fb616c69151f69a3d6697658de3bbc87827729b6
[linux-3.10.git] / drivers / scsi / scsi_lib.c
1 /*
2  *  scsi_lib.c Copyright (C) 1999 Eric Youngdale
3  *
4  *  SCSI queueing library.
5  *      Initial versions: Eric Youngdale (eric@andante.org).
6  *                        Based upon conversations with large numbers
7  *                        of people at Linux Expo.
8  */
9
10 #include <linux/bio.h>
11 #include <linux/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         kmem_cache_t    *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 kmem_cache_t *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 (!q->back_merge_fn(q, rq, bio))
269                 return -EINVAL;
270         else {
271                 rq->biotail->bi_next = bio;
272                 rq->biotail = bio;
273                 rq->hard_nr_sectors += bio_sectors(bio);
274                 rq->nr_sectors = rq->hard_nr_sectors;
275         }
276
277         return 0;
278 }
279
280 static int scsi_bi_endio(struct bio *bio, unsigned int bytes_done, int error)
281 {
282         if (bio->bi_size)
283                 return 1;
284
285         bio_put(bio);
286         return 0;
287 }
288
289 /**
290  * scsi_req_map_sg - map a scatterlist into a request
291  * @rq:         request to fill
292  * @sg:         scatterlist
293  * @nsegs:      number of elements
294  * @bufflen:    len of buffer
295  * @gfp:        memory allocation flags
296  *
297  * scsi_req_map_sg maps a scatterlist into a request so that the
298  * request can be sent to the block layer. We do not trust the scatterlist
299  * sent to use, as some ULDs use that struct to only organize the pages.
300  */
301 static int scsi_req_map_sg(struct request *rq, struct scatterlist *sgl,
302                            int nsegs, unsigned bufflen, gfp_t gfp)
303 {
304         struct request_queue *q = rq->q;
305         int nr_pages = (bufflen + sgl[0].offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
306         unsigned int data_len = 0, len, bytes, off;
307         struct page *page;
308         struct bio *bio = NULL;
309         int i, err, nr_vecs = 0;
310
311         for (i = 0; i < nsegs; i++) {
312                 page = sgl[i].page;
313                 off = sgl[i].offset;
314                 len = sgl[i].length;
315                 data_len += len;
316
317                 while (len > 0) {
318                         bytes = min_t(unsigned int, len, PAGE_SIZE - off);
319
320                         if (!bio) {
321                                 nr_vecs = min_t(int, BIO_MAX_PAGES, nr_pages);
322                                 nr_pages -= nr_vecs;
323
324                                 bio = bio_alloc(gfp, nr_vecs);
325                                 if (!bio) {
326                                         err = -ENOMEM;
327                                         goto free_bios;
328                                 }
329                                 bio->bi_end_io = scsi_bi_endio;
330                         }
331
332                         if (bio_add_pc_page(q, bio, page, bytes, off) !=
333                             bytes) {
334                                 bio_put(bio);
335                                 err = -EINVAL;
336                                 goto free_bios;
337                         }
338
339                         if (bio->bi_vcnt >= nr_vecs) {
340                                 err = scsi_merge_bio(rq, bio);
341                                 if (err) {
342                                         bio_endio(bio, bio->bi_size, 0);
343                                         goto free_bios;
344                                 }
345                                 bio = NULL;
346                         }
347
348                         page++;
349                         len -= bytes;
350                         off = 0;
351                 }
352         }
353
354         rq->buffer = rq->data = NULL;
355         rq->data_len = data_len;
356         return 0;
357
358 free_bios:
359         while ((bio = rq->bio) != NULL) {
360                 rq->bio = bio->bi_next;
361                 /*
362                  * call endio instead of bio_put incase it was bounced
363                  */
364                 bio_endio(bio, bio->bi_size, 0);
365         }
366
367         return err;
368 }
369
370 /**
371  * scsi_execute_async - insert request
372  * @sdev:       scsi device
373  * @cmd:        scsi command
374  * @cmd_len:    length of scsi cdb
375  * @data_direction: data direction
376  * @buffer:     data buffer (this can be a kernel buffer or scatterlist)
377  * @bufflen:    len of buffer
378  * @use_sg:     if buffer is a scatterlist this is the number of elements
379  * @timeout:    request timeout in seconds
380  * @retries:    number of times to retry request
381  * @flags:      or into request flags
382  **/
383 int scsi_execute_async(struct scsi_device *sdev, const unsigned char *cmd,
384                        int cmd_len, int data_direction, void *buffer, unsigned bufflen,
385                        int use_sg, int timeout, int retries, void *privdata,
386                        void (*done)(void *, char *, int, int), gfp_t gfp)
387 {
388         struct request *req;
389         struct scsi_io_context *sioc;
390         int err = 0;
391         int write = (data_direction == DMA_TO_DEVICE);
392
393         sioc = kmem_cache_alloc(scsi_io_context_cache, gfp);
394         if (!sioc)
395                 return DRIVER_ERROR << 24;
396         memset(sioc, 0, sizeof(*sioc));
397
398         req = blk_get_request(sdev->request_queue, write, gfp);
399         if (!req)
400                 goto free_sense;
401         req->cmd_type = REQ_TYPE_BLOCK_PC;
402         req->cmd_flags |= REQ_QUIET;
403
404         if (use_sg)
405                 err = scsi_req_map_sg(req, buffer, use_sg, bufflen, gfp);
406         else if (bufflen)
407                 err = blk_rq_map_kern(req->q, req, buffer, bufflen, gfp);
408
409         if (err)
410                 goto free_req;
411
412         req->cmd_len = cmd_len;
413         memset(req->cmd, 0, BLK_MAX_CDB); /* ATAPI hates garbage after CDB */
414         memcpy(req->cmd, cmd, req->cmd_len);
415         req->sense = sioc->sense;
416         req->sense_len = 0;
417         req->timeout = timeout;
418         req->retries = retries;
419         req->end_io_data = sioc;
420
421         sioc->data = privdata;
422         sioc->done = done;
423
424         blk_execute_rq_nowait(req->q, NULL, req, 1, scsi_end_async);
425         return 0;
426
427 free_req:
428         blk_put_request(req);
429 free_sense:
430         kmem_cache_free(scsi_io_context_cache, sioc);
431         return DRIVER_ERROR << 24;
432 }
433 EXPORT_SYMBOL_GPL(scsi_execute_async);
434
435 /*
436  * Function:    scsi_init_cmd_errh()
437  *
438  * Purpose:     Initialize cmd fields related to error handling.
439  *
440  * Arguments:   cmd     - command that is ready to be queued.
441  *
442  * Notes:       This function has the job of initializing a number of
443  *              fields related to error handling.   Typically this will
444  *              be called once for each command, as required.
445  */
446 static void scsi_init_cmd_errh(struct scsi_cmnd *cmd)
447 {
448         cmd->serial_number = 0;
449         memset(cmd->sense_buffer, 0, sizeof cmd->sense_buffer);
450         if (cmd->cmd_len == 0)
451                 cmd->cmd_len = COMMAND_SIZE(cmd->cmnd[0]);
452 }
453
454 void scsi_device_unbusy(struct scsi_device *sdev)
455 {
456         struct Scsi_Host *shost = sdev->host;
457         unsigned long flags;
458
459         spin_lock_irqsave(shost->host_lock, flags);
460         shost->host_busy--;
461         if (unlikely(scsi_host_in_recovery(shost) &&
462                      (shost->host_failed || shost->host_eh_scheduled)))
463                 scsi_eh_wakeup(shost);
464         spin_unlock(shost->host_lock);
465         spin_lock(sdev->request_queue->queue_lock);
466         sdev->device_busy--;
467         spin_unlock_irqrestore(sdev->request_queue->queue_lock, flags);
468 }
469
470 /*
471  * Called for single_lun devices on IO completion. Clear starget_sdev_user,
472  * and call blk_run_queue for all the scsi_devices on the target -
473  * including current_sdev first.
474  *
475  * Called with *no* scsi locks held.
476  */
477 static void scsi_single_lun_run(struct scsi_device *current_sdev)
478 {
479         struct Scsi_Host *shost = current_sdev->host;
480         struct scsi_device *sdev, *tmp;
481         struct scsi_target *starget = scsi_target(current_sdev);
482         unsigned long flags;
483
484         spin_lock_irqsave(shost->host_lock, flags);
485         starget->starget_sdev_user = NULL;
486         spin_unlock_irqrestore(shost->host_lock, flags);
487
488         /*
489          * Call blk_run_queue for all LUNs on the target, starting with
490          * current_sdev. We race with others (to set starget_sdev_user),
491          * but in most cases, we will be first. Ideally, each LU on the
492          * target would get some limited time or requests on the target.
