Merge branch 'master' of master.kernel.org:/pub/scm/linux/kernel/git/davem/net-2.6
[linux-2.6] / drivers / scsi / scsi_lib.c
1 /*
2  *  scsi_lib.c Copyright (C) 1999 Eric Youngdale
3  *
4  *  SCSI queueing library.
5  *      Initial versions: Eric Youngdale (eric@andante.org).
6  *                        Based upon conversations with large numbers
7  *                        of people at Linux Expo.
8  */
9
10 #include <linux/bio.h>
11 #include <linux/bitops.h>
12 #include <linux/blkdev.h>
13 #include <linux/completion.h>
14 #include <linux/kernel.h>
15 #include <linux/mempool.h>
16 #include <linux/slab.h>
17 #include <linux/init.h>
18 #include <linux/pci.h>
19 #include <linux/delay.h>
20 #include <linux/hardirq.h>
21 #include <linux/scatterlist.h>
22
23 #include <scsi/scsi.h>
24 #include <scsi/scsi_cmnd.h>
25 #include <scsi/scsi_dbg.h>
26 #include <scsi/scsi_device.h>
27 #include <scsi/scsi_driver.h>
28 #include <scsi/scsi_eh.h>
29 #include <scsi/scsi_host.h>
30
31 #include "scsi_priv.h"
32 #include "scsi_logging.h"
33
34
35 #define SG_MEMPOOL_NR           ARRAY_SIZE(scsi_sg_pools)
36 #define SG_MEMPOOL_SIZE         2
37
38 struct scsi_host_sg_pool {
39         size_t          size;
40         char            *name;
41         struct kmem_cache       *slab;
42         mempool_t       *pool;
43 };
44
45 #define SP(x) { x, "sgpool-" __stringify(x) }
46 #if (SCSI_MAX_SG_SEGMENTS < 32)
47 #error SCSI_MAX_SG_SEGMENTS is too small (must be 32 or greater)
48 #endif
49 static struct scsi_host_sg_pool scsi_sg_pools[] = {
50         SP(8),
51         SP(16),
52 #if (SCSI_MAX_SG_SEGMENTS > 32)
53         SP(32),
54 #if (SCSI_MAX_SG_SEGMENTS > 64)
55         SP(64),
56 #if (SCSI_MAX_SG_SEGMENTS > 128)
57         SP(128),
58 #if (SCSI_MAX_SG_SEGMENTS > 256)
59 #error SCSI_MAX_SG_SEGMENTS is too large (256 MAX)
60 #endif
61 #endif
62 #endif
63 #endif
64         SP(SCSI_MAX_SG_SEGMENTS)
65 };
66 #undef SP
67
68 struct kmem_cache *scsi_sdb_cache;
69
70 static void scsi_run_queue(struct request_queue *q);
71
72 /*
73  * Function:    scsi_unprep_request()
74  *
75  * Purpose:     Remove all preparation done for a request, including its
76  *              associated scsi_cmnd, so that it can be requeued.
77  *
78  * Arguments:   req     - request to unprepare
79  *
80  * Lock status: Assumed that no locks are held upon entry.
81  *
82  * Returns:     Nothing.
83  */
84 static void scsi_unprep_request(struct request *req)
85 {
86         struct scsi_cmnd *cmd = req->special;
87
88         req->cmd_flags &= ~REQ_DONTPREP;
89         req->special = NULL;
90
91         scsi_put_command(cmd);
92 }
93
94 /*
95  * Function:    scsi_queue_insert()
96  *
97  * Purpose:     Insert a command in the midlevel queue.
98  *
99  * Arguments:   cmd    - command that we are adding to queue.
100  *              reason - why we are inserting command to queue.
101  *
102  * Lock status: Assumed that lock is not held upon entry.
103  *
104  * Returns:     Nothing.
105  *
106  * Notes:       We do this for one of two cases.  Either the host is busy
107  *              and it cannot accept any more commands for the time being,
108  *              or the device returned QUEUE_FULL and can accept no more
109  *              commands.
110  * Notes:       This could be called either from an interrupt context or a
111  *              normal process context.
112  */
113 int scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
114 {
115         struct Scsi_Host *host = cmd->device->host;
116         struct scsi_device *device = cmd->device;
117         struct scsi_target *starget = scsi_target(device);
118         struct request_queue *q = device->request_queue;
119         unsigned long flags;
120
121         SCSI_LOG_MLQUEUE(1,
122                  printk("Inserting command %p into mlqueue\n", cmd));
123
124         /*
125          * Set the appropriate busy bit for the device/host.
126          *
127          * If the host/device isn't busy, assume that something actually
128          * completed, and that we should be able to queue a command now.
129          *
130          * Note that the prior mid-layer assumption that any host could
131          * always queue at least one command is now broken.  The mid-layer
132          * will implement a user specifiable stall (see
133          * scsi_host.max_host_blocked and scsi_device.max_device_blocked)
134          * if a command is requeued with no other commands outstanding
135          * either for the device or for the host.
136          */
137         switch (reason) {
138         case SCSI_MLQUEUE_HOST_BUSY:
139                 host->host_blocked = host->max_host_blocked;
140                 break;
141         case SCSI_MLQUEUE_DEVICE_BUSY:
142                 device->device_blocked = device->max_device_blocked;
143                 break;
144         case SCSI_MLQUEUE_TARGET_BUSY:
145                 starget->target_blocked = starget->max_target_blocked;
146                 break;
147         }
148
149         /*
150          * Decrement the counters, since these commands are no longer
151          * active on the host/device.
152          */
153         scsi_device_unbusy(device);
154
155         /*
156          * Requeue this command.  It will go before all other commands
157          * that are already in the queue.
158          *
159          * NOTE: there is magic here about the way the queue is plugged if
160          * we have no outstanding commands.
161          * 
162          * Although we *don't* plug the queue, we call the request
163          * function.  The SCSI request function detects the blocked condition
164          * and plugs the queue appropriately.
165          */
166         spin_lock_irqsave(q->queue_lock, flags);
167         blk_requeue_request(q, cmd->request);
168         spin_unlock_irqrestore(q->queue_lock, flags);
169
170         scsi_run_queue(q);
171
172         return 0;
173 }
174
175 /**
176  * scsi_execute - insert request and wait for the result
177  * @sdev:       scsi device
178  * @cmd:        scsi command
179  * @data_direction: data direction
180  * @buffer:     data buffer
181  * @bufflen:    len of buffer
182  * @sense:      optional sense buffer
183  * @timeout:    request timeout in seconds
184  * @retries:    number of times to retry request
185  * @flags:      or into request flags;
186  *
187  * returns the req->errors value which is the scsi_cmnd result
188  * field.
189  */
190 int scsi_execute(struct scsi_device *sdev, const unsigned char *cmd,
191                  int data_direction, void *buffer, unsigned bufflen,
192                  unsigned char *sense, int timeout, int retries, int flags)
193 {
194         struct request *req;
195         int write = (data_direction == DMA_TO_DEVICE);
196         int ret = DRIVER_ERROR << 24;
197
198         req = blk_get_request(sdev->request_queue, write, __GFP_WAIT);
199
200         if (bufflen &&  blk_rq_map_kern(sdev->request_queue, req,
201                                         buffer, bufflen, __GFP_WAIT))
202                 goto out;
203
204         req->cmd_len = COMMAND_SIZE(cmd[0]);
205         memcpy(req->cmd, cmd, req->cmd_len);
206         req->sense = sense;
207         req->sense_len = 0;
208         req->retries = retries;
209         req->timeout = timeout;
210         req->cmd_type = REQ_TYPE_BLOCK_PC;
211         req->cmd_flags |= flags | REQ_QUIET | REQ_PREEMPT;
212
213         /*
214          * head injection *required* here otherwise quiesce won't work
215          */
216         blk_execute_rq(req->q, NULL, req, 1);
217
218         /*
219          * Some devices (USB mass-storage in particular) may transfer
220          * garbage data together with a residue indicating that the data
221          * is invalid.  Prevent the garbage from being misinterpreted
222          * and prevent security leaks by zeroing out the excess data.
223          */
224         if (unlikely(req->data_len > 0 && req->data_len <= bufflen))
225                 memset(buffer + (bufflen - req->data_len), 0, req->data_len);
226
227         ret = req->errors;
228  out:
229         blk_put_request(req);
230
231         return ret;
232 }
233 EXPORT_SYMBOL(scsi_execute);
234
235
236 int scsi_execute_req(struct scsi_device *sdev, const unsigned char *cmd,
237                      int data_direction, void *buffer, unsigned bufflen,
238                      struct scsi_sense_hdr *sshdr, int timeout, int retries)
239 {
240         char *sense = NULL;
241         int result;
242         
243         if (sshdr) {
244                 sense = kzalloc(SCSI_SENSE_BUFFERSIZE, GFP_NOIO);
245                 if (!sense)
246                         return DRIVER_ERROR << 24;
247         }
248         result = scsi_execute(sdev, cmd, data_direction, buffer, bufflen,
249                               sense, timeout, retries, 0);
250         if (sshdr)
251                 scsi_normalize_sense(sense, SCSI_SENSE_BUFFERSIZE, sshdr);
252
253         kfree(sense);
254         return result;
255 }
256 EXPORT_SYMBOL(scsi_execute_req);
257
258 struct scsi_io_context {
259         void *data;
260         void (*done)(void *data, char *sense, int result, int resid);
261         char sense[SCSI_SENSE_BUFFERSIZE];
262 };
263
264 static struct kmem_cache *scsi_io_context_cache;
265
266 static void scsi_end_async(struct request *req, int uptodate)
267 {
268         struct scsi_io_context *sioc = req->end_io_data;
269
270         if (sioc->done)
271                 sioc->done(sioc->data, sioc->sense, req->errors, req->data_len);
272
273         kmem_cache_free(scsi_io_context_cache, sioc);
274         __blk_put_request(req->q, req);
275 }
276
277 static int scsi_merge_bio(struct request *rq, struct bio *bio)
278 {
279         struct request_queue *q = rq->q;
280
281         bio->bi_flags &= ~(1 << BIO_SEG_VALID);
282         if (rq_data_dir(rq) == WRITE)
283                 bio->bi_rw |= (1 << BIO_RW);
284         blk_queue_bounce(q, &bio);
285
286         return blk_rq_append_bio(q, rq, bio);
287 }
288
289 static void scsi_bi_endio(struct bio *bio, int error)
290 {
291         bio_put(bio);
292 }
293
294 /**
295  * scsi_req_map_sg - map a scatterlist into a request
296  * @rq:         request to fill
297  * @sgl:        scatterlist
298  * @nsegs:      number of elements
299  * @bufflen:    len of buffer
300  * @gfp:        memory allocation flags
301  *
302  * scsi_req_map_sg maps a scatterlist into a request so that the
303  * request can be sent to the block layer. We do not trust the scatterlist
304  * sent to use, as some ULDs use that struct to only organize the pages.
