Merge tag 'isci-for-3.5' into misc
[linux-3.10.git] / drivers / scsi / scsi_lib.c
1 /*
2  *  scsi_lib.c Copyright (C) 1999 Eric Youngdale
3  *
4  *  SCSI queueing library.
5  *      Initial versions: Eric Youngdale (eric@andante.org).
6  *                        Based upon conversations with large numbers
7  *                        of people at Linux Expo.
8  */
9
10 #include <linux/bio.h>
11 #include <linux/bitops.h>
12 #include <linux/blkdev.h>
13 #include <linux/completion.h>
14 #include <linux/kernel.h>
15 #include <linux/export.h>
16 #include <linux/mempool.h>
17 #include <linux/slab.h>
18 #include <linux/init.h>
19 #include <linux/pci.h>
20 #include <linux/delay.h>
21 #include <linux/hardirq.h>
22 #include <linux/scatterlist.h>
23
24 #include <scsi/scsi.h>
25 #include <scsi/scsi_cmnd.h>
26 #include <scsi/scsi_dbg.h>
27 #include <scsi/scsi_device.h>
28 #include <scsi/scsi_driver.h>
29 #include <scsi/scsi_eh.h>
30 #include <scsi/scsi_host.h>
31
32 #include "scsi_priv.h"
33 #include "scsi_logging.h"
34
35
36 #define SG_MEMPOOL_NR           ARRAY_SIZE(scsi_sg_pools)
37 #define SG_MEMPOOL_SIZE         2
38
39 struct scsi_host_sg_pool {
40         size_t          size;
41         char            *name;
42         struct kmem_cache       *slab;
43         mempool_t       *pool;
44 };
45
46 #define SP(x) { x, "sgpool-" __stringify(x) }
47 #if (SCSI_MAX_SG_SEGMENTS < 32)
48 #error SCSI_MAX_SG_SEGMENTS is too small (must be 32 or greater)
49 #endif
50 static struct scsi_host_sg_pool scsi_sg_pools[] = {
51         SP(8),
52         SP(16),
53 #if (SCSI_MAX_SG_SEGMENTS > 32)
54         SP(32),
55 #if (SCSI_MAX_SG_SEGMENTS > 64)
56         SP(64),
57 #if (SCSI_MAX_SG_SEGMENTS > 128)
58         SP(128),
59 #if (SCSI_MAX_SG_SEGMENTS > 256)
60 #error SCSI_MAX_SG_SEGMENTS is too large (256 MAX)
61 #endif
62 #endif
63 #endif
64 #endif
65         SP(SCSI_MAX_SG_SEGMENTS)
66 };
67 #undef SP
68
69 struct kmem_cache *scsi_sdb_cache;
70
71 /*
72  * When to reinvoke queueing after a resource shortage. It's 3 msecs to
73  * not change behaviour from the previous unplug mechanism, experimentation
74  * may prove this needs changing.
75  */
76 #define SCSI_QUEUE_DELAY        3
77
78 /*
79  * Function:    scsi_unprep_request()
80  *
81  * Purpose:     Remove all preparation done for a request, including its
82  *              associated scsi_cmnd, so that it can be requeued.
83  *
84  * Arguments:   req     - request to unprepare
85  *
86  * Lock status: Assumed that no locks are held upon entry.
87  *
88  * Returns:     Nothing.
89  */
90 static void scsi_unprep_request(struct request *req)
91 {
92         struct scsi_cmnd *cmd = req->special;
93
94         blk_unprep_request(req);
95         req->special = NULL;
96
97         scsi_put_command(cmd);
98 }
99
100 /**
101  * __scsi_queue_insert - private queue insertion
102  * @cmd: The SCSI command being requeued
103  * @reason:  The reason for the requeue
104  * @unbusy: Whether the queue should be unbusied
105  *
106  * This is a private queue insertion.  The public interface
107  * scsi_queue_insert() always assumes the queue should be unbusied
108  * because it's always called before the completion.  This function is
109  * for a requeue after completion, which should only occur in this
110  * file.
111  */
112 static int __scsi_queue_insert(struct scsi_cmnd *cmd, int reason, int unbusy)
113 {
114         struct Scsi_Host *host = cmd->device->host;
115         struct scsi_device *device = cmd->device;
116         struct scsi_target *starget = scsi_target(device);
117         struct request_queue *q = device->request_queue;
118         unsigned long flags;
119
120         SCSI_LOG_MLQUEUE(1,
121                  printk("Inserting command %p into mlqueue\n", cmd));
122
123         /*
124          * Set the appropriate busy bit for the device/host.
125          *
126          * If the host/device isn't busy, assume that something actually
127          * completed, and that we should be able to queue a command now.
128          *
129          * Note that the prior mid-layer assumption that any host could
130          * always queue at least one command is now broken.  The mid-layer
131          * will implement a user specifiable stall (see
132          * scsi_host.max_host_blocked and scsi_device.max_device_blocked)
133          * if a command is requeued with no other commands outstanding
134          * either for the device or for the host.
135          */
136         switch (reason) {
137         case SCSI_MLQUEUE_HOST_BUSY:
138                 host->host_blocked = host->max_host_blocked;
139                 break;
140         case SCSI_MLQUEUE_DEVICE_BUSY:
141         case SCSI_MLQUEUE_EH_RETRY:
142                 device->device_blocked = device->max_device_blocked;
143                 break;
144         case SCSI_MLQUEUE_TARGET_BUSY:
145                 starget->target_blocked = starget->max_target_blocked;
146                 break;
147         }
148
149         /*
150          * Decrement the counters, since these commands are no longer
151          * active on the host/device.
152          */
153         if (unbusy)
154                 scsi_device_unbusy(device);
155
156         /*
157          * Requeue this command.  It will go before all other commands
158          * that are already in the queue.
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         kblockd_schedule_work(q, &device->requeue_work);
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         if (!req)
219                 return ret;
220
221         if (bufflen &&  blk_rq_map_kern(sdev->request_queue, req,
222                                         buffer, bufflen, __GFP_WAIT))
223                 goto out;
224
225         req->cmd_len = COMMAND_SIZE(cmd[0]);
226         memcpy(req->cmd, cmd, req->cmd_len);
227         req->sense = sense;
228         req->sense_len = 0;
229         req->retries = retries;
230         req->timeout = timeout;
231         req->cmd_type = REQ_TYPE_BLOCK_PC;
232         req->cmd_flags |= flags | REQ_QUIET | REQ_PREEMPT;
233
234         /*
235          * head injection *required* here otherwise quiesce won't work
236          */
237         blk_execute_rq(req->q, NULL, req, 1);
238
239         /*
240          * Some devices (USB mass-storage in particular) may transfer
241          * garbage data together with a residue indicating that the data
242          * is invalid.  Prevent the garbage from being misinterpreted
243          * and prevent security leaks by zeroing out the excess data.
244          */
245         if (unlikely(req->resid_len > 0 && req->resid_len <= bufflen))
246                 memset(buffer + (bufflen - req->resid_len), 0, req->resid_len);
247
248         if (resid)
249                 *resid = req->resid_len;
250         ret = req->errors;
251  out:
252         blk_put_request(req);
253
254         return ret;
255 }
256 EXPORT_SYMBOL(scsi_execute);
257
258
259 int scsi_execute_req(struct scsi_device *sdev, const unsigned char *cmd,
260                      int data_direction, void *buffer, unsigned bufflen,
261                      struct scsi_sense_hdr *sshdr, int timeout, int retries,
262                      int *resid)
263 {
264         char *sense = NULL;
265         int result;
266         
267         if (sshdr) {
268                 sense = kzalloc(SCSI_SENSE_BUFFERSIZE, GFP_NOIO);
269                 if (!sense)
270                         return DRIVER_ERROR << 24;
271         }
272         result = scsi_execute(sdev, cmd, data_direction, buffer, bufflen,
273                               sense, timeout, retries, 0, resid);
274         if (sshdr)
275                 scsi_normalize_sense(sense, SCSI_SENSE_BUFFERSIZE, sshdr);
276
277         kfree(sense);
278         return result;
279 }
280 EXPORT_SYMBOL(scsi_execute_req);
281
282 /*
283  * Function:    scsi_init_cmd_errh()
284  *
285  * Purpose:     Initialize cmd fields related to error handling.
286  *
287  * Arguments:   cmd     - command that is ready to be queued.
288  *
289  * Notes:       This function has the job of initializing a number of
290  *              fields related to error handling.   Typically this will
291  *              be called once for each command, as required.
292  */
293 static void scsi_init_cmd_errh(struct scsi_cmnd *cmd)
294 {
295         cmd->serial_number = 0;
296         scsi_set_resid(cmd, 0);
297         memset(cmd->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
298         if (cmd->cmd_len == 0)
299                 cmd->cmd_len = scsi_command_size(cmd->cmnd);
300 }
301
302 void scsi_device_unbusy(struct scsi_device *sdev)
303 {
304         struct Scsi_Host *shost = sdev->host;
305         struct scsi_target *starget = scsi_target(sdev);
306         unsigned long flags;
307
308         spin_lock_irqsave(shost->host_lock, flags);
309         shost->host_busy--;
310         starget->target_busy--;
311         if (unlikely(scsi_host_in_recovery(shost) &&
312                      (shost->host_failed || shost->host_eh_scheduled)))
313                 scsi_eh_wakeup(shost);
314         spin_unlock(shost->host_lock);
315         spin_lock(sdev->request_queue->queue_lock);
316         sdev->device_busy--;
317         spin_unlock_irqrestore(sdev->request_queue->queue_lock, flags);
318 }
319
320 /*
321  * Called for single_lun devices on IO completion. Clear starget_sdev_user,
322  * and call blk_run_queue for all the scsi_devices on the target -
323  * including current_sdev first.
324  *
325  * Called with *no* scsi locks held.
326  */
327 static void scsi_single_lun_run(struct scsi_device *current_sdev)
328 {
329         struct Scsi_Host *shost = current_sdev->host;
330         struct scsi_device *sdev, *tmp;
331         struct scsi_target *starget = scsi_target(current_sdev);
332         unsigned long flags;
333
334         spin_lock_irqsave(shost->host_lock, flags);
335         starget->starget_sdev_user = NULL;
336         spin_unlock_irqrestore(shost->host_lock, flags);
337
338         /*
339          * Call blk_run_queue for all LUNs on the target, starting with
340          * current_sdev. We race with others (to set starget_sdev_user),
341          * but in most cases, we will be first. Ideally, each LU on the
342          * target would get some limited time or requests on the target.
