1e7c9e9ef772e10c40789ebf44eb428360ac7f77
[linux-2.6.git] / drivers / scsi / scsi_lib.c
1 /*
2  *  scsi_lib.c Copyright (C) 1999 Eric Youngdale
3  *
4  *  SCSI queueing library.
5  *      Initial versions: Eric Youngdale (eric@andante.org).
6  *                        Based upon conversations with large numbers
7  *                        of people at Linux Expo.
8  */
9
10 #include <linux/bio.h>
11 #include <linux/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 *starved_head = NULL, *sdev = q->queuedata;
571         struct Scsi_Host *shost = sdev->host;
572         unsigned long flags;
573
574         if (scsi_target(sdev)->single_lun)
575                 scsi_single_lun_run(sdev);
576
577         spin_lock_irqsave(shost->host_lock, flags);
578         while (!list_empty(&shost->starved_list) && !scsi_host_is_busy(shost)) {
579                 int flagset;
580
581                 /*
582                  * As long as shost is accepting commands and we have
583                  * starved queues, call blk_run_queue. scsi_request_fn
584                  * drops the queue_lock and can add us back to the
585                  * starved_list.
586                  *
587                  * host_lock protects the starved_list and starved_entry.
588                  * scsi_request_fn must get the host_lock before checking
589                  * or modifying starved_list or starved_entry.
590                  */
591                 sdev = list_entry(shost->starved_list.next,
592                                           struct scsi_device, starved_entry);
593                 /*
594                  * The *queue_ready functions can add a device back onto the
595                  * starved list's tail, so we must check for a infinite loop.
596                  */
597                 if (sdev == starved_head)
598                         break;
599                 if (!starved_head)
600                         starved_head = sdev;
601
602                 if (scsi_target_is_busy(scsi_target(sdev))) {
603                         list_move_tail(&sdev->starved_entry,
604                                        &shost->starved_list);
605                         continue;
606                 }
607
608                 list_del_init(&sdev->starved_entry);
609                 spin_unlock(shost->host_lock);
610
611                 spin_lock(sdev->request_queue->queue_lock);
612                 flagset = test_bit(QUEUE_FLAG_REENTER, &q->queue_flags) &&
613                                 !test_bit(QUEUE_FLAG_REENTER,
614                                         &sdev->request_queue->queue_flags);
615                 if (flagset)
616                         queue_flag_set(QUEUE_FLAG_REENTER, sdev->request_queue);
617                 __blk_run_queue(sdev->request_queue);
618                 if (flagset)
619                         queue_flag_clear(QUEUE_FLAG_REENTER, sdev->request_queue);
620                 spin_unlock(sdev->request_queue->queue_lock);
621
622                 spin_lock(shost->host_lock);
623         }
624         spin_unlock_irqrestore(shost->host_lock, flags);
625
626         blk_run_queue(q);
627 }
628
629 /*
630  * Function:    scsi_requeue_command()
631  *
632  * Purpose:     Handle post-processing of completed commands.
633  *
634  * Arguments:   q       - queue to operate on
635  *              cmd     - command that may need to be requeued.
636  *
637  * Returns:     Nothing
638  *
639  * Notes:       After command completion, there may be blocks left
640  *              over which weren't finished by the previous command
641  *              this can be for a number of reasons - the main one is
642  *              I/O errors in the middle of the request, in which case
643  *              we need to request the blocks that come after the bad
644  *              sector.
645  * Notes:       Upon return, cmd is a stale pointer.
646  */
647 static void scsi_requeue_command(struct request_queue *q, struct scsi_cmnd *cmd)
648 {
649         struct request *req = cmd->request;
650         unsigned long flags;
651
652         scsi_unprep_request(req);
653         spin_lock_irqsave(q->queue_lock, flags);
654         blk_requeue_request(q, req);
655         spin_unlock_irqrestore(q->queue_lock, flags);
656
657         scsi_run_queue(q);
658 }
659
660 void scsi_next_command(struct scsi_cmnd *cmd)
661 {
662         struct scsi_device *sdev = cmd->device;
663         struct request_queue *q = sdev->request_queue;
664
665         /* need to hold a reference on the device before we let go of the cmd */
666         get_device(&sdev->sdev_gendev);
667
668         scsi_put_command(cmd);
669         scsi_run_queue(q);
670
671         /* ok to remove device now */
672         put_device(&sdev->sdev_gendev);
673 }
674
675 void scsi_run_host_queues(struct Scsi_Host *shost)
676 {
677         struct scsi_device *sdev;
678
679         shost_for_each_device(sdev, shost)
680                 scsi_run_queue(sdev->request_queue);
681 }
682
683 /*
684  * Function:    scsi_end_request()
685  *
686  * Purpose:     Post-processing of completed commands (usually invoked at end
687  *              of upper level post-processing and scsi_io_completion).
688  *
689  * Arguments:   cmd      - command that is complete.
690  *              error    - 0 if I/O indicates success, < 0 for I/O error.
691  *              bytes    - number of bytes of completed I/O
692  *              requeue  - indicates whether we should requeue leftovers.
693  *
694  * Lock status: Assumed that lock is not held upon entry.
695  *
696  * Returns:     cmd if requeue required, NULL otherwise.
697  *
698  * Notes:       This is called for block device requests in order to
699  *              mark some number of sectors as complete.
700  * 
701  *              We are guaranteeing that the request queue will be goosed
702  *              at some point during this call.
703  * Notes:       If cmd was requeued, upon return it will be a stale pointer.
704  */
705 static struct scsi_cmnd *scsi_end_request(struct scsi_cmnd *cmd, int error,
706                                           int bytes, int requeue)
707 {
708         struct request_queue *q = cmd->device->request_queue;
709         struct request *req = cmd->request;
710
711         /*
712          * If there are blocks left over at the end, set up the command
713          * to queue the remainder of them.
714          */
715         if (blk_end_request(req, error, bytes)) {
716                 int leftover = (req->hard_nr_sectors << 9);
717
718                 if (blk_pc_request(req))
719                         leftover = req->data_len;
720
721                 /* kill remainder if no retrys */
722                 if (error && scsi_noretry_cmd(cmd))
723                         blk_end_request(req, error, leftover);
724                 else {
725                         if (requeue) {
726                                 /*
727                                  * Bleah.  Leftovers again.  Stick the
728                                  * leftovers in the front of the
729                                  * queue, and goose the queue again.
730                                  */
731                                 scsi_requeue_command(q, cmd);
732                                 cmd = NULL;
733                         }
734                         return cmd;
735                 }
736         }
737
738         /*
739          * This will goose the queue request function at the end, so we don't
740          * need to worry about launching another command.
741          */
742         scsi_next_command(cmd);
743         return NULL;
744 }
745
746 static inline unsigned int scsi_sgtable_index(unsigned short nents)
747 {
748         unsigned int index;
749
750         BUG_ON(nents > SCSI_MAX_SG_SEGMENTS);
751
752         if (nents <= 8)
753                 index = 0;
754         else
755                 index = get_count_order(nents) - 3;
756
757         return index;
758 }
759
760 static void scsi_sg_free(struct scatterlist *sgl, unsigned int nents)
761 {
762         struct scsi_host_sg_pool *sgp;
763
764         sgp = scsi_sg_pools + scsi_sgtable_index(nents);
765         mempool_free(sgl, sgp->pool);
766 }
767
768 static struct scatterlist *scsi_sg_alloc(unsigned int nents, gfp_t gfp_mask)
769 {
770         struct scsi_host_sg_pool *sgp;
771
772         sgp = scsi_sg_pools + scsi_sgtable_index(nents);
773         return mempool_alloc(sgp->pool, gfp_mask);
774 }
775
776 static int scsi_alloc_sgtable(struct scsi_data_buffer *sdb, int nents,
777                               gfp_t gfp_mask)
778 {
779         int ret;
780
781         BUG_ON(!nents);
782
783         ret = __sg_alloc_table(&sdb->table, nents, SCSI_MAX_SG_SEGMENTS,
784                                gfp_mask, scsi_sg_alloc);
785         if (unlikely(ret))
786                 __sg_free_table(&sdb->table, SCSI_MAX_SG_SEGMENTS,
787                                 scsi_sg_free);
788
789         return ret;
790 }
791
792 static void scsi_free_sgtable(struct scsi_data_buffer *sdb)
793 {
794         __sg_free_table(&sdb->table, SCSI_MAX_SG_SEGMENTS, scsi_sg_free);
795 }
796
797 /*
798  * Function:    scsi_release_buffers()
799  *
800  * Purpose:     Completion processing for block device I/O requests.