493          */
494         blk_run_queue(current_sdev->request_queue);
495
496         spin_lock_irqsave(shost->host_lock, flags);
497         if (starget->starget_sdev_user)
498                 goto out;
499         list_for_each_entry_safe(sdev, tmp, &starget->devices,
500                         same_target_siblings) {
501                 if (sdev == current_sdev)
502                         continue;
503                 if (scsi_device_get(sdev))
504                         continue;
505
506                 spin_unlock_irqrestore(shost->host_lock, flags);
507                 blk_run_queue(sdev->request_queue);
508                 spin_lock_irqsave(shost->host_lock, flags);
509         
510                 scsi_device_put(sdev);
511         }
512  out:
513         spin_unlock_irqrestore(shost->host_lock, flags);
514 }
515
516 /*
517  * Function:    scsi_run_queue()
518  *
519  * Purpose:     Select a proper request queue to serve next
520  *
521  * Arguments:   q       - last request's queue
522  *
523  * Returns:     Nothing
524  *
525  * Notes:       The previous command was completely finished, start
526  *              a new one if possible.
527  */
528 static void scsi_run_queue(struct request_queue *q)
529 {
530         struct scsi_device *sdev = q->queuedata;
531         struct Scsi_Host *shost = sdev->host;
532         unsigned long flags;
533
534         if (sdev->single_lun)
535                 scsi_single_lun_run(sdev);
536
537         spin_lock_irqsave(shost->host_lock, flags);
538         while (!list_empty(&shost->starved_list) &&
539                !shost->host_blocked && !shost->host_self_blocked &&
540                 !((shost->can_queue > 0) &&
541                   (shost->host_busy >= shost->can_queue))) {
542                 /*
543                  * As long as shost is accepting commands and we have
544                  * starved queues, call blk_run_queue. scsi_request_fn
545                  * drops the queue_lock and can add us back to the
546                  * starved_list.
547                  *
548                  * host_lock protects the starved_list and starved_entry.
549                  * scsi_request_fn must get the host_lock before checking
550                  * or modifying starved_list or starved_entry.
551                  */
552                 sdev = list_entry(shost->starved_list.next,
553                                           struct scsi_device, starved_entry);
554                 list_del_init(&sdev->starved_entry);
555                 spin_unlock_irqrestore(shost->host_lock, flags);
556
557
558                 if (test_bit(QUEUE_FLAG_REENTER, &q->queue_flags) &&
559                     !test_and_set_bit(QUEUE_FLAG_REENTER,
560                                       &sdev->request_queue->queue_flags)) {
561                         blk_run_queue(sdev->request_queue);
562                         clear_bit(QUEUE_FLAG_REENTER,
563                                   &sdev->request_queue->queue_flags);
564                 } else
565                         blk_run_queue(sdev->request_queue);
566
567                 spin_lock_irqsave(shost->host_lock, flags);
568                 if (unlikely(!list_empty(&sdev->starved_entry)))
569                         /*
570                          * sdev lost a race, and was put back on the
571                          * starved list. This is unlikely but without this
572                          * in theory we could loop forever.
573                          */
574                         break;
575         }
576         spin_unlock_irqrestore(shost->host_lock, flags);
577
578         blk_run_queue(q);
579 }
580
581 /*
582  * Function:    scsi_requeue_command()
583  *
584  * Purpose:     Handle post-processing of completed commands.
585  *
586  * Arguments:   q       - queue to operate on
587  *              cmd     - command that may need to be requeued.
588  *
589  * Returns:     Nothing
590  *
591  * Notes:       After command completion, there may be blocks left
592  *              over which weren't finished by the previous command
593  *              this can be for a number of reasons - the main one is
594  *              I/O errors in the middle of the request, in which case
595  *              we need to request the blocks that come after the bad
596  *              sector.
597  * Notes:       Upon return, cmd is a stale pointer.
598  */
599 static void scsi_requeue_command(struct request_queue *q, struct scsi_cmnd *cmd)
600 {
601         struct request *req = cmd->request;
602         unsigned long flags;
603
604         scsi_unprep_request(req);
605         spin_lock_irqsave(q->queue_lock, flags);
606         blk_requeue_request(q, req);
607         spin_unlock_irqrestore(q->queue_lock, flags);
608
609         scsi_run_queue(q);
610 }
611
612 void scsi_next_command(struct scsi_cmnd *cmd)
613 {
614         struct scsi_device *sdev = cmd->device;
615         struct request_queue *q = sdev->request_queue;
616
617         /* need to hold a reference on the device before we let go of the cmd */
618         get_device(&sdev->sdev_gendev);
619
620         scsi_put_command(cmd);
621         scsi_run_queue(q);
622
623         /* ok to remove device now */
624         put_device(&sdev->sdev_gendev);
625 }
626
627 void scsi_run_host_queues(struct Scsi_Host *shost)
628 {
629         struct scsi_device *sdev;
630
631         shost_for_each_device(sdev, shost)
632                 scsi_run_queue(sdev->request_queue);
633 }
634
635 /*
636  * Function:    scsi_end_request()
637  *
638  * Purpose:     Post-processing of completed commands (usually invoked at end
639  *              of upper level post-processing and scsi_io_completion).
640  *
641  * Arguments:   cmd      - command that is complete.
642  *              uptodate - 1 if I/O indicates success, <= 0 for I/O error.
643  *              bytes    - number of bytes of completed I/O
644  *              requeue  - indicates whether we should requeue leftovers.
645  *
646  * Lock status: Assumed that lock is not held upon entry.
647  *
648  * Returns:     cmd if requeue required, NULL otherwise.
649  *
650  * Notes:       This is called for block device requests in order to
651  *              mark some number of sectors as complete.
652  * 
653  *              We are guaranteeing that the request queue will be goosed
654  *              at some point during this call.
655  * Notes:       If cmd was requeued, upon return it will be a stale pointer.
656  */
657 static struct scsi_cmnd *scsi_end_request(struct scsi_cmnd *cmd, int uptodate,
658                                           int bytes, int requeue)
659 {
660         request_queue_t *q = cmd->device->request_queue;
661         struct request *req = cmd->request;
662         unsigned long flags;
663
664         /*
665          * If there are blocks left over at the end, set up the command
666          * to queue the remainder of them.
667          */
668         if (end_that_request_chunk(req, uptodate, bytes)) {
669                 int leftover = (req->hard_nr_sectors << 9);
670
671                 if (blk_pc_request(req))
672                         leftover = req->data_len;
673
674                 /* kill remainder if no retrys */
675                 if (!uptodate && blk_noretry_request(req))
676                         end_that_request_chunk(req, 0, leftover);
677                 else {
678                         if (requeue) {
679                                 /*
680                                  * Bleah.  Leftovers again.  Stick the
681                                  * leftovers in the front of the
682                                  * queue, and goose the queue again.
683                                  */
684                                 scsi_requeue_command(q, cmd);
685                                 cmd = NULL;
686                         }
687                         return cmd;
688                 }
689         }
690
691         add_disk_randomness(req->rq_disk);
692
693         spin_lock_irqsave(q->queue_lock, flags);
694         if (blk_rq_tagged(req))
695                 blk_queue_end_tag(q, req);
696         end_that_request_last(req, uptodate);
697         spin_unlock_irqrestore(q->queue_lock, flags);
698
699         /*
700          * This will goose the queue request function at the end, so we don't
701          * need to worry about launching another command.
702          */
703         scsi_next_command(cmd);
704         return NULL;
705 }
706
707 struct scatterlist *scsi_alloc_sgtable(struct scsi_cmnd *cmd, gfp_t gfp_mask)
708 {
709         struct scsi_host_sg_pool *sgp;
710         struct scatterlist *sgl;
711
712         BUG_ON(!cmd->use_sg);
713
714         switch (cmd->use_sg) {
715         case 1 ... 8:
716                 cmd->sglist_len = 0;
717                 break;
718         case 9 ... 16:
719                 cmd->sglist_len = 1;
720                 break;
721         case 17 ... 32:
722                 cmd->sglist_len = 2;
723                 break;
724 #if (SCSI_MAX_PHYS_SEGMENTS > 32)
725         case 33 ... 64:
726                 cmd->sglist_len = 3;
727                 break;
728 #if (SCSI_MAX_PHYS_SEGMENTS > 64)
729         case 65 ... 128:
730                 cmd->sglist_len = 4;
731                 break;
732 #if (SCSI_MAX_PHYS_SEGMENTS  > 128)
733         case 129 ... 256:
734                 cmd->sglist_len = 5;
735                 break;
736 #endif
737 #endif
738 #endif
739         default:
740                 return NULL;
741         }
742
743         sgp = scsi_sg_pools + cmd->sglist_len;
744         sgl = mempool_alloc(sgp->pool, gfp_mask);
745         return sgl;
746 }
747
748 EXPORT_SYMBOL(scsi_alloc_sgtable);
749
750 void scsi_free_sgtable(struct scatterlist *sgl, int index)
751 {
752         struct scsi_host_sg_pool *sgp;
753
754         BUG_ON(index >= SG_MEMPOOL_NR);
755
756         sgp = scsi_sg_pools + index;
757         mempool_free(sgl, sgp->pool);
758 }
759
760 EXPORT_SYMBOL(scsi_free_sgtable);
761
762 /*
763  * Function:    scsi_release_buffers()
764  *
765  * Purpose:     Completion processing for block device I/O requests.