343          */
344         blk_run_queue(current_sdev->request_queue);
345
346         spin_lock_irqsave(shost->host_lock, flags);
347         if (starget->starget_sdev_user)
348                 goto out;
349         list_for_each_entry_safe(sdev, tmp, &starget->devices,
350                         same_target_siblings) {
351                 if (sdev == current_sdev)
352                         continue;
353                 if (scsi_device_get(sdev))
354                         continue;
355
356                 spin_unlock_irqrestore(shost->host_lock, flags);
357                 blk_run_queue(sdev->request_queue);
358                 spin_lock_irqsave(shost->host_lock, flags);
359         
360                 scsi_device_put(sdev);
361         }
362  out:
363         spin_unlock_irqrestore(shost->host_lock, flags);
364 }
365
366 static inline int scsi_device_is_busy(struct scsi_device *sdev)
367 {
368         if (sdev->device_busy >= sdev->queue_depth || sdev->device_blocked)
369                 return 1;
370
371         return 0;
372 }
373
374 static inline int scsi_target_is_busy(struct scsi_target *starget)
375 {
376         return ((starget->can_queue > 0 &&
377                  starget->target_busy >= starget->can_queue) ||
378                  starget->target_blocked);
379 }
380
381 static inline int scsi_host_is_busy(struct Scsi_Host *shost)
382 {
383         if ((shost->can_queue > 0 && shost->host_busy >= shost->can_queue) ||
384             shost->host_blocked || shost->host_self_blocked)
385                 return 1;
386
387         return 0;
388 }
389
390 /*
391  * Function:    scsi_run_queue()
392  *
393  * Purpose:     Select a proper request queue to serve next
394  *
395  * Arguments:   q       - last request's queue
396  *
397  * Returns:     Nothing
398  *
399  * Notes:       The previous command was completely finished, start
400  *              a new one if possible.
401  */
402 static void scsi_run_queue(struct request_queue *q)
403 {
404         struct scsi_device *sdev = q->queuedata;
405         struct Scsi_Host *shost;
406         LIST_HEAD(starved_list);
407         unsigned long flags;
408
409         /* if the device is dead, sdev will be NULL, so no queue to run */
410         if (!sdev)
411                 return;
412
413         shost = sdev->host;
414         if (scsi_target(sdev)->single_lun)
415                 scsi_single_lun_run(sdev);
416
417         spin_lock_irqsave(shost->host_lock, flags);
418         list_splice_init(&shost->starved_list, &starved_list);
419
420         while (!list_empty(&starved_list)) {
421                 /*
422                  * As long as shost is accepting commands and we have
423                  * starved queues, call blk_run_queue. scsi_request_fn
424                  * drops the queue_lock and can add us back to the
425                  * starved_list.
426                  *
427                  * host_lock protects the starved_list and starved_entry.
428                  * scsi_request_fn must get the host_lock before checking
429                  * or modifying starved_list or starved_entry.
430                  */
431                 if (scsi_host_is_busy(shost))
432                         break;
433
434                 sdev = list_entry(starved_list.next,
435                                   struct scsi_device, starved_entry);
436                 list_del_init(&sdev->starved_entry);
437                 if (scsi_target_is_busy(scsi_target(sdev))) {
438                         list_move_tail(&sdev->starved_entry,
439                                        &shost->starved_list);
440                         continue;
441                 }
442
443                 spin_unlock(shost->host_lock);
444                 spin_lock(sdev->request_queue->queue_lock);
445                 __blk_run_queue(sdev->request_queue);
446                 spin_unlock(sdev->request_queue->queue_lock);
447                 spin_lock(shost->host_lock);
448         }
449         /* put any unprocessed entries back */
450         list_splice(&starved_list, &shost->starved_list);
451         spin_unlock_irqrestore(shost->host_lock, flags);
452
453         blk_run_queue(q);
454 }
455
456 void scsi_requeue_run_queue(struct work_struct *work)
457 {
458         struct scsi_device *sdev;
459         struct request_queue *q;
460
461         sdev = container_of(work, struct scsi_device, requeue_work);
462         q = sdev->request_queue;
463         scsi_run_queue(q);
464 }
465
466 /*
467  * Function:    scsi_requeue_command()
468  *
469  * Purpose:     Handle post-processing of completed commands.
470  *
471  * Arguments:   q       - queue to operate on
472  *              cmd     - command that may need to be requeued.
473  *
474  * Returns:     Nothing
475  *
476  * Notes:       After command completion, there may be blocks left
477  *              over which weren't finished by the previous command
478  *              this can be for a number of reasons - the main one is
479  *              I/O errors in the middle of the request, in which case
480  *              we need to request the blocks that come after the bad
481  *              sector.
482  * Notes:       Upon return, cmd is a stale pointer.
483  */
484 static void scsi_requeue_command(struct request_queue *q, struct scsi_cmnd *cmd)
485 {
486         struct request *req = cmd->request;
487         unsigned long flags;
488
489         spin_lock_irqsave(q->queue_lock, flags);
490         scsi_unprep_request(req);
491         blk_requeue_request(q, req);
492         spin_unlock_irqrestore(q->queue_lock, flags);
493
494         scsi_run_queue(q);
495 }
496
497 void scsi_next_command(struct scsi_cmnd *cmd)
498 {
499         struct scsi_device *sdev = cmd->device;
500         struct request_queue *q = sdev->request_queue;
501
502         /* need to hold a reference on the device before we let go of the cmd */
503         get_device(&sdev->sdev_gendev);
504
505         scsi_put_command(cmd);
506         scsi_run_queue(q);
507
508         /* ok to remove device now */
509         put_device(&sdev->sdev_gendev);
510 }
511
512 void scsi_run_host_queues(struct Scsi_Host *shost)
513 {
514         struct scsi_device *sdev;
515
516         shost_for_each_device(sdev, shost)
517                 scsi_run_queue(sdev->request_queue);
518 }
519
520 static void __scsi_release_buffers(struct scsi_cmnd *, int);
521
522 /*
523  * Function:    scsi_end_request()
524  *
525  * Purpose:     Post-processing of completed commands (usually invoked at end
526  *              of upper level post-processing and scsi_io_completion).
527  *
528  * Arguments:   cmd      - command that is complete.
529  *              error    - 0 if I/O indicates success, < 0 for I/O error.
530  *              bytes    - number of bytes of completed I/O
531  *              requeue  - indicates whether we should requeue leftovers.
532  *
533  * Lock status: Assumed that lock is not held upon entry.
534  *
535  * Returns:     cmd if requeue required, NULL otherwise.
536  *
537  * Notes:       This is called for block device requests in order to
538  *              mark some number of sectors as complete.
539  * 
540  *              We are guaranteeing that the request queue will be goosed
541  *              at some point during this call.
542  * Notes:       If cmd was requeued, upon return it will be a stale pointer.
543  */
544 static struct scsi_cmnd *scsi_end_request(struct scsi_cmnd *cmd, int error,
545                                           int bytes, int requeue)
546 {
547         struct request_queue *q = cmd->device->request_queue;
548         struct request *req = cmd->request;
549
550         /*
551          * If there are blocks left over at the end, set up the command
552          * to queue the remainder of them.
553          */
554         if (blk_end_request(req, error, bytes)) {
555                 /* kill remainder if no retrys */
556                 if (error && scsi_noretry_cmd(cmd))
557                         blk_end_request_all(req, error);
558                 else {
559                         if (requeue) {
560                                 /*
561                                  * Bleah.  Leftovers again.  Stick the
562                                  * leftovers in the front of the
563                                  * queue, and goose the queue again.
564                                  */
565                                 scsi_release_buffers(cmd);
566                                 scsi_requeue_command(q, cmd);
567                                 cmd = NULL;
568                         }
569                         return cmd;
570                 }
571         }
572
573         /*
574          * This will goose the queue request function at the end, so we don't
575          * need to worry about launching another command.
576          */
577         __scsi_release_buffers(cmd, 0);
578         scsi_next_command(cmd);
579         return NULL;
580 }
581
582 static inline unsigned int scsi_sgtable_index(unsigned short nents)
583 {
584         unsigned int index;
585
586         BUG_ON(nents > SCSI_MAX_SG_SEGMENTS);
587
588         if (nents <= 8)
589                 index = 0;
590         else
591                 index = get_count_order(nents) - 3;
592
593         return index;
594 }
595
596 static void scsi_sg_free(struct scatterlist *sgl, unsigned int nents)
597 {
598         struct scsi_host_sg_pool *sgp;
599
600         sgp = scsi_sg_pools + scsi_sgtable_index(nents);
601         mempool_free(sgl, sgp->pool);
602 }
603
604 static struct scatterlist *scsi_sg_alloc(unsigned int nents, gfp_t gfp_mask)
605 {
606         struct scsi_host_sg_pool *sgp;
607
608         sgp = scsi_sg_pools + scsi_sgtable_index(nents);
609         return mempool_alloc(sgp->pool, gfp_mask);
610 }
611
612 static int scsi_alloc_sgtable(struct scsi_data_buffer *sdb, int nents,
613                               gfp_t gfp_mask)
614 {
615         int ret;
616
617         BUG_ON(!nents);
618
619         ret = __sg_alloc_table(&sdb->table, nents, SCSI_MAX_SG_SEGMENTS,
620                                gfp_mask, scsi_sg_alloc);
621         if (unlikely(ret))
622                 __sg_free_table(&sdb->table, SCSI_MAX_SG_SEGMENTS,
623                                 scsi_sg_free);
624
625         return ret;
626 }
627
628 static void scsi_free_sgtable(struct scsi_data_buffer *sdb)
629 {
630         __sg_free_table(&sdb->table, SCSI_MAX_SG_SEGMENTS, scsi_sg_free);
631 }
632
633 static void __scsi_release_buffers(struct scsi_cmnd *cmd, int do_bidi_check)
634 {
635
636         if (cmd->sdb.table.nents)
637                 scsi_free_sgtable(&cmd->sdb);
638
639         memset(&cmd->sdb, 0, sizeof(cmd->sdb));
640
641         if (do_bidi_check && scsi_bidi_cmnd(cmd)) {
642                 struct scsi_data_buffer *bidi_sdb =
643                         cmd->request->next_rq->special;
644                 scsi_free_sgtable(bidi_sdb);
645                 kmem_cache_free(scsi_sdb_cache, bidi_sdb);
646                 cmd->request->next_rq->special = NULL;
647         }
648
649         if (scsi_prot_sg_count(cmd))
650                 scsi_free_sgtable(cmd->prot_sdb);
651 }
652
653 /*
654  * Function:    scsi_release_buffers()
655  *
656  * Purpose:     Completion processing for block device I/O requests.
657  *
658  * Arguments:   cmd     - command that we are bailing.
659  *
660  * Lock status: Assumed that no lock is held upon entry.
661  *
662  * Returns:     Nothing
663  *
664  * Notes:       In the event that an upper level driver rejects a
665  *              command, we must release resources allocated during
666  *              the __init_io() function.  Primarily this would involve
667  *              the scatter-gather table, and potentially any bounce
668  *              buffers.