801  *
802  * Arguments:   cmd     - command that we are bailing.
803  *
804  * Lock status: Assumed that no lock is held upon entry.
805  *
806  * Returns:     Nothing
807  *
808  * Notes:       In the event that an upper level driver rejects a
809  *              command, we must release resources allocated during
810  *              the __init_io() function.  Primarily this would involve
811  *              the scatter-gather table, and potentially any bounce
812  *              buffers.
813  */
814 void scsi_release_buffers(struct scsi_cmnd *cmd)
815 {
816         if (cmd->sdb.table.nents)
817                 scsi_free_sgtable(&cmd->sdb);
818
819         memset(&cmd->sdb, 0, sizeof(cmd->sdb));
820
821         if (scsi_bidi_cmnd(cmd)) {
822                 struct scsi_data_buffer *bidi_sdb =
823                         cmd->request->next_rq->special;
824                 scsi_free_sgtable(bidi_sdb);
825                 kmem_cache_free(scsi_sdb_cache, bidi_sdb);
826                 cmd->request->next_rq->special = NULL;
827         }
828
829         if (scsi_prot_sg_count(cmd))
830                 scsi_free_sgtable(cmd->prot_sdb);
831 }
832 EXPORT_SYMBOL(scsi_release_buffers);
833
834 /*
835  * Bidi commands Must be complete as a whole, both sides at once.
836  * If part of the bytes were written and lld returned
837  * scsi_in()->resid and/or scsi_out()->resid this information will be left
838  * in req->data_len and req->next_rq->data_len. The upper-layer driver can
839  * decide what to do with this information.
840  */
841 static void scsi_end_bidi_request(struct scsi_cmnd *cmd)
842 {
843         struct request *req = cmd->request;
844         unsigned int dlen = req->data_len;
845         unsigned int next_dlen = req->next_rq->data_len;
846
847         req->data_len = scsi_out(cmd)->resid;
848         req->next_rq->data_len = scsi_in(cmd)->resid;
849
850         /* The req and req->next_rq have not been completed */
851         BUG_ON(blk_end_bidi_request(req, 0, dlen, next_dlen));
852
853         scsi_release_buffers(cmd);
854
855         /*
856          * This will goose the queue request function at the end, so we don't
857          * need to worry about launching another command.
858          */
859         scsi_next_command(cmd);
860 }
861
862 /*
863  * Function:    scsi_io_completion()
864  *
865  * Purpose:     Completion processing for block device I/O requests.
866  *
867  * Arguments:   cmd   - command that is finished.
868  *
869  * Lock status: Assumed that no lock is held upon entry.
870  *
871  * Returns:     Nothing
872  *
873  * Notes:       This function is matched in terms of capabilities to
874  *              the function that created the scatter-gather list.
875  *              In other words, if there are no bounce buffers
876  *              (the normal case for most drivers), we don't need
877  *              the logic to deal with cleaning up afterwards.
878  *
879  *              We must do one of several things here:
880  *
881  *              a) Call scsi_end_request.  This will finish off the
882  *                 specified number of sectors.  If we are done, the
883  *                 command block will be released, and the queue
884  *                 function will be goosed.  If we are not done, then
885  *                 scsi_end_request will directly goose the queue.
886  *
887  *              b) We can just use scsi_requeue_command() here.  This would
888  *                 be used if we just wanted to retry, for example.
889  */
890 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
891 {
892         int result = cmd->result;
893         int this_count;
894         struct request_queue *q = cmd->device->request_queue;
895         struct request *req = cmd->request;
896         int error = 0;
897         struct scsi_sense_hdr sshdr;
898         int sense_valid = 0;
899         int sense_deferred = 0;
900
901         if (result) {
902                 sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
903                 if (sense_valid)
904                         sense_deferred = scsi_sense_is_deferred(&sshdr);
905         }
906
907         if (blk_pc_request(req)) { /* SG_IO ioctl from block level */
908                 req->errors = result;
909                 if (result) {
910                         if (sense_valid && req->sense) {
911                                 /*
912                                  * SG_IO wants current and deferred errors
913                                  */
914                                 int len = 8 + cmd->sense_buffer[7];
915
916                                 if (len > SCSI_SENSE_BUFFERSIZE)
917                                         len = SCSI_SENSE_BUFFERSIZE;
918                                 memcpy(req->sense, cmd->sense_buffer,  len);
919                                 req->sense_len = len;
920                         }
921                         if (!sense_deferred)
922                                 error = -EIO;
923                 }
924                 if (scsi_bidi_cmnd(cmd)) {
925                         /* will also release_buffers */
926                         scsi_end_bidi_request(cmd);
927                         return;
928                 }
929                 req->data_len = scsi_get_resid(cmd);
930         }
931
932         BUG_ON(blk_bidi_rq(req)); /* bidi not support for !blk_pc_request yet */
933         scsi_release_buffers(cmd);
934
935         /*
936          * Next deal with any sectors which we were able to correctly
937          * handle.
938          */
939         SCSI_LOG_HLCOMPLETE(1, printk("%ld sectors total, "
940                                       "%d bytes done.\n",
941                                       req->nr_sectors, good_bytes));
942
943         /* A number of bytes were successfully read.  If there
944          * are leftovers and there is some kind of error
945          * (result != 0), retry the rest.
946          */
947         if (scsi_end_request(cmd, error, good_bytes, result == 0) == NULL)
948                 return;
949         this_count = blk_rq_bytes(req);
950
951         /* good_bytes = 0, or (inclusive) there were leftovers and
952          * result = 0, so scsi_end_request couldn't retry.
953          */
954         if (sense_valid && !sense_deferred) {
955                 switch (sshdr.sense_key) {
956                 case UNIT_ATTENTION:
957                         if (cmd->device->removable) {
958                                 /* Detected disc change.  Set a bit
959                                  * and quietly refuse further access.
960                                  */
961                                 cmd->device->changed = 1;
962                                 scsi_end_request(cmd, -EIO, this_count, 1);
963                                 return;
964                         } else {
965                                 /* Must have been a power glitch, or a
966                                  * bus reset.  Could not have been a
967                                  * media change, so we just retry the
968                                  * request and see what happens.
969                                  */
970                                 scsi_requeue_command(q, cmd);
971                                 return;
972                         }
973                         break;
974                 case ILLEGAL_REQUEST:
975                         /* If we had an ILLEGAL REQUEST returned, then
976                          * we may have performed an unsupported
977                          * command.  The only thing this should be
978                          * would be a ten byte read where only a six
979                          * byte read was supported.  Also, on a system
980                          * where READ CAPACITY failed, we may have
981                          * read past the end of the disk.
982                          */
983                         if ((cmd->device->use_10_for_rw &&
984                             sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
985                             (cmd->cmnd[0] == READ_10 ||
986                              cmd->cmnd[0] == WRITE_10)) {
987                                 cmd->device->use_10_for_rw = 0;
988                                 /* This will cause a retry with a
989                                  * 6-byte command.
990                                  */
991                                 scsi_requeue_command(q, cmd);
992                         } else if (sshdr.asc == 0x10) /* DIX */
993                                 scsi_end_request(cmd, -EIO, this_count, 0);
994                         else
995                                 scsi_end_request(cmd, -EIO, this_count, 1);
996                         return;
997                 case ABORTED_COMMAND:
998                         if (sshdr.asc == 0x10) { /* DIF */
999                                 scsi_end_request(cmd, -EIO, this_count, 0);
1000                                 return;
1001                         }
1002                         break;
1003                 case NOT_READY:
1004                         /* If the device is in the process of becoming
1005                          * ready, or has a temporary blockage, retry.