305  */
306 static int scsi_req_map_sg(struct request *rq, struct scatterlist *sgl,
307                            int nsegs, unsigned bufflen, gfp_t gfp)
308 {
309         struct request_queue *q = rq->q;
310         int nr_pages = (bufflen + sgl[0].offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
311         unsigned int data_len = bufflen, len, bytes, off;
312         struct scatterlist *sg;
313         struct page *page;
314         struct bio *bio = NULL;
315         int i, err, nr_vecs = 0;
316
317         for_each_sg(sgl, sg, nsegs, i) {
318                 page = sg_page(sg);
319                 off = sg->offset;
320                 len = sg->length;
321
322                 while (len > 0 && data_len > 0) {
323                         /*
324                          * sg sends a scatterlist that is larger than
325                          * the data_len it wants transferred for certain
326                          * IO sizes
327                          */
328                         bytes = min_t(unsigned int, len, PAGE_SIZE - off);
329                         bytes = min(bytes, data_len);
330
331                         if (!bio) {
332                                 nr_vecs = min_t(int, BIO_MAX_PAGES, nr_pages);
333                                 nr_pages -= nr_vecs;
334
335                                 bio = bio_alloc(gfp, nr_vecs);
336                                 if (!bio) {
337                                         err = -ENOMEM;
338                                         goto free_bios;
339                                 }
340                                 bio->bi_end_io = scsi_bi_endio;
341                         }
342
343                         if (bio_add_pc_page(q, bio, page, bytes, off) !=
344                             bytes) {
345                                 bio_put(bio);
346                                 err = -EINVAL;
347                                 goto free_bios;
348                         }
349
350                         if (bio->bi_vcnt >= nr_vecs) {
351                                 err = scsi_merge_bio(rq, bio);
352                                 if (err) {
353                                         bio_endio(bio, 0);
354                                         goto free_bios;
355                                 }
356                                 bio = NULL;
357                         }
358
359                         page++;
360                         len -= bytes;
361                         data_len -=bytes;
362                         off = 0;
363                 }
364         }
365
366         rq->buffer = rq->data = NULL;
367         rq->data_len = bufflen;
368         return 0;
369
370 free_bios:
371         while ((bio = rq->bio) != NULL) {
372                 rq->bio = bio->bi_next;
373                 /*
374                  * call endio instead of bio_put incase it was bounced
375                  */
376                 bio_endio(bio, 0);
377         }
378
379         return err;
380 }
381
382 /**
383  * scsi_execute_async - insert request
384  * @sdev:       scsi device
385  * @cmd:        scsi command
386  * @cmd_len:    length of scsi cdb
387  * @data_direction: DMA_TO_DEVICE, DMA_FROM_DEVICE, or DMA_NONE
388  * @buffer:     data buffer (this can be a kernel buffer or scatterlist)
389  * @bufflen:    len of buffer
390  * @use_sg:     if buffer is a scatterlist this is the number of elements
391  * @timeout:    request timeout in seconds
392  * @retries:    number of times to retry request
393  * @privdata:   data passed to done()
394  * @done:       callback function when done
395  * @gfp:        memory allocation flags
396  */
397 int scsi_execute_async(struct scsi_device *sdev, const unsigned char *cmd,
398                        int cmd_len, int data_direction, void *buffer, unsigned bufflen,
399                        int use_sg, int timeout, int retries, void *privdata,
400                        void (*done)(void *, char *, int, int), gfp_t gfp)
401 {
402         struct request *req;
403         struct scsi_io_context *sioc;
404         int err = 0;
405         int write = (data_direction == DMA_TO_DEVICE);
406
407         sioc = kmem_cache_zalloc(scsi_io_context_cache, gfp);
408         if (!sioc)
409                 return DRIVER_ERROR << 24;
410
411         req = blk_get_request(sdev->request_queue, write, gfp);
412         if (!req)
413                 goto free_sense;
414         req->cmd_type = REQ_TYPE_BLOCK_PC;
415         req->cmd_flags |= REQ_QUIET;
416
417         if (use_sg)
418                 err = scsi_req_map_sg(req, buffer, use_sg, bufflen, gfp);
419         else if (bufflen)
420                 err = blk_rq_map_kern(req->q, req, buffer, bufflen, gfp);
421
422         if (err)
423                 goto free_req;
424
425         req->cmd_len = cmd_len;
426         memset(req->cmd, 0, BLK_MAX_CDB); /* ATAPI hates garbage after CDB */
427         memcpy(req->cmd, cmd, req->cmd_len);
428         req->sense = sioc->sense;
429         req->sense_len = 0;
430         req->timeout = timeout;
431         req->retries = retries;
432         req->end_io_data = sioc;
433
434         sioc->data = privdata;
435         sioc->done = done;
436
437         blk_execute_rq_nowait(req->q, NULL, req, 1, scsi_end_async);
438         return 0;
439
440 free_req:
441         blk_put_request(req);
442 free_sense:
443         kmem_cache_free(scsi_io_context_cache, sioc);
444         return DRIVER_ERROR << 24;
445 }
446 EXPORT_SYMBOL_GPL(scsi_execute_async);
447
448 /*
449  * Function:    scsi_init_cmd_errh()
450  *
451  * Purpose:     Initialize cmd fields related to error handling.
452  *
453  * Arguments:   cmd     - command that is ready to be queued.
454  *
455  * Notes:       This function has the job of initializing a number of
456  *              fields related to error handling.   Typically this will
457  *              be called once for each command, as required.
458  */
459 static void scsi_init_cmd_errh(struct scsi_cmnd *cmd)
460 {
461         cmd->serial_number = 0;
462         scsi_set_resid(cmd, 0);
463         memset(cmd->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
464         if (cmd->cmd_len == 0)
465                 cmd->cmd_len = scsi_command_size(cmd->cmnd);
466 }
467
468 void scsi_device_unbusy(struct scsi_device *sdev)
469 {
470         struct Scsi_Host *shost = sdev->host;
471         struct scsi_target *starget = scsi_target(sdev);
472         unsigned long flags;
473
474         spin_lock_irqsave(shost->host_lock, flags);
475         shost->host_busy--;
476         starget->target_busy--;
477         if (unlikely(scsi_host_in_recovery(shost) &&
478                      (shost->host_failed || shost->host_eh_scheduled)))
479                 scsi_eh_wakeup(shost);
480         spin_unlock(shost->host_lock);
481         spin_lock(sdev->request_queue->queue_lock);
482         sdev->device_busy--;
483         spin_unlock_irqrestore(sdev->request_queue->queue_lock, flags);
484 }
485
486 /*
487  * Called for single_lun devices on IO completion. Clear starget_sdev_user,
488  * and call blk_run_queue for all the scsi_devices on the target -
489  * including current_sdev first.
490  *
491  * Called with *no* scsi locks held.
492  */
493 static void scsi_single_lun_run(struct scsi_device *current_sdev)
494 {
495         struct Scsi_Host *shost = current_sdev->host;
496         struct scsi_device *sdev, *tmp;
497         struct scsi_target *starget = scsi_target(current_sdev);
498         unsigned long flags;
499
500         spin_lock_irqsave(shost->host_lock, flags);
501         starget->starget_sdev_user = NULL;
502         spin_unlock_irqrestore(shost->host_lock, flags);
503
504         /*
505          * Call blk_run_queue for all LUNs on the target, starting with
506          * current_sdev. We race with others (to set starget_sdev_user),
507          * but in most cases, we will be first. Ideally, each LU on the
508          * target would get some limited time or requests on the target.
509          */
510         blk_run_queue(current_sdev->request_queue);
511
512         spin_lock_irqsave(shost->host_lock, flags);
513         if (starget->starget_sdev_user)
514                 goto out;
515         list_for_each_entry_safe(sdev, tmp, &starget->devices,
516                         same_target_siblings) {
517                 if (sdev == current_sdev)
518                         continue;
519                 if (scsi_device_get(sdev))
520                         continue;
521
522                 spin_unlock_irqrestore(shost->host_lock, flags);
523                 blk_run_queue(sdev->request_queue);
524                 spin_lock_irqsave(shost->host_lock, flags);
525         
526                 scsi_device_put(sdev);
527         }
528  out:
529         spin_unlock_irqrestore(shost->host_lock, flags);
530 }
531
532 static inline int scsi_device_is_busy(struct scsi_device *sdev)
533 {
534         if (sdev->device_busy >= sdev->queue_depth || sdev->device_blocked)
535                 return 1;
536
537         return 0;
538 }
539
540 static inline int scsi_target_is_busy(struct scsi_target *starget)
541 {
542         return ((starget->can_queue > 0 &&
543                  starget->target_busy >= starget->can_queue) ||
544                  starget->target_blocked);
545 }
546
547 static inline int scsi_host_is_busy(struct Scsi_Host *shost)
548 {
549         if ((shost->can_queue > 0 && shost->host_busy >= shost->can_queue) ||
550             shost->host_blocked || shost->host_self_blocked)
551                 return 1;
552
553         return 0;
554 }
555
556 /*
557  * Function:    scsi_run_queue()
558  *
559  * Purpose:     Select a proper request queue to serve next
560  *
561  * Arguments:   q       - last request's queue
562  *
563  * Returns:     Nothing
564  *
565  * Notes:       The previous command was completely finished, start
566  *              a new one if possible.
567  */
568 static void scsi_run_queue(struct request_queue *q)
569 {
570         struct scsi_device *sdev = q->queuedata;
571         struct Scsi_Host *shost = sdev->host;
572         LIST_HEAD(starved_list);
573         unsigned long flags;
574
575         if (scsi_target(sdev)->single_lun)
576                 scsi_single_lun_run(sdev);
577
578         spin_lock_irqsave(shost->host_lock, flags);
579         list_splice_init(&shost->starved_list, &starved_list);
580
581         while (!list_empty(&starved_list)) {
582                 int flagset;
583
584                 /*
585                  * As long as shost is accepting commands and we have
586                  * starved queues, call blk_run_queue. scsi_request_fn
587                  * drops the queue_lock and can add us back to the
588                  * starved_list.
589                  *
590                  * host_lock protects the starved_list and starved_entry.
591                  * scsi_request_fn must get the host_lock before checking
592                  * or modifying starved_list or starved_entry.
593                  */
594                 if (scsi_host_is_busy(shost))
595                         break;
596
597                 sdev = list_entry(starved_list.next,
598                                   struct scsi_device, starved_entry);
599                 list_del_init(&sdev->starved_entry);
600                 if (scsi_target_is_busy(scsi_target(sdev))) {
601                         list_move_tail(&sdev->starved_entry,
602                                        &shost->starved_list);
603                         continue;
604                 }
605
606                 spin_unlock(shost->host_lock);
607
608                 spin_lock(sdev->request_queue->queue_lock);
609                 flagset = test_bit(QUEUE_FLAG_REENTER, &q->queue_flags) &&
610                                 !test_bit(QUEUE_FLAG_REENTER,
611                                         &sdev->request_queue->queue_flags);
612                 if (flagset)
613                         queue_flag_set(QUEUE_FLAG_REENTER, sdev->request_queue);
614                 __blk_run_queue(sdev->request_queue);
615                 if (flagset)
616                         queue_flag_clear(QUEUE_FLAG_REENTER, sdev->request_queue);
617                 spin_unlock(sdev->request_queue->queue_lock);
618
619                 spin_lock(shost->host_lock);
620         }
621         /* put any unprocessed entries back */
622         list_splice(&starved_list, &shost->starved_list);
623         spin_unlock_irqrestore(shost->host_lock, flags);
624
625         blk_run_queue(q);
626 }
627
628 /*
629  * Function:    scsi_requeue_command()
630  *
631  * Purpose:     Handle post-processing of completed commands.
632  *
633  * Arguments:   q       - queue to operate on
634  *              cmd     - command that may need to be requeued.
635  *
636  * Returns:     Nothing
637  *
638  * Notes:       After command completion, there may be blocks left
639  *              over which weren't finished by the previous command
640  *              this can be for a number of reasons - the main one is
641  *              I/O errors in the middle of the request, in which case
642  *              we need to request the blocks that come after the bad
643  *              sector.
644  * Notes:       Upon return, cmd is a stale pointer.
