[SCSI] aacraid: rework packet support code
[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/blkdev.h>
12 #include <linux/completion.h>
13 #include <linux/kernel.h>
14 #include <linux/mempool.h>
15 #include <linux/slab.h>
16 #include <linux/init.h>
17 #include <linux/pci.h>
18 #include <linux/delay.h>
19 #include <linux/hardirq.h>
20
21 #include <scsi/scsi.h>
22 #include <scsi/scsi_cmnd.h>
23 #include <scsi/scsi_dbg.h>
24 #include <scsi/scsi_device.h>
25 #include <scsi/scsi_driver.h>
26 #include <scsi/scsi_eh.h>
27 #include <scsi/scsi_host.h>
28
29 #include "scsi_priv.h"
30 #include "scsi_logging.h"
31
32
33 #define SG_MEMPOOL_NR           ARRAY_SIZE(scsi_sg_pools)
34 #define SG_MEMPOOL_SIZE         32
35
36 struct scsi_host_sg_pool {
37         size_t          size;
38         char            *name; 
39         struct kmem_cache       *slab;
40         mempool_t       *pool;
41 };
42
43 #if (SCSI_MAX_PHYS_SEGMENTS < 32)
44 #error SCSI_MAX_PHYS_SEGMENTS is too small
45 #endif
46
47 #define SP(x) { x, "sgpool-" #x } 
48 static struct scsi_host_sg_pool scsi_sg_pools[] = {
49         SP(8),
50         SP(16),
51         SP(32),
52 #if (SCSI_MAX_PHYS_SEGMENTS > 32)
53         SP(64),
54 #if (SCSI_MAX_PHYS_SEGMENTS > 64)
55         SP(128),
56 #if (SCSI_MAX_PHYS_SEGMENTS > 128)
57         SP(256),
58 #if (SCSI_MAX_PHYS_SEGMENTS > 256)
59 #error SCSI_MAX_PHYS_SEGMENTS is too large
60 #endif
61 #endif
62 #endif
63 #endif
64 };      
65 #undef SP
66
67 static void scsi_run_queue(struct request_queue *q);
68
69 /*
70  * Function:    scsi_unprep_request()
71  *
72  * Purpose:     Remove all preparation done for a request, including its
73  *              associated scsi_cmnd, so that it can be requeued.
74  *
75  * Arguments:   req     - request to unprepare
76  *
77  * Lock status: Assumed that no locks are held upon entry.
78  *
79  * Returns:     Nothing.
80  */
81 static void scsi_unprep_request(struct request *req)
82 {
83         struct scsi_cmnd *cmd = req->special;
84
85         req->cmd_flags &= ~REQ_DONTPREP;
86         req->special = NULL;
87
88         scsi_put_command(cmd);
89 }
90
91 /*
92  * Function:    scsi_queue_insert()
93  *
94  * Purpose:     Insert a command in the midlevel queue.
95  *
96  * Arguments:   cmd    - command that we are adding to queue.
97  *              reason - why we are inserting command to queue.
98  *
99  * Lock status: Assumed that lock is not held upon entry.
100  *
101  * Returns:     Nothing.
102  *
103  * Notes:       We do this for one of two cases.  Either the host is busy
104  *              and it cannot accept any more commands for the time being,
105  *              or the device returned QUEUE_FULL and can accept no more
106  *              commands.
107  * Notes:       This could be called either from an interrupt context or a
108  *              normal process context.
109  */
110 int scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
111 {
112         struct Scsi_Host *host = cmd->device->host;
113         struct scsi_device *device = cmd->device;
114         struct request_queue *q = device->request_queue;
115         unsigned long flags;
116
117         SCSI_LOG_MLQUEUE(1,
118                  printk("Inserting command %p into mlqueue\n", cmd));
119
120         /*
121          * Set the appropriate busy bit for the device/host.
122          *
123          * If the host/device isn't busy, assume that something actually
124          * completed, and that we should be able to queue a command now.
125          *
126          * Note that the prior mid-layer assumption that any host could
127          * always queue at least one command is now broken.  The mid-layer
128          * will implement a user specifiable stall (see
129          * scsi_host.max_host_blocked and scsi_device.max_device_blocked)
130          * if a command is requeued with no other commands outstanding
131          * either for the device or for the host.
132          */
133         if (reason == SCSI_MLQUEUE_HOST_BUSY)
134                 host->host_blocked = host->max_host_blocked;
135         else if (reason == SCSI_MLQUEUE_DEVICE_BUSY)
136                 device->device_blocked = device->max_device_blocked;
137
138         /*
139          * Decrement the counters, since these commands are no longer
140          * active on the host/device.
141          */
142         scsi_device_unbusy(device);
143
144         /*
145          * Requeue this command.  It will go before all other commands
146          * that are already in the queue.
147          *
148          * NOTE: there is magic here about the way the queue is plugged if
149          * we have no outstanding commands.
150          * 
151          * Although we *don't* plug the queue, we call the request
152          * function.  The SCSI request function detects the blocked condition
153          * and plugs the queue appropriately.
154          */
155         spin_lock_irqsave(q->queue_lock, flags);
156         blk_requeue_request(q, cmd->request);
157         spin_unlock_irqrestore(q->queue_lock, flags);
158
159         scsi_run_queue(q);
160
161         return 0;
162 }
163
164 /**
165  * scsi_execute - insert request and wait for the result
166  * @sdev:       scsi device
167  * @cmd:        scsi command
168  * @data_direction: data direction
169  * @buffer:     data buffer
170  * @bufflen:    len of buffer
171  * @sense:      optional sense buffer
172  * @timeout:    request timeout in seconds
173  * @retries:    number of times to retry request
174  * @flags:      or into request flags;
175  *
176  * returns the req->errors value which is the the scsi_cmnd result
177  * field.
178  **/
179 int scsi_execute(struct scsi_device *sdev, const unsigned char *cmd,
180                  int data_direction, void *buffer, unsigned bufflen,
181                  unsigned char *sense, int timeout, int retries, int flags)
182 {
183         struct request *req;
184         int write = (data_direction == DMA_TO_DEVICE);
185         int ret = DRIVER_ERROR << 24;
186
187         req = blk_get_request(sdev->request_queue, write, __GFP_WAIT);
188
189         if (bufflen &&  blk_rq_map_kern(sdev->request_queue, req,
190                                         buffer, bufflen, __GFP_WAIT))
191                 goto out;
192
193         req->cmd_len = COMMAND_SIZE(cmd[0]);
194         memcpy(req->cmd, cmd, req->cmd_len);
195         req->sense = sense;
196         req->sense_len = 0;
197         req->retries = retries;
198         req->timeout = timeout;
199         req->cmd_type = REQ_TYPE_BLOCK_PC;
200         req->cmd_flags |= flags | REQ_QUIET | REQ_PREEMPT;
201
202         /*
203          * head injection *required* here otherwise quiesce won't work
204          */
205         blk_execute_rq(req->q, NULL, req, 1);
206
207         ret = req->errors;
208  out:
209         blk_put_request(req);
210
211         return ret;
212 }
213 EXPORT_SYMBOL(scsi_execute);
214
215
216 int scsi_execute_req(struct scsi_device *sdev, const unsigned char *cmd,
217                      int data_direction, void *buffer, unsigned bufflen,
218                      struct scsi_sense_hdr *sshdr, int timeout, int retries)
219 {
220         char *sense = NULL;
221         int result;
222         
223         if (sshdr) {
224                 sense = kzalloc(SCSI_SENSE_BUFFERSIZE, GFP_NOIO);
225                 if (!sense)
226                         return DRIVER_ERROR << 24;
227         }
228         result = scsi_execute(sdev, cmd, data_direction, buffer, bufflen,
229                               sense, timeout, retries, 0);
230         if (sshdr)
231                 scsi_normalize_sense(sense, SCSI_SENSE_BUFFERSIZE, sshdr);
232
233         kfree(sense);
234         return result;
235 }
236 EXPORT_SYMBOL(scsi_execute_req);
237
238 struct scsi_io_context {
239         void *data;
240         void (*done)(void *data, char *sense, int result, int resid);
241         char sense[SCSI_SENSE_BUFFERSIZE];
242 };
243
244 static struct kmem_cache *scsi_io_context_cache;
245
246 static void scsi_end_async(struct request *req, int uptodate)
247 {
248         struct scsi_io_context *sioc = req->end_io_data;
249
250         if (sioc->done)
251                 sioc->done(sioc->data, sioc->sense, req->errors, req->data_len);
252
253         kmem_cache_free(scsi_io_context_cache, sioc);
254         __blk_put_request(req->q, req);
255 }
256
257 static int scsi_merge_bio(struct request *rq, struct bio *bio)
258 {
259         struct request_queue *q = rq->q;
260
261         bio->bi_flags &= ~(1 << BIO_SEG_VALID);
262         if (rq_data_dir(rq) == WRITE)
263                 bio->bi_rw |= (1 << BIO_RW);
264         blk_queue_bounce(q, &bio);
265
266         if (!rq->bio)
267                 blk_rq_bio_prep(q, rq, bio);
268         else if (!ll_back_merge_fn(q, rq, bio))
269                 return -EINVAL;
270         else {
271                 rq->biotail->bi_next = bio;
272                 rq->biotail = bio;
273         }
274
275         return 0;
276 }
277
278 static int scsi_bi_endio(struct bio *bio, unsigned int bytes_done, int error)
279 {
280         if (bio->bi_size)
281                 return 1;
282
283         bio_put(bio);
284         return 0;
285 }
286
287 /**
288  * scsi_req_map_sg - map a scatterlist into a request
289  * @rq:         request to fill
290  * @sg:         scatterlist
291  * @nsegs:      number of elements
292  * @bufflen:    len of buffer
293  * @gfp:        memory allocation flags
294  *
295  * scsi_req_map_sg maps a scatterlist into a request so that the
296  * request can be sent to the block layer. We do not trust the scatterlist
297  * sent to use, as some ULDs use that struct to only organize the pages.