669  */
670 void scsi_release_buffers(struct scsi_cmnd *cmd)
671 {
672         __scsi_release_buffers(cmd, 1);
673 }
674 EXPORT_SYMBOL(scsi_release_buffers);
675
676 static int __scsi_error_from_host_byte(struct scsi_cmnd *cmd, int result)
677 {
678         int error = 0;
679
680         switch(host_byte(result)) {
681         case DID_TRANSPORT_FAILFAST:
682                 error = -ENOLINK;
683                 break;
684         case DID_TARGET_FAILURE:
685                 set_host_byte(cmd, DID_OK);
686                 error = -EREMOTEIO;
687                 break;
688         case DID_NEXUS_FAILURE:
689                 set_host_byte(cmd, DID_OK);
690                 error = -EBADE;
691                 break;
692         default:
693                 error = -EIO;
694                 break;
695         }
696
697         return error;
698 }
699
700 /*
701  * Function:    scsi_io_completion()
702  *
703  * Purpose:     Completion processing for block device I/O requests.
704  *
705  * Arguments:   cmd   - command that is finished.
706  *
707  * Lock status: Assumed that no lock is held upon entry.
708  *
709  * Returns:     Nothing
710  *
711  * Notes:       This function is matched in terms of capabilities to
712  *              the function that created the scatter-gather list.
713  *              In other words, if there are no bounce buffers
714  *              (the normal case for most drivers), we don't need
715  *              the logic to deal with cleaning up afterwards.
716  *
717  *              We must call scsi_end_request().  This will finish off
718  *              the specified number of sectors.  If we are done, the
719  *              command block will be released and the queue function
720  *              will be goosed.  If we are not done then we have to
721  *              figure out what to do next:
722  *
723  *              a) We can call scsi_requeue_command().  The request
724  *                 will be unprepared and put back on the queue.  Then
725  *                 a new command will be created for it.  This should
726  *                 be used if we made forward progress, or if we want
727  *                 to switch from READ(10) to READ(6) for example.
728  *
729  *              b) We can call scsi_queue_insert().  The request will
730  *                 be put back on the queue and retried using the same
731  *                 command as before, possibly after a delay.
732  *
733  *              c) We can call blk_end_request() with -EIO to fail
734  *                 the remainder of the request.
735  */
736 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
737 {
738         int result = cmd->result;
739         struct request_queue *q = cmd->device->request_queue;
740         struct request *req = cmd->request;
741         int error = 0;
742         struct scsi_sense_hdr sshdr;
743         int sense_valid = 0;
744         int sense_deferred = 0;
745         enum {ACTION_FAIL, ACTION_REPREP, ACTION_RETRY,
746               ACTION_DELAYED_RETRY} action;
747         char *description = NULL;
748
749         if (result) {
750                 sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
751                 if (sense_valid)
752                         sense_deferred = scsi_sense_is_deferred(&sshdr);
753         }
754
755         if (req->cmd_type == REQ_TYPE_BLOCK_PC) { /* SG_IO ioctl from block level */
756                 req->errors = result;
757                 if (result) {
758                         if (sense_valid && req->sense) {
759                                 /*
760                                  * SG_IO wants current and deferred errors
761                                  */
762                                 int len = 8 + cmd->sense_buffer[7];
763
764                                 if (len > SCSI_SENSE_BUFFERSIZE)
765                                         len = SCSI_SENSE_BUFFERSIZE;
766                                 memcpy(req->sense, cmd->sense_buffer,  len);
767                                 req->sense_len = len;
768                         }
769                         if (!sense_deferred)
770                                 error = __scsi_error_from_host_byte(cmd, result);
771                 }
772
773                 req->resid_len = scsi_get_resid(cmd);
774
775                 if (scsi_bidi_cmnd(cmd)) {
776                         /*
777                          * Bidi commands Must be complete as a whole,
778                          * both sides at once.
779                          */
780                         req->next_rq->resid_len = scsi_in(cmd)->resid;
781
782                         scsi_release_buffers(cmd);
783                         blk_end_request_all(req, 0);
784
785                         scsi_next_command(cmd);
786                         return;
787                 }
788         }
789
790         /* no bidi support for !REQ_TYPE_BLOCK_PC yet */
791         BUG_ON(blk_bidi_rq(req));
792
793         /*
794          * Next deal with any sectors which we were able to correctly
795          * handle.
796          */
797         SCSI_LOG_HLCOMPLETE(1, printk("%u sectors total, "
798                                       "%d bytes done.\n",
799                                       blk_rq_sectors(req), good_bytes));
800
801         /*
802          * Recovered errors need reporting, but they're always treated
803          * as success, so fiddle the result code here.  For BLOCK_PC
804          * we already took a copy of the original into rq->errors which
805          * is what gets returned to the user
806          */
807         if (sense_valid && (sshdr.sense_key == RECOVERED_ERROR)) {
808                 /* if ATA PASS-THROUGH INFORMATION AVAILABLE skip
809                  * print since caller wants ATA registers. Only occurs on
810                  * SCSI ATA PASS_THROUGH commands when CK_COND=1
811                  */
812                 if ((sshdr.asc == 0x0) && (sshdr.ascq == 0x1d))
813                         ;
814                 else if (!(req->cmd_flags & REQ_QUIET))
815                         scsi_print_sense("", cmd);
816                 result = 0;
817                 /* BLOCK_PC may have set error */
818                 error = 0;
819         }
820
821         /*
822          * A number of bytes were successfully read.  If there
823          * are leftovers and there is some kind of error
824          * (result != 0), retry the rest.
825          */
826         if (scsi_end_request(cmd, error, good_bytes, result == 0) == NULL)
827                 return;
828
829         error = __scsi_error_from_host_byte(cmd, result);
830
831         if (host_byte(result) == DID_RESET) {
832                 /* Third party bus reset or reset for error recovery
833                  * reasons.  Just retry the command and see what
834                  * happens.
835                  */
836                 action = ACTION_RETRY;
837         } else if (sense_valid && !sense_deferred) {
838                 switch (sshdr.sense_key) {
839                 case UNIT_ATTENTION:
840                         if (cmd->device->removable) {
841                                 /* Detected disc change.  Set a bit
842                                  * and quietly refuse further access.
843                                  */
844                                 cmd->device->changed = 1;
845                                 description = "Media Changed";
846                                 action = ACTION_FAIL;
847                         } else {
848                                 /* Must have been a power glitch, or a
849                                  * bus reset.  Could not have been a
850                                  * media change, so we just retry the
851                                  * command and see what happens.
852                                  */
853                                 action = ACTION_RETRY;
854                         }
855                         break;
856                 case ILLEGAL_REQUEST:
857                         /* If we had an ILLEGAL REQUEST returned, then
858                          * we may have performed an unsupported
859                          * command.  The only thing this should be
860                          * would be a ten byte read where only a six
861                          * byte read was supported.  Also, on a system
862                          * where READ CAPACITY failed, we may have
863                          * read past the end of the disk.
864                          */
865                         if ((cmd->device->use_10_for_rw &&
866                             sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
867                             (cmd->cmnd[0] == READ_10 ||
868                              cmd->cmnd[0] == WRITE_10)) {
869                                 /* This will issue a new 6-byte command. */
870                                 cmd->device->use_10_for_rw = 0;
871                                 action = ACTION_REPREP;
872                         } else if (sshdr.asc == 0x10) /* DIX */ {
873                                 description = "Host Data Integrity Failure";
874                                 action = ACTION_FAIL;
875                                 error = -EILSEQ;
876                         /* INVALID COMMAND OPCODE or INVALID FIELD IN CDB */
877                         } else if ((sshdr.asc == 0x20 || sshdr.asc == 0x24) &&
878                                    (cmd->cmnd[0] == UNMAP ||
879                                     cmd->cmnd[0] == WRITE_SAME_16 ||
880                                     cmd->cmnd[0] == WRITE_SAME)) {
881                                 description = "Discard failure";
882                                 action = ACTION_FAIL;
883                                 error = -EREMOTEIO;
884                         } else
885                                 action = ACTION_FAIL;
886                         break;
887                 case ABORTED_COMMAND:
888                         action = ACTION_FAIL;
889                         if (sshdr.asc == 0x10) { /* DIF */
890                                 description = "Target Data Integrity Failure";
891                                 error = -EILSEQ;
892                         }
893                         break;
894                 case NOT_READY:
895                         /* If the device is in the process of becoming
896                          * ready, or has a temporary blockage, retry.
897                          */
898                         if (sshdr.asc == 0x04) {
899                                 switch (sshdr.ascq) {
900                                 case 0x01: /* becoming ready */
901                                 case 0x04: /* format in progress */
902                                 case 0x05: /* rebuild in progress */
903                                 case 0x06: /* recalculation in progress */
904                                 case 0x07: /* operation in progress */
905                                 case 0x08: /* Long write in progress */
906                                 case 0x09: /* self test in progress */
907                                 case 0x14: /* space allocation in progress */
908                                         action = ACTION_DELAYED_RETRY;
909                                         break;
910                                 default:
911                                         description = "Device not ready";
912                                         action = ACTION_FAIL;
913                                         break;
914                                 }
915                         } else {
916                                 description = "Device not ready";
917                                 action = ACTION_FAIL;
918                         }
919                         break;
920                 case VOLUME_OVERFLOW:
921                         /* See SSC3rXX or current. */
922                         action = ACTION_FAIL;
923                         break;
924                 default:
925                         description = "Unhandled sense code";
926                         action = ACTION_FAIL;
927                         break;
928                 }
929         } else {
930                 description = "Unhandled error code";
931                 action = ACTION_FAIL;
932         }
933
934         switch (action) {
935         case ACTION_FAIL:
936                 /* Give up and fail the remainder of the request */
937                 scsi_release_buffers(cmd);
938                 if (!(req->cmd_flags & REQ_QUIET)) {
939                         if (description)
940                                 scmd_printk(KERN_INFO, cmd, "%s\n",
941                                             description);
942                         scsi_print_result(cmd);
943                         if (driver_byte(result) & DRIVER_SENSE)
944                                 scsi_print_sense("", cmd);
945                         scsi_print_command(cmd);
946                 }
947                 if (blk_end_request_err(req, error))
948                         scsi_requeue_command(q, cmd);
949                 else
950                         scsi_next_command(cmd);
951                 break;
952         case ACTION_REPREP:
953                 /* Unprep the request and put it back at the head of the queue.
954                  * A new command will be prepared and issued.
955                  */
956                 scsi_release_buffers(cmd);
957                 scsi_requeue_command(q, cmd);
958                 break;
959         case ACTION_RETRY:
960                 /* Retry the same command immediately */
961                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY, 0);
962                 break;
963         case ACTION_DELAYED_RETRY:
964                 /* Retry the same command after a delay */
965                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY, 0);
966                 break;
967         }
968 }
969
970 static int scsi_init_sgtable(struct request *req, struct scsi_data_buffer *sdb,
971                              gfp_t gfp_mask)
972 {
973         int count;
974
975         /*
976          * If sg table allocation fails, requeue request later.