1006                          */
1007                         if (sshdr.asc == 0x04) {
1008                                 switch (sshdr.ascq) {
1009                                 case 0x01: /* becoming ready */
1010                                 case 0x04: /* format in progress */
1011                                 case 0x05: /* rebuild in progress */
1012                                 case 0x06: /* recalculation in progress */
1013                                 case 0x07: /* operation in progress */
1014                                 case 0x08: /* Long write in progress */
1015                                 case 0x09: /* self test in progress */
1016                                         scsi_requeue_command(q, cmd);
1017                                         return;
1018                                 default:
1019                                         break;
1020                                 }
1021                         }
1022                         if (!(req->cmd_flags & REQ_QUIET))
1023                                 scsi_cmd_print_sense_hdr(cmd,
1024                                                          "Device not ready",
1025                                                          &sshdr);
1026
1027                         scsi_end_request(cmd, -EIO, this_count, 1);
1028                         return;
1029                 case VOLUME_OVERFLOW:
1030                         if (!(req->cmd_flags & REQ_QUIET)) {
1031                                 scmd_printk(KERN_INFO, cmd,
1032                                             "Volume overflow, CDB: ");
1033                                 __scsi_print_command(cmd->cmnd);
1034                                 scsi_print_sense("", cmd);
1035                         }
1036                         /* See SSC3rXX or current. */
1037                         scsi_end_request(cmd, -EIO, this_count, 1);
1038                         return;
1039                 default:
1040                         break;
1041                 }
1042         }
1043         if (host_byte(result) == DID_RESET) {
1044                 /* Third party bus reset or reset for error recovery
1045                  * reasons.  Just retry the request and see what
1046                  * happens.
1047                  */
1048                 scsi_requeue_command(q, cmd);
1049                 return;
1050         }
1051         if (result) {
1052                 if (!(req->cmd_flags & REQ_QUIET)) {
1053                         scsi_print_result(cmd);
1054                         if (driver_byte(result) & DRIVER_SENSE)
1055                                 scsi_print_sense("", cmd);
1056                 }
1057         }
1058         scsi_end_request(cmd, -EIO, this_count, !result);
1059 }
1060
1061 static int scsi_init_sgtable(struct request *req, struct scsi_data_buffer *sdb,
1062                              gfp_t gfp_mask)
1063 {
1064         int count;
1065
1066         /*
1067          * If sg table allocation fails, requeue request later.
1068          */
1069         if (unlikely(scsi_alloc_sgtable(sdb, req->nr_phys_segments,
1070                                         gfp_mask))) {
1071                 return BLKPREP_DEFER;
1072         }
1073
1074         req->buffer = NULL;
1075
1076         /* 
1077          * Next, walk the list, and fill in the addresses and sizes of
1078          * each segment.
1079          */
1080         count = blk_rq_map_sg(req->q, req, sdb->table.sgl);
1081         BUG_ON(count > sdb->table.nents);
1082         sdb->table.nents = count;
1083         if (blk_pc_request(req))
1084                 sdb->length = req->data_len;
1085         else
1086                 sdb->length = req->nr_sectors << 9;
1087         return BLKPREP_OK;
1088 }
1089
1090 /*
1091  * Function:    scsi_init_io()
1092  *
1093  * Purpose:     SCSI I/O initialize function.
1094  *
1095  * Arguments:   cmd   - Command descriptor we wish to initialize
1096  *
1097  * Returns:     0 on success
1098  *              BLKPREP_DEFER if the failure is retryable
1099  *              BLKPREP_KILL if the failure is fatal
1100  */
1101 int scsi_init_io(struct scsi_cmnd *cmd, gfp_t gfp_mask)
1102 {
1103         int error = scsi_init_sgtable(cmd->request, &cmd->sdb, gfp_mask);
1104         if (error)
1105                 goto err_exit;
1106
1107         if (blk_bidi_rq(cmd->request)) {
1108                 struct scsi_data_buffer *bidi_sdb = kmem_cache_zalloc(
1109                         scsi_sdb_cache, GFP_ATOMIC);
1110                 if (!bidi_sdb) {
1111                         error = BLKPREP_DEFER;
1112                         goto err_exit;
1113                 }
1114
1115                 cmd->request->next_rq->special = bidi_sdb;
1116                 error = scsi_init_sgtable(cmd->request->next_rq, bidi_sdb,
1117                                                                     GFP_ATOMIC);
1118                 if (error)
1119                         goto err_exit;
1120         }
1121
1122         if (blk_integrity_rq(cmd->request)) {
1123                 struct scsi_data_buffer *prot_sdb = cmd->prot_sdb;
1124                 int ivecs, count;
1125
1126                 BUG_ON(prot_sdb == NULL);
1127                 ivecs = blk_rq_count_integrity_sg(cmd->request);
1128
1129                 if (scsi_alloc_sgtable(prot_sdb, ivecs, gfp_mask)) {
1130                         error = BLKPREP_DEFER;
1131                         goto err_exit;
1132                 }
1133
1134                 count = blk_rq_map_integrity_sg(cmd->request,
1135                                                 prot_sdb->table.sgl);
1136                 BUG_ON(unlikely(count > ivecs));
1137
1138                 cmd->prot_sdb = prot_sdb;
1139                 cmd->prot_sdb->table.nents = count;
1140         }
1141
1142         return BLKPREP_OK ;
1143
1144 err_exit:
1145         scsi_release_buffers(cmd);
1146         if (error == BLKPREP_KILL)
1147                 scsi_put_command(cmd);
1148         else /* BLKPREP_DEFER */
1149                 scsi_unprep_request(cmd->request);
1150
1151         return error;
1152 }
1153 EXPORT_SYMBOL(scsi_init_io);
1154
1155 static struct scsi_cmnd *scsi_get_cmd_from_req(struct scsi_device *sdev,
1156                 struct request *req)
1157 {
1158         struct scsi_cmnd *cmd;
1159
1160         if (!req->special) {
1161                 cmd = scsi_get_command(sdev, GFP_ATOMIC);
1162                 if (unlikely(!cmd))
1163                         return NULL;
1164                 req->special = cmd;
1165         } else {
1166                 cmd = req->special;
1167         }
1168
1169         /* pull a tag out of the request if we have one */
1170         cmd->tag = req->tag;
1171         cmd->request = req;
1172
1173         cmd->cmnd = req->cmd;
1174
1175         return cmd;
1176 }
1177
1178 int scsi_setup_blk_pc_cmnd(struct scsi_device *sdev, struct request *req)
1179 {
1180         struct scsi_cmnd *cmd;
1181         int ret = scsi_prep_state_check(sdev, req);
1182
1183         if (ret != BLKPREP_OK)
1184                 return ret;
1185
1186         cmd = scsi_get_cmd_from_req(sdev, req);
1187         if (unlikely(!cmd))
1188                 return BLKPREP_DEFER;
1189
1190         /*
1191          * BLOCK_PC requests may transfer data, in which case they must
1192          * a bio attached to them.  Or they might contain a SCSI command
1193          * that does not transfer data, in which case they may optionally
1194          * submit a request without an attached bio.
1195          */
1196         if (req->bio) {
1197                 int ret;
1198
1199                 BUG_ON(!req->nr_phys_segments);
1200
1201                 ret = scsi_init_io(cmd, GFP_ATOMIC);
1202                 if (unlikely(ret))
1203                         return ret;
1204         } else {
1205                 BUG_ON(req->data_len);
1206                 BUG_ON(req->data);
1207
1208                 memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1209                 req->buffer = NULL;
1210         }
1211
1212         cmd->cmd_len = req->cmd_len;
1213         if (!req->data_len)
1214                 cmd->sc_data_direction = DMA_NONE;
1215         else if (rq_data_dir(req) == WRITE)
1216                 cmd->sc_data_direction = DMA_TO_DEVICE;
1217         else
1218                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1219         
1220         cmd->transfersize = req->data_len;
1221         cmd->allowed = req->retries;
1222         return BLKPREP_OK;
1223 }
1224 EXPORT_SYMBOL(scsi_setup_blk_pc_cmnd);
1225
1226 /*
1227  * Setup a REQ_TYPE_FS command.  These are simple read/write request
1228  * from filesystems that still need to be translated to SCSI CDBs from
1229  * the ULD.