298  */
299 static int scsi_req_map_sg(struct request *rq, struct scatterlist *sgl,
300                            int nsegs, unsigned bufflen, gfp_t gfp)
301 {
302         struct request_queue *q = rq->q;
303         int nr_pages = (bufflen + sgl[0].offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
304         unsigned int data_len = 0, len, bytes, off;
305         struct page *page;
306         struct bio *bio = NULL;
307         int i, err, nr_vecs = 0;
308
309         for (i = 0; i < nsegs; i++) {
310                 page = sgl[i].page;
311                 off = sgl[i].offset;
312                 len = sgl[i].length;
313                 data_len += len;
314
315                 while (len > 0) {
316                         bytes = min_t(unsigned int, len, PAGE_SIZE - off);
317
318                         if (!bio) {
319                                 nr_vecs = min_t(int, BIO_MAX_PAGES, nr_pages);
320                                 nr_pages -= nr_vecs;
321
322                                 bio = bio_alloc(gfp, nr_vecs);
323                                 if (!bio) {
324                                         err = -ENOMEM;
325                                         goto free_bios;
326                                 }
327                                 bio->bi_end_io = scsi_bi_endio;
328                         }
329
330                         if (bio_add_pc_page(q, bio, page, bytes, off) !=
331                             bytes) {
332                                 bio_put(bio);
333                                 err = -EINVAL;
334                                 goto free_bios;
335                         }
336
337                         if (bio->bi_vcnt >= nr_vecs) {
338                                 err = scsi_merge_bio(rq, bio);
339                                 if (err) {
340                                         bio_endio(bio, bio->bi_size, 0);
341                                         goto free_bios;
342                                 }
343                                 bio = NULL;
344                         }
345
346                         page++;
347                         len -= bytes;
348                         off = 0;
349                 }
350         }
351
352         rq->buffer = rq->data = NULL;
353         rq->data_len = data_len;
354         return 0;
355
356 free_bios:
357         while ((bio = rq->bio) != NULL) {
358                 rq->bio = bio->bi_next;
359                 /*
360                  * call endio instead of bio_put incase it was bounced
361                  */
362                 bio_endio(bio, bio->bi_size, 0);
363         }
364
365         return err;
366 }
367
368 /**
369  * scsi_execute_async - insert request
370  * @sdev:       scsi device
371  * @cmd:        scsi command
372  * @cmd_len:    length of scsi cdb
373  * @data_direction: data direction
374  * @buffer:     data buffer (this can be a kernel buffer or scatterlist)
375  * @bufflen:    len of buffer
376  * @use_sg:     if buffer is a scatterlist this is the number of elements
377  * @timeout:    request timeout in seconds
378  * @retries:    number of times to retry request
379  * @flags:      or into request flags
380  **/
381 int scsi_execute_async(struct scsi_device *sdev, const unsigned char *cmd,
382                        int cmd_len, int data_direction, void *buffer, unsigned bufflen,
383                        int use_sg, int timeout, int retries, void *privdata,
384                        void (*done)(void *, char *, int, int), gfp_t gfp)
385 {
386         struct request *req;
387         struct scsi_io_context *sioc;
388         int err = 0;
389         int write = (data_direction == DMA_TO_DEVICE);
390
391         sioc = kmem_cache_alloc(scsi_io_context_cache, gfp);
392         if (!sioc)
393                 return DRIVER_ERROR << 24;
394         memset(sioc, 0, sizeof(*sioc));
395
396         req = blk_get_request(sdev->request_queue, write, gfp);
397         if (!req)
398                 goto free_sense;
399         req->cmd_type = REQ_TYPE_BLOCK_PC;
400         req->cmd_flags |= REQ_QUIET;
401
402         if (use_sg)
403                 err = scsi_req_map_sg(req, buffer, use_sg, bufflen, gfp);
404         else if (bufflen)
405                 err = blk_rq_map_kern(req->q, req, buffer, bufflen, gfp);
406
407         if (err)
408                 goto free_req;
409
410         req->cmd_len = cmd_len;
411         memset(req->cmd, 0, BLK_MAX_CDB); /* ATAPI hates garbage after CDB */
412         memcpy(req->cmd, cmd, req->cmd_len);
413         req->sense = sioc->sense;
414         req->sense_len = 0;
415         req->timeout = timeout;
416         req->retries = retries;
417         req->end_io_data = sioc;
418
419         sioc->data = privdata;
420         sioc->done = done;
421
422         blk_execute_rq_nowait(req->q, NULL, req, 1, scsi_end_async);
423         return 0;
424
425 free_req:
426         blk_put_request(req);
427 free_sense:
428         kmem_cache_free(scsi_io_context_cache, sioc);
429         return DRIVER_ERROR << 24;
430 }
431 EXPORT_SYMBOL_GPL(scsi_execute_async);
432
433 /*
434  * Function:    scsi_init_cmd_errh()
435  *
436  * Purpose:     Initialize cmd fields related to error handling.
437  *
438  * Arguments:   cmd     - command that is ready to be queued.
439  *
440  * Notes:       This function has the job of initializing a number of
441  *              fields related to error handling.   Typically this will
442  *              be called once for each command, as required.
443  */
444 static void scsi_init_cmd_errh(struct scsi_cmnd *cmd)
445 {
446         cmd->serial_number = 0;
447         memset(cmd->sense_buffer, 0, sizeof cmd->sense_buffer);
448         if (cmd->cmd_len == 0)
449                 cmd->cmd_len = COMMAND_SIZE(cmd->cmnd[0]);
450 }
451
452 void scsi_device_unbusy(struct scsi_device *sdev)
453 {
454         struct Scsi_Host *shost = sdev->host;
455         unsigned long flags;
456
457         spin_lock_irqsave(shost->host_lock, flags);
458         shost->host_busy--;
459         if (unlikely(scsi_host_in_recovery(shost) &&
460                      (shost->host_failed || shost->host_eh_scheduled)))
461                 scsi_eh_wakeup(shost);
462         spin_unlock(shost->host_lock);
463         spin_lock(sdev->request_queue->queue_lock);
464         sdev->device_busy--;
465         spin_unlock_irqrestore(sdev->request_queue->queue_lock, flags);
466 }
467
468 /*
469  * Called for single_lun devices on IO completion. Clear starget_sdev_user,
470  * and call blk_run_queue for all the scsi_devices on the target -
471  * including current_sdev first.
472  *
473  * Called with *no* scsi locks held.
474  */
475 static void scsi_single_lun_run(struct scsi_device *current_sdev)
476 {
477         struct Scsi_Host *shost = current_sdev->host;
478         struct scsi_device *sdev, *tmp;
479         struct scsi_target *starget = scsi_target(current_sdev);
480         unsigned long flags;
481
482         spin_lock_irqsave(shost->host_lock, flags);
483         starget->starget_sdev_user = NULL;
484         spin_unlock_irqrestore(shost->host_lock, flags);
485
486         /*
487          * Call blk_run_queue for all LUNs on the target, starting with
488          * current_sdev. We race with others (to set starget_sdev_user),
489          * but in most cases, we will be first. Ideally, each LU on the
490          * target would get some limited time or requests on the target.
491          */
492         blk_run_queue(current_sdev->request_queue);
493
494         spin_lock_irqsave(shost->host_lock, flags);
495         if (starget->starget_sdev_user)
496                 goto out;
497         list_for_each_entry_safe(sdev, tmp, &starget->devices,
498                         same_target_siblings) {
499                 if (sdev == current_sdev)
500                         continue;
501                 if (scsi_device_get(sdev))
502                         continue;
503
504                 spin_unlock_irqrestore(shost->host_lock, flags);
505                 blk_run_queue(sdev->request_queue);
506                 spin_lock_irqsave(shost->host_lock, flags);
507         
508                 scsi_device_put(sdev);
509         }
510  out:
511         spin_unlock_irqrestore(shost->host_lock, flags);
512 }
513
514 /*
515  * Function:    scsi_run_queue()
516  *
517  * Purpose:     Select a proper request queue to serve next
518  *
519  * Arguments:   q       - last request's queue
520  *
521  * Returns:     Nothing
522  *
523  * Notes:       The previous command was completely finished, start
524  *              a new one if possible.
525  */
526 static void scsi_run_queue(struct request_queue *q)
527 {
528         struct scsi_device *sdev = q->queuedata;
529         struct Scsi_Host *shost = sdev->host;
530         unsigned long flags;
531
532         if (sdev->single_lun)
533                 scsi_single_lun_run(sdev);
534
535         spin_lock_irqsave(shost->host_lock, flags);
536         while (!list_empty(&shost->starved_list) &&
537                !shost->host_blocked && !shost->host_self_blocked &&
538                 !((shost->can_queue > 0) &&
539                   (shost->host_busy >= shost->can_queue))) {
540                 /*
541                  * As long as shost is accepting commands and we have
542                  * starved queues, call blk_run_queue. scsi_request_fn
543                  * drops the queue_lock and can add us back to the
544                  * starved_list.
545                  *
546                  * host_lock protects the starved_list and starved_entry.
547                  * scsi_request_fn must get the host_lock before checking
548                  * or modifying starved_list or starved_entry.
549                  */
550                 sdev = list_entry(shost->starved_list.next,
551                                           struct scsi_device, starved_entry);
552                 list_del_init(&sdev->starved_entry);
553                 spin_unlock_irqrestore(shost->host_lock, flags);
554
555
556                 if (test_bit(QUEUE_FLAG_REENTER, &q->queue_flags) &&
557                     !test_and_set_bit(QUEUE_FLAG_REENTER,
558                                       &sdev->request_queue->queue_flags)) {
559                         blk_run_queue(sdev->request_queue);
560                         clear_bit(QUEUE_FLAG_REENTER,
561                                   &sdev->request_queue->queue_flags);
562                 } else
563                         blk_run_queue(sdev->request_queue);
564
565                 spin_lock_irqsave(shost->host_lock, flags);
566                 if (unlikely(!list_empty(&sdev->starved_entry)))
567                         /*
568                          * sdev lost a race, and was put back on the
569                          * starved list. This is unlikely but without this
570                          * in theory we could loop forever.