977          */
978         if (unlikely(scsi_alloc_sgtable(sdb, req->nr_phys_segments,
979                                         gfp_mask))) {
980                 return BLKPREP_DEFER;
981         }
982
983         req->buffer = NULL;
984
985         /* 
986          * Next, walk the list, and fill in the addresses and sizes of
987          * each segment.
988          */
989         count = blk_rq_map_sg(req->q, req, sdb->table.sgl);
990         BUG_ON(count > sdb->table.nents);
991         sdb->table.nents = count;
992         sdb->length = blk_rq_bytes(req);
993         return BLKPREP_OK;
994 }
995
996 /*
997  * Function:    scsi_init_io()
998  *
999  * Purpose:     SCSI I/O initialize function.
1000  *
1001  * Arguments:   cmd   - Command descriptor we wish to initialize
1002  *
1003  * Returns:     0 on success
1004  *              BLKPREP_DEFER if the failure is retryable
1005  *              BLKPREP_KILL if the failure is fatal
1006  */
1007 int scsi_init_io(struct scsi_cmnd *cmd, gfp_t gfp_mask)
1008 {
1009         struct request *rq = cmd->request;
1010
1011         int error = scsi_init_sgtable(rq, &cmd->sdb, gfp_mask);
1012         if (error)
1013                 goto err_exit;
1014
1015         if (blk_bidi_rq(rq)) {
1016                 struct scsi_data_buffer *bidi_sdb = kmem_cache_zalloc(
1017                         scsi_sdb_cache, GFP_ATOMIC);
1018                 if (!bidi_sdb) {
1019                         error = BLKPREP_DEFER;
1020                         goto err_exit;
1021                 }
1022
1023                 rq->next_rq->special = bidi_sdb;
1024                 error = scsi_init_sgtable(rq->next_rq, bidi_sdb, GFP_ATOMIC);
1025                 if (error)
1026                         goto err_exit;
1027         }
1028
1029         if (blk_integrity_rq(rq)) {
1030                 struct scsi_data_buffer *prot_sdb = cmd->prot_sdb;
1031                 int ivecs, count;
1032
1033                 BUG_ON(prot_sdb == NULL);
1034                 ivecs = blk_rq_count_integrity_sg(rq->q, rq->bio);
1035
1036                 if (scsi_alloc_sgtable(prot_sdb, ivecs, gfp_mask)) {
1037                         error = BLKPREP_DEFER;
1038                         goto err_exit;
1039                 }
1040
1041                 count = blk_rq_map_integrity_sg(rq->q, rq->bio,
1042                                                 prot_sdb->table.sgl);
1043                 BUG_ON(unlikely(count > ivecs));
1044                 BUG_ON(unlikely(count > queue_max_integrity_segments(rq->q)));
1045
1046                 cmd->prot_sdb = prot_sdb;
1047                 cmd->prot_sdb->table.nents = count;
1048         }
1049
1050         return BLKPREP_OK ;
1051
1052 err_exit:
1053         scsi_release_buffers(cmd);
1054         cmd->request->special = NULL;
1055         scsi_put_command(cmd);
1056         return error;
1057 }
1058 EXPORT_SYMBOL(scsi_init_io);
1059
1060 static struct scsi_cmnd *scsi_get_cmd_from_req(struct scsi_device *sdev,
1061                 struct request *req)
1062 {
1063         struct scsi_cmnd *cmd;
1064
1065         if (!req->special) {
1066                 cmd = scsi_get_command(sdev, GFP_ATOMIC);
1067                 if (unlikely(!cmd))
1068                         return NULL;
1069                 req->special = cmd;
1070         } else {
1071                 cmd = req->special;
1072         }
1073
1074         /* pull a tag out of the request if we have one */
1075         cmd->tag = req->tag;
1076         cmd->request = req;
1077
1078         cmd->cmnd = req->cmd;
1079         cmd->prot_op = SCSI_PROT_NORMAL;
1080
1081         return cmd;
1082 }
1083
1084 int scsi_setup_blk_pc_cmnd(struct scsi_device *sdev, struct request *req)
1085 {
1086         struct scsi_cmnd *cmd;
1087         int ret = scsi_prep_state_check(sdev, req);
1088
1089         if (ret != BLKPREP_OK)
1090                 return ret;
1091
1092         cmd = scsi_get_cmd_from_req(sdev, req);
1093         if (unlikely(!cmd))
1094                 return BLKPREP_DEFER;
1095
1096         /*
1097          * BLOCK_PC requests may transfer data, in which case they must
1098          * a bio attached to them.  Or they might contain a SCSI command
1099          * that does not transfer data, in which case they may optionally
1100          * submit a request without an attached bio.
1101          */
1102         if (req->bio) {
1103                 int ret;
1104
1105                 BUG_ON(!req->nr_phys_segments);
1106
1107                 ret = scsi_init_io(cmd, GFP_ATOMIC);
1108                 if (unlikely(ret))
1109                         return ret;
1110         } else {
1111                 BUG_ON(blk_rq_bytes(req));
1112
1113                 memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1114                 req->buffer = NULL;
1115         }
1116
1117         cmd->cmd_len = req->cmd_len;
1118         if (!blk_rq_bytes(req))
1119                 cmd->sc_data_direction = DMA_NONE;
1120         else if (rq_data_dir(req) == WRITE)
1121                 cmd->sc_data_direction = DMA_TO_DEVICE;
1122         else
1123                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1124         
1125         cmd->transfersize = blk_rq_bytes(req);
1126         cmd->allowed = req->retries;
1127         return BLKPREP_OK;
1128 }
1129 EXPORT_SYMBOL(scsi_setup_blk_pc_cmnd);
1130
1131 /*
1132  * Setup a REQ_TYPE_FS command.  These are simple read/write request
1133  * from filesystems that still need to be translated to SCSI CDBs from
1134  * the ULD.
1135  */
1136 int scsi_setup_fs_cmnd(struct scsi_device *sdev, struct request *req)
1137 {
1138         struct scsi_cmnd *cmd;
1139         int ret = scsi_prep_state_check(sdev, req);
1140
1141         if (ret != BLKPREP_OK)
1142                 return ret;
1143
1144         if (unlikely(sdev->scsi_dh_data && sdev->scsi_dh_data->scsi_dh
1145                          && sdev->scsi_dh_data->scsi_dh->prep_fn)) {
1146                 ret = sdev->scsi_dh_data->scsi_dh->prep_fn(sdev, req);
1147                 if (ret != BLKPREP_OK)
1148                         return ret;
1149         }
1150
1151         /*
1152          * Filesystem requests must transfer data.
1153          */
1154         BUG_ON(!req->nr_phys_segments);
1155
1156         cmd = scsi_get_cmd_from_req(sdev, req);
1157         if (unlikely(!cmd))
1158                 return BLKPREP_DEFER;
1159
1160         memset(cmd->cmnd, 0, BLK_MAX_CDB);
1161         return scsi_init_io(cmd, GFP_ATOMIC);
1162 }
1163 EXPORT_SYMBOL(scsi_setup_fs_cmnd);
1164
1165 int scsi_prep_state_check(struct scsi_device *sdev, struct request *req)
1166 {
1167         int ret = BLKPREP_OK;
1168
1169         /*
1170          * If the device is not in running state we will reject some
1171          * or all commands.
1172          */
1173         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1174                 switch (sdev->sdev_state) {
1175                 case SDEV_OFFLINE:
1176                         /*
1177                          * If the device is offline we refuse to process any
1178                          * commands.  The device must be brought online
1179                          * before trying any recovery commands.
1180                          */
1181                         sdev_printk(KERN_ERR, sdev,
1182                                     "rejecting I/O to offline device\n");
1183                         ret = BLKPREP_KILL;
1184                         break;
1185                 case SDEV_DEL:
1186                         /*
1187                          * If the device is fully deleted, we refuse to
1188                          * process any commands as well.
1189                          */
1190                         sdev_printk(KERN_ERR, sdev,
1191                                     "rejecting I/O to dead device\n");
1192                         ret = BLKPREP_KILL;
1193                         break;
1194                 case SDEV_QUIESCE:
1195                 case SDEV_BLOCK:
1196                 case SDEV_CREATED_BLOCK:
1197                         /*
1198                          * If the devices is blocked we defer normal commands.
1199                          */
1200                         if (!(req->cmd_flags & REQ_PREEMPT))
1201                                 ret = BLKPREP_DEFER;
1202                         break;
1203                 default:
1204                         /*
1205                          * For any other not fully online state we only allow
1206                          * special commands.  In particular any user initiated
1207                          * command is not allowed.
1208                          */
1209                         if (!(req->cmd_flags & REQ_PREEMPT))
1210                                 ret = BLKPREP_KILL;
1211                         break;
1212                 }
1213         }
1214         return ret;
1215 }
1216 EXPORT_SYMBOL(scsi_prep_state_check);
1217
1218 int scsi_prep_return(struct request_queue *q, struct request *req, int ret)
1219 {
1220         struct scsi_device *sdev = q->queuedata;
1221
1222         switch (ret) {
1223         case BLKPREP_KILL:
1224                 req->errors = DID_NO_CONNECT << 16;
1225                 /* release the command and kill it */
1226                 if (req->special) {
1227                         struct scsi_cmnd *cmd = req->special;
1228                         scsi_release_buffers(cmd);
1229                         scsi_put_command(cmd);
1230                         req->special = NULL;
1231                 }
1232                 break;
1233         case BLKPREP_DEFER:
1234                 /*
1235                  * If we defer, the blk_peek_request() returns NULL, but the
1236                  * queue must be restarted, so we schedule a callback to happen
1237                  * shortly.
1238                  */
1239                 if (sdev->device_busy == 0)
1240                         blk_delay_queue(q, SCSI_QUEUE_DELAY);
1241                 break;
1242         default:
1243                 req->cmd_flags |= REQ_DONTPREP;
1244         }
1245
1246         return ret;
1247 }
1248 EXPORT_SYMBOL(scsi_prep_return);
1249
1250 int scsi_prep_fn(struct request_queue *q, struct request *req)
1251 {
1252         struct scsi_device *sdev = q->queuedata;
1253         int ret = BLKPREP_KILL;
1254
1255         if (req->cmd_type == REQ_TYPE_BLOCK_PC)
1256                 ret = scsi_setup_blk_pc_cmnd(sdev, req);
1257         return scsi_prep_return(q, req, ret);
1258 }
1259 EXPORT_SYMBOL(scsi_prep_fn);
1260
1261 /*
1262  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1263  * return 0.
1264  *
1265  * Called with the queue_lock held.