645  */
646 static void scsi_requeue_command(struct request_queue *q, struct scsi_cmnd *cmd)
647 {
648         struct request *req = cmd->request;
649         unsigned long flags;
650
651         scsi_unprep_request(req);
652         spin_lock_irqsave(q->queue_lock, flags);
653         blk_requeue_request(q, req);
654         spin_unlock_irqrestore(q->queue_lock, flags);
655
656         scsi_run_queue(q);
657 }
658
659 void scsi_next_command(struct scsi_cmnd *cmd)
660 {
661         struct scsi_device *sdev = cmd->device;
662         struct request_queue *q = sdev->request_queue;
663
664         /* need to hold a reference on the device before we let go of the cmd */
665         get_device(&sdev->sdev_gendev);
666
667         scsi_put_command(cmd);
668         scsi_run_queue(q);
669
670         /* ok to remove device now */
671         put_device(&sdev->sdev_gendev);
672 }
673
674 void scsi_run_host_queues(struct Scsi_Host *shost)
675 {
676         struct scsi_device *sdev;
677
678         shost_for_each_device(sdev, shost)
679                 scsi_run_queue(sdev->request_queue);
680 }
681
682 /*
683  * Function:    scsi_end_request()
684  *
685  * Purpose:     Post-processing of completed commands (usually invoked at end
686  *              of upper level post-processing and scsi_io_completion).
687  *
688  * Arguments:   cmd      - command that is complete.
689  *              error    - 0 if I/O indicates success, < 0 for I/O error.
690  *              bytes    - number of bytes of completed I/O
691  *              requeue  - indicates whether we should requeue leftovers.
692  *
693  * Lock status: Assumed that lock is not held upon entry.
694  *
695  * Returns:     cmd if requeue required, NULL otherwise.
696  *
697  * Notes:       This is called for block device requests in order to
698  *              mark some number of sectors as complete.
699  * 
700  *              We are guaranteeing that the request queue will be goosed
701  *              at some point during this call.
702  * Notes:       If cmd was requeued, upon return it will be a stale pointer.
703  */
704 static struct scsi_cmnd *scsi_end_request(struct scsi_cmnd *cmd, int error,
705                                           int bytes, int requeue)
706 {
707         struct request_queue *q = cmd->device->request_queue;
708         struct request *req = cmd->request;
709
710         /*
711          * If there are blocks left over at the end, set up the command
712          * to queue the remainder of them.
713          */
714         if (blk_end_request(req, error, bytes)) {
715                 int leftover = (req->hard_nr_sectors << 9);
716
717                 if (blk_pc_request(req))
718                         leftover = req->data_len;
719
720                 /* kill remainder if no retrys */
721                 if (error && scsi_noretry_cmd(cmd))
722                         blk_end_request(req, error, leftover);
723                 else {
724                         if (requeue) {
725                                 /*
726                                  * Bleah.  Leftovers again.  Stick the
727                                  * leftovers in the front of the
728                                  * queue, and goose the queue again.
729                                  */
730                                 scsi_requeue_command(q, cmd);
731                                 cmd = NULL;
732                         }
733                         return cmd;
734                 }
735         }
736
737         /*
738          * This will goose the queue request function at the end, so we don't
739          * need to worry about launching another command.
740          */
741         scsi_next_command(cmd);
742         return NULL;
743 }
744
745 static inline unsigned int scsi_sgtable_index(unsigned short nents)
746 {
747         unsigned int index;
748
749         BUG_ON(nents > SCSI_MAX_SG_SEGMENTS);
750
751         if (nents <= 8)
752                 index = 0;
753         else
754                 index = get_count_order(nents) - 3;
755
756         return index;
757 }
758
759 static void scsi_sg_free(struct scatterlist *sgl, unsigned int nents)
760 {
761         struct scsi_host_sg_pool *sgp;
762
763         sgp = scsi_sg_pools + scsi_sgtable_index(nents);
764         mempool_free(sgl, sgp->pool);
765 }
766
767 static struct scatterlist *scsi_sg_alloc(unsigned int nents, gfp_t gfp_mask)
768 {
769         struct scsi_host_sg_pool *sgp;
770
771         sgp = scsi_sg_pools + scsi_sgtable_index(nents);
772         return mempool_alloc(sgp->pool, gfp_mask);
773 }
774
775 static int scsi_alloc_sgtable(struct scsi_data_buffer *sdb, int nents,
776                               gfp_t gfp_mask)
777 {
778         int ret;
779
780         BUG_ON(!nents);
781
782         ret = __sg_alloc_table(&sdb->table, nents, SCSI_MAX_SG_SEGMENTS,
783                                gfp_mask, scsi_sg_alloc);
784         if (unlikely(ret))
785                 __sg_free_table(&sdb->table, SCSI_MAX_SG_SEGMENTS,
786                                 scsi_sg_free);
787
788         return ret;
789 }
790
791 static void scsi_free_sgtable(struct scsi_data_buffer *sdb)
792 {
793         __sg_free_table(&sdb->table, SCSI_MAX_SG_SEGMENTS, scsi_sg_free);
794 }
795
796 /*
797  * Function:    scsi_release_buffers()
798  *
799  * Purpose:     Completion processing for block device I/O requests.
800  *
801  * Arguments:   cmd     - command that we are bailing.
802  *
803  * Lock status: Assumed that no lock is held upon entry.
804  *
805  * Returns:     Nothing
806  *
807  * Notes:       In the event that an upper level driver rejects a
808  *              command, we must release resources allocated during
809  *              the __init_io() function.  Primarily this would involve
810  *              the scatter-gather table, and potentially any bounce
811  *              buffers.
812  */
813 void scsi_release_buffers(struct scsi_cmnd *cmd)
814 {
815         if (cmd->sdb.table.nents)
816                 scsi_free_sgtable(&cmd->sdb);
817
818         memset(&cmd->sdb, 0, sizeof(cmd->sdb));
819
820         if (scsi_bidi_cmnd(cmd)) {
821                 struct scsi_data_buffer *bidi_sdb =
822                         cmd->request->next_rq->special;
823                 scsi_free_sgtable(bidi_sdb);
824                 kmem_cache_free(scsi_sdb_cache, bidi_sdb);
825                 cmd->request->next_rq->special = NULL;
826         }
827
828         if (scsi_prot_sg_count(cmd))
829                 scsi_free_sgtable(cmd->prot_sdb);
830 }
831 EXPORT_SYMBOL(scsi_release_buffers);
832
833 /*
834  * Bidi commands Must be complete as a whole, both sides at once.
835  * If part of the bytes were written and lld returned
836  * scsi_in()->resid and/or scsi_out()->resid this information will be left
837  * in req->data_len and req->next_rq->data_len. The upper-layer driver can
838  * decide what to do with this information.
839  */
840 static void scsi_end_bidi_request(struct scsi_cmnd *cmd)
841 {
842         struct request *req = cmd->request;
843         unsigned int dlen = req->data_len;
844         unsigned int next_dlen = req->next_rq->data_len;
845
846         req->data_len = scsi_out(cmd)->resid;
847         req->next_rq->data_len = scsi_in(cmd)->resid;
848
849         /* The req and req->next_rq have not been completed */
850         BUG_ON(blk_end_bidi_request(req, 0, dlen, next_dlen));
851
852         scsi_release_buffers(cmd);
853
854         /*
855          * This will goose the queue request function at the end, so we don't
856          * need to worry about launching another command.
857          */
858         scsi_next_command(cmd);
859 }
860
861 /*
862  * Function:    scsi_io_completion()
863  *
864  * Purpose:     Completion processing for block device I/O requests.
865  *
866  * Arguments:   cmd   - command that is finished.
867  *
868  * Lock status: Assumed that no lock is held upon entry.
869  *
870  * Returns:     Nothing
871  *
872  * Notes:       This function is matched in terms of capabilities to
873  *              the function that created the scatter-gather list.
874  *              In other words, if there are no bounce buffers
875  *              (the normal case for most drivers), we don't need
876  *              the logic to deal with cleaning up afterwards.
877  *
878  *              We must do one of several things here:
879  *
880  *              a) Call scsi_end_request.  This will finish off the
881  *                 specified number of sectors.  If we are done, the
882  *                 command block will be released, and the queue
883  *                 function will be goosed.  If we are not done, then
884  *                 scsi_end_request will directly goose the queue.
885  *
886  *              b) We can just use scsi_requeue_command() here.  This would
887  *                 be used if we just wanted to retry, for example.
888  */
889 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
890 {
891         int result = cmd->result;
892         int this_count;
893         struct request_queue *q = cmd->device->request_queue;
894         struct request *req = cmd->request;
895         int error = 0;
896         struct scsi_sense_hdr sshdr;
897         int sense_valid = 0;
898         int sense_deferred = 0;
899
900         if (result) {
901                 sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
902                 if (sense_valid)
903                         sense_deferred = scsi_sense_is_deferred(&sshdr);
904         }
905
906         if (blk_pc_request(req)) { /* SG_IO ioctl from block level */
907                 req->errors = result;
908                 if (result) {
909                         if (sense_valid && req->sense) {
910                                 /*
911                                  * SG_IO wants current and deferred errors
912                                  */
913                                 int len = 8 + cmd->sense_buffer[7];
914
915                                 if (len > SCSI_SENSE_BUFFERSIZE)
916                                         len = SCSI_SENSE_BUFFERSIZE;
917                                 memcpy(req->sense, cmd->sense_buffer,  len);
918                                 req->sense_len = len;
919                         }
920                         if (!sense_deferred)
921                                 error = -EIO;
922                 }
923                 if (scsi_bidi_cmnd(cmd)) {
924                         /* will also release_buffers */
925                         scsi_end_bidi_request(cmd);
926                         return;
927                 }
928                 req->data_len = scsi_get_resid(cmd);
929         }
930
931         BUG_ON(blk_bidi_rq(req)); /* bidi not support for !blk_pc_request yet */
932         scsi_release_buffers(cmd);
933
934         /*
935          * Next deal with any sectors which we were able to correctly
936          * handle.
937          */
938         SCSI_LOG_HLCOMPLETE(1, printk("%ld sectors total, "
939                                       "%d bytes done.\n",
940                                       req->nr_sectors, good_bytes));
941
942         /* A number of bytes were successfully read.  If there
943          * are leftovers and there is some kind of error
944          * (result != 0), retry the rest.
945          */
946         if (scsi_end_request(cmd, error, good_bytes, result == 0) == NULL)
947                 return;
948         this_count = blk_rq_bytes(req);
949
950         /* good_bytes = 0, or (inclusive) there were leftovers and
951          * result = 0, so scsi_end_request couldn't retry.
952          */
953         if (sense_valid && !sense_deferred) {
954                 switch (sshdr.sense_key) {
955                 case UNIT_ATTENTION:
956                         if (cmd->device->removable) {
957                                 /* Detected disc change.  Set a bit
958                                  * and quietly refuse further access.
959                                  */
960                                 cmd->device->changed = 1;
961                                 scsi_end_request(cmd, -EIO, this_count, 1);
962                                 return;
963                         } else {
964                                 /* Must have been a power glitch, or a
965                                  * bus reset.  Could not have been a
966                                  * media change, so we just retry the
967                                  * request and see what happens.
968                                  */
969                                 scsi_requeue_command(q, cmd);
970                                 return;
971                         }
972                         break;
973                 case ILLEGAL_REQUEST:
974                         /* If we had an ILLEGAL REQUEST returned, then
975                          * we may have performed an unsupported
976                          * command.  The only thing this should be
977                          * would be a ten byte read where only a six
978                          * byte read was supported.  Also, on a system
979                          * where READ CAPACITY failed, we may have
980                          * read past the end of the disk.
981                          */
982                         if ((cmd->device->use_10_for_rw &&
983                             sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
984                             (cmd->cmnd[0] == READ_10 ||
985                              cmd->cmnd[0] == WRITE_10)) {
986                                 cmd->device->use_10_for_rw = 0;
987                                 /* This will cause a retry with a
988                                  * 6-byte command.