571                          */
572                         break;
573         }
574         spin_unlock_irqrestore(shost->host_lock, flags);
575
576         blk_run_queue(q);
577 }
578
579 /*
580  * Function:    scsi_requeue_command()
581  *
582  * Purpose:     Handle post-processing of completed commands.
583  *
584  * Arguments:   q       - queue to operate on
585  *              cmd     - command that may need to be requeued.
586  *
587  * Returns:     Nothing
588  *
589  * Notes:       After command completion, there may be blocks left
590  *              over which weren't finished by the previous command
591  *              this can be for a number of reasons - the main one is
592  *              I/O errors in the middle of the request, in which case
593  *              we need to request the blocks that come after the bad
594  *              sector.
595  * Notes:       Upon return, cmd is a stale pointer.
596  */
597 static void scsi_requeue_command(struct request_queue *q, struct scsi_cmnd *cmd)
598 {
599         struct request *req = cmd->request;
600         unsigned long flags;
601
602         scsi_unprep_request(req);
603         spin_lock_irqsave(q->queue_lock, flags);
604         blk_requeue_request(q, req);
605         spin_unlock_irqrestore(q->queue_lock, flags);
606
607         scsi_run_queue(q);
608 }
609
610 void scsi_next_command(struct scsi_cmnd *cmd)
611 {
612         struct scsi_device *sdev = cmd->device;
613         struct request_queue *q = sdev->request_queue;
614
615         /* need to hold a reference on the device before we let go of the cmd */
616         get_device(&sdev->sdev_gendev);
617
618         scsi_put_command(cmd);
619         scsi_run_queue(q);
620
621         /* ok to remove device now */
622         put_device(&sdev->sdev_gendev);
623 }
624
625 void scsi_run_host_queues(struct Scsi_Host *shost)
626 {
627         struct scsi_device *sdev;
628
629         shost_for_each_device(sdev, shost)
630                 scsi_run_queue(sdev->request_queue);
631 }
632
633 /*
634  * Function:    scsi_end_request()
635  *
636  * Purpose:     Post-processing of completed commands (usually invoked at end
637  *              of upper level post-processing and scsi_io_completion).
638  *
639  * Arguments:   cmd      - command that is complete.
640  *              uptodate - 1 if I/O indicates success, <= 0 for I/O error.
641  *              bytes    - number of bytes of completed I/O
642  *              requeue  - indicates whether we should requeue leftovers.
643  *
644  * Lock status: Assumed that lock is not held upon entry.
645  *
646  * Returns:     cmd if requeue required, NULL otherwise.
647  *
648  * Notes:       This is called for block device requests in order to
649  *              mark some number of sectors as complete.
650  * 
651  *              We are guaranteeing that the request queue will be goosed
652  *              at some point during this call.
653  * Notes:       If cmd was requeued, upon return it will be a stale pointer.
654  */
655 static struct scsi_cmnd *scsi_end_request(struct scsi_cmnd *cmd, int uptodate,
656                                           int bytes, int requeue)
657 {
658         request_queue_t *q = cmd->device->request_queue;
659         struct request *req = cmd->request;
660         unsigned long flags;
661
662         /*
663          * If there are blocks left over at the end, set up the command
664          * to queue the remainder of them.
665          */
666         if (end_that_request_chunk(req, uptodate, bytes)) {
667                 int leftover = (req->hard_nr_sectors << 9);
668
669                 if (blk_pc_request(req))
670                         leftover = req->data_len;
671
672                 /* kill remainder if no retrys */
673                 if (!uptodate && blk_noretry_request(req))
674                         end_that_request_chunk(req, 0, leftover);
675                 else {
676                         if (requeue) {
677                                 /*
678                                  * Bleah.  Leftovers again.  Stick the
679                                  * leftovers in the front of the
680                                  * queue, and goose the queue again.
681                                  */
682                                 scsi_requeue_command(q, cmd);
683                                 cmd = NULL;
684                         }
685                         return cmd;
686                 }
687         }
688
689         add_disk_randomness(req->rq_disk);
690
691         spin_lock_irqsave(q->queue_lock, flags);
692         if (blk_rq_tagged(req))
693                 blk_queue_end_tag(q, req);
694         end_that_request_last(req, uptodate);
695         spin_unlock_irqrestore(q->queue_lock, flags);
696
697         /*
698          * This will goose the queue request function at the end, so we don't
699          * need to worry about launching another command.
700          */
701         scsi_next_command(cmd);
702         return NULL;
703 }
704
705 struct scatterlist *scsi_alloc_sgtable(struct scsi_cmnd *cmd, gfp_t gfp_mask)
706 {
707         struct scsi_host_sg_pool *sgp;
708         struct scatterlist *sgl;
709
710         BUG_ON(!cmd->use_sg);
711
712         switch (cmd->use_sg) {
713         case 1 ... 8:
714                 cmd->sglist_len = 0;
715                 break;
716         case 9 ... 16:
717                 cmd->sglist_len = 1;
718                 break;
719         case 17 ... 32:
720                 cmd->sglist_len = 2;
721                 break;
722 #if (SCSI_MAX_PHYS_SEGMENTS > 32)
723         case 33 ... 64:
724                 cmd->sglist_len = 3;
725                 break;
726 #if (SCSI_MAX_PHYS_SEGMENTS > 64)
727         case 65 ... 128:
728                 cmd->sglist_len = 4;
729                 break;
730 #if (SCSI_MAX_PHYS_SEGMENTS  > 128)
731         case 129 ... 256:
732                 cmd->sglist_len = 5;
733                 break;
734 #endif
735 #endif
736 #endif
737         default:
738                 return NULL;
739         }
740
741         sgp = scsi_sg_pools + cmd->sglist_len;
742         sgl = mempool_alloc(sgp->pool, gfp_mask);
743         return sgl;
744 }
745
746 EXPORT_SYMBOL(scsi_alloc_sgtable);
747
748 void scsi_free_sgtable(struct scatterlist *sgl, int index)
749 {
750         struct scsi_host_sg_pool *sgp;
751
752         BUG_ON(index >= SG_MEMPOOL_NR);
753
754         sgp = scsi_sg_pools + index;
755         mempool_free(sgl, sgp->pool);
756 }
757
758 EXPORT_SYMBOL(scsi_free_sgtable);
759
760 /*
761  * Function:    scsi_release_buffers()
762  *
763  * Purpose:     Completion processing for block device I/O requests.
764  *
765  * Arguments:   cmd     - command that we are bailing.
766  *
767  * Lock status: Assumed that no lock is held upon entry.
768  *
769  * Returns:     Nothing
770  *
771  * Notes:       In the event that an upper level driver rejects a
772  *              command, we must release resources allocated during
773  *              the __init_io() function.  Primarily this would involve
774  *              the scatter-gather table, and potentially any bounce
775  *              buffers.
776  */
777 static void scsi_release_buffers(struct scsi_cmnd *cmd)
778 {
779         if (cmd->use_sg)
780                 scsi_free_sgtable(cmd->request_buffer, cmd->sglist_len);
781
782         /*
783          * Zero these out.  They now point to freed memory, and it is
784          * dangerous to hang onto the pointers.
785          */
786         cmd->request_buffer = NULL;
787         cmd->request_bufflen = 0;
788 }
789
790 /*
791  * Function:    scsi_io_completion()
792  *
793  * Purpose:     Completion processing for block device I/O requests.
794  *
795  * Arguments:   cmd   - command that is finished.
796  *
797  * Lock status: Assumed that no lock is held upon entry.
798  *
799  * Returns:     Nothing
800  *
801  * Notes:       This function is matched in terms of capabilities to
802  *              the function that created the scatter-gather list.
803  *              In other words, if there are no bounce buffers
804  *              (the normal case for most drivers), we don't need
805  *              the logic to deal with cleaning up afterwards.
806  *
807  *              We must do one of several things here:
808  *
809  *              a) Call scsi_end_request.  This will finish off the
810  *                 specified number of sectors.  If we are done, the
811  *                 command block will be released, and the queue
812  *                 function will be goosed.  If we are not done, then
813  *                 scsi_end_request will directly goose the queue.
814  *
815  *              b) We can just use scsi_requeue_command() here.  This would
816  *                 be used if we just wanted to retry, for example.
817  */
818 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
819 {
820         int result = cmd->result;
821         int this_count = cmd->request_bufflen;
822         request_queue_t *q = cmd->device->request_queue;
823         struct request *req = cmd->request;
824         int clear_errors = 1;
825         struct scsi_sense_hdr sshdr;
826         int sense_valid = 0;
827         int sense_deferred = 0;
828
829         scsi_release_buffers(cmd);
830
831         if (result) {
832                 sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
833                 if (sense_valid)
834                         sense_deferred = scsi_sense_is_deferred(&sshdr);
835         }
836
837         if (blk_pc_request(req)) { /* SG_IO ioctl from block level */
838                 req->errors = result;
839                 if (result) {
840                         clear_errors = 0;
841                         if (sense_valid && req->sense) {
842                                 /*
843                                  * SG_IO wants current and deferred errors
844                                  */
845                                 int len = 8 + cmd->sense_buffer[7];
846
847                                 if (len > SCSI_SENSE_BUFFERSIZE)
848                                         len = SCSI_SENSE_BUFFERSIZE;
849                                 memcpy(req->sense, cmd->sense_buffer,  len);
850                                 req->sense_len = len;
851                         }
852                 } else
853                         req->data_len = cmd->resid;
854         }
855
856         /*
857          * Next deal with any sectors which we were able to correctly
858          * handle.