1266  */
1267 static inline int scsi_dev_queue_ready(struct request_queue *q,
1268                                   struct scsi_device *sdev)
1269 {
1270         if (sdev->device_busy == 0 && sdev->device_blocked) {
1271                 /*
1272                  * unblock after device_blocked iterates to zero
1273                  */
1274                 if (--sdev->device_blocked == 0) {
1275                         SCSI_LOG_MLQUEUE(3,
1276                                    sdev_printk(KERN_INFO, sdev,
1277                                    "unblocking device at zero depth\n"));
1278                 } else {
1279                         blk_delay_queue(q, SCSI_QUEUE_DELAY);
1280                         return 0;
1281                 }
1282         }
1283         if (scsi_device_is_busy(sdev))
1284                 return 0;
1285
1286         return 1;
1287 }
1288
1289
1290 /*
1291  * scsi_target_queue_ready: checks if there we can send commands to target
1292  * @sdev: scsi device on starget to check.
1293  *
1294  * Called with the host lock held.
1295  */
1296 static inline int scsi_target_queue_ready(struct Scsi_Host *shost,
1297                                            struct scsi_device *sdev)
1298 {
1299         struct scsi_target *starget = scsi_target(sdev);
1300
1301         if (starget->single_lun) {
1302                 if (starget->starget_sdev_user &&
1303                     starget->starget_sdev_user != sdev)
1304                         return 0;
1305                 starget->starget_sdev_user = sdev;
1306         }
1307
1308         if (starget->target_busy == 0 && starget->target_blocked) {
1309                 /*
1310                  * unblock after target_blocked iterates to zero
1311                  */
1312                 if (--starget->target_blocked == 0) {
1313                         SCSI_LOG_MLQUEUE(3, starget_printk(KERN_INFO, starget,
1314                                          "unblocking target at zero depth\n"));
1315                 } else
1316                         return 0;
1317         }
1318
1319         if (scsi_target_is_busy(starget)) {
1320                 list_move_tail(&sdev->starved_entry, &shost->starved_list);
1321                 return 0;
1322         }
1323
1324         return 1;
1325 }
1326
1327 /*
1328  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1329  * return 0. We must end up running the queue again whenever 0 is
1330  * returned, else IO can hang.
1331  *
1332  * Called with host_lock held.
1333  */
1334 static inline int scsi_host_queue_ready(struct request_queue *q,
1335                                    struct Scsi_Host *shost,
1336                                    struct scsi_device *sdev)
1337 {
1338         if (scsi_host_in_recovery(shost))
1339                 return 0;
1340         if (shost->host_busy == 0 && shost->host_blocked) {
1341                 /*
1342                  * unblock after host_blocked iterates to zero
1343                  */
1344                 if (--shost->host_blocked == 0) {
1345                         SCSI_LOG_MLQUEUE(3,
1346                                 printk("scsi%d unblocking host at zero depth\n",
1347                                         shost->host_no));
1348                 } else {
1349                         return 0;
1350                 }
1351         }
1352         if (scsi_host_is_busy(shost)) {
1353                 if (list_empty(&sdev->starved_entry))
1354                         list_add_tail(&sdev->starved_entry, &shost->starved_list);
1355                 return 0;
1356         }
1357
1358         /* We're OK to process the command, so we can't be starved */
1359         if (!list_empty(&sdev->starved_entry))
1360                 list_del_init(&sdev->starved_entry);
1361
1362         return 1;
1363 }
1364
1365 /*
1366  * Busy state exporting function for request stacking drivers.
1367  *
1368  * For efficiency, no lock is taken to check the busy state of
1369  * shost/starget/sdev, since the returned value is not guaranteed and
1370  * may be changed after request stacking drivers call the function,
1371  * regardless of taking lock or not.
1372  *
1373  * When scsi can't dispatch I/Os anymore and needs to kill I/Os
1374  * (e.g. !sdev), scsi needs to return 'not busy'.
1375  * Otherwise, request stacking drivers may hold requests forever.
1376  */
1377 static int scsi_lld_busy(struct request_queue *q)
1378 {
1379         struct scsi_device *sdev = q->queuedata;
1380         struct Scsi_Host *shost;
1381         struct scsi_target *starget;
1382
1383         if (!sdev)
1384                 return 0;
1385
1386         shost = sdev->host;
1387         starget = scsi_target(sdev);
1388
1389         if (scsi_host_in_recovery(shost) || scsi_host_is_busy(shost) ||
1390             scsi_target_is_busy(starget) || scsi_device_is_busy(sdev))
1391                 return 1;
1392
1393         return 0;
1394 }
1395
1396 /*
1397  * Kill a request for a dead device
1398  */
1399 static void scsi_kill_request(struct request *req, struct request_queue *q)
1400 {
1401         struct scsi_cmnd *cmd = req->special;
1402         struct scsi_device *sdev;
1403         struct scsi_target *starget;
1404         struct Scsi_Host *shost;
1405
1406         blk_start_request(req);
1407
1408         scmd_printk(KERN_INFO, cmd, "killing request\n");
1409
1410         sdev = cmd->device;
1411         starget = scsi_target(sdev);
1412         shost = sdev->host;
1413         scsi_init_cmd_errh(cmd);
1414         cmd->result = DID_NO_CONNECT << 16;
1415         atomic_inc(&cmd->device->iorequest_cnt);
1416
1417         /*
1418          * SCSI request completion path will do scsi_device_unbusy(),
1419          * bump busy counts.  To bump the counters, we need to dance
1420          * with the locks as normal issue path does.
1421          */
1422         sdev->device_busy++;
1423         spin_unlock(sdev->request_queue->queue_lock);
1424         spin_lock(shost->host_lock);
1425         shost->host_busy++;
1426         starget->target_busy++;
1427         spin_unlock(shost->host_lock);
1428         spin_lock(sdev->request_queue->queue_lock);
1429
1430         blk_complete_request(req);
1431 }
1432
1433 static void scsi_softirq_done(struct request *rq)
1434 {
1435         struct scsi_cmnd *cmd = rq->special;
1436         unsigned long wait_for = (cmd->allowed + 1) * rq->timeout;
1437         int disposition;
1438
1439         INIT_LIST_HEAD(&cmd->eh_entry);
1440
1441         atomic_inc(&cmd->device->iodone_cnt);
1442         if (cmd->result)
1443                 atomic_inc(&cmd->device->ioerr_cnt);
1444
1445         disposition = scsi_decide_disposition(cmd);
1446         if (disposition != SUCCESS &&
1447             time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1448                 sdev_printk(KERN_ERR, cmd->device,
1449                             "timing out command, waited %lus\n",
1450                             wait_for/HZ);
1451                 disposition = SUCCESS;
1452         }
1453                         
1454         scsi_log_completion(cmd, disposition);
1455
1456         switch (disposition) {
1457                 case SUCCESS:
1458                         scsi_finish_command(cmd);
1459                         break;
1460                 case NEEDS_RETRY:
1461                         scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1462                         break;
1463                 case ADD_TO_MLQUEUE:
1464                         scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1465                         break;
1466                 default:
1467                         if (!scsi_eh_scmd_add(cmd, 0))
1468                                 scsi_finish_command(cmd);
1469         }
1470 }
1471
1472 /*
1473  * Function:    scsi_request_fn()
1474  *
1475  * Purpose:     Main strategy routine for SCSI.
1476  *
1477  * Arguments:   q       - Pointer to actual queue.
1478  *
1479  * Returns:     Nothing
1480  *
1481  * Lock status: IO request lock assumed to be held when called.
1482  */
1483 static void scsi_request_fn(struct request_queue *q)
1484 {
1485         struct scsi_device *sdev = q->queuedata;
1486         struct Scsi_Host *shost;
1487         struct scsi_cmnd *cmd;
1488         struct request *req;
1489
1490         if (!sdev) {
1491                 while ((req = blk_peek_request(q)) != NULL)
1492                         scsi_kill_request(req, q);
1493                 return;
1494         }
1495
1496         if(!get_device(&sdev->sdev_gendev))
1497                 /* We must be tearing the block queue down already */
1498                 return;
1499
1500         /*
1501          * To start with, we keep looping until the queue is empty, or until
1502          * the host is no longer able to accept any more requests.
1503          */
1504         shost = sdev->host;
1505         for (;;) {
1506                 int rtn;
1507                 /*
1508                  * get next queueable request.  We do this early to make sure
1509                  * that the request is fully prepared even if we cannot 
1510                  * accept it.
1511                  */
1512                 req = blk_peek_request(q);
1513                 if (!req || !scsi_dev_queue_ready(q, sdev))
1514                         break;
1515
1516                 if (unlikely(!scsi_device_online(sdev))) {
1517                         sdev_printk(KERN_ERR, sdev,
1518                                     "rejecting I/O to offline device\n");
1519                         scsi_kill_request(req, q);
1520                         continue;
1521                 }
1522
1523
1524                 /*
1525                  * Remove the request from the request list.
1526                  */
1527                 if (!(blk_queue_tagged(q) && !blk_queue_start_tag(q, req)))
1528                         blk_start_request(req);
1529                 sdev->device_busy++;
1530
1531                 spin_unlock(q->queue_lock);
1532                 cmd = req->special;
1533                 if (unlikely(cmd == NULL)) {
1534                         printk(KERN_CRIT "impossible request in %s.\n"
1535                                          "please mail a stack trace to "
1536                                          "linux-scsi@vger.kernel.org\n",
1537                                          __func__);
1538                         blk_dump_rq_flags(req, "foo");
1539                         BUG();
1540                 }
1541                 spin_lock(shost->host_lock);
1542
1543                 /*
1544                  * We hit this when the driver is using a host wide
1545                  * tag map. For device level tag maps the queue_depth check
1546                  * in the device ready fn would prevent us from trying
1547                  * to allocate a tag. Since the map is a shared host resource
1548                  * we add the dev to the starved list so it eventually gets
1549                  * a run when a tag is freed.
1550                  */
1551                 if (blk_queue_tagged(q) && !blk_rq_tagged(req)) {
1552                         if (list_empty(&sdev->starved_entry))
1553                                 list_add_tail(&sdev->starved_entry,
1554                                               &shost->starved_list);
1555                         goto not_ready;
1556                 }
1557
1558                 if (!scsi_target_queue_ready(shost, sdev))
1559                         goto not_ready;
1560
1561                 if (!scsi_host_queue_ready(q, shost, sdev))
1562                         goto not_ready;
1563
1564                 scsi_target(sdev)->target_busy++;
1565                 shost->host_busy++;
1566
1567                 /*
1568                  * XXX(hch): This is rather suboptimal, scsi_dispatch_cmd will
1569                  *              take the lock again.
1570                  */
1571                 spin_unlock_irq(shost->host_lock);
1572
1573                 /*
1574                  * Finally, initialize any error handling parameters, and set up
1575                  * the timers for timeouts.
