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