scsi_device_unblock or device_block_tmo fires.
2396  *      This routine assumes the host_lock is held on entry.
2397  */
2398 int
2399 scsi_internal_device_block(struct scsi_device *sdev)
2400 {
2401         struct request_queue *q = sdev->request_queue;
2402         unsigned long flags;
2403         int err = 0;
2404
2405         err = scsi_device_set_state(sdev, SDEV_BLOCK);
2406         if (err) {
2407                 err = scsi_device_set_state(sdev, SDEV_CREATED_BLOCK);
2408
2409                 if (err)
2410                         return err;
2411         }
2412
2413         /* 
2414          * The device has transitioned to SDEV_BLOCK.  Stop the
2415          * block layer from calling the midlayer with this device's
2416          * request queue. 
2417          */
2418         spin_lock_irqsave(q->queue_lock, flags);
2419         blk_stop_queue(q);
2420         spin_unlock_irqrestore(q->queue_lock, flags);
2421
2422         return 0;
2423 }
2424 EXPORT_SYMBOL_GPL(scsi_internal_device_block);
2425  
2426 /**
2427  * scsi_internal_device_unblock - resume a device after a block request
2428  * @sdev:       device to resume
2429  *
2430  * Called by scsi lld's or the midlayer to restart the device queue
2431  * for the previously suspended scsi device.  Called from interrupt or
2432  * normal process context.
2433  *
2434  * Returns zero if successful or error if not.
2435  *
2436  * Notes:       
2437  *      This routine transitions the device to the SDEV_RUNNING state
2438  *      (which must be a legal transition) allowing the midlayer to
2439  *      goose the queue for this device.  This routine assumes the 
2440  *      host_lock is held upon entry.
2441  */
2442 int
2443 scsi_internal_device_unblock(struct scsi_device *sdev)
2444 {
2445         struct request_queue *q = sdev->request_queue; 
2446         int err;
2447         unsigned long flags;
2448         
2449         /* 
2450          * Try to transition the scsi device to SDEV_RUNNING
2451          * and goose the device queue if successful.  
2452          */
2453         err = scsi_device_set_state(sdev, SDEV_RUNNING);
2454         if (err) {
2455                 err = scsi_device_set_state(sdev, SDEV_CREATED);
2456
2457                 if (err)
2458                         return err;
2459         }
2460
2461         spin_lock_irqsave(q->queue_lock, flags);
2462         blk_start_queue(q);
2463         spin_unlock_irqrestore(q->queue_lock, flags);
2464
2465         return 0;
2466 }
2467 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock);
2468
2469 static void
2470 device_block(struct scsi_device *sdev, void *data)
2471 {
2472         scsi_internal_device_block(sdev);
2473 }
2474
2475 static int
2476 target_block(struct device *dev, void *data)
2477 {
2478         if (scsi_is_target_device(dev))
2479                 starget_for_each_device(to_scsi_target(dev), NULL,
2480                                         device_block);
2481         return 0;
2482 }
2483
2484 void
2485 scsi_target_block(struct device *dev)
2486 {
2487         if (scsi_is_target_device(dev))
2488                 starget_for_each_device(to_scsi_target(dev), NULL,
2489                                         device_block);
2490         else
2491                 device_for_each_child(dev, NULL, target_block);
2492 }
2493 EXPORT_SYMBOL_GPL(scsi_target_block);
2494
2495 static void
2496 device_unblock(struct scsi_device *sdev, void *data)
2497 {
2498         scsi_internal_device_unblock(sdev);
2499 }
2500
2501 static int
2502 target_unblock(struct device *dev, void *data)
2503 {
2504         if (scsi_is_target_device(dev))
2505                 starget_for_each_device(to_scsi_target(dev), NULL,
2506                                         device_unblock);
2507         return 0;
2508 }
2509
2510 void
2511 scsi_target_unblock(struct device *dev)
2512 {
2513         if (scsi_is_target_device(dev))
2514                 starget_for_each_device(to_scsi_target(dev), NULL,
2515                                         device_unblock);
2516         else
2517                 device_for_each_child(dev, NULL, target_unblock);
2518 }
2519 EXPORT_SYMBOL_GPL(scsi_target_unblock);
2520
2521 /**
2522  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2523  * @sgl:        scatter-gather list
2524  * @sg_count:   number of segments in sg
2525  * @offset:     offset in bytes into sg, on return offset into the mapped area
2526  * @len:        bytes to map, on return number of bytes mapped
2527  *
2528  * Returns virtual address of the start of the mapped page
2529  */
2530 void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
2531                           size_t *offset, size_t *len)
2532 {
2533         int i;
2534         size_t sg_len = 0, len_complete = 0;
2535         struct scatterlist *sg;
2536         struct page *page;
2537
2538         WARN_ON(!irqs_disabled());
2539
2540         for_each_sg(sgl, sg, sg_count, i) {
2541                 len_complete = sg_len; /* Complete sg-entries */
2542                 sg_len += sg->length;
2543                 if (sg_len > *offset)
2544                         break;
2545         }
2546
2547         if (unlikely(i == sg_count)) {
2548                 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2549                         "elements %d\n",
2550                        __func__, sg_len, *offset, sg_count);
2551                 WARN_ON(1);
2552                 return NULL;
2553         }
2554
2555         /* Offset starting from the beginning of first page in this sg-entry */
2556         *offset = *offset - len_complete + sg->offset;
2557
2558         /* Assumption: contiguous pages can be accessed as "page + i" */
2559         page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
2560         *offset &= ~PAGE_MASK;
2561
2562         /* Bytes in this sg-entry from *offset to the end of the page */
2563         sg_len = PAGE_SIZE - *offset;
2564         if (*len > sg_len)
2565                 *len = sg_len;
2566
2567         return kmap_atomic(page, KM_BIO_SRC_IRQ);
2568 }
2569 EXPORT_SYMBOL(scsi_kmap_atomic_sg);
2570
2571 /**
2572  * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
2573  * @virt:       virtual address to be unmapped
2574  */
2575 void scsi_kunmap_atomic_sg(void *virt)
2576 {
2577         kunmap_atomic(virt, KM_BIO_SRC_IRQ);
2578 }
2579 EXPORT_SYMBOL(scsi_kunmap_atomic_sg);