1576                  */
1577                 scsi_init_cmd_errh(cmd);
1578
1579                 /*
1580                  * Dispatch the command to the low-level driver.
1581                  */
1582                 rtn = scsi_dispatch_cmd(cmd);
1583                 spin_lock_irq(q->queue_lock);
1584                 if (rtn)
1585                         goto out_delay;
1586         }
1587
1588         goto out;
1589
1590  not_ready:
1591         spin_unlock_irq(shost->host_lock);
1592
1593         /*
1594          * lock q, handle tag, requeue req, and decrement device_busy. We
1595          * must return with queue_lock held.
1596          *
1597          * Decrementing device_busy without checking it is OK, as all such
1598          * cases (host limits or settings) should run the queue at some
1599          * later time.
1600          */
1601         spin_lock_irq(q->queue_lock);
1602         blk_requeue_request(q, req);
1603         sdev->device_busy--;
1604 out_delay:
1605         if (sdev->device_busy == 0)
1606                 blk_delay_queue(q, SCSI_QUEUE_DELAY);
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->dma_dev;
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_segments(q, min_t(unsigned short, shost->sg_tablesize,
1651                                         SCSI_MAX_SG_CHAIN_SEGMENTS));
1652
1653         if (scsi_host_prot_dma(shost)) {
1654                 shost->sg_prot_tablesize =
1655                         min_not_zero(shost->sg_prot_tablesize,
1656                                      (unsigned short)SCSI_MAX_PROT_SG_SEGMENTS);
1657                 BUG_ON(shost->sg_prot_tablesize < shost->sg_tablesize);
1658                 blk_queue_max_integrity_segments(q, shost->sg_prot_tablesize);
1659         }
1660
1661         blk_queue_max_hw_sectors(q, shost->max_sectors);
1662         blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
1663         blk_queue_segment_boundary(q, shost->dma_boundary);
1664         dma_set_seg_boundary(dev, shost->dma_boundary);
1665
1666         blk_queue_max_segment_size(q, dma_get_max_seg_size(dev));
1667
1668         if (!shost->use_clustering)
1669                 q->limits.cluster = 0;
1670
1671         /*
1672          * set a reasonable default alignment on word boundaries: the
1673          * host and device may alter it using
1674          * blk_queue_update_dma_alignment() later.
1675          */
1676         blk_queue_dma_alignment(q, 0x03);
1677
1678         return q;
1679 }
1680 EXPORT_SYMBOL(__scsi_alloc_queue);
1681
1682 struct request_queue *scsi_alloc_queue(struct scsi_device *sdev)
1683 {
1684         struct request_queue *q;
1685
1686         q = __scsi_alloc_queue(sdev->host, scsi_request_fn);
1687         if (!q)
1688                 return NULL;
1689
1690         blk_queue_prep_rq(q, scsi_prep_fn);
1691         blk_queue_softirq_done(q, scsi_softirq_done);
1692         blk_queue_rq_timed_out(q, scsi_times_out);
1693         blk_queue_lld_busy(q, scsi_lld_busy);
1694         return q;
1695 }
1696
1697 void scsi_free_queue(struct request_queue *q)
1698 {
1699         unsigned long flags;
1700
1701         WARN_ON(q->queuedata);
1702
1703         /* cause scsi_request_fn() to kill all non-finished requests */
1704         spin_lock_irqsave(q->queue_lock, flags);
1705         q->request_fn(q);
1706         spin_unlock_irqrestore(q->queue_lock, flags);
1707
1708         blk_cleanup_queue(q);
1709 }
1710
1711 /*
1712  * Function:    scsi_block_requests()
1713  *
1714  * Purpose:     Utility function used by low-level drivers to prevent further
1715  *              commands from being queued to the device.
1716  *
1717  * Arguments:   shost       - Host in question
1718  *
1719  * Returns:     Nothing
1720  *
1721  * Lock status: No locks are assumed held.
1722  *
1723  * Notes:       There is no timer nor any other means by which the requests
1724  *              get unblocked other than the low-level driver calling
1725  *              scsi_unblock_requests().
1726  */
1727 void scsi_block_requests(struct Scsi_Host *shost)
1728 {
1729         shost->host_self_blocked = 1;
1730 }
1731 EXPORT_SYMBOL(scsi_block_requests);
1732
1733 /*
1734  * Function:    scsi_unblock_requests()
1735  *
1736  * Purpose:     Utility function used by low-level drivers to allow further
1737  *              commands from being queued to the device.
1738  *
1739  * Arguments:   shost       - Host in question
1740  *
1741  * Returns:     Nothing
1742  *
1743  * Lock status: No locks are assumed held.
1744  *
1745  * Notes:       There is no timer nor any other means by which the requests
1746  *              get unblocked other than the low-level driver calling
1747  *              scsi_unblock_requests().
1748  *
1749  *              This is done as an API function so that changes to the
1750  *              internals of the scsi mid-layer won't require wholesale
1751  *              changes to drivers that use this feature.
1752  */
1753 void scsi_unblock_requests(struct Scsi_Host *shost)
1754 {
1755         shost->host_self_blocked = 0;
1756         scsi_run_host_queues(shost);
1757 }
1758 EXPORT_SYMBOL(scsi_unblock_requests);
1759
1760 int __init scsi_init_queue(void)
1761 {
1762         int i;
1763
1764         scsi_sdb_cache = kmem_cache_create("scsi_data_buffer",
1765                                            sizeof(struct scsi_data_buffer),
1766                                            0, 0, NULL);
1767         if (!scsi_sdb_cache) {
1768                 printk(KERN_ERR "SCSI: can't init scsi sdb cache\n");
1769                 return -ENOMEM;
1770         }
1771
1772         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1773                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1774                 int size = sgp->size * sizeof(struct scatterlist);
1775
1776                 sgp->slab = kmem_cache_create(sgp->name, size, 0,
1777                                 SLAB_HWCACHE_ALIGN, NULL);
1778                 if (!sgp->slab) {
1779                         printk(KERN_ERR "SCSI: can't init sg slab %s\n",
1780                                         sgp->name);
1781                         goto cleanup_sdb;
1782                 }
1783
1784                 sgp->pool = mempool_create_slab_pool(SG_MEMPOOL_SIZE,
1785                                                      sgp->slab);
1786                 if (!sgp->pool) {
1787                         printk(KERN_ERR "SCSI: can't init sg mempool %s\n",
1788                                         sgp->name);
1789                         goto cleanup_sdb;
1790                 }
1791         }
1792
1793         return 0;
1794
1795 cleanup_sdb:
1796         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1797                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1798                 if (sgp->pool)
1799                         mempool_destroy(sgp->pool);
1800                 if (sgp->slab)
1801                         kmem_cache_destroy(sgp->slab);
1802         }
1803         kmem_cache_destroy(scsi_sdb_cache);
1804
1805         return -ENOMEM;
1806 }
1807
1808 void scsi_exit_queue(void)
1809 {
1810         int i;
1811
1812         kmem_cache_destroy(scsi_sdb_cache);
1813
1814         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1815                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1816                 mempool_destroy(sgp->pool);
1817                 kmem_cache_destroy(sgp->slab);
1818         }
1819 }
1820
1821 /**
1822  *      scsi_mode_select - issue a mode select
1823  *      @sdev:  SCSI device to be queried
1824  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
1825  *      @sp:    Save page bit (0 == don't save, 1 == save)
1826  *      @modepage: mode page being requested
1827  *      @buffer: request buffer (may not be smaller than eight bytes)
1828  *      @len:   length of request buffer.
1829  *      @timeout: command timeout
1830  *      @retries: number of retries before failing
1831  *      @data: returns a structure abstracting the mode header data
1832  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
1833  *              must be SCSI_SENSE_BUFFERSIZE big.
1834  *
1835  *      Returns zero if successful; negative error number or scsi
1836  *      status on error
1837  *
1838  */
1839 int
1840 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
1841                  unsigned char *buffer, int len, int timeout, int retries,
1842                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1843 {
1844         unsigned char cmd[10];
1845         unsigned char *real_buffer;
1846         int ret;
1847
1848         memset(cmd, 0, sizeof(cmd));
1849         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
1850
1851         if (sdev->use_10_for_ms) {
1852                 if (len > 65535)
1853                         return -EINVAL;
1854                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
1855                 if (!real_buffer)
1856                         return -ENOMEM;
1857                 memcpy(real_buffer + 8, buffer, len);
1858                 len += 8;
1859                 real_buffer[0] = 0;
1860                 real_buffer[1] = 0;
1861                 real_buffer[2] = data->medium_type;
1862                 real_buffer[3] = data->device_specific;
1863                 real_buffer[4] = data->longlba ? 0x01 : 0;
1864                 real_buffer[5] = 0;
1865                 real_buffer[6] = data->block_descriptor_length >> 8;
1866                 real_buffer[7] = data->block_descriptor_length;
1867
1868                 cmd[0] = MODE_SELECT_10;
1869                 cmd[7] = len >> 8;
1870                 cmd[8] = len;
1871         } else {
1872                 if (len > 255 || data->block_descriptor_length > 255 ||
1873                     data->longlba)
1874                         return -EINVAL;
1875
1876                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
1877                 if (!real_buffer)
1878                         return -ENOMEM;
1879                 memcpy(real_buffer + 4, buffer, len);
1880                 len += 4;
1881                 real_buffer[0] = 0;
1882                 real_buffer[1] = data->medium_type;
1883                 real_buffer[2] = data->device_specific;
1884                 real_buffer[3] = data->block_descriptor_length;
1885                 
1886
1887                 cmd[0] = MODE_SELECT;
1888                 cmd[4] = len;
1889         }
1890
1891         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
1892                                sshdr, timeout, retries, NULL);
1893         kfree(real_buffer);
1894         return ret;
1895 }
1896 EXPORT_SYMBOL_GPL(scsi_mode_select);
1897
1898 /**
1899  *      scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
1900  *      @sdev:  SCSI device to be queried
1901  *      @dbd:   set if mode sense will allow block descriptors to be returned
1902  *      @modepage: mode page being requested
1903  *      @buffer: request buffer (may not be smaller than eight bytes)
1904  *      @len:   length of request buffer.
1905  *      @timeout: command timeout
1906  *      @retries: number of retries before failing
1907  *      @data: returns a structure abstracting the mode header data
1908  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
1909  *              must be SCSI_SENSE_BUFFERSIZE big.
1910  *
1911  *      Returns zero if unsuccessful, or the header offset (either 4
1912  *      or 8 depending on whether a six or ten byte command was
1913  *      issued) if successful.