859          */
860         SCSI_LOG_HLCOMPLETE(1, printk("%ld sectors total, "
861                                       "%d bytes done.\n",
862                                       req->nr_sectors, good_bytes));
863         SCSI_LOG_HLCOMPLETE(1, printk("use_sg is %d\n", cmd->use_sg));
864
865         if (clear_errors)
866                 req->errors = 0;
867
868         /* A number of bytes were successfully read.  If there
869          * are leftovers and there is some kind of error
870          * (result != 0), retry the rest.
871          */
872         if (scsi_end_request(cmd, 1, good_bytes, result == 0) == NULL)
873                 return;
874
875         /* good_bytes = 0, or (inclusive) there were leftovers and
876          * result = 0, so scsi_end_request couldn't retry.
877          */
878         if (sense_valid && !sense_deferred) {
879                 switch (sshdr.sense_key) {
880                 case UNIT_ATTENTION:
881                         if (cmd->device->removable) {
882                                 /* Detected disc change.  Set a bit
883                                  * and quietly refuse further access.
884                                  */
885                                 cmd->device->changed = 1;
886                                 scsi_end_request(cmd, 0, this_count, 1);
887                                 return;
888                         } else {
889                                 /* Must have been a power glitch, or a
890                                  * bus reset.  Could not have been a
891                                  * media change, so we just retry the
892                                  * request and see what happens.
893                                  */
894                                 scsi_requeue_command(q, cmd);
895                                 return;
896                         }
897                         break;
898                 case ILLEGAL_REQUEST:
899                         /* If we had an ILLEGAL REQUEST returned, then
900                          * we may have performed an unsupported
901                          * command.  The only thing this should be
902                          * would be a ten byte read where only a six
903                          * byte read was supported.  Also, on a system
904                          * where READ CAPACITY failed, we may have
905                          * read past the end of the disk.
906                          */
907                         if ((cmd->device->use_10_for_rw &&
908                             sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
909                             (cmd->cmnd[0] == READ_10 ||
910                              cmd->cmnd[0] == WRITE_10)) {
911                                 cmd->device->use_10_for_rw = 0;
912                                 /* This will cause a retry with a
913                                  * 6-byte command.
914                                  */
915                                 scsi_requeue_command(q, cmd);
916                                 return;
917                         } else {
918                                 scsi_end_request(cmd, 0, this_count, 1);
919                                 return;
920                         }
921                         break;
922                 case NOT_READY:
923                         /* If the device is in the process of becoming
924                          * ready, or has a temporary blockage, retry.
925                          */
926                         if (sshdr.asc == 0x04) {
927                                 switch (sshdr.ascq) {
928                                 case 0x01: /* becoming ready */
929                                 case 0x04: /* format in progress */
930                                 case 0x05: /* rebuild in progress */
931                                 case 0x06: /* recalculation in progress */
932                                 case 0x07: /* operation in progress */
933                                 case 0x08: /* Long write in progress */
934                                 case 0x09: /* self test in progress */
935                                         scsi_requeue_command(q, cmd);
936                                         return;
937                                 default:
938                                         break;
939                                 }
940                         }
941                         if (!(req->cmd_flags & REQ_QUIET)) {
942                                 scmd_printk(KERN_INFO, cmd,
943                                             "Device not ready: ");
944                                 scsi_print_sense_hdr("", &sshdr);
945                         }
946                         scsi_end_request(cmd, 0, this_count, 1);
947                         return;
948                 case VOLUME_OVERFLOW:
949                         if (!(req->cmd_flags & REQ_QUIET)) {
950                                 scmd_printk(KERN_INFO, cmd,
951                                             "Volume overflow, CDB: ");
952                                 __scsi_print_command(cmd->cmnd);
953                                 scsi_print_sense("", cmd);
954                         }
955                         /* See SSC3rXX or current. */
956                         scsi_end_request(cmd, 0, this_count, 1);
957                         return;
958                 default:
959                         break;
960                 }
961         }
962         if (host_byte(result) == DID_RESET) {
963                 /* Third party bus reset or reset for error recovery
964                  * reasons.  Just retry the request and see what
965                  * happens.
966                  */
967                 scsi_requeue_command(q, cmd);
968                 return;
969         }
970         if (result) {
971                 if (!(req->cmd_flags & REQ_QUIET)) {
972                         scmd_printk(KERN_INFO, cmd,
973                                     "SCSI error: return code = 0x%08x\n",
974                                     result);
975                         if (driver_byte(result) & DRIVER_SENSE)
976                                 scsi_print_sense("", cmd);
977                 }
978         }
979         scsi_end_request(cmd, 0, this_count, !result);
980 }
981 EXPORT_SYMBOL(scsi_io_completion);
982
983 /*
984  * Function:    scsi_init_io()
985  *
986  * Purpose:     SCSI I/O initialize function.
987  *
988  * Arguments:   cmd   - Command descriptor we wish to initialize
989  *
990  * Returns:     0 on success
991  *              BLKPREP_DEFER if the failure is retryable
992  *              BLKPREP_KILL if the failure is fatal
993  */
994 static int scsi_init_io(struct scsi_cmnd *cmd)
995 {
996         struct request     *req = cmd->request;
997         struct scatterlist *sgpnt;
998         int                count;
999
1000         /*
1001          * We used to not use scatter-gather for single segment request,
1002          * but now we do (it makes highmem I/O easier to support without
1003          * kmapping pages)
1004          */
1005         cmd->use_sg = req->nr_phys_segments;
1006
1007         /*
1008          * If sg table allocation fails, requeue request later.
1009          */
1010         sgpnt = scsi_alloc_sgtable(cmd, GFP_ATOMIC);
1011         if (unlikely(!sgpnt)) {
1012                 scsi_unprep_request(req);
1013                 return BLKPREP_DEFER;
1014         }
1015
1016         req->buffer = NULL;
1017         cmd->request_buffer = (char *) sgpnt;
1018         if (blk_pc_request(req))
1019                 cmd->request_bufflen = req->data_len;
1020         else
1021                 cmd->request_bufflen = req->nr_sectors << 9;
1022
1023         /* 
1024          * Next, walk the list, and fill in the addresses and sizes of
1025          * each segment.
1026          */
1027         count = blk_rq_map_sg(req->q, req, cmd->request_buffer);
1028         if (likely(count <= cmd->use_sg)) {
1029                 cmd->use_sg = count;
1030                 return BLKPREP_OK;
1031         }
1032
1033         printk(KERN_ERR "Incorrect number of segments after building list\n");
1034         printk(KERN_ERR "counted %d, received %d\n", count, cmd->use_sg);
1035         printk(KERN_ERR "req nr_sec %lu, cur_nr_sec %u\n", req->nr_sectors,
1036                         req->current_nr_sectors);
1037
1038         /* release the command and kill it */
1039         scsi_release_buffers(cmd);
1040         scsi_put_command(cmd);
1041         return BLKPREP_KILL;
1042 }
1043
1044 static int scsi_issue_flush_fn(request_queue_t *q, struct gendisk *disk,
1045                                sector_t *error_sector)
1046 {
1047         struct scsi_device *sdev = q->queuedata;
1048         struct scsi_driver *drv;
1049
1050         if (sdev->sdev_state != SDEV_RUNNING)
1051                 return -ENXIO;
1052
1053         drv = *(struct scsi_driver **) disk->private_data;
1054         if (drv->issue_flush)
1055                 return drv->issue_flush(&sdev->sdev_gendev, error_sector);
1056
1057         return -EOPNOTSUPP;
1058 }
1059
1060 static struct scsi_cmnd *scsi_get_cmd_from_req(struct scsi_device *sdev,
1061                 struct request *req)
1062 {
1063         struct scsi_cmnd *cmd;
1064
1065         if (!req->special) {
1066                 cmd = scsi_get_command(sdev, GFP_ATOMIC);
1067                 if (unlikely(!cmd))
1068                         return NULL;
1069                 req->special = cmd;
1070         } else {
1071                 cmd = req->special;
1072         }
1073
1074         /* pull a tag out of the request if we have one */
1075         cmd->tag = req->tag;
1076         cmd->request = req;
1077
1078         return cmd;
1079 }
1080
1081 static void scsi_blk_pc_done(struct scsi_cmnd *cmd)
1082 {
1083         BUG_ON(!blk_pc_request(cmd->request));
1084         /*
1085          * This will complete the whole command with uptodate=1 so
1086          * as far as the block layer is concerned the command completed
1087          * successfully. Since this is a REQ_BLOCK_PC command the
1088          * caller should check the request's errors value
1089          */
1090         scsi_io_completion(cmd, cmd->request_bufflen);
1091 }
1092
1093 static int scsi_setup_blk_pc_cmnd(struct scsi_device *sdev, struct request *req)
1094 {
1095         struct scsi_cmnd *cmd;
1096
1097         cmd = scsi_get_cmd_from_req(sdev, req);
1098         if (unlikely(!cmd))
1099                 return BLKPREP_DEFER;
1100
1101         /*
1102          * BLOCK_PC requests may transfer data, in which case they must
1103          * a bio attached to them.  Or they might contain a SCSI command
1104          * that does not transfer data, in which case they may optionally
1105          * submit a request without an attached bio.
1106          */
1107         if (req->bio) {
1108                 int ret;
1109
1110                 BUG_ON(!req->nr_phys_segments);
1111
1112                 ret = scsi_init_io(cmd);
1113                 if (unlikely(ret))
1114                         return ret;
1115         } else {
1116                 BUG_ON(req->data_len);
1117                 BUG_ON(req->data);
1118
1119                 cmd->request_bufflen = 0;
1120                 cmd->request_buffer = NULL;
1121                 cmd->use_sg = 0;
1122                 req->buffer = NULL;
1123         }
1124
1125         BUILD_BUG_ON(sizeof(req->cmd) > sizeof(cmd->cmnd));
1126         memcpy(cmd->cmnd, req->cmd, sizeof(cmd->cmnd));
1127         cmd->cmd_len = req->cmd_len;
1128         if (!req->data_len)
1129                 cmd->sc_data_direction = DMA_NONE;
1130         else if (rq_data_dir(req) == WRITE)
1131                 cmd->sc_data_direction = DMA_TO_DEVICE;
1132         else
1133                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1134         
1135         cmd->transfersize = req->data_len;
1136         cmd->allowed = req->retries;
1137         cmd->timeout_per_command = req->timeout;
1138         cmd->done = scsi_blk_pc_done;
1139         return BLKPREP_OK;
1140 }
1141
1142 /*
1143  * Setup a REQ_TYPE_FS command.  These are simple read/write request
1144  * from filesystems that still need to be translated to SCSI CDBs from
1145  * the ULD.