766  *
767  * Arguments:   cmd     - command that we are bailing.
768  *
769  * Lock status: Assumed that no lock is held upon entry.
770  *
771  * Returns:     Nothing
772  *
773  * Notes:       In the event that an upper level driver rejects a
774  *              command, we must release resources allocated during
775  *              the __init_io() function.  Primarily this would involve
776  *              the scatter-gather table, and potentially any bounce
777  *              buffers.
778  */
779 static void scsi_release_buffers(struct scsi_cmnd *cmd)
780 {
781         if (cmd->use_sg)
782                 scsi_free_sgtable(cmd->request_buffer, cmd->sglist_len);
783
784         /*
785          * Zero these out.  They now point to freed memory, and it is
786          * dangerous to hang onto the pointers.
787          */
788         cmd->request_buffer = NULL;
789         cmd->request_bufflen = 0;
790 }
791
792 /*
793  * Function:    scsi_io_completion()
794  *
795  * Purpose:     Completion processing for block device I/O requests.
796  *
797  * Arguments:   cmd   - command that is finished.
798  *
799  * Lock status: Assumed that no lock is held upon entry.
800  *
801  * Returns:     Nothing
802  *
803  * Notes:       This function is matched in terms of capabilities to
804  *              the function that created the scatter-gather list.
805  *              In other words, if there are no bounce buffers
806  *              (the normal case for most drivers), we don't need
807  *              the logic to deal with cleaning up afterwards.
808  *
809  *              We must do one of several things here:
810  *
811  *              a) Call scsi_end_request.  This will finish off the
812  *                 specified number of sectors.  If we are done, the
813  *                 command block will be released, and the queue
814  *                 function will be goosed.  If we are not done, then
815  *                 scsi_end_request will directly goose the queue.
816  *
817  *              b) We can just use scsi_requeue_command() here.  This would
818  *                 be used if we just wanted to retry, for example.
819  */
820 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
821 {
822         int result = cmd->result;
823         int this_count = cmd->request_bufflen;
824         request_queue_t *q = cmd->device->request_queue;
825         struct request *req = cmd->request;
826         int clear_errors = 1;
827         struct scsi_sense_hdr sshdr;
828         int sense_valid = 0;
829         int sense_deferred = 0;
830
831         scsi_release_buffers(cmd);
832
833         if (result) {
834                 sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
835                 if (sense_valid)
836                         sense_deferred = scsi_sense_is_deferred(&sshdr);
837         }
838
839         if (blk_pc_request(req)) { /* SG_IO ioctl from block level */
840                 req->errors = result;
841                 if (result) {
842                         clear_errors = 0;
843                         if (sense_valid && req->sense) {
844                                 /*
845                                  * SG_IO wants current and deferred errors
846                                  */
847                                 int len = 8 + cmd->sense_buffer[7];
848
849                                 if (len > SCSI_SENSE_BUFFERSIZE)
850                                         len = SCSI_SENSE_BUFFERSIZE;
851                                 memcpy(req->sense, cmd->sense_buffer,  len);
852                                 req->sense_len = len;
853                         }
854                 } else
855                         req->data_len = cmd->resid;
856         }
857
858         /*
859          * Next deal with any sectors which we were able to correctly
860          * handle.
861          */
862         SCSI_LOG_HLCOMPLETE(1, printk("%ld sectors total, "
863                                       "%d bytes done.\n",
864                                       req->nr_sectors, good_bytes));
865         SCSI_LOG_HLCOMPLETE(1, printk("use_sg is %d\n", cmd->use_sg));
866
867         if (clear_errors)
868                 req->errors = 0;
869
870         /* A number of bytes were successfully read.  If there
871          * are leftovers and there is some kind of error
872          * (result != 0), retry the rest.
873          */
874         if (scsi_end_request(cmd, 1, good_bytes, result == 0) == NULL)
875                 return;
876
877         /* good_bytes = 0, or (inclusive) there were leftovers and
878          * result = 0, so scsi_end_request couldn't retry.
879          */
880         if (sense_valid && !sense_deferred) {
881                 switch (sshdr.sense_key) {
882                 case UNIT_ATTENTION:
883                         if (cmd->device->removable) {
884                                 /* Detected disc change.  Set a bit
885                                  * and quietly refuse further access.
886                                  */
887                                 cmd->device->changed = 1;
888                                 scsi_end_request(cmd, 0, this_count, 1);
889                                 return;
890                         } else {
891                                 /* Must have been a power glitch, or a
892                                  * bus reset.  Could not have been a
893                                  * media change, so we just retry the
894                                  * request and see what happens.
895                                  */
896                                 scsi_requeue_command(q, cmd);
897                                 return;
898                         }
899                         break;
900                 case ILLEGAL_REQUEST:
901                         /* If we had an ILLEGAL REQUEST returned, then
902                          * we may have performed an unsupported
903                          * command.  The only thing this should be
904                          * would be a ten byte read where only a six
905                          * byte read was supported.  Also, on a system
906                          * where READ CAPACITY failed, we may have
907                          * read past the end of the disk.
908                          */
909                         if ((cmd->device->use_10_for_rw &&
910                             sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
911                             (cmd->cmnd[0] == READ_10 ||
912                              cmd->cmnd[0] == WRITE_10)) {
913                                 cmd->device->use_10_for_rw = 0;
914                                 /* This will cause a retry with a
915                                  * 6-byte command.
916                                  */
917                                 scsi_requeue_command(q, cmd);
918                                 return;
919                         } else {
920                                 scsi_end_request(cmd, 0, this_count, 1);
921                                 return;
922                         }
923                         break;
924                 case NOT_READY:
925                         /* If the device is in the process of becoming
926                          * ready, or has a temporary blockage, retry.
927                          */
928                         if (sshdr.asc == 0x04) {
929                                 switch (sshdr.ascq) {
930                                 case 0x01: /* becoming ready */
931                                 case 0x04: /* format in progress */
932                                 case 0x05: /* rebuild in progress */
933                                 case 0x06: /* recalculation in progress */
934                                 case 0x07: /* operation in progress */
935                                 case 0x08: /* Long write in progress */
936                                 case 0x09: /* self test in progress */
937                                         scsi_requeue_command(q, cmd);
938                                         return;
939                                 default:
940                                         break;
941                                 }
942                         }
943                         if (!(req->cmd_flags & REQ_QUIET)) {
944                                 scmd_printk(KERN_INFO, cmd,
945                                             "Device not ready: ");
946                                 scsi_print_sense_hdr("", &sshdr);
947                         }
948                         scsi_end_request(cmd, 0, this_count, 1);
949                         return;
950                 case VOLUME_OVERFLOW:
951                         if (!(req->cmd_flags & REQ_QUIET)) {
952                                 scmd_printk(KERN_INFO, cmd,
953                                             "Volume overflow, CDB: ");
954                                 __scsi_print_command(cmd->cmnd);
955                                 scsi_print_sense("", cmd);
956                         }
957                         /* See SSC3rXX or current. */
958                         scsi_end_request(cmd, 0, this_count, 1);
959                         return;
960                 default:
961                         break;
962                 }
963         }
964         if (host_byte(result) == DID_RESET) {
965                 /* Third party bus reset or reset for error recovery
966                  * reasons.  Just retry the request and see what
967                  * happens.
968                  */
969                 scsi_requeue_command(q, cmd);
970                 return;
971         }
972         if (result) {
973                 if (!(req->cmd_flags & REQ_QUIET)) {
974                         scmd_printk(KERN_INFO, cmd,
975                                     "SCSI error: return code = 0x%08x\n",
976                                     result);
977                         if (driver_byte(result) & DRIVER_SENSE)
978                                 scsi_print_sense("", cmd);
979                 }
980         }
981         scsi_end_request(cmd, 0, this_count, !result);
982 }
983 EXPORT_SYMBOL(scsi_io_completion);
984
985 /*
986  * Function:    scsi_init_io()
987  *
988  * Purpose:     SCSI I/O initialize function.
989  *
990  * Arguments:   cmd   - Command descriptor we wish to initialize
991  *
992  * Returns:     0 on success
993  *              BLKPREP_DEFER if the failure is retryable
994  *              BLKPREP_KILL if the failure is fatal
995  */
996 static int scsi_init_io(struct scsi_cmnd *cmd)
997 {
998         struct request     *req = cmd->request;
999         struct scatterlist *sgpnt;
1000         int                count;
1001
1002         /*
1003          * We used to not use scatter-gather for single segment request,
1004          * but now we do (it makes highmem I/O easier to support without
1005          * kmapping pages)
1006          */
1007         cmd->use_sg = req->nr_phys_segments;
1008
1009         /*
1010          * If sg table allocation fails, requeue request later.