1230  */
1231 int scsi_setup_fs_cmnd(struct scsi_device *sdev, struct request *req)
1232 {
1233         struct scsi_cmnd *cmd;
1234         int ret = scsi_prep_state_check(sdev, req);
1235
1236         if (ret != BLKPREP_OK)
1237                 return ret;
1238
1239         if (unlikely(sdev->scsi_dh_data && sdev->scsi_dh_data->scsi_dh
1240                          && sdev->scsi_dh_data->scsi_dh->prep_fn)) {
1241                 ret = sdev->scsi_dh_data->scsi_dh->prep_fn(sdev, req);
1242                 if (ret != BLKPREP_OK)
1243                         return ret;
1244         }
1245
1246         /*
1247          * Filesystem requests must transfer data.
1248          */
1249         BUG_ON(!req->nr_phys_segments);
1250
1251         cmd = scsi_get_cmd_from_req(sdev, req);
1252         if (unlikely(!cmd))
1253                 return BLKPREP_DEFER;
1254
1255         memset(cmd->cmnd, 0, BLK_MAX_CDB);
1256         return scsi_init_io(cmd, GFP_ATOMIC);
1257 }
1258 EXPORT_SYMBOL(scsi_setup_fs_cmnd);
1259
1260 int scsi_prep_state_check(struct scsi_device *sdev, struct request *req)
1261 {
1262         int ret = BLKPREP_OK;
1263
1264         /*
1265          * If the device is not in running state we will reject some
1266          * or all commands.
1267          */
1268         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1269                 switch (sdev->sdev_state) {
1270                 case SDEV_OFFLINE:
1271                         /*
1272                          * If the device is offline we refuse to process any
1273                          * commands.  The device must be brought online
1274                          * before trying any recovery commands.
1275                          */
1276                         sdev_printk(KERN_ERR, sdev,
1277                                     "rejecting I/O to offline device\n");
1278                         ret = BLKPREP_KILL;
1279                         break;
1280                 case SDEV_DEL:
1281                         /*
1282                          * If the device is fully deleted, we refuse to
1283                          * process any commands as well.
1284                          */
1285                         sdev_printk(KERN_ERR, sdev,
1286                                     "rejecting I/O to dead device\n");
1287                         ret = BLKPREP_KILL;
1288                         break;
1289                 case SDEV_QUIESCE:
1290                 case SDEV_BLOCK:
1291                 case SDEV_CREATED_BLOCK:
1292                         /*
1293                          * If the devices is blocked we defer normal commands.
1294                          */
1295                         if (!(req->cmd_flags & REQ_PREEMPT))
1296                                 ret = BLKPREP_DEFER;
1297                         break;
1298                 default:
1299                         /*
1300                          * For any other not fully online state we only allow
1301                          * special commands.  In particular any user initiated
1302                          * command is not allowed.
1303                          */
1304                         if (!(req->cmd_flags & REQ_PREEMPT))
1305                                 ret = BLKPREP_KILL;
1306                         break;
1307                 }
1308         }
1309         return ret;
1310 }
1311 EXPORT_SYMBOL(scsi_prep_state_check);
1312
1313 int scsi_prep_return(struct request_queue *q, struct request *req, int ret)
1314 {
1315         struct scsi_device *sdev = q->queuedata;
1316
1317         switch (ret) {
1318         case BLKPREP_KILL:
1319                 req->errors = DID_NO_CONNECT << 16;
1320                 /* release the command and kill it */
1321                 if (req->special) {
1322                         struct scsi_cmnd *cmd = req->special;
1323                         scsi_release_buffers(cmd);
1324                         scsi_put_command(cmd);
1325                         req->special = NULL;
1326                 }
1327                 break;
1328         case BLKPREP_DEFER:
1329                 /*
1330                  * If we defer, the elv_next_request() returns NULL, but the
1331                  * queue must be restarted, so we plug here if no returning
1332                  * command will automatically do that.
1333                  */
1334                 if (sdev->device_busy == 0)
1335                         blk_plug_device(q);
1336                 break;
1337         default:
1338                 req->cmd_flags |= REQ_DONTPREP;
1339         }
1340
1341         return ret;
1342 }
1343 EXPORT_SYMBOL(scsi_prep_return);
1344
1345 int scsi_prep_fn(struct request_queue *q, struct request *req)
1346 {
1347         struct scsi_device *sdev = q->queuedata;
1348         int ret = BLKPREP_KILL;
1349
1350         if (req->cmd_type == REQ_TYPE_BLOCK_PC)
1351                 ret = scsi_setup_blk_pc_cmnd(sdev, req);
1352         return scsi_prep_return(q, req, ret);
1353 }
1354
1355 /*
1356  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1357  * return 0.
1358  *
1359  * Called with the queue_lock held.
1360  */
1361 static inline int scsi_dev_queue_ready(struct request_queue *q,
1362                                   struct scsi_device *sdev)
1363 {
1364         if (sdev->device_busy == 0 && sdev->device_blocked) {
1365                 /*
1366                  * unblock after device_blocked iterates to zero
1367                  */
1368                 if (--sdev->device_blocked == 0) {
1369                         SCSI_LOG_MLQUEUE(3,
1370                                    sdev_printk(KERN_INFO, sdev,
1371                                    "unblocking device at zero depth\n"));
1372                 } else {
1373                         blk_plug_device(q);
1374                         return 0;
1375                 }
1376         }
1377         if (scsi_device_is_busy(sdev))
1378                 return 0;
1379
1380         return 1;
1381 }
1382
1383
1384 /*
1385  * scsi_target_queue_ready: checks if there we can send commands to target
1386  * @sdev: scsi device on starget to check.
1387  *
1388  * Called with the host lock held.
1389  */
1390 static inline int scsi_target_queue_ready(struct Scsi_Host *shost,
1391                                            struct scsi_device *sdev)
1392 {
1393         struct scsi_target *starget = scsi_target(sdev);
1394
1395         if (starget->single_lun) {
1396                 if (starget->starget_sdev_user &&
1397                     starget->starget_sdev_user != sdev)
1398                         return 0;
1399                 starget->starget_sdev_user = sdev;
1400         }
1401
1402         if (starget->target_busy == 0 && starget->target_blocked) {
1403                 /*
1404                  * unblock after target_blocked iterates to zero
1405                  */
1406                 if (--starget->target_blocked == 0) {
1407                         SCSI_LOG_MLQUEUE(3, starget_printk(KERN_INFO, starget,
1408                                          "unblocking target at zero depth\n"));
1409                 } else {
1410                         blk_plug_device(sdev->request_queue);
1411                         return 0;
1412                 }
1413         }
1414
1415         if (scsi_target_is_busy(starget)) {
1416                 if (list_empty(&sdev->starved_entry)) {
1417                         list_add_tail(&sdev->starved_entry,
1418                                       &shost->starved_list);
1419                         return 0;
1420                 }
1421         }
1422
1423         /* We're OK to process the command, so we can't be starved */
1424         if (!list_empty(&sdev->starved_entry))
1425                 list_del_init(&sdev->starved_entry);
1426         return 1;
1427 }
1428
1429 /*
1430  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1431  * return 0. We must end up running the queue again whenever 0 is
1432  * returned, else IO can hang.
1433  *
1434  * Called with host_lock held.