1914  */
1915 int
1916 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
1917                   unsigned char *buffer, int len, int timeout, int retries,
1918                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1919 {
1920         unsigned char cmd[12];
1921         int use_10_for_ms;
1922         int header_length;
1923         int result;
1924         struct scsi_sense_hdr my_sshdr;
1925
1926         memset(data, 0, sizeof(*data));
1927         memset(&cmd[0], 0, 12);
1928         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
1929         cmd[2] = modepage;
1930
1931         /* caller might not be interested in sense, but we need it */
1932         if (!sshdr)
1933                 sshdr = &my_sshdr;
1934
1935  retry:
1936         use_10_for_ms = sdev->use_10_for_ms;
1937
1938         if (use_10_for_ms) {
1939                 if (len < 8)
1940                         len = 8;
1941
1942                 cmd[0] = MODE_SENSE_10;
1943                 cmd[8] = len;
1944                 header_length = 8;
1945         } else {
1946                 if (len < 4)
1947                         len = 4;
1948
1949                 cmd[0] = MODE_SENSE;
1950                 cmd[4] = len;
1951                 header_length = 4;
1952         }
1953
1954         memset(buffer, 0, len);
1955
1956         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
1957                                   sshdr, timeout, retries, NULL);
1958
1959         /* This code looks awful: what it's doing is making sure an
1960          * ILLEGAL REQUEST sense return identifies the actual command
1961          * byte as the problem.  MODE_SENSE commands can return
1962          * ILLEGAL REQUEST if the code page isn't supported */
1963
1964         if (use_10_for_ms && !scsi_status_is_good(result) &&
1965             (driver_byte(result) & DRIVER_SENSE)) {
1966                 if (scsi_sense_valid(sshdr)) {
1967                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
1968                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
1969                                 /* 
1970                                  * Invalid command operation code
1971                                  */
1972                                 sdev->use_10_for_ms = 0;
1973                                 goto retry;
1974                         }
1975                 }
1976         }
1977
1978         if(scsi_status_is_good(result)) {
1979                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
1980                              (modepage == 6 || modepage == 8))) {
1981                         /* Initio breakage? */
1982                         header_length = 0;
1983                         data->length = 13;
1984                         data->medium_type = 0;
1985                         data->device_specific = 0;
1986                         data->longlba = 0;
1987                         data->block_descriptor_length = 0;
1988                 } else if(use_10_for_ms) {
1989                         data->length = buffer[0]*256 + buffer[1] + 2;
1990                         data->medium_type = buffer[2];
1991                         data->device_specific = buffer[3];
1992                         data->longlba = buffer[4] & 0x01;
1993                         data->block_descriptor_length = buffer[6]*256
1994                                 + buffer[7];
1995                 } else {
1996                         data->length = buffer[0] + 1;
1997                         data->medium_type = buffer[1];
1998                         data->device_specific = buffer[2];
1999                         data->block_descriptor_length = buffer[3];
2000                 }
2001                 data->header_length = header_length;
2002         }
2003
2004         return result;
2005 }
2006 EXPORT_SYMBOL(scsi_mode_sense);
2007
2008 /**
2009  *      scsi_test_unit_ready - test if unit is ready
2010  *      @sdev:  scsi device to change the state of.
2011  *      @timeout: command timeout
2012  *      @retries: number of retries before failing
2013  *      @sshdr_external: Optional pointer to struct scsi_sense_hdr for
2014  *              returning sense. Make sure that this is cleared before passing
2015  *              in.
2016  *
2017  *      Returns zero if unsuccessful or an error if TUR failed.  For
2018  *      removable media, UNIT_ATTENTION sets ->changed flag.
2019  **/
2020 int
2021 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
2022                      struct scsi_sense_hdr *sshdr_external)
2023 {
2024         char cmd[] = {
2025                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
2026         };
2027         struct scsi_sense_hdr *sshdr;
2028         int result;
2029
2030         if (!sshdr_external)
2031                 sshdr = kzalloc(sizeof(*sshdr), GFP_KERNEL);
2032         else
2033                 sshdr = sshdr_external;
2034
2035         /* try to eat the UNIT_ATTENTION if there are enough retries */
2036         do {
2037                 result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
2038                                           timeout, retries, NULL);
2039                 if (sdev->removable && scsi_sense_valid(sshdr) &&
2040                     sshdr->sense_key == UNIT_ATTENTION)
2041                         sdev->changed = 1;
2042         } while (scsi_sense_valid(sshdr) &&
2043                  sshdr->sense_key == UNIT_ATTENTION && --retries);
2044
2045         if (!sshdr_external)
2046                 kfree(sshdr);
2047         return result;
2048 }
2049 EXPORT_SYMBOL(scsi_test_unit_ready);
2050
2051 /**
2052  *      scsi_device_set_state - Take the given device through the device state model.
2053  *      @sdev:  scsi device to change the state of.
2054  *      @state: state to change to.
2055  *
2056  *      Returns zero if unsuccessful or an error if the requested 
2057  *      transition is illegal.
2058  */
2059 int
2060 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2061 {
2062         enum scsi_device_state oldstate = sdev->sdev_state;
2063
2064         if (state == oldstate)
2065                 return 0;
2066
2067         switch (state) {
2068         case SDEV_CREATED:
2069                 switch (oldstate) {
2070                 case SDEV_CREATED_BLOCK:
2071                         break;
2072                 default:
2073                         goto illegal;
2074                 }
2075                 break;
2076                         
2077         case SDEV_RUNNING:
2078                 switch (oldstate) {
2079                 case SDEV_CREATED:
2080                 case SDEV_OFFLINE:
2081                 case SDEV_QUIESCE:
2082                 case SDEV_BLOCK:
2083                         break;
2084                 default:
2085                         goto illegal;
2086                 }
2087                 break;
2088
2089         case SDEV_QUIESCE:
2090                 switch (oldstate) {
2091                 case SDEV_RUNNING:
2092                 case SDEV_OFFLINE:
2093                         break;
2094                 default:
2095                         goto illegal;
2096                 }
2097                 break;
2098
2099         case SDEV_OFFLINE:
2100                 switch (oldstate) {
2101                 case SDEV_CREATED:
2102                 case SDEV_RUNNING:
2103                 case SDEV_QUIESCE:
2104                 case SDEV_BLOCK:
2105                         break;
2106                 default:
2107                         goto illegal;
2108                 }
2109                 break;
2110
2111         case SDEV_BLOCK:
2112                 switch (oldstate) {
2113                 case SDEV_RUNNING:
2114                 case SDEV_CREATED_BLOCK:
2115                         break;
2116                 default:
2117                         goto illegal;
2118                 }
2119                 break;
2120
2121         case SDEV_CREATED_BLOCK:
2122                 switch (oldstate) {
2123                 case SDEV_CREATED:
2124                         break;
2125                 default:
2126                         goto illegal;
2127                 }
2128                 break;
2129
2130         case SDEV_CANCEL:
2131                 switch (oldstate) {
2132                 case SDEV_CREATED:
2133                 case SDEV_RUNNING:
2134                 case SDEV_QUIESCE:
2135                 case SDEV_OFFLINE:
2136                 case SDEV_BLOCK:
2137                         break;
2138                 default:
2139                         goto illegal;
2140                 }
2141                 break;
2142
2143         case SDEV_DEL:
2144                 switch (oldstate) {
2145                 case SDEV_CREATED:
2146                 case SDEV_RUNNING:
2147                 case SDEV_OFFLINE:
2148                 case SDEV_CANCEL:
2149                         break;
2150                 default:
2151                         goto illegal;
2152                 }
2153                 break;
2154
2155         }
2156         sdev->sdev_state = state;
2157         return 0;
2158
2159  illegal:
2160         SCSI_LOG_ERROR_RECOVERY(1, 
2161                                 sdev_printk(KERN_ERR, sdev,
2162                                             "Illegal state transition %s->%s\n",
2163                                             scsi_device_state_name(oldstate),
2164                                             scsi_device_state_name(state))
2165                                 );
2166         return -EINVAL;
2167 }
2168 EXPORT_SYMBOL(scsi_device_set_state);
2169
2170 /**
2171  *      sdev_evt_emit - emit a single SCSI device uevent
2172  *      @sdev: associated SCSI device
2173  *      @evt: event to emit
2174  *
2175  *      Send a single uevent (scsi_event) to the associated scsi_device.
2176  */
2177 static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2178 {
2179         int idx = 0;
2180         char *envp[3];
2181
2182         switch (evt->evt_type) {
2183         case SDEV_EVT_MEDIA_CHANGE:
2184                 envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2185                 break;
2186
2187         default:
2188                 /* do nothing */
2189                 break;
2190         }
2191
2192         envp[idx++] = NULL;
2193
2194         kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2195 }
2196
2197 /**
2198  *      sdev_evt_thread - send a uevent for each scsi event
2199  *      @work: work struct for scsi_device
2200  *
2201  *      Dispatch queued events to their associated scsi_device kobjects
2202  *      as uevents.
2203  */
2204 void scsi_evt_thread(struct work_struct *work)
2205 {
2206         struct scsi_device *sdev;
2207         LIST_HEAD(event_list);
2208
2209         sdev = container_of(work, struct scsi_device, event_work);
2210
2211         while (1) {
2212                 struct scsi_event *evt;
2213                 struct list_head *this, *tmp;
2214                 unsigned long flags;
2215
2216                 spin_lock_irqsave(&sdev->list_lock, flags);
2217                 list_splice_init(&sdev->event_list, &event_list);
2218                 spin_unlock_irqrestore(&sdev->list_lock, flags);
2219
2220                 if (list_empty(&event_list))
2221                         break;
2222
2223                 list_for_each_safe(this, tmp, &event_list) {
2224                         evt = list_entry(this, struct scsi_event, node);
2225                         list_del(&evt->node);
2226                         scsi_evt_emit(sdev, evt);
2227                         kfree(evt);
2228                 }
2229         }
2230 }
2231
2232 /**
2233  *      sdev_evt_send - send asserted event to uevent thread
2234  *      @sdev: scsi_device event occurred on
2235  *      @evt: event to send
2236  *
2237  *      Assert scsi device event asynchronously.
2238  */
2239 void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2240 {
2241         unsigned long flags;
2242
2243 #if 0
2244         /* FIXME: currently this check eliminates all media change events
2245          * for polled devices.  Need to update to discriminate between AN
2246          * and polled events */
2247         if (!test_bit(evt->evt_type, sdev->supported_events)) {
2248                 kfree(evt);
2249                 return;
2250         }
2251 #endif
2252
2253         spin_lock_irqsave(&sdev->list_lock, flags);
2254         list_add_tail(&evt->node, &sdev->event_list);
2255         schedule_work(&sdev->event_work);
2256         spin_unlock_irqrestore(&sdev->list_lock, flags);
2257 }
2258 EXPORT_SYMBOL_GPL(sdev_evt_send);
2259
2260 /**
2261  *      sdev_evt_alloc - allocate a new scsi event
2262  *      @evt_type: type of event to allocate
2263  *      @gfpflags: GFP flags for allocation
2264  *
2265  *      Allocates and returns a new scsi_event.