1011          */
1012         sgpnt = scsi_alloc_sgtable(cmd, GFP_ATOMIC);
1013         if (unlikely(!sgpnt)) {
1014                 scsi_unprep_request(req);
1015                 return BLKPREP_DEFER;
1016         }
1017
1018         req->buffer = NULL;
1019         cmd->request_buffer = (char *) sgpnt;
1020         if (blk_pc_request(req))
1021                 cmd->request_bufflen = req->data_len;
1022         else
1023                 cmd->request_bufflen = req->nr_sectors << 9;
1024
1025         /* 
1026          * Next, walk the list, and fill in the addresses and sizes of
1027          * each segment.
1028          */
1029         count = blk_rq_map_sg(req->q, req, cmd->request_buffer);
1030         if (likely(count <= cmd->use_sg)) {
1031                 cmd->use_sg = count;
1032                 return BLKPREP_OK;
1033         }
1034
1035         printk(KERN_ERR "Incorrect number of segments after building list\n");
1036         printk(KERN_ERR "counted %d, received %d\n", count, cmd->use_sg);
1037         printk(KERN_ERR "req nr_sec %lu, cur_nr_sec %u\n", req->nr_sectors,
1038                         req->current_nr_sectors);
1039
1040         /* release the command and kill it */
1041         scsi_release_buffers(cmd);
1042         scsi_put_command(cmd);
1043         return BLKPREP_KILL;
1044 }
1045
1046 static int scsi_issue_flush_fn(request_queue_t *q, struct gendisk *disk,
1047                                sector_t *error_sector)
1048 {
1049         struct scsi_device *sdev = q->queuedata;
1050         struct scsi_driver *drv;
1051
1052         if (sdev->sdev_state != SDEV_RUNNING)
1053                 return -ENXIO;
1054
1055         drv = *(struct scsi_driver **) disk->private_data;
1056         if (drv->issue_flush)
1057                 return drv->issue_flush(&sdev->sdev_gendev, error_sector);
1058
1059         return -EOPNOTSUPP;
1060 }
1061
1062 static struct scsi_cmnd *scsi_get_cmd_from_req(struct scsi_device *sdev,
1063                 struct request *req)
1064 {
1065         struct scsi_cmnd *cmd;
1066
1067         if (!req->special) {
1068                 cmd = scsi_get_command(sdev, GFP_ATOMIC);
1069                 if (unlikely(!cmd))
1070                         return NULL;
1071                 req->special = cmd;
1072         } else {
1073                 cmd = req->special;
1074         }
1075
1076         /* pull a tag out of the request if we have one */
1077         cmd->tag = req->tag;
1078         cmd->request = req;
1079
1080         return cmd;
1081 }
1082
1083 static void scsi_blk_pc_done(struct scsi_cmnd *cmd)
1084 {
1085         BUG_ON(!blk_pc_request(cmd->request));
1086         /*
1087          * This will complete the whole command with uptodate=1 so
1088          * as far as the block layer is concerned the command completed
1089          * successfully. Since this is a REQ_BLOCK_PC command the
1090          * caller should check the request's errors value
1091          */
1092         scsi_io_completion(cmd, cmd->request_bufflen);
1093 }
1094
1095 static int scsi_setup_blk_pc_cmnd(struct scsi_device *sdev, struct request *req)
1096 {
1097         struct scsi_cmnd *cmd;
1098
1099         cmd = scsi_get_cmd_from_req(sdev, req);
1100         if (unlikely(!cmd))
1101                 return BLKPREP_DEFER;
1102
1103         /*
1104          * BLOCK_PC requests may transfer data, in which case they must
1105          * a bio attached to them.  Or they might contain a SCSI command
1106          * that does not transfer data, in which case they may optionally
1107          * submit a request without an attached bio.
1108          */
1109         if (req->bio) {
1110                 int ret;
1111
1112                 BUG_ON(!req->nr_phys_segments);
1113
1114                 ret = scsi_init_io(cmd);
1115                 if (unlikely(ret))
1116                         return ret;
1117         } else {
1118                 BUG_ON(req->data_len);
1119                 BUG_ON(req->data);
1120
1121                 cmd->request_bufflen = 0;
1122                 cmd->request_buffer = NULL;
1123                 cmd->use_sg = 0;
1124                 req->buffer = NULL;
1125         }
1126
1127         BUILD_BUG_ON(sizeof(req->cmd) > sizeof(cmd->cmnd));
1128         memcpy(cmd->cmnd, req->cmd, sizeof(cmd->cmnd));
1129         cmd->cmd_len = req->cmd_len;
1130         if (!req->data_len)
1131                 cmd->sc_data_direction = DMA_NONE;
1132         else if (rq_data_dir(req) == WRITE)
1133                 cmd->sc_data_direction = DMA_TO_DEVICE;
1134         else
1135                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1136         
1137         cmd->transfersize = req->data_len;
1138         cmd->allowed = req->retries;
1139         cmd->timeout_per_command = req->timeout;
1140         cmd->done = scsi_blk_pc_done;
1141         return BLKPREP_OK;
1142 }
1143
1144 /*
1145  * Setup a REQ_TYPE_FS command.  These are simple read/write request
1146  * from filesystems that still need to be translated to SCSI CDBs from
1147  * the ULD.
1148  */
1149 static int scsi_setup_fs_cmnd(struct scsi_device *sdev, struct request *req)
1150 {
1151         struct scsi_cmnd *cmd;
1152         struct scsi_driver *drv;
1153         int ret;
1154
1155         /*
1156          * Filesystem requests must transfer data.
1157          */
1158         BUG_ON(!req->nr_phys_segments);
1159
1160         cmd = scsi_get_cmd_from_req(sdev, req);
1161         if (unlikely(!cmd))
1162                 return BLKPREP_DEFER;
1163
1164         ret = scsi_init_io(cmd);
1165         if (unlikely(ret))
1166                 return ret;
1167
1168         /*
1169          * Initialize the actual SCSI command for this request.
1170          */
1171         drv = *(struct scsi_driver **)req->rq_disk->private_data;
1172         if (unlikely(!drv->init_command(cmd))) {
1173                 scsi_release_buffers(cmd);
1174                 scsi_put_command(cmd);
1175                 return BLKPREP_KILL;
1176         }
1177
1178         return BLKPREP_OK;
1179 }
1180
1181 static int scsi_prep_fn(struct request_queue *q, struct request *req)
1182 {
1183         struct scsi_device *sdev = q->queuedata;
1184         int ret = BLKPREP_OK;
1185
1186         /*
1187          * If the device is not in running state we will reject some
1188          * or all commands.
1189          */
1190         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1191                 switch (sdev->sdev_state) {
1192                 case SDEV_OFFLINE:
1193                         /*
1194                          * If the device is offline we refuse to process any
1195                          * commands.  The device must be brought online
1196                          * before trying any recovery commands.
1197                          */
1198                         sdev_printk(KERN_ERR, sdev,
1199                                     "rejecting I/O to offline device\n");
1200                         ret = BLKPREP_KILL;
1201                         break;
1202                 case SDEV_DEL:
1203                         /*
1204                          * If the device is fully deleted, we refuse to
1205                          * process any commands as well.
1206                          */
1207                         sdev_printk(KERN_ERR, sdev,
1208                                     "rejecting I/O to dead device\n");
1209                         ret = BLKPREP_KILL;
1210                         break;
1211                 case SDEV_QUIESCE:
1212                 case SDEV_BLOCK:
1213                         /*
1214                          * If the devices is blocked we defer normal commands.
1215                          */
1216                         if (!(req->cmd_flags & REQ_PREEMPT))
1217                                 ret = BLKPREP_DEFER;
1218                         break;
1219                 default:
1220                         /*
1221                          * For any other not fully online state we only allow
1222                          * special commands.  In particular any user initiated
1223                          * command is not allowed.
1224                          */
1225                         if (!(req->cmd_flags & REQ_PREEMPT))
1226                                 ret = BLKPREP_KILL;
1227                         break;
1228                 }
1229
1230                 if (ret != BLKPREP_OK)
1231                         goto out;
1232         }
1233
1234         switch (req->cmd_type) {
1235         case REQ_TYPE_BLOCK_PC:
1236                 ret = scsi_setup_blk_pc_cmnd(sdev, req);
1237                 break;
1238         case REQ_TYPE_FS:
1239                 ret = scsi_setup_fs_cmnd(sdev, req);
1240                 break;
1241         default:
1242                 /*
1243                  * All other command types are not supported.
1244                  *
1245                  * Note that these days the SCSI subsystem does not use
1246                  * REQ_TYPE_SPECIAL requests anymore.  These are only used
1247                  * (directly or via blk_insert_request) by non-SCSI drivers.