1435  */
1436 static inline int scsi_host_queue_ready(struct request_queue *q,
1437                                    struct Scsi_Host *shost,
1438                                    struct scsi_device *sdev)
1439 {
1440         if (scsi_host_in_recovery(shost))
1441                 return 0;
1442         if (shost->host_busy == 0 && shost->host_blocked) {
1443                 /*
1444                  * unblock after host_blocked iterates to zero
1445                  */
1446                 if (--shost->host_blocked == 0) {
1447                         SCSI_LOG_MLQUEUE(3,
1448                                 printk("scsi%d unblocking host at zero depth\n",
1449                                         shost->host_no));
1450                 } else {
1451                         return 0;
1452                 }
1453         }
1454         if (scsi_host_is_busy(shost)) {
1455                 if (list_empty(&sdev->starved_entry))
1456                         list_add_tail(&sdev->starved_entry, &shost->starved_list);
1457                 return 0;
1458         }
1459
1460         /* We're OK to process the command, so we can't be starved */
1461         if (!list_empty(&sdev->starved_entry))
1462                 list_del_init(&sdev->starved_entry);
1463
1464         return 1;
1465 }
1466
1467 /*
1468  * Kill a request for a dead device
1469  */
1470 static void scsi_kill_request(struct request *req, struct request_queue *q)
1471 {
1472         struct scsi_cmnd *cmd = req->special;
1473         struct scsi_device *sdev = cmd->device;
1474         struct scsi_target *starget = scsi_target(sdev);
1475         struct Scsi_Host *shost = sdev->host;
1476
1477         blkdev_dequeue_request(req);
1478
1479         if (unlikely(cmd == NULL)) {
1480                 printk(KERN_CRIT "impossible request in %s.\n",
1481                                  __func__);
1482                 BUG();
1483         }
1484
1485         scsi_init_cmd_errh(cmd);
1486         cmd->result = DID_NO_CONNECT << 16;
1487         atomic_inc(&cmd->device->iorequest_cnt);
1488
1489         /*
1490          * SCSI request completion path will do scsi_device_unbusy(),
1491          * bump busy counts.  To bump the counters, we need to dance
1492          * with the locks as normal issue path does.
1493          */
1494         sdev->device_busy++;
1495         spin_unlock(sdev->request_queue->queue_lock);
1496         spin_lock(shost->host_lock);
1497         shost->host_busy++;
1498         starget->target_busy++;
1499         spin_unlock(shost->host_lock);
1500         spin_lock(sdev->request_queue->queue_lock);
1501
1502         blk_complete_request(req);
1503 }
1504
1505 static void scsi_softirq_done(struct request *rq)
1506 {
1507         struct scsi_cmnd *cmd = rq->special;
1508         unsigned long wait_for = (cmd->allowed + 1) * rq->timeout;
1509         int disposition;
1510
1511         INIT_LIST_HEAD(&cmd->eh_entry);
1512
1513         /*
1514          * Set the serial numbers back to zero
1515          */
1516         cmd->serial_number = 0;
1517
1518         atomic_inc(&cmd->device->iodone_cnt);
1519         if (cmd->result)
1520                 atomic_inc(&cmd->device->ioerr_cnt);
1521
1522         disposition = scsi_decide_disposition(cmd);
1523         if (disposition != SUCCESS &&
1524             time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1525                 sdev_printk(KERN_ERR, cmd->device,
1526                             "timing out command, waited %lus\n",
1527                             wait_for/HZ);
1528                 disposition = SUCCESS;
1529         }
1530                         
1531         scsi_log_completion(cmd, disposition);
1532
1533         switch (disposition) {
1534                 case SUCCESS:
1535                         scsi_finish_command(cmd);
1536                         break;
1537                 case NEEDS_RETRY:
1538                         scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1539                         break;
1540                 case ADD_TO_MLQUEUE:
1541                         scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1542                         break;
1543                 default:
1544                         if (!scsi_eh_scmd_add(cmd, 0))
1545                                 scsi_finish_command(cmd);
1546         }
1547 }
1548
1549 /*
1550  * Function:    scsi_request_fn()
1551  *
1552  * Purpose:     Main strategy routine for SCSI.
1553  *
1554  * Arguments:   q       - Pointer to actual queue.
1555  *
1556  * Returns:     Nothing
1557  *
1558  * Lock status: IO request lock assumed to be held when called.
1559  */
1560 static void scsi_request_fn(struct request_queue *q)
1561 {
1562         struct scsi_device *sdev = q->queuedata;
1563         struct Scsi_Host *shost;
1564         struct scsi_cmnd *cmd;
1565         struct request *req;
1566
1567         if (!sdev) {
1568                 printk("scsi: killing requests for dead queue\n");
1569                 while ((req = elv_next_request(q)) != NULL)
1570                         scsi_kill_request(req, q);
1571                 return;
1572         }
1573
1574         if(!get_device(&sdev->sdev_gendev))
1575                 /* We must be tearing the block queue down already */
1576                 return;
1577
1578         /*
1579          * To start with, we keep looping until the queue is empty, or until
1580          * the host is no longer able to accept any more requests.
1581          */
1582         shost = sdev->host;
1583         while (!blk_queue_plugged(q)) {
1584                 int rtn;
1585                 /*
1586                  * get next queueable request.  We do this early to make sure
1587                  * that the request is fully prepared even if we cannot 
1588                  * accept it.
1589                  */
1590                 req = elv_next_request(q);
1591                 if (!req || !scsi_dev_queue_ready(q, sdev))
1592                         break;
1593
1594                 if (unlikely(!scsi_device_online(sdev))) {
1595                         sdev_printk(KERN_ERR, sdev,
1596                                     "rejecting I/O to offline device\n");
1597                         scsi_kill_request(req, q);
1598                         continue;
1599                 }
1600
1601
1602                 /*
1603                  * Remove the request from the request list.
1604                  */
1605                 if (!(blk_queue_tagged(q) && !blk_queue_start_tag(q, req)))
1606                         blkdev_dequeue_request(req);
1607                 sdev->device_busy++;
1608
1609                 spin_unlock(q->queue_lock);
1610                 cmd = req->special;
1611                 if (unlikely(cmd == NULL)) {
1612                         printk(KERN_CRIT "impossible request in %s.\n"
1613                                          "please mail a stack trace to "
1614                                          "linux-scsi@vger.kernel.org\n",
1615                                          __func__);
1616                         blk_dump_rq_flags(req, "foo");
1617                         BUG();
1618                 }
1619                 spin_lock(shost->host_lock);
1620
1621                 /*
1622                  * We hit this when the driver is using a host wide
1623                  * tag map. For device level tag maps the queue_depth check
1624                  * in the device ready fn would prevent us from trying
1625                  * to allocate a tag. Since the map is a shared host resource
1626                  * we add the dev to the starved list so it eventually gets
1627                  * a run when a tag is freed.
1628                  */
1629                 if (blk_queue_tagged(q) && !blk_rq_tagged(req)) {
1630                         if (list_empty(&sdev->starved_entry))
1631                                 list_add_tail(&sdev->starved_entry,
1632                                               &shost->starved_list);
1633                         goto not_ready;
1634                 }
1635
1636                 if (!scsi_target_queue_ready(shost, sdev))
1637                         goto not_ready;
1638
1639                 if (!scsi_host_queue_ready(q, shost, sdev))
1640                         goto not_ready;
1641
1642                 scsi_target(sdev)->target_busy++;
1643                 shost->host_busy++;
1644
1645                 /*
1646                  * XXX(hch): This is rather suboptimal, scsi_dispatch_cmd will
1647                  *              take the lock again.
1648                  */
1649                 spin_unlock_irq(shost->host_lock);
1650
1651                 /*
1652                  * Finally, initialize any error handling parameters, and set up
1653                  * the timers for timeouts.
1654                  */
1655                 scsi_init_cmd_errh(cmd);
1656
1657                 /*
1658                  * Dispatch the command to the low-level driver.
1659                  */
1660                 rtn = scsi_dispatch_cmd(cmd);
1661                 spin_lock_irq(q->queue_lock);
1662                 if(rtn) {
1663                         /* we're refusing the command; because of
1664                          * the way locks get dropped, we need to 
1665                          * check here if plugging is required */
1666                         if(sdev->device_busy == 0)
1667                                 blk_plug_device(q);
1668
1669                         break;
1670                 }
1671         }
1672
1673         goto out;
1674
1675  not_ready:
1676         spin_unlock_irq(shost->host_lock);
1677
1678         /*
1679          * lock q, handle tag, requeue req, and decrement device_busy. We
1680          * must return with queue_lock held.
1681          *
1682          * Decrementing device_busy without checking it is OK, as all such
1683          * cases (host limits or settings) should run the queue at some
1684          * later time.