2266  */
2267 struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2268                                   gfp_t gfpflags)
2269 {
2270         struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2271         if (!evt)
2272                 return NULL;
2273
2274         evt->evt_type = evt_type;
2275         INIT_LIST_HEAD(&evt->node);
2276
2277         /* evt_type-specific initialization, if any */
2278         switch (evt_type) {
2279         case SDEV_EVT_MEDIA_CHANGE:
2280         default:
2281                 /* do nothing */
2282                 break;
2283         }
2284
2285         return evt;
2286 }
2287 EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2288
2289 /**
2290  *      sdev_evt_send_simple - send asserted event to uevent thread
2291  *      @sdev: scsi_device event occurred on
2292  *      @evt_type: type of event to send
2293  *      @gfpflags: GFP flags for allocation
2294  *
2295  *      Assert scsi device event asynchronously, given an event type.
2296  */
2297 void sdev_evt_send_simple(struct scsi_device *sdev,
2298                           enum scsi_device_event evt_type, gfp_t gfpflags)
2299 {
2300         struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2301         if (!evt) {
2302                 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2303                             evt_type);
2304                 return;
2305         }
2306
2307         sdev_evt_send(sdev, evt);
2308 }
2309 EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2310
2311 /**
2312  *      scsi_device_quiesce - Block user issued commands.
2313  *      @sdev:  scsi device to quiesce.
2314  *
2315  *      This works by trying to transition to the SDEV_QUIESCE state
2316  *      (which must be a legal transition).  When the device is in this
2317  *      state, only special requests will be accepted, all others will
2318  *      be deferred.  Since special requests may also be requeued requests,
2319  *      a successful return doesn't guarantee the device will be 
2320  *      totally quiescent.
2321  *
2322  *      Must be called with user context, may sleep.
2323  *
2324  *      Returns zero if unsuccessful or an error if not.
2325  */
2326 int
2327 scsi_device_quiesce(struct scsi_device *sdev)
2328 {
2329         int err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2330         if (err)
2331                 return err;
2332
2333         scsi_run_queue(sdev->request_queue);
2334         while (sdev->device_busy) {
2335                 msleep_interruptible(200);
2336                 scsi_run_queue(sdev->request_queue);
2337         }
2338         return 0;
2339 }
2340 EXPORT_SYMBOL(scsi_device_quiesce);
2341
2342 /**
2343  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2344  *      @sdev:  scsi device to resume.
2345  *
2346  *      Moves the device from quiesced back to running and restarts the
2347  *      queues.
2348  *
2349  *      Must be called with user context, may sleep.
2350  */
2351 void scsi_device_resume(struct scsi_device *sdev)
2352 {
2353         /* check if the device state was mutated prior to resume, and if
2354          * so assume the state is being managed elsewhere (for example
2355          * device deleted during suspend)
2356          */
2357         if (sdev->sdev_state != SDEV_QUIESCE ||
2358             scsi_device_set_state(sdev, SDEV_RUNNING))
2359                 return;
2360         scsi_run_queue(sdev->request_queue);
2361 }
2362 EXPORT_SYMBOL(scsi_device_resume);
2363
2364 static void
2365 device_quiesce_fn(struct scsi_device *sdev, void *data)
2366 {
2367         scsi_device_quiesce(sdev);
2368 }
2369
2370 void
2371 scsi_target_quiesce(struct scsi_target *starget)
2372 {
2373         starget_for_each_device(starget, NULL, device_quiesce_fn);
2374 }
2375 EXPORT_SYMBOL(scsi_target_quiesce);
2376
2377 static void
2378 device_resume_fn(struct scsi_device *sdev, void *data)
2379 {
2380         scsi_device_resume(sdev);
2381 }
2382
2383 void
2384 scsi_target_resume(struct scsi_target *starget)
2385 {
2386         starget_for_each_device(starget, NULL, device_resume_fn);
2387 }
2388 EXPORT_SYMBOL(scsi_target_resume);
2389
2390 /**
2391  * scsi_internal_device_block - internal function to put a device temporarily into the SDEV_BLOCK state
2392  * @sdev:       device to block
2393  *
2394  * Block request made by scsi lld's to temporarily stop all
2395  * scsi commands on the specified device.  Called from interrupt
2396  * or normal process context.
2397  *
2398  * Returns zero if successful or error if not
2399  *
2400  * Notes:       
2401  *      This routine transitions the device to the SDEV_BLOCK state
2402  *      (which must be a legal transition).  When the device is in this
2403  *      state, all commands are deferred until the scsi lld reenables
2404  *      the device with scsi_device_unblock or device_block_tmo fires.
2405  *      This routine assumes the host_lock is held on entry.
2406  */
2407 int
2408 scsi_internal_device_block(struct scsi_device *sdev)
2409 {
2410         struct request_queue *q = sdev->request_queue;
2411         unsigned long flags;
2412         int err = 0;
2413
2414         err = scsi_device_set_state(sdev, SDEV_BLOCK);
2415         if (err) {
2416                 err = scsi_device_set_state(sdev, SDEV_CREATED_BLOCK);
2417
2418                 if (err)
2419                         return err;
2420         }
2421
2422         /* 
2423          * The device has transitioned to SDEV_BLOCK.  Stop the
2424          * block layer from calling the midlayer with this device's
2425          * request queue. 
2426          */
2427         spin_lock_irqsave(q->queue_lock, flags);
2428         blk_stop_queue(q);
2429         spin_unlock_irqrestore(q->queue_lock, flags);
2430
2431         return 0;
2432 }
2433 EXPORT_SYMBOL_GPL(scsi_internal_device_block);
2434  
2435 /**
2436  * scsi_internal_device_unblock - resume a device after a block request
2437  * @sdev:       device to resume
2438  *
2439  * Called by scsi lld's or the midlayer to restart the device queue
2440  * for the previously suspended scsi device.  Called from interrupt or
2441  * normal process context.
2442  *
2443  * Returns zero if successful or error if not.
2444  *
2445  * Notes:       
2446  *      This routine transitions the device to the SDEV_RUNNING state
2447  *      (which must be a legal transition) allowing the midlayer to
2448  *      goose the queue for this device.  This routine assumes the 
2449  *      host_lock is held upon entry.
2450  */
2451 int
2452 scsi_internal_device_unblock(struct scsi_device *sdev)
2453 {
2454         struct request_queue *q = sdev->request_queue; 
2455         unsigned long flags;
2456         
2457         /* 
2458          * Try to transition the scsi device to SDEV_RUNNING
2459          * and goose the device queue if successful.  
2460          */
2461         if (sdev->sdev_state == SDEV_BLOCK)
2462                 sdev->sdev_state = SDEV_RUNNING;
2463         else if (sdev->sdev_state == SDEV_CREATED_BLOCK)
2464                 sdev->sdev_state = SDEV_CREATED;
2465         else if (sdev->sdev_state != SDEV_CANCEL &&
2466                  sdev->sdev_state != SDEV_OFFLINE)
2467                 return -EINVAL;
2468
2469         spin_lock_irqsave(q->queue_lock, flags);
2470         blk_start_queue(q);
2471         spin_unlock_irqrestore(q->queue_lock, flags);
2472
2473         return 0;
2474 }
2475 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock);
2476
2477 static void
2478 device_block(struct scsi_device *sdev, void *data)
2479 {
2480         scsi_internal_device_block(sdev);
2481 }
2482
2483 static int
2484 target_block(struct device *dev, void *data)
2485 {
2486         if (scsi_is_target_device(dev))
2487                 starget_for_each_device(to_scsi_target(dev), NULL,
2488                                         device_block);
2489         return 0;
2490 }
2491
2492 void
2493 scsi_target_block(struct device *dev)
2494 {
2495         if (scsi_is_target_device(dev))
2496                 starget_for_each_device(to_scsi_target(dev), NULL,
2497                                         device_block);
2498         else
2499                 device_for_each_child(dev, NULL, target_block);
2500 }
2501 EXPORT_SYMBOL_GPL(scsi_target_block);
2502
2503 static void
2504 device_unblock(struct scsi_device *sdev, void *data)
2505 {
2506         scsi_internal_device_unblock(sdev);
2507 }
2508
2509 static int
2510 target_unblock(struct device *dev, void *data)
2511 {
2512         if (scsi_is_target_device(dev))
2513                 starget_for_each_device(to_scsi_target(dev), NULL,
2514                                         device_unblock);
2515         return 0;
2516 }
2517
2518 void
2519 scsi_target_unblock(struct device *dev)
2520 {
2521         if (scsi_is_target_device(dev))
2522                 starget_for_each_device(to_scsi_target(dev), NULL,
2523                                         device_unblock);
2524         else
2525                 device_for_each_child(dev, NULL, target_unblock);
2526 }
2527 EXPORT_SYMBOL_GPL(scsi_target_unblock);
2528
2529 /**
2530  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2531  * @sgl:        scatter-gather list
2532  * @sg_count:   number of segments in sg
2533  * @offset:     offset in bytes into sg, on return offset into the mapped area
2534  * @len:        bytes to map, on return number of bytes mapped
2535  *
2536  * Returns virtual address of the start of the mapped page
2537  */
2538 void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
2539                           size_t *offset, size_t *len)
2540 {
2541         int i;
2542         size_t sg_len = 0, len_complete = 0;
2543         struct scatterlist *sg;
2544         struct page *page;
2545
2546         WARN_ON(!irqs_disabled());
2547
2548         for_each_sg(sgl, sg, sg_count, i) {
2549                 len_complete = sg_len; /* Complete sg-entries */
2550                 sg_len += sg->length;
2551                 if (sg_len > *offset)
2552                         break;
2553         }
2554
2555         if (unlikely(i == sg_count)) {
2556                 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2557                         "elements %d\n",
2558                        __func__, sg_len, *offset, sg_count);
2559                 WARN_ON(1);
2560                 return NULL;
2561         }
2562
2563         /* Offset starting from the beginning of first page in this sg-entry */
2564         *offset = *offset - len_complete + sg->offset;
2565
2566         /* Assumption: contiguous pages can be accessed as "page + i" */
2567         page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
2568         *offset &= ~PAGE_MASK;
2569
2570         /* Bytes in this sg-entry from *offset to the end of the page */
2571         sg_len = PAGE_SIZE - *offset;
2572         if (*len > sg_len)
2573                 *len = sg_len;
2574
2575         return kmap_atomic(page);
2576 }
2577 EXPORT_SYMBOL(scsi_kmap_atomic_sg);
2578
2579 /**
2580  * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
2581  * @virt:       virtual address to be unmapped
2582  */
2583 void scsi_kunmap_atomic_sg(void *virt)
2584 {
2585         kunmap_atomic(virt);
2586 }
2587 EXPORT_SYMBOL(scsi_kunmap_atomic_sg);