1248                  */
1249                 blk_dump_rq_flags(req, "SCSI bad req");
1250                 ret = BLKPREP_KILL;
1251                 break;
1252         }
1253
1254  out:
1255         switch (ret) {
1256         case BLKPREP_KILL:
1257                 req->errors = DID_NO_CONNECT << 16;
1258                 break;
1259         case BLKPREP_DEFER:
1260                 /*
1261                  * If we defer, the elv_next_request() returns NULL, but the
1262                  * queue must be restarted, so we plug here if no returning
1263                  * command will automatically do that.
1264                  */
1265                 if (sdev->device_busy == 0)
1266                         blk_plug_device(q);
1267                 break;
1268         default:
1269                 req->cmd_flags |= REQ_DONTPREP;
1270         }
1271
1272         return ret;
1273 }
1274
1275 /*
1276  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1277  * return 0.
1278  *
1279  * Called with the queue_lock held.
1280  */
1281 static inline int scsi_dev_queue_ready(struct request_queue *q,
1282                                   struct scsi_device *sdev)
1283 {
1284         if (sdev->device_busy >= sdev->queue_depth)
1285                 return 0;
1286         if (sdev->device_busy == 0 && sdev->device_blocked) {
1287                 /*
1288                  * unblock after device_blocked iterates to zero
1289                  */
1290                 if (--sdev->device_blocked == 0) {
1291                         SCSI_LOG_MLQUEUE(3,
1292                                    sdev_printk(KERN_INFO, sdev,
1293                                    "unblocking device at zero depth\n"));
1294                 } else {
1295                         blk_plug_device(q);
1296                         return 0;
1297                 }
1298         }
1299         if (sdev->device_blocked)
1300                 return 0;
1301
1302         return 1;
1303 }
1304
1305 /*
1306  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1307  * return 0. We must end up running the queue again whenever 0 is
1308  * returned, else IO can hang.
1309  *
1310  * Called with host_lock held.
1311  */
1312 static inline int scsi_host_queue_ready(struct request_queue *q,
1313                                    struct Scsi_Host *shost,
1314                                    struct scsi_device *sdev)
1315 {
1316         if (scsi_host_in_recovery(shost))
1317                 return 0;
1318         if (shost->host_busy == 0 && shost->host_blocked) {
1319                 /*
1320                  * unblock after host_blocked iterates to zero
1321                  */
1322                 if (--shost->host_blocked == 0) {
1323                         SCSI_LOG_MLQUEUE(3,
1324                                 printk("scsi%d unblocking host at zero depth\n",
1325                                         shost->host_no));
1326                 } else {
1327                         blk_plug_device(q);
1328                         return 0;
1329                 }
1330         }
1331         if ((shost->can_queue > 0 && shost->host_busy >= shost->can_queue) ||
1332             shost->host_blocked || shost->host_self_blocked) {
1333                 if (list_empty(&sdev->starved_entry))
1334                         list_add_tail(&sdev->starved_entry, &shost->starved_list);
1335                 return 0;
1336         }
1337
1338         /* We're OK to process the command, so we can't be starved */
1339         if (!list_empty(&sdev->starved_entry))
1340                 list_del_init(&sdev->starved_entry);
1341
1342         return 1;
1343 }
1344
1345 /*
1346  * Kill a request for a dead device
1347  */
1348 static void scsi_kill_request(struct request *req, request_queue_t *q)
1349 {
1350         struct scsi_cmnd *cmd = req->special;
1351         struct scsi_device *sdev = cmd->device;
1352         struct Scsi_Host *shost = sdev->host;
1353
1354         blkdev_dequeue_request(req);
1355
1356         if (unlikely(cmd == NULL)) {
1357                 printk(KERN_CRIT "impossible request in %s.\n",
1358                                  __FUNCTION__);
1359                 BUG();
1360         }
1361
1362         scsi_init_cmd_errh(cmd);
1363         cmd->result = DID_NO_CONNECT << 16;
1364         atomic_inc(&cmd->device->iorequest_cnt);
1365
1366         /*
1367          * SCSI request completion path will do scsi_device_unbusy(),
1368          * bump busy counts.  To bump the counters, we need to dance
1369          * with the locks as normal issue path does.
1370          */
1371         sdev->device_busy++;
1372         spin_unlock(sdev->request_queue->queue_lock);
1373         spin_lock(shost->host_lock);
1374         shost->host_busy++;
1375         spin_unlock(shost->host_lock);
1376         spin_lock(sdev->request_queue->queue_lock);
1377
1378         __scsi_done(cmd);
1379 }
1380
1381 static void scsi_softirq_done(struct request *rq)
1382 {
1383         struct scsi_cmnd *cmd = rq->completion_data;
1384         unsigned long wait_for = (cmd->allowed + 1) * cmd->timeout_per_command;
1385         int disposition;
1386
1387         INIT_LIST_HEAD(&cmd->eh_entry);
1388
1389         disposition = scsi_decide_disposition(cmd);
1390         if (disposition != SUCCESS &&
1391             time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1392                 sdev_printk(KERN_ERR, cmd->device,
1393                             "timing out command, waited %lus\n",
1394                             wait_for/HZ);
1395                 disposition = SUCCESS;
1396         }
1397                         
1398         scsi_log_completion(cmd, disposition);
1399
1400         switch (disposition) {
1401                 case SUCCESS:
1402                         scsi_finish_command(cmd);
1403                         break;
1404                 case NEEDS_RETRY:
1405                         scsi_retry_command(cmd);
1406                         break;
1407                 case ADD_TO_MLQUEUE:
1408                         scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1409                         break;
1410                 default:
1411                         if (!scsi_eh_scmd_add(cmd, 0))
1412                                 scsi_finish_command(cmd);
1413         }
1414 }
1415
1416 /*
1417  * Function:    scsi_request_fn()
1418  *
1419  * Purpose:     Main strategy routine for SCSI.
1420  *
1421  * Arguments:   q       - Pointer to actual queue.
1422  *
1423  * Returns:     Nothing
1424  *
1425  * Lock status: IO request lock assumed to be held when called.
1426  */
1427 static void scsi_request_fn(struct request_queue *q)
1428 {
1429         struct scsi_device *sdev = q->queuedata;
1430         struct Scsi_Host *shost;
1431         struct scsi_cmnd *cmd;
1432         struct request *req;
1433
1434         if (!sdev) {
1435                 printk("scsi: killing requests for dead queue\n");
1436                 while ((req = elv_next_request(q)) != NULL)
1437                         scsi_kill_request(req, q);
1438                 return;
1439         }
1440
1441         if(!get_device(&sdev->sdev_gendev))
1442                 /* We must be tearing the block queue down already */
1443                 return;
1444
1445         /*
1446          * To start with, we keep looping until the queue is empty, or until
1447          * the host is no longer able to accept any more requests.
1448          */
1449         shost = sdev->host;
1450         while (!blk_queue_plugged(q)) {
1451                 int rtn;
1452                 /*
1453                  * get next queueable request.  We do this early to make sure
1454                  * that the request is fully prepared even if we cannot 
1455                  * accept it.
1456                  */
1457                 req = elv_next_request(q);
1458                 if (!req || !scsi_dev_queue_ready(q, sdev))
1459                         break;
1460
1461                 if (unlikely(!scsi_device_online(sdev))) {
1462                         sdev_printk(KERN_ERR, sdev,
1463                                     "rejecting I/O to offline device\n");
1464                         scsi_kill_request(req, q);
1465                         continue;
1466                 }
1467
1468
1469                 /*
1470                  * Remove the request from the request list.
1471                  */
1472                 if (!(blk_queue_tagged(q) && !blk_queue_start_tag(q, req)))
1473                         blkdev_dequeue_request(req);
1474                 sdev->device_busy++;
1475
1476                 spin_unlock(q->queue_lock);
1477                 cmd = req->special;
1478                 if (unlikely(cmd == NULL)) {
1479                         printk(KERN_CRIT "impossible request in %s.\n"
1480                                          "please mail a stack trace to "
1481                                          "linux-scsi@vger.kernel.org\n",
1482                                          __FUNCTION__);
1483                         blk_dump_rq_flags(req, "foo");
1484                         BUG();
1485                 }
1486                 spin_lock(shost->host_lock);
1487
1488                 if (!scsi_host_queue_ready(q, shost, sdev))
1489                         goto not_ready;
1490                 if (sdev->single_lun) {
1491                         if (scsi_target(sdev)->starget_sdev_user &&
1492                             scsi_target(sdev)->starget_sdev_user != sdev)
1493                                 goto not_ready;
1494                         scsi_target(sdev)->starget_sdev_user = sdev;
1495                 }
1496                 shost->host_busy++;
1497
1498                 /*
1499                  * XXX(hch): This is rather suboptimal, scsi_dispatch_cmd will
1500                  *              take the lock again.
1501                  */
1502                 spin_unlock_irq(shost->host_lock);
1503
1504                 /*
1505                  * Finally, initialize any error handling parameters, and set up
1506                  * the timers for timeouts.