1146  */
1147 static int scsi_setup_fs_cmnd(struct scsi_device *sdev, struct request *req)
1148 {
1149         struct scsi_cmnd *cmd;
1150         struct scsi_driver *drv;
1151         int ret;
1152
1153         /*
1154          * Filesystem requests must transfer data.
1155          */
1156         BUG_ON(!req->nr_phys_segments);
1157
1158         cmd = scsi_get_cmd_from_req(sdev, req);
1159         if (unlikely(!cmd))
1160                 return BLKPREP_DEFER;
1161
1162         ret = scsi_init_io(cmd);
1163         if (unlikely(ret))
1164                 return ret;
1165
1166         /*
1167          * Initialize the actual SCSI command for this request.
1168          */
1169         drv = *(struct scsi_driver **)req->rq_disk->private_data;
1170         if (unlikely(!drv->init_command(cmd))) {
1171                 scsi_release_buffers(cmd);
1172                 scsi_put_command(cmd);
1173                 return BLKPREP_KILL;
1174         }
1175
1176         return BLKPREP_OK;
1177 }
1178
1179 static int scsi_prep_fn(struct request_queue *q, struct request *req)
1180 {
1181         struct scsi_device *sdev = q->queuedata;
1182         int ret = BLKPREP_OK;
1183
1184         /*
1185          * If the device is not in running state we will reject some
1186          * or all commands.
1187          */
1188         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1189                 switch (sdev->sdev_state) {
1190                 case SDEV_OFFLINE:
1191                         /*
1192                          * If the device is offline we refuse to process any
1193                          * commands.  The device must be brought online
1194                          * before trying any recovery commands.
1195                          */
1196                         sdev_printk(KERN_ERR, sdev,
1197                                     "rejecting I/O to offline device\n");
1198                         ret = BLKPREP_KILL;
1199                         break;
1200                 case SDEV_DEL:
1201                         /*
1202                          * If the device is fully deleted, we refuse to
1203                          * process any commands as well.
1204                          */
1205                         sdev_printk(KERN_ERR, sdev,
1206                                     "rejecting I/O to dead device\n");
1207                         ret = BLKPREP_KILL;
1208                         break;
1209                 case SDEV_QUIESCE:
1210                 case SDEV_BLOCK:
1211                         /*
1212                          * If the devices is blocked we defer normal commands.
1213                          */
1214                         if (!(req->cmd_flags & REQ_PREEMPT))
1215                                 ret = BLKPREP_DEFER;
1216                         break;
1217                 default:
1218                         /*
1219                          * For any other not fully online state we only allow
1220                          * special commands.  In particular any user initiated
1221                          * command is not allowed.
1222                          */
1223                         if (!(req->cmd_flags & REQ_PREEMPT))
1224                                 ret = BLKPREP_KILL;
1225                         break;
1226                 }
1227
1228                 if (ret != BLKPREP_OK)
1229                         goto out;
1230         }
1231
1232         switch (req->cmd_type) {
1233         case REQ_TYPE_BLOCK_PC:
1234                 ret = scsi_setup_blk_pc_cmnd(sdev, req);
1235                 break;
1236         case REQ_TYPE_FS:
1237                 ret = scsi_setup_fs_cmnd(sdev, req);
1238                 break;
1239         default:
1240                 /*
1241                  * All other command types are not supported.
1242                  *
1243                  * Note that these days the SCSI subsystem does not use
1244                  * REQ_TYPE_SPECIAL requests anymore.  These are only used
1245                  * (directly or via blk_insert_request) by non-SCSI drivers.
1246                  */
1247                 blk_dump_rq_flags(req, "SCSI bad req");
1248                 ret = BLKPREP_KILL;
1249                 break;
1250         }
1251
1252  out:
1253         switch (ret) {
1254         case BLKPREP_KILL:
1255                 req->errors = DID_NO_CONNECT << 16;
1256                 break;
1257         case BLKPREP_DEFER:
1258                 /*
1259                  * If we defer, the elv_next_request() returns NULL, but the
1260                  * queue must be restarted, so we plug here if no returning
1261                  * command will automatically do that.
1262                  */
1263                 if (sdev->device_busy == 0)
1264                         blk_plug_device(q);
1265                 break;
1266         default:
1267                 req->cmd_flags |= REQ_DONTPREP;
1268         }
1269
1270         return ret;
1271 }
1272
1273 /*
1274  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1275  * return 0.
1276  *
1277  * Called with the queue_lock held.
1278  */
1279 static inline int scsi_dev_queue_ready(struct request_queue *q,
1280                                   struct scsi_device *sdev)
1281 {
1282         if (sdev->device_busy >= sdev->queue_depth)
1283                 return 0;
1284         if (sdev->device_busy == 0 && sdev->device_blocked) {
1285                 /*
1286                  * unblock after device_blocked iterates to zero
1287                  */
1288                 if (--sdev->device_blocked == 0) {
1289                         SCSI_LOG_MLQUEUE(3,
1290                                    sdev_printk(KERN_INFO, sdev,
1291                                    "unblocking device at zero depth\n"));
1292                 } else {
1293                         blk_plug_device(q);
1294                         return 0;
1295                 }
1296         }
1297         if (sdev->device_blocked)
1298                 return 0;
1299
1300         return 1;
1301 }
1302
1303 /*
1304  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1305  * return 0. We must end up running the queue again whenever 0 is
1306  * returned, else IO can hang.
1307  *
1308  * Called with host_lock held.
1309  */
1310 static inline int scsi_host_queue_ready(struct request_queue *q,
1311                                    struct Scsi_Host *shost,
1312                                    struct scsi_device *sdev)
1313 {
1314         if (scsi_host_in_recovery(shost))
1315                 return 0;
1316         if (shost->host_busy == 0 && shost->host_blocked) {
1317                 /*
1318                  * unblock after host_blocked iterates to zero
1319                  */
1320                 if (--shost->host_blocked == 0) {
1321                         SCSI_LOG_MLQUEUE(3,
1322                                 printk("scsi%d unblocking host at zero depth\n",
1323                                         shost->host_no));
1324                 } else {
1325                         blk_plug_device(q);
1326                         return 0;
1327                 }
1328         }
1329         if ((shost->can_queue > 0 && shost->host_busy >= shost->can_queue) ||
1330             shost->host_blocked || shost->host_self_blocked) {
1331                 if (list_empty(&sdev->starved_entry))
1332                         list_add_tail(&sdev->starved_entry, &shost->starved_list);
1333                 return 0;
1334         }
1335
1336         /* We're OK to process the command, so we can't be starved */
1337         if (!list_empty(&sdev->starved_entry))
1338                 list_del_init(&sdev->starved_entry);
1339
1340         return 1;
1341 }
1342
1343 /*
1344  * Kill a request for a dead device
1345  */
1346 static void scsi_kill_request(struct request *req, request_queue_t *q)
1347 {
1348         struct scsi_cmnd *cmd = req->special;
1349         struct scsi_device *sdev = cmd->device;
1350         struct Scsi_Host *shost = sdev->host;
1351
1352         blkdev_dequeue_request(req);
1353
1354         if (unlikely(cmd == NULL)) {
1355                 printk(KERN_CRIT "impossible request in %s.\n",
1356                                  __FUNCTION__);
1357                 BUG();
1358         }
1359
1360         scsi_init_cmd_errh(cmd);
1361         cmd->result = DID_NO_CONNECT << 16;
1362         atomic_inc(&cmd->device->iorequest_cnt);
1363
1364         /*
1365          * SCSI request completion path will do scsi_device_unbusy(),
1366          * bump busy counts.  To bump the counters, we need to dance
1367          * with the locks as normal issue path does.
1368          */
1369         sdev->device_busy++;
1370         spin_unlock(sdev->request_queue->queue_lock);
1371         spin_lock(shost->host_lock);
1372         shost->host_busy++;
1373         spin_unlock(shost->host_lock);
1374         spin_lock(sdev->request_queue->queue_lock);
1375
1376         __scsi_done(cmd);
1377 }
1378
1379 static void scsi_softirq_done(struct request *rq)
1380 {
1381         struct scsi_cmnd *cmd = rq->completion_data;
1382         unsigned long wait_for = (cmd->allowed + 1) * cmd->timeout_per_command;
1383         int disposition;
1384
1385         INIT_LIST_HEAD(&cmd->eh_entry);
1386
1387         disposition = scsi_decide_disposition(cmd);
1388         if (disposition != SUCCESS &&
1389             time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1390                 sdev_printk(KERN_ERR, cmd->device,
1391                             "timing out command, waited %lus\n",
1392                             wait_for/HZ);
1393                 disposition = SUCCESS;
1394         }
1395                         
1396         scsi_log_completion(cmd, disposition);
1397
1398         switch (disposition) {
1399                 case SUCCESS:
1400                         scsi_finish_command(cmd);
1401                         break;
1402                 case NEEDS_RETRY:
1403                         scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1404                         break;
1405                 case ADD_TO_MLQUEUE:
1406                         scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1407                         break;
1408                 default:
1409                         if (!scsi_eh_scmd_add(cmd, 0))
1410                                 scsi_finish_command(cmd);
1411         }
1412 }
1413
1414 /*
1415  * Function:    scsi_request_fn()
1416  *
1417  * Purpose:     Main strategy routine for SCSI.
1418  *
1419  * Arguments:   q       - Pointer to actual queue.