1685          */
1686         spin_lock_irq(q->queue_lock);
1687         blk_requeue_request(q, req);
1688         sdev->device_busy--;
1689         if(sdev->device_busy == 0)
1690                 blk_plug_device(q);
1691  out:
1692         /* must be careful here...if we trigger the ->remove() function
1693          * we cannot be holding the q lock */
1694         spin_unlock_irq(q->queue_lock);
1695         put_device(&sdev->sdev_gendev);
1696         spin_lock_irq(q->queue_lock);
1697 }
1698
1699 u64 scsi_calculate_bounce_limit(struct Scsi_Host *shost)
1700 {
1701         struct device *host_dev;
1702         u64 bounce_limit = 0xffffffff;
1703
1704         if (shost->unchecked_isa_dma)
1705                 return BLK_BOUNCE_ISA;
1706         /*
1707          * Platforms with virtual-DMA translation
1708          * hardware have no practical limit.
1709          */
1710         if (!PCI_DMA_BUS_IS_PHYS)
1711                 return BLK_BOUNCE_ANY;
1712
1713         host_dev = scsi_get_device(shost);
1714         if (host_dev && host_dev->dma_mask)
1715                 bounce_limit = *host_dev->dma_mask;
1716
1717         return bounce_limit;
1718 }
1719 EXPORT_SYMBOL(scsi_calculate_bounce_limit);
1720
1721 struct request_queue *__scsi_alloc_queue(struct Scsi_Host *shost,
1722                                          request_fn_proc *request_fn)
1723 {
1724         struct request_queue *q;
1725         struct device *dev = shost->shost_gendev.parent;
1726
1727         q = blk_init_queue(request_fn, NULL);
1728         if (!q)
1729                 return NULL;
1730
1731         /*
1732          * this limit is imposed by hardware restrictions
1733          */
1734         blk_queue_max_hw_segments(q, shost->sg_tablesize);
1735         blk_queue_max_phys_segments(q, SCSI_MAX_SG_CHAIN_SEGMENTS);
1736
1737         blk_queue_max_sectors(q, shost->max_sectors);
1738         blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
1739         blk_queue_segment_boundary(q, shost->dma_boundary);
1740         dma_set_seg_boundary(dev, shost->dma_boundary);
1741
1742         blk_queue_max_segment_size(q, dma_get_max_seg_size(dev));
1743
1744         /* New queue, no concurrency on queue_flags */
1745         if (!shost->use_clustering)
1746                 queue_flag_clear_unlocked(QUEUE_FLAG_CLUSTER, q);
1747
1748         /*
1749          * set a reasonable default alignment on word boundaries: the
1750          * host and device may alter it using
1751          * blk_queue_update_dma_alignment() later.
1752          */
1753         blk_queue_dma_alignment(q, 0x03);
1754
1755         return q;
1756 }
1757 EXPORT_SYMBOL(__scsi_alloc_queue);
1758
1759 struct request_queue *scsi_alloc_queue(struct scsi_device *sdev)
1760 {
1761         struct request_queue *q;
1762
1763         q = __scsi_alloc_queue(sdev->host, scsi_request_fn);
1764         if (!q)
1765                 return NULL;
1766
1767         blk_queue_prep_rq(q, scsi_prep_fn);
1768         blk_queue_softirq_done(q, scsi_softirq_done);
1769         blk_queue_rq_timed_out(q, scsi_times_out);
1770         return q;
1771 }
1772
1773 void scsi_free_queue(struct request_queue *q)
1774 {
1775         blk_cleanup_queue(q);
1776 }
1777
1778 /*
1779  * Function:    scsi_block_requests()
1780  *
1781  * Purpose:     Utility function used by low-level drivers to prevent further
1782  *              commands from being queued to the device.
1783  *
1784  * Arguments:   shost       - Host in question
1785  *
1786  * Returns:     Nothing
1787  *
1788  * Lock status: No locks are assumed held.
1789  *
1790  * Notes:       There is no timer nor any other means by which the requests
1791  *              get unblocked other than the low-level driver calling
1792  *              scsi_unblock_requests().
1793  */
1794 void scsi_block_requests(struct Scsi_Host *shost)
1795 {
1796         shost->host_self_blocked = 1;
1797 }
1798 EXPORT_SYMBOL(scsi_block_requests);
1799
1800 /*
1801  * Function:    scsi_unblock_requests()
1802  *
1803  * Purpose:     Utility function used by low-level drivers to allow further
1804  *              commands from being queued to the device.
1805  *
1806  * Arguments:   shost       - Host in question
1807  *
1808  * Returns:     Nothing
1809  *
1810  * Lock status: No locks are assumed held.
1811  *
1812  * Notes:       There is no timer nor any other means by which the requests
1813  *              get unblocked other than the low-level driver calling
1814  *              scsi_unblock_requests().
1815  *
1816  *              This is done as an API function so that changes to the
1817  *              internals of the scsi mid-layer won't require wholesale
1818  *              changes to drivers that use this feature.
1819  */
1820 void scsi_unblock_requests(struct Scsi_Host *shost)
1821 {
1822         shost->host_self_blocked = 0;
1823         scsi_run_host_queues(shost);
1824 }
1825 EXPORT_SYMBOL(scsi_unblock_requests);
1826
1827 int __init scsi_init_queue(void)
1828 {
1829         int i;
1830
1831         scsi_io_context_cache = kmem_cache_create("scsi_io_context",
1832                                         sizeof(struct scsi_io_context),
1833                                         0, 0, NULL);
1834         if (!scsi_io_context_cache) {
1835                 printk(KERN_ERR "SCSI: can't init scsi io context cache\n");
1836                 return -ENOMEM;
1837         }
1838
1839         scsi_sdb_cache = kmem_cache_create("scsi_data_buffer",
1840                                            sizeof(struct scsi_data_buffer),
1841                                            0, 0, NULL);
1842         if (!scsi_sdb_cache) {
1843                 printk(KERN_ERR "SCSI: can't init scsi sdb cache\n");
1844                 goto cleanup_io_context;
1845         }
1846
1847         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1848                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1849                 int size = sgp->size * sizeof(struct scatterlist);
1850
1851                 sgp->slab = kmem_cache_create(sgp->name, size, 0,
1852                                 SLAB_HWCACHE_ALIGN, NULL);
1853                 if (!sgp->slab) {
1854                         printk(KERN_ERR "SCSI: can't init sg slab %s\n",
1855                                         sgp->name);
1856                         goto cleanup_sdb;
1857                 }
1858
1859                 sgp->pool = mempool_create_slab_pool(SG_MEMPOOL_SIZE,
1860                                                      sgp->slab);
1861                 if (!sgp->pool) {
1862                         printk(KERN_ERR "SCSI: can't init sg mempool %s\n",
1863                                         sgp->name);
1864                         goto cleanup_sdb;
1865                 }
1866         }
1867
1868         return 0;
1869
1870 cleanup_sdb:
1871         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1872                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1873                 if (sgp->pool)
1874                         mempool_destroy(sgp->pool);
1875                 if (sgp->slab)
1876                         kmem_cache_destroy(sgp->slab);
1877         }
1878         kmem_cache_destroy(scsi_sdb_cache);
1879 cleanup_io_context:
1880         kmem_cache_destroy(scsi_io_context_cache);
1881
1882         return -ENOMEM;
1883 }
1884
1885 void scsi_exit_queue(void)
1886 {
1887         int i;
1888
1889         kmem_cache_destroy(scsi_io_context_cache);
1890         kmem_cache_destroy(scsi_sdb_cache);
1891
1892         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1893                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1894                 mempool_destroy(sgp->pool);
1895                 kmem_cache_destroy(sgp->slab);
1896         }
1897 }
1898
1899 /**
1900  *      scsi_mode_select - issue a mode select
1901  *      @sdev:  SCSI device to be queried
1902  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
1903  *      @sp:    Save page bit (0 == don't save, 1 == save)
1904  *      @modepage: mode page being requested
1905  *      @buffer: request buffer (may not be smaller than eight bytes)
1906  *      @len:   length of request buffer.
1907  *      @timeout: command timeout
1908  *      @retries: number of retries before failing
1909  *      @data: returns a structure abstracting the mode header data
1910  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
1911  *              must be SCSI_SENSE_BUFFERSIZE big.