989                                  */
990                                 scsi_requeue_command(q, cmd);
991                         } else if (sshdr.asc == 0x10) /* DIX */
992                                 scsi_end_request(cmd, -EIO, this_count, 0);
993                         else
994                                 scsi_end_request(cmd, -EIO, this_count, 1);
995                         return;
996                 case ABORTED_COMMAND:
997                         if (sshdr.asc == 0x10) { /* DIF */
998                                 scsi_end_request(cmd, -EIO, this_count, 0);
999                                 return;
1000                         }
1001                         break;
1002                 case NOT_READY:
1003                         /* If the device is in the process of becoming
1004                          * ready, or has a temporary blockage, retry.
1005                          */
1006                         if (sshdr.asc == 0x04) {
1007                                 switch (sshdr.ascq) {
1008                                 case 0x01: /* becoming ready */
1009                                 case 0x04: /* format in progress */
1010                                 case 0x05: /* rebuild in progress */
1011                                 case 0x06: /* recalculation in progress */
1012                                 case 0x07: /* operation in progress */
1013                                 case 0x08: /* Long write in progress */
1014                                 case 0x09: /* self test in progress */
1015                                         scsi_requeue_command(q, cmd);
1016                                         return;
1017                                 default:
1018                                         break;
1019                                 }
1020                         }
1021                         if (!(req->cmd_flags & REQ_QUIET))
1022                                 scsi_cmd_print_sense_hdr(cmd,
1023                                                          "Device not ready",
1024                                                          &sshdr);
1025
1026                         scsi_end_request(cmd, -EIO, this_count, 1);
1027                         return;
1028                 case VOLUME_OVERFLOW:
1029                         if (!(req->cmd_flags & REQ_QUIET)) {
1030                                 scmd_printk(KERN_INFO, cmd,
1031                                             "Volume overflow, CDB: ");
1032                                 __scsi_print_command(cmd->cmnd);
1033                                 scsi_print_sense("", cmd);
1034                         }
1035                         /* See SSC3rXX or current. */
1036                         scsi_end_request(cmd, -EIO, this_count, 1);
1037                         return;
1038                 default:
1039                         break;
1040                 }
1041         }
1042         if (host_byte(result) == DID_RESET) {
1043                 /* Third party bus reset or reset for error recovery
1044                  * reasons.  Just retry the request and see what
1045                  * happens.
1046                  */
1047                 scsi_requeue_command(q, cmd);
1048                 return;
1049         }
1050         if (result) {
1051                 if (!(req->cmd_flags & REQ_QUIET)) {
1052                         scsi_print_result(cmd);
1053                         if (driver_byte(result) & DRIVER_SENSE)
1054                                 scsi_print_sense("", cmd);
1055                 }
1056         }
1057         scsi_end_request(cmd, -EIO, this_count, !result);
1058 }
1059
1060 static int scsi_init_sgtable(struct request *req, struct scsi_data_buffer *sdb,
1061                              gfp_t gfp_mask)
1062 {
1063         int count;
1064
1065         /*
1066          * If sg table allocation fails, requeue request later.
1067          */
1068         if (unlikely(scsi_alloc_sgtable(sdb, req->nr_phys_segments,
1069                                         gfp_mask))) {
1070                 return BLKPREP_DEFER;
1071         }
1072
1073         req->buffer = NULL;
1074
1075         /* 
1076          * Next, walk the list, and fill in the addresses and sizes of
1077          * each segment.
1078          */
1079         count = blk_rq_map_sg(req->q, req, sdb->table.sgl);
1080         BUG_ON(count > sdb->table.nents);
1081         sdb->table.nents = count;
1082         if (blk_pc_request(req))
1083                 sdb->length = req->data_len;
1084         else
1085                 sdb->length = req->nr_sectors << 9;
1086         return BLKPREP_OK;
1087 }
1088
1089 /*
1090  * Function:    scsi_init_io()
1091  *
1092  * Purpose:     SCSI I/O initialize function.
1093  *
1094  * Arguments:   cmd   - Command descriptor we wish to initialize
1095  *
1096  * Returns:     0 on success
1097  *              BLKPREP_DEFER if the failure is retryable
1098  *              BLKPREP_KILL if the failure is fatal
1099  */
1100 int scsi_init_io(struct scsi_cmnd *cmd, gfp_t gfp_mask)
1101 {
1102         int error = scsi_init_sgtable(cmd->request, &cmd->sdb, gfp_mask);
1103         if (error)
1104                 goto err_exit;
1105
1106         if (blk_bidi_rq(cmd->request)) {
1107                 struct scsi_data_buffer *bidi_sdb = kmem_cache_zalloc(
1108                         scsi_sdb_cache, GFP_ATOMIC);
1109                 if (!bidi_sdb) {
1110                         error = BLKPREP_DEFER;
1111                         goto err_exit;
1112                 }
1113
1114                 cmd->request->next_rq->special = bidi_sdb;
1115                 error = scsi_init_sgtable(cmd->request->next_rq, bidi_sdb,
1116                                                                     GFP_ATOMIC);
1117                 if (error)
1118                         goto err_exit;
1119         }
1120
1121         if (blk_integrity_rq(cmd->request)) {
1122                 struct scsi_data_buffer *prot_sdb = cmd->prot_sdb;
1123                 int ivecs, count;
1124
1125                 BUG_ON(prot_sdb == NULL);
1126                 ivecs = blk_rq_count_integrity_sg(cmd->request);
1127
1128                 if (scsi_alloc_sgtable(prot_sdb, ivecs, gfp_mask)) {
1129                         error = BLKPREP_DEFER;
1130                         goto err_exit;
1131                 }
1132
1133                 count = blk_rq_map_integrity_sg(cmd->request,
1134                                                 prot_sdb->table.sgl);
1135                 BUG_ON(unlikely(count > ivecs));
1136
1137                 cmd->prot_sdb = prot_sdb;
1138                 cmd->prot_sdb->table.nents = count;
1139         }
1140
1141         return BLKPREP_OK ;
1142
1143 err_exit:
1144         scsi_release_buffers(cmd);
1145         if (error == BLKPREP_KILL)
1146                 scsi_put_command(cmd);
1147         else /* BLKPREP_DEFER */
1148                 scsi_unprep_request(cmd->request);
1149
1150         return error;
1151 }
1152 EXPORT_SYMBOL(scsi_init_io);
1153
1154 static struct scsi_cmnd *scsi_get_cmd_from_req(struct scsi_device *sdev,
1155                 struct request *req)
1156 {
1157         struct scsi_cmnd *cmd;
1158
1159         if (!req->special) {
1160                 cmd = scsi_get_command(sdev, GFP_ATOMIC);
1161                 if (unlikely(!cmd))
1162                         return NULL;
1163                 req->special = cmd;
1164         } else {
1165                 cmd = req->special;
1166         }
1167
1168         /* pull a tag out of the request if we have one */
1169         cmd->tag = req->tag;
1170         cmd->request = req;
1171
1172         cmd->cmnd = req->cmd;
1173
1174         return cmd;
1175 }
1176
1177 int scsi_setup_blk_pc_cmnd(struct scsi_device *sdev, struct request *req)
1178 {
1179         struct scsi_cmnd *cmd;
1180         int ret = scsi_prep_state_check(sdev, req);
1181
1182         if (ret != BLKPREP_OK)
1183                 return ret;
1184
1185         cmd = scsi_get_cmd_from_req(sdev, req);
1186         if (unlikely(!cmd))
1187                 return BLKPREP_DEFER;
1188
1189         /*
1190          * BLOCK_PC requests may transfer data, in which case they must
1191          * a bio attached to them.  Or they might contain a SCSI command
1192          * that does not transfer data, in which case they may optionally
1193          * submit a request without an attached bio.
1194          */
1195         if (req->bio) {
1196                 int ret;
1197
1198                 BUG_ON(!req->nr_phys_segments);
1199
1200                 ret = scsi_init_io(cmd, GFP_ATOMIC);
1201                 if (unlikely(ret))
1202                         return ret;
1203         } else {
1204                 BUG_ON(req->data_len);
1205                 BUG_ON(req->data);
1206
1207                 memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1208                 req->buffer = NULL;
1209         }
1210
1211         cmd->cmd_len = req->cmd_len;
1212         if (!req->data_len)
1213                 cmd->sc_data_direction = DMA_NONE;
1214         else if (rq_data_dir(req) == WRITE)
1215                 cmd->sc_data_direction = DMA_TO_DEVICE;
1216         else
1217                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1218         
1219         cmd->transfersize = req->data_len;
1220         cmd->allowed = req->retries;
1221         return BLKPREP_OK;
1222 }
1223 EXPORT_SYMBOL(scsi_setup_blk_pc_cmnd);
1224
1225 /*
1226  * Setup a REQ_TYPE_FS command.  These are simple read/write request
1227  * from filesystems that still need to be translated to SCSI CDBs from
1228  * the ULD.
1229  */
1230 int scsi_setup_fs_cmnd(struct scsi_device *sdev, struct request *req)
1231 {
1232         struct scsi_cmnd *cmd;
1233         int ret = scsi_prep_state_check(sdev, req);
1234
1235         if (ret != BLKPREP_OK)
1236                 return ret;
1237
1238         if (unlikely(sdev->scsi_dh_data && sdev->scsi_dh_data->scsi_dh
1239                          && sdev->scsi_dh_data->scsi_dh->prep_fn)) {
1240                 ret = sdev->scsi_dh_data->scsi_dh->prep_fn(sdev, req);
1241                 if (ret != BLKPREP_OK)
1242                         return ret;
1243         }
1244
1245         /*
1246          * Filesystem requests must transfer data.
1247          */
1248         BUG_ON(!req->nr_phys_segments);
1249
1250         cmd = scsi_get_cmd_from_req(sdev, req);
1251         if (unlikely(!cmd))
1252                 return BLKPREP_DEFER;
1253
1254         memset(cmd->cmnd, 0, BLK_MAX_CDB);
1255         return scsi_init_io(cmd, GFP_ATOMIC);
1256 }
1257 EXPORT_SYMBOL(scsi_setup_fs_cmnd);
1258
1259 int scsi_prep_state_check(struct scsi_device *sdev, struct request *req)
1260 {
1261         int ret = BLKPREP_OK;
1262
1263         /*
1264          * If the device is not in running state we will reject some
1265          * or all commands.
1266          */
1267         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1268                 switch (sdev->sdev_state) {
1269                 case SDEV_OFFLINE:
1270                         /*
1271                          * If the device is offline we refuse to process any
1272                          * commands.  The device must be brought online
1273                          * before trying any recovery commands.
1274                          */
1275                         sdev_printk(KERN_ERR, sdev,
1276                                     "rejecting I/O to offline device\n");
1277                         ret = BLKPREP_KILL;
1278                         break;
1279                 case SDEV_DEL:
1280                         /*
1281                          * If the device is fully deleted, we refuse to
1282                          * process any commands as well.
1283                          */
1284                         sdev_printk(KERN_ERR, sdev,
1285                                     "rejecting I/O to dead device\n");
1286                         ret = BLKPREP_KILL;
1287                         break;
1288                 case SDEV_QUIESCE:
1289                 case SDEV_BLOCK:
1290                 case SDEV_CREATED_BLOCK:
1291                         /*
1292                          * If the devices is blocked we defer normal commands.
1293                          */
1294                         if (!(req->cmd_flags & REQ_PREEMPT))
1295                                 ret = BLKPREP_DEFER;
1296                         break;
1297                 default:
1298                         /*
1299                          * For any other not fully online state we only allow
1300                          * special commands.  In particular any user initiated
1301                          * command is not allowed.