1420  *
1421  * Returns:     Nothing
1422  *
1423  * Lock status: IO request lock assumed to be held when called.
1424  */
1425 static void scsi_request_fn(struct request_queue *q)
1426 {
1427         struct scsi_device *sdev = q->queuedata;
1428         struct Scsi_Host *shost;
1429         struct scsi_cmnd *cmd;
1430         struct request *req;
1431
1432         if (!sdev) {
1433                 printk("scsi: killing requests for dead queue\n");
1434                 while ((req = elv_next_request(q)) != NULL)
1435                         scsi_kill_request(req, q);
1436                 return;
1437         }
1438
1439         if(!get_device(&sdev->sdev_gendev))
1440                 /* We must be tearing the block queue down already */
1441                 return;
1442
1443         /*
1444          * To start with, we keep looping until the queue is empty, or until
1445          * the host is no longer able to accept any more requests.
1446          */
1447         shost = sdev->host;
1448         while (!blk_queue_plugged(q)) {
1449                 int rtn;
1450                 /*
1451                  * get next queueable request.  We do this early to make sure
1452                  * that the request is fully prepared even if we cannot 
1453                  * accept it.
1454                  */
1455                 req = elv_next_request(q);
1456                 if (!req || !scsi_dev_queue_ready(q, sdev))
1457                         break;
1458
1459                 if (unlikely(!scsi_device_online(sdev))) {
1460                         sdev_printk(KERN_ERR, sdev,
1461                                     "rejecting I/O to offline device\n");
1462                         scsi_kill_request(req, q);
1463                         continue;
1464                 }
1465
1466
1467                 /*
1468                  * Remove the request from the request list.
1469                  */
1470                 if (!(blk_queue_tagged(q) && !blk_queue_start_tag(q, req)))
1471                         blkdev_dequeue_request(req);
1472                 sdev->device_busy++;
1473
1474                 spin_unlock(q->queue_lock);
1475                 cmd = req->special;
1476                 if (unlikely(cmd == NULL)) {
1477                         printk(KERN_CRIT "impossible request in %s.\n"
1478                                          "please mail a stack trace to "
1479                                          "linux-scsi@vger.kernel.org\n",
1480                                          __FUNCTION__);
1481                         blk_dump_rq_flags(req, "foo");
1482                         BUG();
1483                 }
1484                 spin_lock(shost->host_lock);
1485
1486                 if (!scsi_host_queue_ready(q, shost, sdev))
1487                         goto not_ready;
1488                 if (sdev->single_lun) {
1489                         if (scsi_target(sdev)->starget_sdev_user &&
1490                             scsi_target(sdev)->starget_sdev_user != sdev)
1491                                 goto not_ready;
1492                         scsi_target(sdev)->starget_sdev_user = sdev;
1493                 }
1494                 shost->host_busy++;
1495
1496                 /*
1497                  * XXX(hch): This is rather suboptimal, scsi_dispatch_cmd will
1498                  *              take the lock again.
1499                  */
1500                 spin_unlock_irq(shost->host_lock);
1501
1502                 /*
1503                  * Finally, initialize any error handling parameters, and set up
1504                  * the timers for timeouts.
1505                  */
1506                 scsi_init_cmd_errh(cmd);
1507
1508                 /*
1509                  * Dispatch the command to the low-level driver.
1510                  */
1511                 rtn = scsi_dispatch_cmd(cmd);
1512                 spin_lock_irq(q->queue_lock);
1513                 if(rtn) {
1514                         /* we're refusing the command; because of
1515                          * the way locks get dropped, we need to 
1516                          * check here if plugging is required */
1517                         if(sdev->device_busy == 0)
1518                                 blk_plug_device(q);
1519
1520                         break;
1521                 }
1522         }
1523
1524         goto out;
1525
1526  not_ready:
1527         spin_unlock_irq(shost->host_lock);
1528
1529         /*
1530          * lock q, handle tag, requeue req, and decrement device_busy. We
1531          * must return with queue_lock held.
1532          *
1533          * Decrementing device_busy without checking it is OK, as all such
1534          * cases (host limits or settings) should run the queue at some
1535          * later time.
1536          */
1537         spin_lock_irq(q->queue_lock);
1538         blk_requeue_request(q, req);
1539         sdev->device_busy--;
1540         if(sdev->device_busy == 0)
1541                 blk_plug_device(q);
1542  out:
1543         /* must be careful here...if we trigger the ->remove() function
1544          * we cannot be holding the q lock */
1545         spin_unlock_irq(q->queue_lock);
1546         put_device(&sdev->sdev_gendev);
1547         spin_lock_irq(q->queue_lock);
1548 }
1549
1550 u64 scsi_calculate_bounce_limit(struct Scsi_Host *shost)
1551 {
1552         struct device *host_dev;
1553         u64 bounce_limit = 0xffffffff;
1554
1555         if (shost->unchecked_isa_dma)
1556                 return BLK_BOUNCE_ISA;
1557         /*
1558          * Platforms with virtual-DMA translation
1559          * hardware have no practical limit.
1560          */
1561         if (!PCI_DMA_BUS_IS_PHYS)
1562                 return BLK_BOUNCE_ANY;
1563
1564         host_dev = scsi_get_device(shost);
1565         if (host_dev && host_dev->dma_mask)
1566                 bounce_limit = *host_dev->dma_mask;
1567
1568         return bounce_limit;
1569 }
1570 EXPORT_SYMBOL(scsi_calculate_bounce_limit);
1571
1572 struct request_queue *__scsi_alloc_queue(struct Scsi_Host *shost,
1573                                          request_fn_proc *request_fn)
1574 {
1575         struct request_queue *q;
1576
1577         q = blk_init_queue(request_fn, NULL);
1578         if (!q)
1579                 return NULL;
1580
1581         blk_queue_max_hw_segments(q, shost->sg_tablesize);
1582         blk_queue_max_phys_segments(q, SCSI_MAX_PHYS_SEGMENTS);
1583         blk_queue_max_sectors(q, shost->max_sectors);
1584         blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
1585         blk_queue_segment_boundary(q, shost->dma_boundary);
1586
1587         if (!shost->use_clustering)
1588                 clear_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags);
1589         return q;
1590 }
1591 EXPORT_SYMBOL(__scsi_alloc_queue);
1592
1593 struct request_queue *scsi_alloc_queue(struct scsi_device *sdev)
1594 {
1595         struct request_queue *q;
1596
1597         q = __scsi_alloc_queue(sdev->host, scsi_request_fn);
1598         if (!q)
1599                 return NULL;
1600
1601         blk_queue_prep_rq(q, scsi_prep_fn);
1602         blk_queue_issue_flush_fn(q, scsi_issue_flush_fn);
1603         blk_queue_softirq_done(q, scsi_softirq_done);
1604         return q;
1605 }
1606
1607 void scsi_free_queue(struct request_queue *q)
1608 {
1609         blk_cleanup_queue(q);
1610 }
1611
1612 /*
1613  * Function:    scsi_block_requests()
1614  *
1615  * Purpose:     Utility function used by low-level drivers to prevent further
1616  *              commands from being queued to the device.
1617  *
1618  * Arguments:   shost       - Host in question
1619  *
1620  * Returns:     Nothing
1621  *
1622  * Lock status: No locks are assumed held.
1623  *
1624  * Notes:       There is no timer nor any other means by which the requests
1625  *              get unblocked other than the low-level driver calling
1626  *              scsi_unblock_requests().
1627  */
1628 void scsi_block_requests(struct Scsi_Host *shost)
1629 {
1630         shost->host_self_blocked = 1;
1631 }
1632 EXPORT_SYMBOL(scsi_block_requests);
1633
1634 /*
1635  * Function:    scsi_unblock_requests()
1636  *
1637  * Purpose:     Utility function used by low-level drivers to allow further
1638  *              commands from being queued to the device.
1639  *
1640  * Arguments:   shost       - Host in question
1641  *
1642  * Returns:     Nothing
1643  *
1644  * Lock status: No locks are assumed held.
1645  *
1646  * Notes:       There is no timer nor any other means by which the requests
1647  *              get unblocked other than the low-level driver calling
1648  *              scsi_unblock_requests().
1649  *
1650  *              This is done as an API function so that changes to the
1651  *              internals of the scsi mid-layer won't require wholesale
1652  *              changes to drivers that use this feature.
1653  */
1654 void scsi_unblock_requests(struct Scsi_Host *shost)
1655 {
1656         shost->host_self_blocked = 0;
1657         scsi_run_host_queues(shost);
1658 }
1659 EXPORT_SYMBOL(scsi_unblock_requests);
1660
1661 int __init scsi_init_queue(void)
1662 {
1663         int i;
1664
1665         scsi_io_context_cache = kmem_cache_create("scsi_io_context",
1666                                         sizeof(struct scsi_io_context),
1667                                         0, 0, NULL, NULL);
1668         if (!scsi_io_context_cache) {
1669                 printk(KERN_ERR "SCSI: can't init scsi io context cache\n");
1670                 return -ENOMEM;
1671         }
1672
1673         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1674                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1675                 int size = sgp->size * sizeof(struct scatterlist);
1676
1677                 sgp->slab = kmem_cache_create(sgp->name, size, 0,
1678                                 SLAB_HWCACHE_ALIGN, NULL, NULL);
1679                 if (!sgp->slab) {
1680                         printk(KERN_ERR "SCSI: can't init sg slab %s\n",
1681                                         sgp->name);
1682                 }
1683
1684                 sgp->pool = mempool_create_slab_pool(SG_MEMPOOL_SIZE,
1685                                                      sgp->slab);
1686                 if (!sgp->pool) {
1687                         printk(KERN_ERR "SCSI: can't init sg mempool %s\n",
1688                                         sgp->name);
1689                 }
1690         }
1691
1692         return 0;
1693 }
1694
1695 void scsi_exit_queue(void)
1696 {
1697         int i;
1698
1699         kmem_cache_destroy(scsi_io_context_cache);
1700
1701         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1702                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1703                 mempool_destroy(sgp->pool);
1704                 kmem_cache_destroy(sgp->slab);
1705         }
1706 }
1707
1708 /**
1709  *      scsi_mode_select - issue a mode select
1710  *      @sdev:  SCSI device to be queried
1711  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
1712  *      @sp:    Save page bit (0 == don't save, 1 == save)
1713  *      @modepage: mode page being requested
1714  *      @buffer: request buffer (may not be smaller than eight bytes)
1715  *      @len:   length of request buffer.