1507                  */
1508                 scsi_init_cmd_errh(cmd);
1509
1510                 /*
1511                  * Dispatch the command to the low-level driver.
1512                  */
1513                 rtn = scsi_dispatch_cmd(cmd);
1514                 spin_lock_irq(q->queue_lock);
1515                 if(rtn) {
1516                         /* we're refusing the command; because of
1517                          * the way locks get dropped, we need to 
1518                          * check here if plugging is required */
1519                         if(sdev->device_busy == 0)
1520                                 blk_plug_device(q);
1521
1522                         break;
1523                 }
1524         }
1525
1526         goto out;
1527
1528  not_ready:
1529         spin_unlock_irq(shost->host_lock);
1530
1531         /*
1532          * lock q, handle tag, requeue req, and decrement device_busy. We
1533          * must return with queue_lock held.
1534          *
1535          * Decrementing device_busy without checking it is OK, as all such
1536          * cases (host limits or settings) should run the queue at some
1537          * later time.
1538          */
1539         spin_lock_irq(q->queue_lock);
1540         blk_requeue_request(q, req);
1541         sdev->device_busy--;
1542         if(sdev->device_busy == 0)
1543                 blk_plug_device(q);
1544  out:
1545         /* must be careful here...if we trigger the ->remove() function
1546          * we cannot be holding the q lock */
1547         spin_unlock_irq(q->queue_lock);
1548         put_device(&sdev->sdev_gendev);
1549         spin_lock_irq(q->queue_lock);
1550 }
1551
1552 u64 scsi_calculate_bounce_limit(struct Scsi_Host *shost)
1553 {
1554         struct device *host_dev;
1555         u64 bounce_limit = 0xffffffff;
1556
1557         if (shost->unchecked_isa_dma)
1558                 return BLK_BOUNCE_ISA;
1559         /*
1560          * Platforms with virtual-DMA translation
1561          * hardware have no practical limit.
1562          */
1563         if (!PCI_DMA_BUS_IS_PHYS)
1564                 return BLK_BOUNCE_ANY;
1565
1566         host_dev = scsi_get_device(shost);
1567         if (host_dev && host_dev->dma_mask)
1568                 bounce_limit = *host_dev->dma_mask;
1569
1570         return bounce_limit;
1571 }
1572 EXPORT_SYMBOL(scsi_calculate_bounce_limit);
1573
1574 struct request_queue *__scsi_alloc_queue(struct Scsi_Host *shost,
1575                                          request_fn_proc *request_fn)
1576 {
1577         struct request_queue *q;
1578
1579         q = blk_init_queue(request_fn, NULL);
1580         if (!q)
1581                 return NULL;
1582
1583         blk_queue_max_hw_segments(q, shost->sg_tablesize);
1584         blk_queue_max_phys_segments(q, SCSI_MAX_PHYS_SEGMENTS);
1585         blk_queue_max_sectors(q, shost->max_sectors);
1586         blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
1587         blk_queue_segment_boundary(q, shost->dma_boundary);
1588
1589         if (!shost->use_clustering)
1590                 clear_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags);
1591         return q;
1592 }
1593 EXPORT_SYMBOL(__scsi_alloc_queue);
1594
1595 struct request_queue *scsi_alloc_queue(struct scsi_device *sdev)
1596 {
1597         struct request_queue *q;
1598
1599         q = __scsi_alloc_queue(sdev->host, scsi_request_fn);
1600         if (!q)
1601                 return NULL;
1602
1603         blk_queue_prep_rq(q, scsi_prep_fn);
1604         blk_queue_issue_flush_fn(q, scsi_issue_flush_fn);
1605         blk_queue_softirq_done(q, scsi_softirq_done);
1606         return q;
1607 }
1608
1609 void scsi_free_queue(struct request_queue *q)
1610 {
1611         blk_cleanup_queue(q);
1612 }
1613
1614 /*
1615  * Function:    scsi_block_requests()
1616  *
1617  * Purpose:     Utility function used by low-level drivers to prevent further
1618  *              commands from being queued to the device.
1619  *
1620  * Arguments:   shost       - Host in question
1621  *
1622  * Returns:     Nothing
1623  *
1624  * Lock status: No locks are assumed held.
1625  *
1626  * Notes:       There is no timer nor any other means by which the requests
1627  *              get unblocked other than the low-level driver calling
1628  *              scsi_unblock_requests().
1629  */
1630 void scsi_block_requests(struct Scsi_Host *shost)
1631 {
1632         shost->host_self_blocked = 1;
1633 }
1634 EXPORT_SYMBOL(scsi_block_requests);
1635
1636 /*
1637  * Function:    scsi_unblock_requests()
1638  *
1639  * Purpose:     Utility function used by low-level drivers to allow further
1640  *              commands from being queued to the device.
1641  *
1642  * Arguments:   shost       - Host in question
1643  *
1644  * Returns:     Nothing
1645  *
1646  * Lock status: No locks are assumed held.
1647  *
1648  * Notes:       There is no timer nor any other means by which the requests
1649  *              get unblocked other than the low-level driver calling
1650  *              scsi_unblock_requests().
1651  *
1652  *              This is done as an API function so that changes to the
1653  *              internals of the scsi mid-layer won't require wholesale
1654  *              changes to drivers that use this feature.
1655  */
1656 void scsi_unblock_requests(struct Scsi_Host *shost)
1657 {
1658         shost->host_self_blocked = 0;
1659         scsi_run_host_queues(shost);
1660 }
1661 EXPORT_SYMBOL(scsi_unblock_requests);
1662
1663 int __init scsi_init_queue(void)
1664 {
1665         int i;
1666
1667         scsi_io_context_cache = kmem_cache_create("scsi_io_context",
1668                                         sizeof(struct scsi_io_context),
1669                                         0, 0, NULL, NULL);
1670         if (!scsi_io_context_cache) {
1671                 printk(KERN_ERR "SCSI: can't init scsi io context cache\n");
1672                 return -ENOMEM;
1673         }
1674
1675         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1676                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1677                 int size = sgp->size * sizeof(struct scatterlist);
1678
1679                 sgp->slab = kmem_cache_create(sgp->name, size, 0,
1680                                 SLAB_HWCACHE_ALIGN, NULL, NULL);
1681                 if (!sgp->slab) {
1682                         printk(KERN_ERR "SCSI: can't init sg slab %s\n",
1683                                         sgp->name);
1684                 }
1685
1686                 sgp->pool = mempool_create_slab_pool(SG_MEMPOOL_SIZE,
1687                                                      sgp->slab);
1688                 if (!sgp->pool) {
1689                         printk(KERN_ERR "SCSI: can't init sg mempool %s\n",
1690                                         sgp->name);
1691                 }
1692         }
1693
1694         return 0;
1695 }
1696
1697 void scsi_exit_queue(void)
1698 {
1699         int i;
1700
1701         kmem_cache_destroy(scsi_io_context_cache);
1702
1703         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1704                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1705                 mempool_destroy(sgp->pool);
1706                 kmem_cache_destroy(sgp->slab);
1707         }
1708 }
1709
1710 /**
1711  *      scsi_mode_select - issue a mode select
1712  *      @sdev:  SCSI device to be queried
1713  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
1714  *      @sp:    Save page bit (0 == don't save, 1 == save)
1715  *      @modepage: mode page being requested
1716  *      @buffer: request buffer (may not be smaller than eight bytes)
1717  *      @len:   length of request buffer.
1718  *      @timeout: command timeout
1719  *      @retries: number of retries before failing
1720  *      @data: returns a structure abstracting the mode header data
1721  *      @sense: place to put sense data (or NULL if no sense to be collected).
1722  *              must be SCSI_SENSE_BUFFERSIZE big.
1723  *
1724  *      Returns zero if successful; negative error number or scsi
1725  *      status on error
1726  *
1727  */
1728 int
1729 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
1730                  unsigned char *buffer, int len, int timeout, int retries,
1731                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1732 {
1733         unsigned char cmd[10];
1734         unsigned char *real_buffer;
1735         int ret;
1736
1737         memset(cmd, 0, sizeof(cmd));
1738         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
1739
1740         if (sdev->use_10_for_ms) {
1741                 if (len > 65535)
1742                         return -EINVAL;
1743                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
1744                 if (!real_buffer)
1745                         return -ENOMEM;
1746                 memcpy(real_buffer + 8, buffer, len);
1747                 len += 8;
1748                 real_buffer[0] = 0;
1749                 real_buffer[1] = 0;
1750                 real_buffer[2] = data->medium_type;
1751                 real_buffer[3] = data->device_specific;
1752                 real_buffer[4] = data->longlba ? 0x01 : 0;
1753                 real_buffer[5] = 0;
1754                 real_buffer[6] = data->block_descriptor_length >> 8;
1755                 real_buffer[7] = data->block_descriptor_length;
1756
1757                 cmd[0] = MODE_SELECT_10;
1758                 cmd[7] = len >> 8;
1759                 cmd[8] = len;
1760         } else {
1761                 if (len > 255 || data->block_descriptor_length > 255 ||
1762                     data->longlba)
1763                         return -EINVAL;
1764
1765                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
1766                 if (!real_buffer)
1767                         return -ENOMEM;
1768                 memcpy(real_buffer + 4, buffer, len);
1769                 len += 4;
1770                 real_buffer[0] = 0;
1771                 real_buffer[1] = data->medium_type;
1772                 real_buffer[2] = data->device_specific;
1773                 real_buffer[3] = data->block_descriptor_length;
1774                 
1775
1776                 cmd[0] = MODE_SELECT;
1777                 cmd[4] = len;
1778         }
1779
1780         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
1781                                sshdr, timeout, retries);
1782         kfree(real_buffer);
1783         return ret;
1784 }
1785 EXPORT_SYMBOL_GPL(scsi_mode_select);
1786
1787 /**
1788  *      scsi_mode_sense - issue a mode sense, falling back from 10 to 
1789  *              six bytes if necessary.