1912  *
1913  *      Returns zero if successful; negative error number or scsi
1914  *      status on error
1915  *
1916  */
1917 int
1918 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
1919                  unsigned char *buffer, int len, int timeout, int retries,
1920                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1921 {
1922         unsigned char cmd[10];
1923         unsigned char *real_buffer;
1924         int ret;
1925
1926         memset(cmd, 0, sizeof(cmd));
1927         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
1928
1929         if (sdev->use_10_for_ms) {
1930                 if (len > 65535)
1931                         return -EINVAL;
1932                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
1933                 if (!real_buffer)
1934                         return -ENOMEM;
1935                 memcpy(real_buffer + 8, buffer, len);
1936                 len += 8;
1937                 real_buffer[0] = 0;
1938                 real_buffer[1] = 0;
1939                 real_buffer[2] = data->medium_type;
1940                 real_buffer[3] = data->device_specific;
1941                 real_buffer[4] = data->longlba ? 0x01 : 0;
1942                 real_buffer[5] = 0;
1943                 real_buffer[6] = data->block_descriptor_length >> 8;
1944                 real_buffer[7] = data->block_descriptor_length;
1945
1946                 cmd[0] = MODE_SELECT_10;
1947                 cmd[7] = len >> 8;
1948                 cmd[8] = len;
1949         } else {
1950                 if (len > 255 || data->block_descriptor_length > 255 ||
1951                     data->longlba)
1952                         return -EINVAL;
1953
1954                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
1955                 if (!real_buffer)
1956                         return -ENOMEM;
1957                 memcpy(real_buffer + 4, buffer, len);
1958                 len += 4;
1959                 real_buffer[0] = 0;
1960                 real_buffer[1] = data->medium_type;
1961                 real_buffer[2] = data->device_specific;
1962                 real_buffer[3] = data->block_descriptor_length;
1963                 
1964
1965                 cmd[0] = MODE_SELECT;
1966                 cmd[4] = len;
1967         }
1968
1969         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
1970                                sshdr, timeout, retries);
1971         kfree(real_buffer);
1972         return ret;
1973 }
1974 EXPORT_SYMBOL_GPL(scsi_mode_select);
1975
1976 /**
1977  *      scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
1978  *      @sdev:  SCSI device to be queried
1979  *      @dbd:   set if mode sense will allow block descriptors to be returned
1980  *      @modepage: mode page being requested
1981  *      @buffer: request buffer (may not be smaller than eight bytes)
1982  *      @len:   length of request buffer.
1983  *      @timeout: command timeout
1984  *      @retries: number of retries before failing
1985  *      @data: returns a structure abstracting the mode header data
1986  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
1987  *              must be SCSI_SENSE_BUFFERSIZE big.
1988  *
1989  *      Returns zero if unsuccessful, or the header offset (either 4
1990  *      or 8 depending on whether a six or ten byte command was
1991  *      issued) if successful.
1992  */
1993 int
1994 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
1995                   unsigned char *buffer, int len, int timeout, int retries,
1996                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1997 {
1998         unsigned char cmd[12];
1999         int use_10_for_ms;
2000         int header_length;
2001         int result;
2002         struct scsi_sense_hdr my_sshdr;
2003
2004         memset(data, 0, sizeof(*data));
2005         memset(&cmd[0], 0, 12);
2006         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
2007         cmd[2] = modepage;
2008
2009         /* caller might not be interested in sense, but we need it */
2010         if (!sshdr)
2011                 sshdr = &my_sshdr;
2012
2013  retry:
2014         use_10_for_ms = sdev->use_10_for_ms;
2015
2016         if (use_10_for_ms) {
2017                 if (len < 8)
2018                         len = 8;
2019
2020                 cmd[0] = MODE_SENSE_10;
2021                 cmd[8] = len;
2022                 header_length = 8;
2023         } else {
2024                 if (len < 4)
2025                         len = 4;
2026
2027                 cmd[0] = MODE_SENSE;
2028                 cmd[4] = len;
2029                 header_length = 4;
2030         }
2031
2032         memset(buffer, 0, len);
2033
2034         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
2035                                   sshdr, timeout, retries);
2036
2037         /* This code looks awful: what it's doing is making sure an
2038          * ILLEGAL REQUEST sense return identifies the actual command
2039          * byte as the problem.  MODE_SENSE commands can return
2040          * ILLEGAL REQUEST if the code page isn't supported */
2041
2042         if (use_10_for_ms && !scsi_status_is_good(result) &&
2043             (driver_byte(result) & DRIVER_SENSE)) {
2044                 if (scsi_sense_valid(sshdr)) {
2045                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
2046                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
2047                                 /* 
2048                                  * Invalid command operation code
2049                                  */
2050                                 sdev->use_10_for_ms = 0;
2051                                 goto retry;
2052                         }
2053                 }
2054         }
2055
2056         if(scsi_status_is_good(result)) {
2057                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
2058                              (modepage == 6 || modepage == 8))) {
2059                         /* Initio breakage? */
2060                         header_length = 0;
2061                         data->length = 13;
2062                         data->medium_type = 0;
2063                         data->device_specific = 0;
2064                         data->longlba = 0;
2065                         data->block_descriptor_length = 0;
2066                 } else if(use_10_for_ms) {
2067                         data->length = buffer[0]*256 + buffer[1] + 2;
2068                         data->medium_type = buffer[2];
2069                         data->device_specific = buffer[3];
2070                         data->longlba = buffer[4] & 0x01;
2071                         data->block_descriptor_length = buffer[6]*256
2072                                 + buffer[7];
2073                 } else {
2074                         data->length = buffer[0] + 1;
2075                         data->medium_type = buffer[1];
2076                         data->device_specific = buffer[2];
2077                         data->block_descriptor_length = buffer[3];
2078                 }
2079                 data->header_length = header_length;
2080         }
2081
2082         return result;
2083 }
2084 EXPORT_SYMBOL(scsi_mode_sense);
2085
2086 /**
2087  *      scsi_test_unit_ready - test if unit is ready
2088  *      @sdev:  scsi device to change the state of.
2089  *      @timeout: command timeout
2090  *      @retries: number of retries before failing
2091  *      @sshdr_external: Optional pointer to struct scsi_sense_hdr for
2092  *              returning sense. Make sure that this is cleared before passing
2093  *              in.
2094  *
2095  *      Returns zero if unsuccessful or an error if TUR failed.  For
2096  *      removable media, a return of NOT_READY or UNIT_ATTENTION is
2097  *      translated to success, with the ->changed flag updated.
2098  **/
2099 int
2100 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
2101                      struct scsi_sense_hdr *sshdr_external)
2102 {
2103         char cmd[] = {
2104                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
2105         };
2106         struct scsi_sense_hdr *sshdr;
2107         int result;
2108
2109         if (!sshdr_external)
2110                 sshdr = kzalloc(sizeof(*sshdr), GFP_KERNEL);
2111         else
2112                 sshdr = sshdr_external;
2113
2114         /* try to eat the UNIT_ATTENTION if there are enough retries */
2115         do {
2116                 result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
2117                                           timeout, retries);
2118                 if (sdev->removable && scsi_sense_valid(sshdr) &&
2119                     sshdr->sense_key == UNIT_ATTENTION)
2120                         sdev->changed = 1;
2121         } while (scsi_sense_valid(sshdr) &&
2122                  sshdr->sense_key == UNIT_ATTENTION && --retries);
2123
2124         if (!sshdr)
2125                 /* could not allocate sense buffer, so can't process it */
2126                 return result;
2127
2128         if (sdev->removable && scsi_sense_valid(sshdr) &&
2129             (sshdr->sense_key == UNIT_ATTENTION ||
2130              sshdr->sense_key == NOT_READY)) {
2131                 sdev->changed = 1;
2132                 result = 0;
2133         }
2134         if (!sshdr_external)
2135                 kfree(sshdr);
2136         return result;
2137 }
2138 EXPORT_SYMBOL(scsi_test_unit_ready);
2139
2140 /**
2141  *      scsi_device_set_state - Take the given device through the device state model.
2142  *      @sdev:  scsi device to change the state of.
2143  *      @state: state to change to.
2144  *
2145  *      Returns zero if unsuccessful or an error if the requested 
2146  *      transition is illegal.