1302                          */
1303                         if (!(req->cmd_flags & REQ_PREEMPT))
1304                                 ret = BLKPREP_KILL;
1305                         break;
1306                 }
1307         }
1308         return ret;
1309 }
1310 EXPORT_SYMBOL(scsi_prep_state_check);
1311
1312 int scsi_prep_return(struct request_queue *q, struct request *req, int ret)
1313 {
1314         struct scsi_device *sdev = q->queuedata;
1315
1316         switch (ret) {
1317         case BLKPREP_KILL:
1318                 req->errors = DID_NO_CONNECT << 16;
1319                 /* release the command and kill it */
1320                 if (req->special) {
1321                         struct scsi_cmnd *cmd = req->special;
1322                         scsi_release_buffers(cmd);
1323                         scsi_put_command(cmd);
1324                         req->special = NULL;
1325                 }
1326                 break;
1327         case BLKPREP_DEFER:
1328                 /*
1329                  * If we defer, the elv_next_request() returns NULL, but the
1330                  * queue must be restarted, so we plug here if no returning
1331                  * command will automatically do that.
1332                  */
1333                 if (sdev->device_busy == 0)
1334                         blk_plug_device(q);
1335                 break;
1336         default:
1337                 req->cmd_flags |= REQ_DONTPREP;
1338         }
1339
1340         return ret;
1341 }
1342 EXPORT_SYMBOL(scsi_prep_return);
1343
1344 int scsi_prep_fn(struct request_queue *q, struct request *req)
1345 {
1346         struct scsi_device *sdev = q->queuedata;
1347         int ret = BLKPREP_KILL;
1348
1349         if (req->cmd_type == REQ_TYPE_BLOCK_PC)
1350                 ret = scsi_setup_blk_pc_cmnd(sdev, req);
1351         return scsi_prep_return(q, req, ret);
1352 }
1353
1354 /*
1355  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1356  * return 0.
1357  *
1358  * Called with the queue_lock held.
1359  */
1360 static inline int scsi_dev_queue_ready(struct request_queue *q,
1361                                   struct scsi_device *sdev)
1362 {
1363         if (sdev->device_busy == 0 && sdev->device_blocked) {
1364                 /*
1365                  * unblock after device_blocked iterates to zero
1366                  */
1367                 if (--sdev->device_blocked == 0) {
1368                         SCSI_LOG_MLQUEUE(3,
1369                                    sdev_printk(KERN_INFO, sdev,
1370                                    "unblocking device at zero depth\n"));
1371                 } else {
1372                         blk_plug_device(q);
1373                         return 0;
1374                 }
1375         }
1376         if (scsi_device_is_busy(sdev))
1377                 return 0;
1378
1379         return 1;
1380 }
1381
1382
1383 /*
1384  * scsi_target_queue_ready: checks if there we can send commands to target
1385  * @sdev: scsi device on starget to check.
1386  *
1387  * Called with the host lock held.
1388  */
1389 static inline int scsi_target_queue_ready(struct Scsi_Host *shost,
1390                                            struct scsi_device *sdev)
1391 {
1392         struct scsi_target *starget = scsi_target(sdev);
1393
1394         if (starget->single_lun) {
1395                 if (starget->starget_sdev_user &&
1396                     starget->starget_sdev_user != sdev)
1397                         return 0;
1398                 starget->starget_sdev_user = sdev;
1399         }
1400
1401         if (starget->target_busy == 0 && starget->target_blocked) {
1402                 /*
1403                  * unblock after target_blocked iterates to zero
1404                  */
1405                 if (--starget->target_blocked == 0) {
1406                         SCSI_LOG_MLQUEUE(3, starget_printk(KERN_INFO, starget,
1407                                          "unblocking target at zero depth\n"));
1408                 } else {
1409                         blk_plug_device(sdev->request_queue);
1410                         return 0;
1411                 }
1412         }
1413
1414         if (scsi_target_is_busy(starget)) {
1415                 if (list_empty(&sdev->starved_entry)) {
1416                         list_add_tail(&sdev->starved_entry,
1417                                       &shost->starved_list);
1418                         return 0;
1419                 }
1420         }
1421
1422         /* We're OK to process the command, so we can't be starved */
1423         if (!list_empty(&sdev->starved_entry))
1424                 list_del_init(&sdev->starved_entry);
1425         return 1;
1426 }
1427
1428 /*
1429  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1430  * return 0. We must end up running the queue again whenever 0 is
1431  * returned, else IO can hang.
1432  *
1433  * Called with host_lock held.
1434  */
1435 static inline int scsi_host_queue_ready(struct request_queue *q,
1436                                    struct Scsi_Host *shost,
1437                                    struct scsi_device *sdev)
1438 {
1439         if (scsi_host_in_recovery(shost))
1440                 return 0;
1441         if (shost->host_busy == 0 && shost->host_blocked) {
1442                 /*
1443                  * unblock after host_blocked iterates to zero
1444                  */
1445                 if (--shost->host_blocked == 0) {
1446                         SCSI_LOG_MLQUEUE(3,
1447                                 printk("scsi%d unblocking host at zero depth\n",
1448                                         shost->host_no));
1449                 } else {
1450                         return 0;
1451                 }
1452         }
1453         if (scsi_host_is_busy(shost)) {
1454                 if (list_empty(&sdev->starved_entry))
1455                         list_add_tail(&sdev->starved_entry, &shost->starved_list);
1456                 return 0;
1457         }
1458
1459         /* We're OK to process the command, so we can't be starved */
1460         if (!list_empty(&sdev->starved_entry))
1461                 list_del_init(&sdev->starved_entry);
1462
1463         return 1;
1464 }
1465
1466 /*
1467  * Busy state exporting function for request stacking drivers.
1468  *
1469  * For efficiency, no lock is taken to check the busy state of
1470  * shost/starget/sdev, since the returned value is not guaranteed and
1471  * may be changed after request stacking drivers call the function,
1472  * regardless of taking lock or not.
1473  *
1474  * When scsi can't dispatch I/Os anymore and needs to kill I/Os
1475  * (e.g. !sdev), scsi needs to return 'not busy'.
1476  * Otherwise, request stacking drivers may hold requests forever.
1477  */
1478 static int scsi_lld_busy(struct request_queue *q)
1479 {
1480         struct scsi_device *sdev = q->queuedata;
1481         struct Scsi_Host *shost;
1482         struct scsi_target *starget;
1483
1484         if (!sdev)
1485                 return 0;
1486
1487         shost = sdev->host;
1488         starget = scsi_target(sdev);
1489
1490         if (scsi_host_in_recovery(shost) || scsi_host_is_busy(shost) ||
1491             scsi_target_is_busy(starget) || scsi_device_is_busy(sdev))
1492                 return 1;
1493
1494         return 0;
1495 }
1496
1497 /*
1498  * Kill a request for a dead device
1499  */
1500 static void scsi_kill_request(struct request *req, struct request_queue *q)
1501 {
1502         struct scsi_cmnd *cmd = req->special;
1503         struct scsi_device *sdev = cmd->device;
1504         struct scsi_target *starget = scsi_target(sdev);
1505         struct Scsi_Host *shost = sdev->host;
1506
1507         blkdev_dequeue_request(req);
1508
1509         if (unlikely(cmd == NULL)) {
1510                 printk(KERN_CRIT "impossible request in %s.\n",
1511                                  __func__);
1512                 BUG();
1513         }
1514
1515         scsi_init_cmd_errh(cmd);
1516         cmd->result = DID_NO_CONNECT << 16;
1517         atomic_inc(&cmd->device->iorequest_cnt);
1518
1519         /*
1520          * SCSI request completion path will do scsi_device_unbusy(),
1521          * bump busy counts.  To bump the counters, we need to dance
1522          * with the locks as normal issue path does.
1523          */
1524         sdev->device_busy++;
1525         spin_unlock(sdev->request_queue->queue_lock);
1526         spin_lock(shost->host_lock);
1527         shost->host_busy++;
1528         starget->target_busy++;
1529         spin_unlock(shost->host_lock);
1530         spin_lock(sdev->request_queue->queue_lock);
1531
1532         blk_complete_request(req);
1533 }
1534
1535 static void scsi_softirq_done(struct request *rq)
1536 {
1537         struct scsi_cmnd *cmd = rq->special;
1538         unsigned long wait_for = (cmd->allowed + 1) * rq->timeout;
1539         int disposition;
1540
1541         INIT_LIST_HEAD(&cmd->eh_entry);
1542
1543         /*
1544          * Set the serial numbers back to zero
1545          */
1546         cmd->serial_number = 0;
1547
1548         atomic_inc(&cmd->device->iodone_cnt);
1549         if (cmd->result)
1550                 atomic_inc(&cmd->device->ioerr_cnt);
1551
1552         disposition = scsi_decide_disposition(cmd);
1553         if (disposition != SUCCESS &&
1554             time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1555                 sdev_printk(KERN_ERR, cmd->device,
1556                             "timing out command, waited %lus\n",
1557                             wait_for/HZ);
1558                 disposition = SUCCESS;
1559         }
1560                         
1561         scsi_log_completion(cmd, disposition);
1562
1563         switch (disposition) {
1564                 case SUCCESS:
1565                         scsi_finish_command(cmd);
1566                         break;
1567                 case NEEDS_RETRY:
1568                         scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1569                         break;
1570                 case ADD_TO_MLQUEUE:
1571                         scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1572                         break;
1573                 default:
1574                         if (!scsi_eh_scmd_add(cmd, 0))
1575                                 scsi_finish_command(cmd);
1576         }
1577 }
1578
1579 /*
1580  * Function:    scsi_request_fn()
1581  *
1582  * Purpose:     Main strategy routine for SCSI.
1583  *
1584  * Arguments:   q       - Pointer to actual queue.
1585  *
1586  * Returns:     Nothing
1587  *
1588  * Lock status: IO request lock assumed to be held when called.
1589  */
1590 static void scsi_request_fn(struct request_queue *q)
1591 {
1592         struct scsi_device *sdev = q->queuedata;
1593         struct Scsi_Host *shost;
1594         struct scsi_cmnd *cmd;
1595         struct request *req;
1596
1597         if (!sdev) {
1598                 printk("scsi: killing requests for dead queue\n");
1599                 while ((req = elv_next_request(q)) != NULL)
1600                         scsi_kill_request(req, q);
1601                 return;
1602         }
1603
1604         if(!get_device(&sdev->sdev_gendev))
1605                 /* We must be tearing the block queue down already */
1606                 return;
1607
1608         /*
1609          * To start with, we keep looping until the queue is empty, or until
1610          * the host is no longer able to accept any more requests.
1611          */
1612         shost = sdev->host;
1613         while (!blk_queue_plugged(q)) {
1614                 int rtn;
1615                 /*
1616                  * get next queueable request.  We do this early to make sure
1617                  * that the request is fully prepared even if we cannot 
1618                  * accept it.
1619                  */
1620                 req = elv_next_request(q);
1621                 if (!req || !scsi_dev_queue_ready(q, sdev))
1622                         break;
1623
1624                 if (unlikely(!scsi_device_online(sdev))) {
1625                         sdev_printk(KERN_ERR, sdev,
1626                                     "rejecting I/O to offline device\n");
1627                         scsi_kill_request(req, q);
1628                         continue;
1629                 }
1630
1631
1632                 /*
1633                  * Remove the request from the request list.
1634                  */
1635                 if (!(blk_queue_tagged(q) && !blk_queue_start_tag(q, req)))
1636                         blkdev_dequeue_request(req);
1637                 sdev->device_busy++;
1638
1639                 spin_unlock(q->queue_lock);
1640                 cmd = req->special;
1641                 if (unlikely(cmd == NULL)) {
1642                         printk(KERN_CRIT "impossible request in %s.\n"
1643                                          "please mail a stack trace to "
1644                                          "linux-scsi@vger.kernel.org\n",
1645                                          __func__);
1646                         blk_dump_rq_flags(req, "foo");
1647                         BUG();
1648                 }
1649                 spin_lock(shost->host_lock);
1650
1651                 /*
1652                  * We hit this when the driver is using a host wide
1653                  * tag map. For device level tag maps the queue_depth check
1654                  * in the device ready fn would prevent us from trying
1655                  * to allocate a tag. Since the map is a shared host resource
1656                  * we add the dev to the starved list so it eventually gets
1657                  * a run when a tag is freed.