1790  *      @sdev:  SCSI device to be queried
1791  *      @dbd:   set if mode sense will allow block descriptors to be returned
1792  *      @modepage: mode page being requested
1793  *      @buffer: request buffer (may not be smaller than eight bytes)
1794  *      @len:   length of request buffer.
1795  *      @timeout: command timeout
1796  *      @retries: number of retries before failing
1797  *      @data: returns a structure abstracting the mode header data
1798  *      @sense: place to put sense data (or NULL if no sense to be collected).
1799  *              must be SCSI_SENSE_BUFFERSIZE big.
1800  *
1801  *      Returns zero if unsuccessful, or the header offset (either 4
1802  *      or 8 depending on whether a six or ten byte command was
1803  *      issued) if successful.
1804  **/
1805 int
1806 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
1807                   unsigned char *buffer, int len, int timeout, int retries,
1808                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1809 {
1810         unsigned char cmd[12];
1811         int use_10_for_ms;
1812         int header_length;
1813         int result;
1814         struct scsi_sense_hdr my_sshdr;
1815
1816         memset(data, 0, sizeof(*data));
1817         memset(&cmd[0], 0, 12);
1818         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
1819         cmd[2] = modepage;
1820
1821         /* caller might not be interested in sense, but we need it */
1822         if (!sshdr)
1823                 sshdr = &my_sshdr;
1824
1825  retry:
1826         use_10_for_ms = sdev->use_10_for_ms;
1827
1828         if (use_10_for_ms) {
1829                 if (len < 8)
1830                         len = 8;
1831
1832                 cmd[0] = MODE_SENSE_10;
1833                 cmd[8] = len;
1834                 header_length = 8;
1835         } else {
1836                 if (len < 4)
1837                         len = 4;
1838
1839                 cmd[0] = MODE_SENSE;
1840                 cmd[4] = len;
1841                 header_length = 4;
1842         }
1843
1844         memset(buffer, 0, len);
1845
1846         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
1847                                   sshdr, timeout, retries);
1848
1849         /* This code looks awful: what it's doing is making sure an
1850          * ILLEGAL REQUEST sense return identifies the actual command
1851          * byte as the problem.  MODE_SENSE commands can return
1852          * ILLEGAL REQUEST if the code page isn't supported */
1853
1854         if (use_10_for_ms && !scsi_status_is_good(result) &&
1855             (driver_byte(result) & DRIVER_SENSE)) {
1856                 if (scsi_sense_valid(sshdr)) {
1857                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
1858                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
1859                                 /* 
1860                                  * Invalid command operation code
1861                                  */
1862                                 sdev->use_10_for_ms = 0;
1863                                 goto retry;
1864                         }
1865                 }
1866         }
1867
1868         if(scsi_status_is_good(result)) {
1869                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
1870                              (modepage == 6 || modepage == 8))) {
1871                         /* Initio breakage? */
1872                         header_length = 0;
1873                         data->length = 13;
1874                         data->medium_type = 0;
1875                         data->device_specific = 0;
1876                         data->longlba = 0;
1877                         data->block_descriptor_length = 0;
1878                 } else if(use_10_for_ms) {
1879                         data->length = buffer[0]*256 + buffer[1] + 2;
1880                         data->medium_type = buffer[2];
1881                         data->device_specific = buffer[3];
1882                         data->longlba = buffer[4] & 0x01;
1883                         data->block_descriptor_length = buffer[6]*256
1884                                 + buffer[7];
1885                 } else {
1886                         data->length = buffer[0] + 1;
1887                         data->medium_type = buffer[1];
1888                         data->device_specific = buffer[2];
1889                         data->block_descriptor_length = buffer[3];
1890                 }
1891                 data->header_length = header_length;
1892         }
1893
1894         return result;
1895 }
1896 EXPORT_SYMBOL(scsi_mode_sense);
1897
1898 int
1899 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries)
1900 {
1901         char cmd[] = {
1902                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
1903         };
1904         struct scsi_sense_hdr sshdr;
1905         int result;
1906         
1907         result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, &sshdr,
1908                                   timeout, retries);
1909
1910         if ((driver_byte(result) & DRIVER_SENSE) && sdev->removable) {
1911
1912                 if ((scsi_sense_valid(&sshdr)) &&
1913                     ((sshdr.sense_key == UNIT_ATTENTION) ||
1914                      (sshdr.sense_key == NOT_READY))) {
1915                         sdev->changed = 1;
1916                         result = 0;
1917                 }
1918         }
1919         return result;
1920 }
1921 EXPORT_SYMBOL(scsi_test_unit_ready);
1922
1923 /**
1924  *      scsi_device_set_state - Take the given device through the device
1925  *              state model.
1926  *      @sdev:  scsi device to change the state of.
1927  *      @state: state to change to.
1928  *
1929  *      Returns zero if unsuccessful or an error if the requested 
1930  *      transition is illegal.
1931  **/
1932 int
1933 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
1934 {
1935         enum scsi_device_state oldstate = sdev->sdev_state;
1936
1937         if (state == oldstate)
1938                 return 0;
1939
1940         switch (state) {
1941         case SDEV_CREATED:
1942                 /* There are no legal states that come back to
1943                  * created.  This is the manually initialised start
1944                  * state */
1945                 goto illegal;
1946                         
1947         case SDEV_RUNNING:
1948                 switch (oldstate) {
1949                 case SDEV_CREATED:
1950                 case SDEV_OFFLINE:
1951                 case SDEV_QUIESCE:
1952                 case SDEV_BLOCK:
1953                         break;
1954                 default:
1955                         goto illegal;
1956                 }
1957                 break;
1958
1959         case SDEV_QUIESCE:
1960                 switch (oldstate) {
1961                 case SDEV_RUNNING:
1962                 case SDEV_OFFLINE:
1963                         break;
1964                 default:
1965                         goto illegal;
1966                 }
1967                 break;
1968
1969         case SDEV_OFFLINE:
1970                 switch (oldstate) {
1971                 case SDEV_CREATED:
1972                 case SDEV_RUNNING:
1973                 case SDEV_QUIESCE:
1974                 case SDEV_BLOCK:
1975                         break;
1976                 default:
1977                         goto illegal;
1978                 }
1979                 break;
1980
1981         case SDEV_BLOCK:
1982                 switch (oldstate) {
1983                 case SDEV_CREATED:
1984                 case SDEV_RUNNING:
1985                         break;
1986                 default:
1987                         goto illegal;
1988                 }
1989                 break;
1990
1991         case SDEV_CANCEL:
1992                 switch (oldstate) {
1993                 case SDEV_CREATED:
1994                 case SDEV_RUNNING:
1995                 case SDEV_QUIESCE:
1996                 case SDEV_OFFLINE:
1997                 case SDEV_BLOCK:
1998                         break;
1999                 default:
2000                         goto illegal;
2001                 }
2002                 break;
2003
2004         case SDEV_DEL:
2005                 switch (oldstate) {
2006                 case SDEV_CREATED:
2007                 case SDEV_RUNNING:
2008                 case SDEV_OFFLINE:
2009                 case SDEV_CANCEL:
2010                         break;
2011                 default:
2012                         goto illegal;
2013                 }
2014                 break;
2015
2016         }
2017         sdev->sdev_state = state;
2018         return 0;
2019
2020  illegal:
2021         SCSI_LOG_ERROR_RECOVERY(1, 
2022                                 sdev_printk(KERN_ERR, sdev,
2023                                             "Illegal state transition %s->%s\n",
2024                                             scsi_device_state_name(oldstate),
2025                                             scsi_device_state_name(state))
2026                                 );
2027         return -EINVAL;
2028 }
2029 EXPORT_SYMBOL(scsi_device_set_state);
2030
2031 /**
2032  *      scsi_device_quiesce - Block user issued commands.
2033  *      @sdev:  scsi device to quiesce.