1716  *      @timeout: command timeout
1717  *      @retries: number of retries before failing
1718  *      @data: returns a structure abstracting the mode header data
1719  *      @sense: place to put sense data (or NULL if no sense to be collected).
1720  *              must be SCSI_SENSE_BUFFERSIZE big.
1721  *
1722  *      Returns zero if successful; negative error number or scsi
1723  *      status on error
1724  *
1725  */
1726 int
1727 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
1728                  unsigned char *buffer, int len, int timeout, int retries,
1729                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1730 {
1731         unsigned char cmd[10];
1732         unsigned char *real_buffer;
1733         int ret;
1734
1735         memset(cmd, 0, sizeof(cmd));
1736         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
1737
1738         if (sdev->use_10_for_ms) {
1739                 if (len > 65535)
1740                         return -EINVAL;
1741                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
1742                 if (!real_buffer)
1743                         return -ENOMEM;
1744                 memcpy(real_buffer + 8, buffer, len);
1745                 len += 8;
1746                 real_buffer[0] = 0;
1747                 real_buffer[1] = 0;
1748                 real_buffer[2] = data->medium_type;
1749                 real_buffer[3] = data->device_specific;
1750                 real_buffer[4] = data->longlba ? 0x01 : 0;
1751                 real_buffer[5] = 0;
1752                 real_buffer[6] = data->block_descriptor_length >> 8;
1753                 real_buffer[7] = data->block_descriptor_length;
1754
1755                 cmd[0] = MODE_SELECT_10;
1756                 cmd[7] = len >> 8;
1757                 cmd[8] = len;
1758         } else {
1759                 if (len > 255 || data->block_descriptor_length > 255 ||
1760                     data->longlba)
1761                         return -EINVAL;
1762
1763                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
1764                 if (!real_buffer)
1765                         return -ENOMEM;
1766                 memcpy(real_buffer + 4, buffer, len);
1767                 len += 4;
1768                 real_buffer[0] = 0;
1769                 real_buffer[1] = data->medium_type;
1770                 real_buffer[2] = data->device_specific;
1771                 real_buffer[3] = data->block_descriptor_length;
1772                 
1773
1774                 cmd[0] = MODE_SELECT;
1775                 cmd[4] = len;
1776         }
1777
1778         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
1779                                sshdr, timeout, retries);
1780         kfree(real_buffer);
1781         return ret;
1782 }
1783 EXPORT_SYMBOL_GPL(scsi_mode_select);
1784
1785 /**
1786  *      scsi_mode_sense - issue a mode sense, falling back from 10 to 
1787  *              six bytes if necessary.
1788  *      @sdev:  SCSI device to be queried
1789  *      @dbd:   set if mode sense will allow block descriptors to be returned
1790  *      @modepage: mode page being requested
1791  *      @buffer: request buffer (may not be smaller than eight bytes)
1792  *      @len:   length of request buffer.
1793  *      @timeout: command timeout
1794  *      @retries: number of retries before failing
1795  *      @data: returns a structure abstracting the mode header data
1796  *      @sense: place to put sense data (or NULL if no sense to be collected).
1797  *              must be SCSI_SENSE_BUFFERSIZE big.
1798  *
1799  *      Returns zero if unsuccessful, or the header offset (either 4
1800  *      or 8 depending on whether a six or ten byte command was
1801  *      issued) if successful.
1802  **/
1803 int
1804 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
1805                   unsigned char *buffer, int len, int timeout, int retries,
1806                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1807 {
1808         unsigned char cmd[12];
1809         int use_10_for_ms;
1810         int header_length;
1811         int result;
1812         struct scsi_sense_hdr my_sshdr;
1813
1814         memset(data, 0, sizeof(*data));
1815         memset(&cmd[0], 0, 12);
1816         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
1817         cmd[2] = modepage;
1818
1819         /* caller might not be interested in sense, but we need it */
1820         if (!sshdr)
1821                 sshdr = &my_sshdr;
1822
1823  retry:
1824         use_10_for_ms = sdev->use_10_for_ms;
1825
1826         if (use_10_for_ms) {
1827                 if (len < 8)
1828                         len = 8;
1829
1830                 cmd[0] = MODE_SENSE_10;
1831                 cmd[8] = len;
1832                 header_length = 8;
1833         } else {
1834                 if (len < 4)
1835                         len = 4;
1836
1837                 cmd[0] = MODE_SENSE;
1838                 cmd[4] = len;
1839                 header_length = 4;
1840         }
1841
1842         memset(buffer, 0, len);
1843
1844         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
1845                                   sshdr, timeout, retries);
1846
1847         /* This code looks awful: what it's doing is making sure an
1848          * ILLEGAL REQUEST sense return identifies the actual command
1849          * byte as the problem.  MODE_SENSE commands can return
1850          * ILLEGAL REQUEST if the code page isn't supported */
1851
1852         if (use_10_for_ms && !scsi_status_is_good(result) &&
1853             (driver_byte(result) & DRIVER_SENSE)) {
1854                 if (scsi_sense_valid(sshdr)) {
1855                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
1856                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
1857                                 /* 
1858                                  * Invalid command operation code
1859                                  */
1860                                 sdev->use_10_for_ms = 0;
1861                                 goto retry;
1862                         }
1863                 }
1864         }
1865
1866         if(scsi_status_is_good(result)) {
1867                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
1868                              (modepage == 6 || modepage == 8))) {
1869                         /* Initio breakage? */
1870                         header_length = 0;
1871                         data->length = 13;
1872                         data->medium_type = 0;
1873                         data->device_specific = 0;
1874                         data->longlba = 0;
1875                         data->block_descriptor_length = 0;
1876                 } else if(use_10_for_ms) {
1877                         data->length = buffer[0]*256 + buffer[1] + 2;
1878                         data->medium_type = buffer[2];
1879                         data->device_specific = buffer[3];
1880                         data->longlba = buffer[4] & 0x01;
1881                         data->block_descriptor_length = buffer[6]*256
1882                                 + buffer[7];
1883                 } else {
1884                         data->length = buffer[0] + 1;
1885                         data->medium_type = buffer[1];
1886                         data->device_specific = buffer[2];
1887                         data->block_descriptor_length = buffer[3];
1888                 }
1889                 data->header_length = header_length;
1890         }
1891
1892         return result;
1893 }
1894 EXPORT_SYMBOL(scsi_mode_sense);
1895
1896 int
1897 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries)
1898 {
1899         char cmd[] = {
1900                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
1901         };
1902         struct scsi_sense_hdr sshdr;
1903         int result;
1904         
1905         result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, &sshdr,
1906                                   timeout, retries);
1907
1908         if ((driver_byte(result) & DRIVER_SENSE) && sdev->removable) {
1909
1910                 if ((scsi_sense_valid(&sshdr)) &&
1911                     ((sshdr.sense_key == UNIT_ATTENTION) ||
1912                      (sshdr.sense_key == NOT_READY))) {
1913                         sdev->changed = 1;
1914                         result = 0;
1915                 }
1916         }
1917         return result;
1918 }
1919 EXPORT_SYMBOL(scsi_test_unit_ready);
1920
1921 /**
1922  *      scsi_device_set_state - Take the given device through the device
1923  *              state model.
1924  *      @sdev:  scsi device to change the state of.
1925  *      @state: state to change to.
1926  *
1927  *      Returns zero if unsuccessful or an error if the requested 
1928  *      transition is illegal.
1929  **/
1930 int
1931 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
1932 {
1933         enum scsi_device_state oldstate = sdev->sdev_state;
1934
1935         if (state == oldstate)
1936                 return 0;
1937
1938         switch (state) {
1939         case SDEV_CREATED:
1940                 /* There are no legal states that come back to
1941                  * created.  This is the manually initialised start
1942                  * state */
1943                 goto illegal;
1944                         
1945         case SDEV_RUNNING:
1946                 switch (oldstate) {
1947                 case SDEV_CREATED:
1948                 case SDEV_OFFLINE:
1949                 case SDEV_QUIESCE:
1950                 case SDEV_BLOCK:
1951                         break;
1952                 default:
1953                         goto illegal;
1954                 }
1955                 break;
1956
1957         case SDEV_QUIESCE:
1958                 switch (oldstate) {
1959                 case SDEV_RUNNING:
1960                 case SDEV_OFFLINE:
1961                         break;
1962                 default:
1963                         goto illegal;
1964                 }
1965                 break;
1966
1967         case SDEV_OFFLINE:
1968                 switch (oldstate) {
1969                 case SDEV_CREATED:
1970                 case SDEV_RUNNING:
1971                 case SDEV_QUIESCE:
1972                 case SDEV_BLOCK:
1973                         break;
1974                 default:
1975                         goto illegal;
1976                 }
1977                 break;
1978
1979         case SDEV_BLOCK:
1980                 switch (oldstate) {
1981                 case SDEV_CREATED:
1982                 case SDEV_RUNNING:
1983                         break;
1984                 default:
1985                         goto illegal;
1986                 }
1987                 break;
1988
1989         case SDEV_CANCEL:
1990                 switch (oldstate) {
1991                 case SDEV_CREATED:
1992                 case SDEV_RUNNING:
1993                 case SDEV_QUIESCE:
1994                 case SDEV_OFFLINE:
1995                 case SDEV_BLOCK:
1996                         break;
1997                 default:
1998                         goto illegal;
1999                 }
2000                 break;
2001
2002         case SDEV_DEL:
2003                 switch (oldstate) {
2004                 case SDEV_CREATED:
2005                 case SDEV_RUNNING:
2006                 case SDEV_OFFLINE:
2007                 case SDEV_CANCEL:
2008                         break;
2009                 default:
2010                         goto illegal;
2011                 }
2012                 break;
2013
2014         }
2015         sdev->sdev_state = state;
2016         return 0;
2017
2018  illegal:
2019         SCSI_LOG_ERROR_RECOVERY(1, 
2020                                 sdev_printk(KERN_ERR, sdev,
2021                                             "Illegal state transition %s->%s\n",
2022                                             scsi_device_state_name(oldstate),
2023                                             scsi_device_state_name(state))
2024                                 );
2025         return -EINVAL;
2026 }
2027 EXPORT_SYMBOL(scsi_device_set_state);
2028
2029 /**
2030  *      scsi_device_quiesce - Block user issued commands.