1658                  */
1659                 if (blk_queue_tagged(q) && !blk_rq_tagged(req)) {
1660                         if (list_empty(&sdev->starved_entry))
1661                                 list_add_tail(&sdev->starved_entry,
1662                                               &shost->starved_list);
1663                         goto not_ready;
1664                 }
1665
1666                 if (!scsi_target_queue_ready(shost, sdev))
1667                         goto not_ready;
1668
1669                 if (!scsi_host_queue_ready(q, shost, sdev))
1670                         goto not_ready;
1671
1672                 scsi_target(sdev)->target_busy++;
1673                 shost->host_busy++;
1674
1675                 /*
1676                  * XXX(hch): This is rather suboptimal, scsi_dispatch_cmd will
1677                  *              take the lock again.
1678                  */
1679                 spin_unlock_irq(shost->host_lock);
1680
1681                 /*
1682                  * Finally, initialize any error handling parameters, and set up
1683                  * the timers for timeouts.
1684                  */
1685                 scsi_init_cmd_errh(cmd);
1686
1687                 /*
1688                  * Dispatch the command to the low-level driver.
1689                  */
1690                 rtn = scsi_dispatch_cmd(cmd);
1691                 spin_lock_irq(q->queue_lock);
1692                 if(rtn) {
1693                         /* we're refusing the command; because of
1694                          * the way locks get dropped, we need to 
1695                          * check here if plugging is required */
1696                         if(sdev->device_busy == 0)
1697                                 blk_plug_device(q);
1698
1699                         break;
1700                 }
1701         }
1702
1703         goto out;
1704
1705  not_ready:
1706         spin_unlock_irq(shost->host_lock);
1707
1708         /*
1709          * lock q, handle tag, requeue req, and decrement device_busy. We
1710          * must return with queue_lock held.
1711          *
1712          * Decrementing device_busy without checking it is OK, as all such
1713          * cases (host limits or settings) should run the queue at some
1714          * later time.
1715          */
1716         spin_lock_irq(q->queue_lock);
1717         blk_requeue_request(q, req);
1718         sdev->device_busy--;
1719         if(sdev->device_busy == 0)
1720                 blk_plug_device(q);
1721  out:
1722         /* must be careful here...if we trigger the ->remove() function
1723          * we cannot be holding the q lock */
1724         spin_unlock_irq(q->queue_lock);
1725         put_device(&sdev->sdev_gendev);
1726         spin_lock_irq(q->queue_lock);
1727 }
1728
1729 u64 scsi_calculate_bounce_limit(struct Scsi_Host *shost)
1730 {
1731         struct device *host_dev;
1732         u64 bounce_limit = 0xffffffff;
1733
1734         if (shost->unchecked_isa_dma)
1735                 return BLK_BOUNCE_ISA;
1736         /*
1737          * Platforms with virtual-DMA translation
1738          * hardware have no practical limit.
1739          */
1740         if (!PCI_DMA_BUS_IS_PHYS)
1741                 return BLK_BOUNCE_ANY;
1742
1743         host_dev = scsi_get_device(shost);
1744         if (host_dev && host_dev->dma_mask)
1745                 bounce_limit = *host_dev->dma_mask;
1746
1747         return bounce_limit;
1748 }
1749 EXPORT_SYMBOL(scsi_calculate_bounce_limit);
1750
1751 struct request_queue *__scsi_alloc_queue(struct Scsi_Host *shost,
1752                                          request_fn_proc *request_fn)
1753 {
1754         struct request_queue *q;
1755         struct device *dev = shost->shost_gendev.parent;
1756
1757         q = blk_init_queue(request_fn, NULL);
1758         if (!q)
1759                 return NULL;
1760
1761         /*
1762          * this limit is imposed by hardware restrictions
1763          */
1764         blk_queue_max_hw_segments(q, shost->sg_tablesize);
1765         blk_queue_max_phys_segments(q, SCSI_MAX_SG_CHAIN_SEGMENTS);
1766
1767         blk_queue_max_sectors(q, shost->max_sectors);
1768         blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
1769         blk_queue_segment_boundary(q, shost->dma_boundary);
1770         dma_set_seg_boundary(dev, shost->dma_boundary);
1771
1772         blk_queue_max_segment_size(q, dma_get_max_seg_size(dev));
1773
1774         /* New queue, no concurrency on queue_flags */
1775         if (!shost->use_clustering)
1776                 queue_flag_clear_unlocked(QUEUE_FLAG_CLUSTER, q);
1777
1778         /*
1779          * set a reasonable default alignment on word boundaries: the
1780          * host and device may alter it using
1781          * blk_queue_update_dma_alignment() later.
1782          */
1783         blk_queue_dma_alignment(q, 0x03);
1784
1785         return q;
1786 }
1787 EXPORT_SYMBOL(__scsi_alloc_queue);
1788
1789 struct request_queue *scsi_alloc_queue(struct scsi_device *sdev)
1790 {
1791         struct request_queue *q;
1792
1793         q = __scsi_alloc_queue(sdev->host, scsi_request_fn);
1794         if (!q)
1795                 return NULL;
1796
1797         blk_queue_prep_rq(q, scsi_prep_fn);
1798         blk_queue_softirq_done(q, scsi_softirq_done);
1799         blk_queue_rq_timed_out(q, scsi_times_out);
1800         blk_queue_lld_busy(q, scsi_lld_busy);
1801         return q;
1802 }
1803
1804 void scsi_free_queue(struct request_queue *q)
1805 {
1806         blk_cleanup_queue(q);
1807 }
1808
1809 /*
1810  * Function:    scsi_block_requests()
1811  *
1812  * Purpose:     Utility function used by low-level drivers to prevent further
1813  *              commands from being queued to the device.
1814  *
1815  * Arguments:   shost       - Host in question
1816  *
1817  * Returns:     Nothing
1818  *
1819  * Lock status: No locks are assumed held.
1820  *
1821  * Notes:       There is no timer nor any other means by which the requests
1822  *              get unblocked other than the low-level driver calling
1823  *              scsi_unblock_requests().
1824  */
1825 void scsi_block_requests(struct Scsi_Host *shost)
1826 {
1827         shost->host_self_blocked = 1;
1828 }
1829 EXPORT_SYMBOL(scsi_block_requests);
1830
1831 /*
1832  * Function:    scsi_unblock_requests()
1833  *
1834  * Purpose:     Utility function used by low-level drivers to allow further
1835  *              commands from being queued to the device.
1836  *
1837  * Arguments:   shost       - Host in question
1838  *
1839  * Returns:     Nothing
1840  *
1841  * Lock status: No locks are assumed held.
1842  *
1843  * Notes:       There is no timer nor any other means by which the requests
1844  *              get unblocked other than the low-level driver calling
1845  *              scsi_unblock_requests().
1846  *
1847  *              This is done as an API function so that changes to the
1848  *              internals of the scsi mid-layer won't require wholesale
1849  *              changes to drivers that use this feature.
1850  */
1851 void scsi_unblock_requests(struct Scsi_Host *shost)
1852 {
1853         shost->host_self_blocked = 0;
1854         scsi_run_host_queues(shost);
1855 }
1856 EXPORT_SYMBOL(scsi_unblock_requests);
1857
1858 int __init scsi_init_queue(void)
1859 {
1860         int i;
1861
1862         scsi_io_context_cache = kmem_cache_create("scsi_io_context",
1863                                         sizeof(struct scsi_io_context),
1864                                         0, 0, NULL);
1865         if (!scsi_io_context_cache) {
1866                 printk(KERN_ERR "SCSI: can't init scsi io context cache\n");
1867                 return -ENOMEM;
1868         }
1869
1870         scsi_sdb_cache = kmem_cache_create("scsi_data_buffer",
1871                                            sizeof(struct scsi_data_buffer),
1872                                            0, 0, NULL);
1873         if (!scsi_sdb_cache) {
1874                 printk(KERN_ERR "SCSI: can't init scsi sdb cache\n");
1875                 goto cleanup_io_context;
1876         }
1877
1878         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1879                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1880                 int size = sgp->size * sizeof(struct scatterlist);
1881
1882                 sgp->slab = kmem_cache_create(sgp->name, size, 0,
1883                                 SLAB_HWCACHE_ALIGN, NULL);
1884                 if (!sgp->slab) {
1885                         printk(KERN_ERR "SCSI: can't init sg slab %s\n",
1886                                         sgp->name);
1887                         goto cleanup_sdb;
1888                 }
1889
1890                 sgp->pool = mempool_create_slab_pool(SG_MEMPOOL_SIZE,
1891                                                      sgp->slab);
1892                 if (!sgp->pool) {
1893                         printk(KERN_ERR "SCSI: can't init sg mempool %s\n",
1894                                         sgp->name);
1895                         goto cleanup_sdb;
1896                 }
1897         }
1898
1899         return 0;
1900
1901 cleanup_sdb:
1902         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1903                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1904                 if (sgp->pool)
1905                         mempool_destroy(sgp->pool);
1906                 if (sgp->slab)
1907                         kmem_cache_destroy(sgp->slab);
1908         }
1909         kmem_cache_destroy(scsi_sdb_cache);
1910 cleanup_io_context:
1911         kmem_cache_destroy(scsi_io_context_cache);
1912
1913         return -ENOMEM;
1914 }
1915
1916 void scsi_exit_queue(void)
1917 {
1918         int i;
1919
1920         kmem_cache_destroy(scsi_io_context_cache);
1921         kmem_cache_destroy(scsi_sdb_cache);
1922
1923         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1924                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1925                 mempool_destroy(sgp->pool);
1926                 kmem_cache_destroy(sgp->slab);
1927         }
1928 }
1929
1930 /**
1931  *      scsi_mode_select - issue a mode select
1932  *      @sdev:  SCSI device to be queried
1933  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
1934  *      @sp:    Save page bit (0 == don't save, 1 == save)
1935  *      @modepage: mode page being requested
1936  *      @buffer: request buffer (may not be smaller than eight bytes)
1937  *      @len:   length of request buffer.
1938  *      @timeout: command timeout
1939  *      @retries: number of retries before failing
1940  *      @data: returns a structure abstracting the mode header data
1941  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
1942  *              must be SCSI_SENSE_BUFFERSIZE big.
1943  *
1944  *      Returns zero if successful; negative error number or scsi
1945  *      status on error
1946  *
1947  */
1948 int
1949 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
1950                  unsigned char *buffer, int len, int timeout, int retries,
1951                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1952 {
1953         unsigned char cmd[10];
1954         unsigned char *real_buffer;
1955         int ret;
1956
1957         memset(cmd, 0, sizeof(cmd));
1958         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
1959
1960         if (sdev->use_10_for_ms) {
1961                 if (len > 65535)
1962                         return -EINVAL;
1963                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
1964                 if (!real_buffer)
1965                         return -ENOMEM;
1966                 memcpy(real_buffer + 8, buffer, len);
1967                 len += 8;
1968                 real_buffer[0] = 0;
1969                 real_buffer[1] = 0;
1970                 real_buffer[2] = data->medium_type;
1971                 real_buffer[3] = data->device_specific;
1972                 real_buffer[4] = data->longlba ? 0x01 : 0;
1973                 real_buffer[5] = 0;
1974                 real_buffer[6] = data->block_descriptor_length >> 8;
1975                 real_buffer[7] = data->block_descriptor_length;
1976
1977                 cmd[0] = MODE_SELECT_10;
1978                 cmd[7] = len >> 8;
1979                 cmd[8] = len;
1980         } else {
1981                 if (len > 255 || data->block_descriptor_length > 255 ||
1982                     data->longlba)
1983                         return -EINVAL;
1984
1985                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
1986                 if (!real_buffer)
1987                         return -ENOMEM;
1988                 memcpy(real_buffer + 4, buffer, len);
1989                 len += 4;
1990                 real_buffer[0] = 0;
1991                 real_buffer[1] = data->medium_type;
1992                 real_buffer[2] = data->device_specific;
1993                 real_buffer[3] = data->block_descriptor_length;
1994                 
1995
1996                 cmd[0] = MODE_SELECT;
1997                 cmd[4] = len;
1998         }
1999
2000         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
2001                                sshdr, timeout, retries);
2002         kfree(real_buffer);
2003         return ret;
2004 }
2005 EXPORT_SYMBOL_GPL(scsi_mode_select);
2006
2007 /**
2008  *      scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
2009  *      @sdev:  SCSI device to be queried
2010  *      @dbd:   set if mode sense will allow block descriptors to be returned
2011  *      @modepage: mode page being requested
2012  *      @buffer: request buffer (may not be smaller than eight bytes)
2013  *      @len:   length of request buffer.