2034  *
2035  *      This works by trying to transition to the SDEV_QUIESCE state
2036  *      (which must be a legal transition).  When the device is in this
2037  *      state, only special requests will be accepted, all others will
2038  *      be deferred.  Since special requests may also be requeued requests,
2039  *      a successful return doesn't guarantee the device will be 
2040  *      totally quiescent.
2041  *
2042  *      Must be called with user context, may sleep.
2043  *
2044  *      Returns zero if unsuccessful or an error if not.
2045  **/
2046 int
2047 scsi_device_quiesce(struct scsi_device *sdev)
2048 {
2049         int err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2050         if (err)
2051                 return err;
2052
2053         scsi_run_queue(sdev->request_queue);
2054         while (sdev->device_busy) {
2055                 msleep_interruptible(200);
2056                 scsi_run_queue(sdev->request_queue);
2057         }
2058         return 0;
2059 }
2060 EXPORT_SYMBOL(scsi_device_quiesce);
2061
2062 /**
2063  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2064  *      @sdev:  scsi device to resume.
2065  *
2066  *      Moves the device from quiesced back to running and restarts the
2067  *      queues.
2068  *
2069  *      Must be called with user context, may sleep.
2070  **/
2071 void
2072 scsi_device_resume(struct scsi_device *sdev)
2073 {
2074         if(scsi_device_set_state(sdev, SDEV_RUNNING))
2075                 return;
2076         scsi_run_queue(sdev->request_queue);
2077 }
2078 EXPORT_SYMBOL(scsi_device_resume);
2079
2080 static void
2081 device_quiesce_fn(struct scsi_device *sdev, void *data)
2082 {
2083         scsi_device_quiesce(sdev);
2084 }
2085
2086 void
2087 scsi_target_quiesce(struct scsi_target *starget)
2088 {
2089         starget_for_each_device(starget, NULL, device_quiesce_fn);
2090 }
2091 EXPORT_SYMBOL(scsi_target_quiesce);
2092
2093 static void
2094 device_resume_fn(struct scsi_device *sdev, void *data)
2095 {
2096         scsi_device_resume(sdev);
2097 }
2098
2099 void
2100 scsi_target_resume(struct scsi_target *starget)
2101 {
2102         starget_for_each_device(starget, NULL, device_resume_fn);
2103 }
2104 EXPORT_SYMBOL(scsi_target_resume);
2105
2106 /**
2107  * scsi_internal_device_block - internal function to put a device
2108  *                              temporarily into the SDEV_BLOCK state
2109  * @sdev:       device to block
2110  *
2111  * Block request made by scsi lld's to temporarily stop all
2112  * scsi commands on the specified device.  Called from interrupt
2113  * or normal process context.
2114  *
2115  * Returns zero if successful or error if not
2116  *
2117  * Notes:       
2118  *      This routine transitions the device to the SDEV_BLOCK state
2119  *      (which must be a legal transition).  When the device is in this
2120  *      state, all commands are deferred until the scsi lld reenables
2121  *      the device with scsi_device_unblock or device_block_tmo fires.
2122  *      This routine assumes the host_lock is held on entry.
2123  **/
2124 int
2125 scsi_internal_device_block(struct scsi_device *sdev)
2126 {
2127         request_queue_t *q = sdev->request_queue;
2128         unsigned long flags;
2129         int err = 0;
2130
2131         err = scsi_device_set_state(sdev, SDEV_BLOCK);
2132         if (err)
2133                 return err;
2134
2135         /* 
2136          * The device has transitioned to SDEV_BLOCK.  Stop the
2137          * block layer from calling the midlayer with this device's
2138          * request queue. 
2139          */
2140         spin_lock_irqsave(q->queue_lock, flags);
2141         blk_stop_queue(q);
2142         spin_unlock_irqrestore(q->queue_lock, flags);
2143
2144         return 0;
2145 }
2146 EXPORT_SYMBOL_GPL(scsi_internal_device_block);
2147  
2148 /**
2149  * scsi_internal_device_unblock - resume a device after a block request
2150  * @sdev:       device to resume
2151  *
2152  * Called by scsi lld's or the midlayer to restart the device queue
2153  * for the previously suspended scsi device.  Called from interrupt or
2154  * normal process context.
2155  *
2156  * Returns zero if successful or error if not.
2157  *
2158  * Notes:       
2159  *      This routine transitions the device to the SDEV_RUNNING state
2160  *      (which must be a legal transition) allowing the midlayer to
2161  *      goose the queue for this device.  This routine assumes the 
2162  *      host_lock is held upon entry.
2163  **/
2164 int
2165 scsi_internal_device_unblock(struct scsi_device *sdev)
2166 {
2167         request_queue_t *q = sdev->request_queue; 
2168         int err;
2169         unsigned long flags;
2170         
2171         /* 
2172          * Try to transition the scsi device to SDEV_RUNNING
2173          * and goose the device queue if successful.  
2174          */
2175         err = scsi_device_set_state(sdev, SDEV_RUNNING);
2176         if (err)
2177                 return err;
2178
2179         spin_lock_irqsave(q->queue_lock, flags);
2180         blk_start_queue(q);
2181         spin_unlock_irqrestore(q->queue_lock, flags);
2182
2183         return 0;
2184 }
2185 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock);
2186
2187 static void
2188 device_block(struct scsi_device *sdev, void *data)
2189 {
2190         scsi_internal_device_block(sdev);
2191 }
2192
2193 static int
2194 target_block(struct device *dev, void *data)
2195 {
2196         if (scsi_is_target_device(dev))
2197                 starget_for_each_device(to_scsi_target(dev), NULL,
2198                                         device_block);
2199         return 0;
2200 }
2201
2202 void
2203 scsi_target_block(struct device *dev)
2204 {
2205         if (scsi_is_target_device(dev))
2206                 starget_for_each_device(to_scsi_target(dev), NULL,
2207                                         device_block);
2208         else
2209                 device_for_each_child(dev, NULL, target_block);
2210 }
2211 EXPORT_SYMBOL_GPL(scsi_target_block);
2212
2213 static void
2214 device_unblock(struct scsi_device *sdev, void *data)
2215 {
2216         scsi_internal_device_unblock(sdev);
2217 }
2218
2219 static int
2220 target_unblock(struct device *dev, void *data)
2221 {
2222         if (scsi_is_target_device(dev))
2223                 starget_for_each_device(to_scsi_target(dev), NULL,
2224                                         device_unblock);
2225         return 0;
2226 }
2227
2228 void
2229 scsi_target_unblock(struct device *dev)
2230 {
2231         if (scsi_is_target_device(dev))
2232                 starget_for_each_device(to_scsi_target(dev), NULL,
2233                                         device_unblock);
2234         else
2235                 device_for_each_child(dev, NULL, target_unblock);
2236 }
2237 EXPORT_SYMBOL_GPL(scsi_target_unblock);
2238
2239 /**
2240  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2241  * @sg:         scatter-gather list
2242  * @sg_count:   number of segments in sg
2243  * @offset:     offset in bytes into sg, on return offset into the mapped area
2244  * @len:        bytes to map, on return number of bytes mapped
2245  *
2246  * Returns virtual address of the start of the mapped page
2247  */
2248 void *scsi_kmap_atomic_sg(struct scatterlist *sg, int sg_count,
2249                           size_t *offset, size_t *len)
2250 {
2251         int i;
2252         size_t sg_len = 0, len_complete = 0;
2253         struct page *page;
2254
2255         for (i = 0; i < sg_count; i++) {
2256                 len_complete = sg_len; /* Complete sg-entries */
2257                 sg_len += sg[i].length;
2258                 if (sg_len > *offset)
2259                         break;
2260         }
2261
2262         if (unlikely(i == sg_count)) {
2263                 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2264                         "elements %d\n",
2265                        __FUNCTION__, sg_len, *offset, sg_count);
2266                 WARN_ON(1);
2267                 return NULL;
2268         }
2269
2270         /* Offset starting from the beginning of first page in this sg-entry */
2271         *offset = *offset - len_complete + sg[i].offset;
2272
2273         /* Assumption: contiguous pages can be accessed as "page + i" */
2274         page = nth_page(sg[i].page, (*offset >> PAGE_SHIFT));
2275         *offset &= ~PAGE_MASK;
2276
2277         /* Bytes in this sg-entry from *offset to the end of the page */
2278         sg_len = PAGE_SIZE - *offset;
2279         if (*len > sg_len)
2280                 *len = sg_len;
2281
2282         return kmap_atomic(page, KM_BIO_SRC_IRQ);
2283 }
2284 EXPORT_SYMBOL(scsi_kmap_atomic_sg);
2285
2286 /**
2287  * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously
2288  *                         mapped with scsi_kmap_atomic_sg
2289  * @virt:       virtual address to be unmapped
2290  */
2291 void scsi_kunmap_atomic_sg(void *virt)
2292 {
2293         kunmap_atomic(virt, KM_BIO_SRC_IRQ);
2294 }
2295 EXPORT_SYMBOL(scsi_kunmap_atomic_sg);