2031  *      @sdev:  scsi device to quiesce.
2032  *
2033  *      This works by trying to transition to the SDEV_QUIESCE state
2034  *      (which must be a legal transition).  When the device is in this
2035  *      state, only special requests will be accepted, all others will
2036  *      be deferred.  Since special requests may also be requeued requests,
2037  *      a successful return doesn't guarantee the device will be 
2038  *      totally quiescent.
2039  *
2040  *      Must be called with user context, may sleep.
2041  *
2042  *      Returns zero if unsuccessful or an error if not.
2043  **/
2044 int
2045 scsi_device_quiesce(struct scsi_device *sdev)
2046 {
2047         int err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2048         if (err)
2049                 return err;
2050
2051         scsi_run_queue(sdev->request_queue);
2052         while (sdev->device_busy) {
2053                 msleep_interruptible(200);
2054                 scsi_run_queue(sdev->request_queue);
2055         }
2056         return 0;
2057 }
2058 EXPORT_SYMBOL(scsi_device_quiesce);
2059
2060 /**
2061  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2062  *      @sdev:  scsi device to resume.
2063  *
2064  *      Moves the device from quiesced back to running and restarts the
2065  *      queues.
2066  *
2067  *      Must be called with user context, may sleep.
2068  **/
2069 void
2070 scsi_device_resume(struct scsi_device *sdev)
2071 {
2072         if(scsi_device_set_state(sdev, SDEV_RUNNING))
2073                 return;
2074         scsi_run_queue(sdev->request_queue);
2075 }
2076 EXPORT_SYMBOL(scsi_device_resume);
2077
2078 static void
2079 device_quiesce_fn(struct scsi_device *sdev, void *data)
2080 {
2081         scsi_device_quiesce(sdev);
2082 }
2083
2084 void
2085 scsi_target_quiesce(struct scsi_target *starget)
2086 {
2087         starget_for_each_device(starget, NULL, device_quiesce_fn);
2088 }
2089 EXPORT_SYMBOL(scsi_target_quiesce);
2090
2091 static void
2092 device_resume_fn(struct scsi_device *sdev, void *data)
2093 {
2094         scsi_device_resume(sdev);
2095 }
2096
2097 void
2098 scsi_target_resume(struct scsi_target *starget)
2099 {
2100         starget_for_each_device(starget, NULL, device_resume_fn);
2101 }
2102 EXPORT_SYMBOL(scsi_target_resume);
2103
2104 /**
2105  * scsi_internal_device_block - internal function to put a device
2106  *                              temporarily into the SDEV_BLOCK state
2107  * @sdev:       device to block
2108  *
2109  * Block request made by scsi lld's to temporarily stop all
2110  * scsi commands on the specified device.  Called from interrupt
2111  * or normal process context.
2112  *
2113  * Returns zero if successful or error if not
2114  *
2115  * Notes:       
2116  *      This routine transitions the device to the SDEV_BLOCK state
2117  *      (which must be a legal transition).  When the device is in this
2118  *      state, all commands are deferred until the scsi lld reenables
2119  *      the device with scsi_device_unblock or device_block_tmo fires.
2120  *      This routine assumes the host_lock is held on entry.
2121  **/
2122 int
2123 scsi_internal_device_block(struct scsi_device *sdev)
2124 {
2125         request_queue_t *q = sdev->request_queue;
2126         unsigned long flags;
2127         int err = 0;
2128
2129         err = scsi_device_set_state(sdev, SDEV_BLOCK);
2130         if (err)
2131                 return err;
2132
2133         /* 
2134          * The device has transitioned to SDEV_BLOCK.  Stop the
2135          * block layer from calling the midlayer with this device's
2136          * request queue. 
2137          */
2138         spin_lock_irqsave(q->queue_lock, flags);
2139         blk_stop_queue(q);
2140         spin_unlock_irqrestore(q->queue_lock, flags);
2141
2142         return 0;
2143 }
2144 EXPORT_SYMBOL_GPL(scsi_internal_device_block);
2145  
2146 /**
2147  * scsi_internal_device_unblock - resume a device after a block request
2148  * @sdev:       device to resume
2149  *
2150  * Called by scsi lld's or the midlayer to restart the device queue
2151  * for the previously suspended scsi device.  Called from interrupt or
2152  * normal process context.
2153  *
2154  * Returns zero if successful or error if not.
2155  *
2156  * Notes:       
2157  *      This routine transitions the device to the SDEV_RUNNING state
2158  *      (which must be a legal transition) allowing the midlayer to
2159  *      goose the queue for this device.  This routine assumes the 
2160  *      host_lock is held upon entry.
2161  **/
2162 int
2163 scsi_internal_device_unblock(struct scsi_device *sdev)
2164 {
2165         request_queue_t *q = sdev->request_queue; 
2166         int err;
2167         unsigned long flags;
2168         
2169         /* 
2170          * Try to transition the scsi device to SDEV_RUNNING
2171          * and goose the device queue if successful.  
2172          */
2173         err = scsi_device_set_state(sdev, SDEV_RUNNING);
2174         if (err)
2175                 return err;
2176
2177         spin_lock_irqsave(q->queue_lock, flags);
2178         blk_start_queue(q);
2179         spin_unlock_irqrestore(q->queue_lock, flags);
2180
2181         return 0;
2182 }
2183 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock);
2184
2185 static void
2186 device_block(struct scsi_device *sdev, void *data)
2187 {
2188         scsi_internal_device_block(sdev);
2189 }
2190
2191 static int
2192 target_block(struct device *dev, void *data)
2193 {
2194         if (scsi_is_target_device(dev))
2195                 starget_for_each_device(to_scsi_target(dev), NULL,
2196                                         device_block);
2197         return 0;
2198 }
2199
2200 void
2201 scsi_target_block(struct device *dev)
2202 {
2203         if (scsi_is_target_device(dev))
2204                 starget_for_each_device(to_scsi_target(dev), NULL,
2205                                         device_block);
2206         else
2207                 device_for_each_child(dev, NULL, target_block);
2208 }
2209 EXPORT_SYMBOL_GPL(scsi_target_block);
2210
2211 static void
2212 device_unblock(struct scsi_device *sdev, void *data)
2213 {
2214         scsi_internal_device_unblock(sdev);
2215 }
2216
2217 static int
2218 target_unblock(struct device *dev, void *data)
2219 {
2220         if (scsi_is_target_device(dev))
2221                 starget_for_each_device(to_scsi_target(dev), NULL,
2222                                         device_unblock);
2223         return 0;
2224 }
2225
2226 void
2227 scsi_target_unblock(struct device *dev)
2228 {
2229         if (scsi_is_target_device(dev))
2230                 starget_for_each_device(to_scsi_target(dev), NULL,
2231                                         device_unblock);
2232         else
2233                 device_for_each_child(dev, NULL, target_unblock);
2234 }
2235 EXPORT_SYMBOL_GPL(scsi_target_unblock);
2236
2237 /**
2238  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2239  * @sg:         scatter-gather list
2240  * @sg_count:   number of segments in sg
2241  * @offset:     offset in bytes into sg, on return offset into the mapped area
2242  * @len:        bytes to map, on return number of bytes mapped
2243  *
2244  * Returns virtual address of the start of the mapped page
2245  */
2246 void *scsi_kmap_atomic_sg(struct scatterlist *sg, int sg_count,
2247                           size_t *offset, size_t *len)
2248 {
2249         int i;
2250         size_t sg_len = 0, len_complete = 0;
2251         struct page *page;
2252
2253         for (i = 0; i < sg_count; i++) {
2254                 len_complete = sg_len; /* Complete sg-entries */
2255                 sg_len += sg[i].length;
2256                 if (sg_len > *offset)
2257                         break;
2258         }
2259
2260         if (unlikely(i == sg_count)) {
2261                 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2262                         "elements %d\n",
2263                        __FUNCTION__, sg_len, *offset, sg_count);
2264                 WARN_ON(1);
2265                 return NULL;
2266         }
2267
2268         /* Offset starting from the beginning of first page in this sg-entry */
2269         *offset = *offset - len_complete + sg[i].offset;
2270
2271         /* Assumption: contiguous pages can be accessed as "page + i" */
2272         page = nth_page(sg[i].page, (*offset >> PAGE_SHIFT));
2273         *offset &= ~PAGE_MASK;
2274
2275         /* Bytes in this sg-entry from *offset to the end of the page */
2276         sg_len = PAGE_SIZE - *offset;
2277         if (*len > sg_len)
2278                 *len = sg_len;
2279
2280         return kmap_atomic(page, KM_BIO_SRC_IRQ);
2281 }
2282 EXPORT_SYMBOL(scsi_kmap_atomic_sg);
2283
2284 /**
2285  * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously
2286  *                         mapped with scsi_kmap_atomic_sg
2287  * @virt:       virtual address to be unmapped
2288  */
2289 void scsi_kunmap_atomic_sg(void *virt)
2290 {
2291         kunmap_atomic(virt, KM_BIO_SRC_IRQ);
2292 }
2293 EXPORT_SYMBOL(scsi_kunmap_atomic_sg);