2014  *      @timeout: command timeout
2015  *      @retries: number of retries before failing
2016  *      @data: returns a structure abstracting the mode header data
2017  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
2018  *              must be SCSI_SENSE_BUFFERSIZE big.
2019  *
2020  *      Returns zero if unsuccessful, or the header offset (either 4
2021  *      or 8 depending on whether a six or ten byte command was
2022  *      issued) if successful.
2023  */
2024 int
2025 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
2026                   unsigned char *buffer, int len, int timeout, int retries,
2027                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2028 {
2029         unsigned char cmd[12];
2030         int use_10_for_ms;
2031         int header_length;
2032         int result;
2033         struct scsi_sense_hdr my_sshdr;
2034
2035         memset(data, 0, sizeof(*data));
2036         memset(&cmd[0], 0, 12);
2037         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
2038         cmd[2] = modepage;
2039
2040         /* caller might not be interested in sense, but we need it */
2041         if (!sshdr)
2042                 sshdr = &my_sshdr;
2043
2044  retry:
2045         use_10_for_ms = sdev->use_10_for_ms;
2046
2047         if (use_10_for_ms) {
2048                 if (len < 8)
2049                         len = 8;
2050
2051                 cmd[0] = MODE_SENSE_10;
2052                 cmd[8] = len;
2053                 header_length = 8;
2054         } else {
2055                 if (len < 4)
2056                         len = 4;
2057
2058                 cmd[0] = MODE_SENSE;
2059                 cmd[4] = len;
2060                 header_length = 4;
2061         }
2062
2063         memset(buffer, 0, len);
2064
2065         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
2066                                   sshdr, timeout, retries);
2067
2068         /* This code looks awful: what it's doing is making sure an
2069          * ILLEGAL REQUEST sense return identifies the actual command
2070          * byte as the problem.  MODE_SENSE commands can return
2071          * ILLEGAL REQUEST if the code page isn't supported */
2072
2073         if (use_10_for_ms && !scsi_status_is_good(result) &&
2074             (driver_byte(result) & DRIVER_SENSE)) {
2075                 if (scsi_sense_valid(sshdr)) {
2076                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
2077                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
2078                                 /* 
2079                                  * Invalid command operation code
2080                                  */
2081                                 sdev->use_10_for_ms = 0;
2082                                 goto retry;
2083                         }
2084                 }
2085         }
2086
2087         if(scsi_status_is_good(result)) {
2088                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
2089                              (modepage == 6 || modepage == 8))) {
2090                         /* Initio breakage? */
2091                         header_length = 0;
2092                         data->length = 13;
2093                         data->medium_type = 0;
2094                         data->device_specific = 0;
2095                         data->longlba = 0;
2096                         data->block_descriptor_length = 0;
2097                 } else if(use_10_for_ms) {
2098                         data->length = buffer[0]*256 + buffer[1] + 2;
2099                         data->medium_type = buffer[2];
2100                         data->device_specific = buffer[3];
2101                         data->longlba = buffer[4] & 0x01;
2102                         data->block_descriptor_length = buffer[6]*256
2103                                 + buffer[7];
2104                 } else {
2105                         data->length = buffer[0] + 1;
2106                         data->medium_type = buffer[1];
2107                         data->device_specific = buffer[2];
2108                         data->block_descriptor_length = buffer[3];
2109                 }
2110                 data->header_length = header_length;
2111         }
2112
2113         return result;
2114 }
2115 EXPORT_SYMBOL(scsi_mode_sense);
2116
2117 /**
2118  *      scsi_test_unit_ready - test if unit is ready
2119  *      @sdev:  scsi device to change the state of.
2120  *      @timeout: command timeout
2121  *      @retries: number of retries before failing
2122  *      @sshdr_external: Optional pointer to struct scsi_sense_hdr for
2123  *              returning sense. Make sure that this is cleared before passing
2124  *              in.
2125  *
2126  *      Returns zero if unsuccessful or an error if TUR failed.  For
2127  *      removable media, a return of NOT_READY or UNIT_ATTENTION is
2128  *      translated to success, with the ->changed flag updated.
2129  **/
2130 int
2131 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
2132                      struct scsi_sense_hdr *sshdr_external)
2133 {
2134         char cmd[] = {
2135                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
2136         };
2137         struct scsi_sense_hdr *sshdr;
2138         int result;
2139
2140         if (!sshdr_external)
2141                 sshdr = kzalloc(sizeof(*sshdr), GFP_KERNEL);
2142         else
2143                 sshdr = sshdr_external;
2144
2145         /* try to eat the UNIT_ATTENTION if there are enough retries */
2146         do {
2147                 result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
2148                                           timeout, retries);
2149                 if (sdev->removable && scsi_sense_valid(sshdr) &&
2150                     sshdr->sense_key == UNIT_ATTENTION)
2151                         sdev->changed = 1;
2152         } while (scsi_sense_valid(sshdr) &&
2153                  sshdr->sense_key == UNIT_ATTENTION && --retries);
2154
2155         if (!sshdr)
2156                 /* could not allocate sense buffer, so can't process it */
2157                 return result;
2158
2159         if (sdev->removable && scsi_sense_valid(sshdr) &&
2160             (sshdr->sense_key == UNIT_ATTENTION ||
2161              sshdr->sense_key == NOT_READY)) {
2162                 sdev->changed = 1;
2163                 result = 0;
2164         }
2165         if (!sshdr_external)
2166                 kfree(sshdr);
2167         return result;
2168 }
2169 EXPORT_SYMBOL(scsi_test_unit_ready);
2170
2171 /**
2172  *      scsi_device_set_state - Take the given device through the device state model.
2173  *      @sdev:  scsi device to change the state of.
2174  *      @state: state to change to.
2175  *
2176  *      Returns zero if unsuccessful or an error if the requested 
2177  *      transition is illegal.
2178  */
2179 int
2180 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2181 {
2182         enum scsi_device_state oldstate = sdev->sdev_state;
2183
2184         if (state == oldstate)
2185                 return 0;
2186
2187         switch (state) {
2188         case SDEV_CREATED:
2189                 switch (oldstate) {
2190                 case SDEV_CREATED_BLOCK:
2191                         break;
2192                 default:
2193                         goto illegal;
2194                 }
2195                 break;
2196                         
2197         case SDEV_RUNNING:
2198                 switch (oldstate) {
2199                 case SDEV_CREATED:
2200                 case SDEV_OFFLINE:
2201                 case SDEV_QUIESCE:
2202                 case SDEV_BLOCK:
2203                         break;
2204                 default:
2205                         goto illegal;
2206                 }
2207                 break;
2208
2209         case SDEV_QUIESCE:
2210                 switch (oldstate) {
2211                 case SDEV_RUNNING:
2212                 case SDEV_OFFLINE:
2213                         break;
2214                 default:
2215                         goto illegal;
2216                 }
2217                 break;
2218
2219         case SDEV_OFFLINE:
2220                 switch (oldstate) {
2221                 case SDEV_CREATED:
2222                 case SDEV_RUNNING:
2223                 case SDEV_QUIESCE:
2224                 case SDEV_BLOCK:
2225                         break;
2226                 default:
2227                         goto illegal;
2228                 }
2229                 break;
2230
2231         case SDEV_BLOCK:
2232                 switch (oldstate) {
2233                 case SDEV_RUNNING:
2234                 case SDEV_CREATED_BLOCK:
2235                         break;
2236                 default:
2237                         goto illegal;
2238                 }
2239                 break;
2240
2241         case SDEV_CREATED_BLOCK:
2242                 switch (oldstate) {
2243                 case SDEV_CREATED:
2244                         break;
2245                 default:
2246                         goto illegal;
2247                 }
2248                 break;
2249
2250         case SDEV_CANCEL:
2251                 switch (oldstate) {
2252                 case SDEV_CREATED:
2253                 case SDEV_RUNNING:
2254                 case SDEV_QUIESCE:
2255                 case SDEV_OFFLINE:
2256                 case SDEV_BLOCK:
2257                         break;
2258                 default:
2259                         goto illegal;
2260                 }
2261                 break;
2262
2263         case SDEV_DEL:
2264                 switch (oldstate) {
2265                 case SDEV_CREATED:
2266                 case SDEV_RUNNING:
2267                 case SDEV_OFFLINE:
2268                 case SDEV_CANCEL:
2269                         break;
2270                 default:
2271                         goto illegal;
2272                 }
2273                 break;
2274
2275         }
2276         sdev->sdev_state = state;
2277         return 0;
2278
2279  illegal:
2280         SCSI_LOG_ERROR_RECOVERY(1, 
2281                                 sdev_printk(KERN_ERR, sdev,
2282                                             "Illegal state transition %s->%s\n",
2283                                             scsi_device_state_name(oldstate),
2284                                             scsi_device_state_name(state))
2285                                 );
2286         return -EINVAL;
2287 }
2288 EXPORT_SYMBOL(scsi_device_set_state);
2289
2290 /**
2291  *      sdev_evt_emit - emit a single SCSI device uevent
2292  *      @sdev: associated SCSI device
2293  *      @evt: event to emit
2294  *
2295  *      Send a single uevent (scsi_event) to the associated scsi_device.
2296  */
2297 static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2298 {
2299         int idx = 0;
2300         char *envp[3];
2301
2302         switch (evt->evt_type) {
2303         case SDEV_EVT_MEDIA_CHANGE:
2304                 envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2305                 break;
2306
2307         default:
2308                 /* do nothing */
2309                 break;
2310         }
2311
2312         envp[idx++] = NULL;
2313
2314         kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2315 }
2316
2317 /**
2318  *      sdev_evt_thread - send a uevent for each scsi event
2319  *      @work: work struct for scsi_device
2320  *
2321  *      Dispatch queued events to their associated scsi_device kobjects
2322  *      as uevents.
2323  */
2324 void scsi_evt_thread(struct work_struct *work)
2325 {
2326         struct scsi_device *sdev;
2327         LIST_HEAD(event_list);
2328
2329         sdev = container_of(work, struct scsi_device, event_work);
2330
2331         while (1) {
2332                 struct scsi_event *evt;
2333                 struct list_head *this, *tmp;
2334                 unsigned long flags;
2335
2336                 spin_lock_irqsave(&sdev->list_lock, flags);
2337                 list_splice_init(&sdev->event_list, &event_list);
2338                 spin_unlock_irqrestore(&sdev->list_lock, flags);
2339
2340                 if (list_empty(&event_list))
2341                         break;
2342
2343                 list_for_each_safe(this, tmp, &event_list) {
2344                         evt = list_entry(this, struct scsi_event, node);
2345                         list_del(&evt->node);
2346                         scsi_evt_emit(sdev, evt);
2347                         kfree(evt);
2348                 }
2349         }
2350 }
2351
2352 /**
2353  *      sdev_evt_send - send asserted event to uevent thread
2354  *      @sdev: scsi_device event occurred on
2355  *      @evt: event to send
2356  *
2357  *      Assert scsi device event asynchronously.
