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