2358  */
2359 void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2360 {
2361         unsigned long flags;
2362
2363 #if 0
2364         /* FIXME: currently this check eliminates all media change events
2365          * for polled devices.  Need to update to discriminate between AN
2366          * and polled events */
2367         if (!test_bit(evt->evt_type, sdev->supported_events)) {
2368                 kfree(evt);
2369                 return;
2370         }
2371 #endif
2372
2373         spin_lock_irqsave(&sdev->list_lock, flags);
2374         list_add_tail(&evt->node, &sdev->event_list);
2375         schedule_work(&sdev->event_work);
2376         spin_unlock_irqrestore(&sdev->list_lock, flags);
2377 }
2378 EXPORT_SYMBOL_GPL(sdev_evt_send);
2379
2380 /**
2381  *      sdev_evt_alloc - allocate a new scsi event
2382  *      @evt_type: type of event to allocate
2383  *      @gfpflags: GFP flags for allocation
2384  *
2385  *      Allocates and returns a new scsi_event.
2386  */
2387 struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2388                                   gfp_t gfpflags)
2389 {
2390         struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2391         if (!evt)
2392                 return NULL;
2393
2394         evt->evt_type = evt_type;
2395         INIT_LIST_HEAD(&evt->node);
2396
2397         /* evt_type-specific initialization, if any */
2398         switch (evt_type) {
2399         case SDEV_EVT_MEDIA_CHANGE:
2400         default:
2401                 /* do nothing */
2402                 break;
2403         }
2404
2405         return evt;
2406 }
2407 EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2408
2409 /**
2410  *      sdev_evt_send_simple - send asserted event to uevent thread
2411  *      @sdev: scsi_device event occurred on
2412  *      @evt_type: type of event to send
2413  *      @gfpflags: GFP flags for allocation
2414  *
2415  *      Assert scsi device event asynchronously, given an event type.
2416  */
2417 void sdev_evt_send_simple(struct scsi_device *sdev,
2418                           enum scsi_device_event evt_type, gfp_t gfpflags)
2419 {
2420         struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2421         if (!evt) {
2422                 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2423                             evt_type);
2424                 return;
2425         }
2426
2427         sdev_evt_send(sdev, evt);
2428 }
2429 EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2430
2431 /**
2432  *      scsi_device_quiesce - Block user issued commands.
2433  *      @sdev:  scsi device to quiesce.
2434  *
2435  *      This works by trying to transition to the SDEV_QUIESCE state
2436  *      (which must be a legal transition).  When the device is in this
2437  *      state, only special requests will be accepted, all others will
2438  *      be deferred.  Since special requests may also be requeued requests,
2439  *      a successful return doesn't guarantee the device will be 
2440  *      totally quiescent.
2441  *
2442  *      Must be called with user context, may sleep.
2443  *
2444  *      Returns zero if unsuccessful or an error if not.
2445  */
2446 int
2447 scsi_device_quiesce(struct scsi_device *sdev)
2448 {
2449         int err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2450         if (err)
2451                 return err;
2452
2453         scsi_run_queue(sdev->request_queue);
2454         while (sdev->device_busy) {
2455                 msleep_interruptible(200);
2456                 scsi_run_queue(sdev->request_queue);
2457         }
2458         return 0;
2459 }
2460 EXPORT_SYMBOL(scsi_device_quiesce);
2461
2462 /**
2463  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2464  *      @sdev:  scsi device to resume.
2465  *
2466  *      Moves the device from quiesced back to running and restarts the
2467  *      queues.
2468  *
2469  *      Must be called with user context, may sleep.
2470  */
2471 void
2472 scsi_device_resume(struct scsi_device *sdev)
2473 {
2474         if(scsi_device_set_state(sdev, SDEV_RUNNING))
2475                 return;
2476         scsi_run_queue(sdev->request_queue);
2477 }
2478 EXPORT_SYMBOL(scsi_device_resume);
2479
2480 static void
2481 device_quiesce_fn(struct scsi_device *sdev, void *data)
2482 {
2483         scsi_device_quiesce(sdev);
2484 }
2485
2486 void
2487 scsi_target_quiesce(struct scsi_target *starget)
2488 {
2489         starget_for_each_device(starget, NULL, device_quiesce_fn);
2490 }
2491 EXPORT_SYMBOL(scsi_target_quiesce);
2492
2493 static void
2494 device_resume_fn(struct scsi_device *sdev, void *data)
2495 {
2496         scsi_device_resume(sdev);
2497 }
2498
2499 void
2500 scsi_target_resume(struct scsi_target *starget)
2501 {
2502         starget_for_each_device(starget, NULL, device_resume_fn);
2503 }
2504 EXPORT_SYMBOL(scsi_target_resume);
2505
2506 /**
2507  * scsi_internal_device_block - internal function to put a device temporarily into the SDEV_BLOCK state
2508  * @sdev:       device to block
2509  *
2510  * Block request made by scsi lld's to temporarily stop all
2511  * scsi commands on the specified device.  Called from interrupt
2512  * or normal process context.
2513  *
2514  * Returns zero if successful or error if not
2515  *
2516  * Notes:       
2517  *      This routine transitions the device to the SDEV_BLOCK state
2518  *      (which must be a legal transition).  When the device is in this
2519  *      state, all commands are deferred until the scsi lld reenables
2520  *      the device with scsi_device_unblock or device_block_tmo fires.
2521  *      This routine assumes the host_lock is held on entry.
2522  */
2523 int
2524 scsi_internal_device_block(struct scsi_device *sdev)
2525 {
2526         struct request_queue *q = sdev->request_queue;
2527         unsigned long flags;
2528         int err = 0;
2529
2530         err = scsi_device_set_state(sdev, SDEV_BLOCK);
2531         if (err) {
2532                 err = scsi_device_set_state(sdev, SDEV_CREATED_BLOCK);
2533
2534                 if (err)
2535                         return err;
2536         }
2537
2538         /* 
2539          * The device has transitioned to SDEV_BLOCK.  Stop the
2540          * block layer from calling the midlayer with this device's
2541          * request queue. 
2542          */
2543         spin_lock_irqsave(q->queue_lock, flags);
2544         blk_stop_queue(q);
2545         spin_unlock_irqrestore(q->queue_lock, flags);
2546
2547         return 0;
2548 }
2549 EXPORT_SYMBOL_GPL(scsi_internal_device_block);
2550  
2551 /**
2552  * scsi_internal_device_unblock - resume a device after a block request
2553  * @sdev:       device to resume
2554  *
2555  * Called by scsi lld's or the midlayer to restart the device queue
2556  * for the previously suspended scsi device.  Called from interrupt or
2557  * normal process context.
2558  *
2559  * Returns zero if successful or error if not.
2560  *
2561  * Notes:       
2562  *      This routine transitions the device to the SDEV_RUNNING state
2563  *      (which must be a legal transition) allowing the midlayer to
2564  *      goose the queue for this device.  This routine assumes the 
2565  *      host_lock is held upon entry.
2566  */
2567 int
2568 scsi_internal_device_unblock(struct scsi_device *sdev)
2569 {
2570         struct request_queue *q = sdev->request_queue; 
2571         int err;
2572         unsigned long flags;
2573         
2574         /* 
2575          * Try to transition the scsi device to SDEV_RUNNING
2576          * and goose the device queue if successful.  
2577          */
2578         err = scsi_device_set_state(sdev, SDEV_RUNNING);
2579         if (err) {
2580                 err = scsi_device_set_state(sdev, SDEV_CREATED);
2581
2582                 if (err)
2583                         return err;
2584         }
2585
2586         spin_lock_irqsave(q->queue_lock, flags);
2587         blk_start_queue(q);
2588         spin_unlock_irqrestore(q->queue_lock, flags);
2589
2590         return 0;
2591 }
2592 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock);
2593
2594 static void
2595 device_block(struct scsi_device *sdev, void *data)
2596 {
2597         scsi_internal_device_block(sdev);
2598 }
2599
2600 static int
2601 target_block(struct device *dev, void *data)
2602 {
2603         if (scsi_is_target_device(dev))
2604                 starget_for_each_device(to_scsi_target(dev), NULL,
2605                                         device_block);
2606         return 0;
2607 }
2608
2609 void
2610 scsi_target_block(struct device *dev)
2611 {
2612         if (scsi_is_target_device(dev))
2613                 starget_for_each_device(to_scsi_target(dev), NULL,
2614                                         device_block);
2615         else
2616                 device_for_each_child(dev, NULL, target_block);
2617 }
2618 EXPORT_SYMBOL_GPL(scsi_target_block);
2619
2620 static void
2621 device_unblock(struct scsi_device *sdev, void *data)
2622 {
2623         scsi_internal_device_unblock(sdev);
2624 }
2625
2626 static int
2627 target_unblock(struct device *dev, void *data)
2628 {
2629         if (scsi_is_target_device(dev))
2630                 starget_for_each_device(to_scsi_target(dev), NULL,
2631                                         device_unblock);
2632         return 0;
2633 }
2634
2635 void
2636 scsi_target_unblock(struct device *dev)
2637 {
2638         if (scsi_is_target_device(dev))
2639                 starget_for_each_device(to_scsi_target(dev), NULL,
2640                                         device_unblock);
2641         else
2642                 device_for_each_child(dev, NULL, target_unblock);
2643 }
2644 EXPORT_SYMBOL_GPL(scsi_target_unblock);
2645
2646 /**
2647  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2648  * @sgl:        scatter-gather list
2649  * @sg_count:   number of segments in sg
2650  * @offset:     offset in bytes into sg, on return offset into the mapped area
2651  * @len:        bytes to map, on return number of bytes mapped
2652  *
2653  * Returns virtual address of the start of the mapped page
2654  */
2655 void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
2656                           size_t *offset, size_t *len)
2657 {
2658         int i;
2659         size_t sg_len = 0, len_complete = 0;
2660         struct scatterlist *sg;
2661         struct page *page;
2662
2663         WARN_ON(!irqs_disabled());
2664
2665         for_each_sg(sgl, sg, sg_count, i) {
2666                 len_complete = sg_len; /* Complete sg-entries */
2667                 sg_len += sg->length;
2668                 if (sg_len > *offset)
2669                         break;
2670         }
2671
2672         if (unlikely(i == sg_count)) {
2673                 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2674                         "elements %d\n",
2675                        __func__, sg_len, *offset, sg_count);
2676                 WARN_ON(1);
2677                 return NULL;
2678         }
2679
2680         /* Offset starting from the beginning of first page in this sg-entry */
2681         *offset = *offset - len_complete + sg->offset;
2682
2683         /* Assumption: contiguous pages can be accessed as "page + i" */
2684         page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
2685         *offset &= ~PAGE_MASK;
2686
2687         /* Bytes in this sg-entry from *offset to the end of the page */
2688         sg_len = PAGE_SIZE - *offset;
2689         if (*len > sg_len)
2690                 *len = sg_len;
2691
2692         return kmap_atomic(page, KM_BIO_SRC_IRQ);
2693 }
2694 EXPORT_SYMBOL(scsi_kmap_atomic_sg);
2695
2696 /**
2697  * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
2698  * @virt:       virtual address to be unmapped
2699  */
2700 void scsi_kunmap_atomic_sg(void *virt)
2701 {
2702         kunmap_atomic(virt, KM_BIO_SRC_IRQ);
2703 }
2704 EXPORT_SYMBOL(scsi_kunmap_atomic_sg);