2147  */
2148 int
2149 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2150 {
2151         enum scsi_device_state oldstate = sdev->sdev_state;
2152
2153         if (state == oldstate)
2154                 return 0;
2155
2156         switch (state) {
2157         case SDEV_CREATED:
2158                 switch (oldstate) {
2159                 case SDEV_CREATED_BLOCK:
2160                         break;
2161                 default:
2162                         goto illegal;
2163                 }
2164                 break;
2165                         
2166         case SDEV_RUNNING:
2167                 switch (oldstate) {
2168                 case SDEV_CREATED:
2169                 case SDEV_OFFLINE:
2170                 case SDEV_QUIESCE:
2171                 case SDEV_BLOCK:
2172                         break;
2173                 default:
2174                         goto illegal;
2175                 }
2176                 break;
2177
2178         case SDEV_QUIESCE:
2179                 switch (oldstate) {
2180                 case SDEV_RUNNING:
2181                 case SDEV_OFFLINE:
2182                         break;
2183                 default:
2184                         goto illegal;
2185                 }
2186                 break;
2187
2188         case SDEV_OFFLINE:
2189                 switch (oldstate) {
2190                 case SDEV_CREATED:
2191                 case SDEV_RUNNING:
2192                 case SDEV_QUIESCE:
2193                 case SDEV_BLOCK:
2194                         break;
2195                 default:
2196                         goto illegal;
2197                 }
2198                 break;
2199
2200         case SDEV_BLOCK:
2201                 switch (oldstate) {
2202                 case SDEV_RUNNING:
2203                 case SDEV_CREATED_BLOCK:
2204                         break;
2205                 default:
2206                         goto illegal;
2207                 }
2208                 break;
2209
2210         case SDEV_CREATED_BLOCK:
2211                 switch (oldstate) {
2212                 case SDEV_CREATED:
2213                         break;
2214                 default:
2215                         goto illegal;
2216                 }
2217                 break;
2218
2219         case SDEV_CANCEL:
2220                 switch (oldstate) {
2221                 case SDEV_CREATED:
2222                 case SDEV_RUNNING:
2223                 case SDEV_QUIESCE:
2224                 case SDEV_OFFLINE:
2225                 case SDEV_BLOCK:
2226                         break;
2227                 default:
2228                         goto illegal;
2229                 }
2230                 break;
2231
2232         case SDEV_DEL:
2233                 switch (oldstate) {
2234                 case SDEV_CREATED:
2235                 case SDEV_RUNNING:
2236                 case SDEV_OFFLINE:
2237                 case SDEV_CANCEL:
2238                         break;
2239                 default:
2240                         goto illegal;
2241                 }
2242                 break;
2243
2244         }
2245         sdev->sdev_state = state;
2246         return 0;
2247
2248  illegal:
2249         SCSI_LOG_ERROR_RECOVERY(1, 
2250                                 sdev_printk(KERN_ERR, sdev,
2251                                             "Illegal state transition %s->%s\n",
2252                                             scsi_device_state_name(oldstate),
2253                                             scsi_device_state_name(state))
2254                                 );
2255         return -EINVAL;
2256 }
2257 EXPORT_SYMBOL(scsi_device_set_state);
2258
2259 /**
2260  *      sdev_evt_emit - emit a single SCSI device uevent
2261  *      @sdev: associated SCSI device
2262  *      @evt: event to emit
2263  *
2264  *      Send a single uevent (scsi_event) to the associated scsi_device.
2265  */
2266 static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2267 {
2268         int idx = 0;
2269         char *envp[3];
2270
2271         switch (evt->evt_type) {
2272         case SDEV_EVT_MEDIA_CHANGE:
2273                 envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2274                 break;
2275
2276         default:
2277                 /* do nothing */
2278                 break;
2279         }
2280
2281         envp[idx++] = NULL;
2282
2283         kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2284 }
2285
2286 /**
2287  *      sdev_evt_thread - send a uevent for each scsi event
2288  *      @work: work struct for scsi_device
2289  *
2290  *      Dispatch queued events to their associated scsi_device kobjects
2291  *      as uevents.
2292  */
2293 void scsi_evt_thread(struct work_struct *work)
2294 {
2295         struct scsi_device *sdev;
2296         LIST_HEAD(event_list);
2297
2298         sdev = container_of(work, struct scsi_device, event_work);
2299
2300         while (1) {
2301                 struct scsi_event *evt;
2302                 struct list_head *this, *tmp;
2303                 unsigned long flags;
2304
2305                 spin_lock_irqsave(&sdev->list_lock, flags);
2306                 list_splice_init(&sdev->event_list, &event_list);
2307                 spin_unlock_irqrestore(&sdev->list_lock, flags);
2308
2309                 if (list_empty(&event_list))
2310                         break;
2311
2312                 list_for_each_safe(this, tmp, &event_list) {
2313                         evt = list_entry(this, struct scsi_event, node);
2314                         list_del(&evt->node);
2315                         scsi_evt_emit(sdev, evt);
2316                         kfree(evt);
2317                 }
2318         }
2319 }
2320
2321 /**
2322  *      sdev_evt_send - send asserted event to uevent thread
2323  *      @sdev: scsi_device event occurred on
2324  *      @evt: event to send
2325  *
2326  *      Assert scsi device event asynchronously.
2327  */
2328 void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2329 {
2330         unsigned long flags;
2331
2332 #if 0
2333         /* FIXME: currently this check eliminates all media change events
2334          * for polled devices.  Need to update to discriminate between AN
2335          * and polled events */
2336         if (!test_bit(evt->evt_type, sdev->supported_events)) {
2337                 kfree(evt);
2338                 return;
2339         }
2340 #endif
2341
2342         spin_lock_irqsave(&sdev->list_lock, flags);
2343         list_add_tail(&evt->node, &sdev->event_list);
2344         schedule_work(&sdev->event_work);
2345         spin_unlock_irqrestore(&sdev->list_lock, flags);
2346 }
2347 EXPORT_SYMBOL_GPL(sdev_evt_send);
2348
2349 /**
2350  *      sdev_evt_alloc - allocate a new scsi event
2351  *      @evt_type: type of event to allocate
2352  *      @gfpflags: GFP flags for allocation
2353  *
2354  *      Allocates and returns a new scsi_event.
2355  */
2356 struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2357                                   gfp_t gfpflags)
2358 {
2359         struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2360         if (!evt)
2361                 return NULL;
2362
2363         evt->evt_type = evt_type;
2364         INIT_LIST_HEAD(&evt->node);
2365
2366         /* evt_type-specific initialization, if any */
2367         switch (evt_type) {
2368         case SDEV_EVT_MEDIA_CHANGE:
2369         default:
2370                 /* do nothing */
2371                 break;
2372         }
2373
2374         return evt;
2375 }
2376 EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2377
2378 /**
2379  *      sdev_evt_send_simple - send asserted event to uevent thread
2380  *      @sdev: scsi_device event occurred on
2381  *      @evt_type: type of event to send
2382  *      @gfpflags: GFP flags for allocation
2383  *
2384  *      Assert scsi device event asynchronously, given an event type.
2385  */
2386 void sdev_evt_send_simple(struct scsi_device *sdev,
2387                           enum scsi_device_event evt_type, gfp_t gfpflags)
2388 {
2389         struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2390         if (!evt) {
2391                 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2392                             evt_type);
2393                 return;
2394         }
2395
2396         sdev_evt_send(sdev, evt);
2397 }
2398 EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2399
2400 /**
2401  *      scsi_device_quiesce - Block user issued commands.
2402  *      @sdev:  scsi device to quiesce.
2403  *
2404  *      This works by trying to transition to the SDEV_QUIESCE state
2405  *      (which must be a legal transition).  When the device is in this
2406  *      state, only special requests will be accepted, all others will
2407  *      be deferred.  Since special requests may also be requeued requests,
2408  *      a successful return doesn't guarantee the device will be 
2409  *      totally quiescent.
2410  *
2411  *      Must be called with user context, may sleep.
2412  *
2413  *      Returns zero if unsuccessful or an error if not.
2414  */
2415 int
2416 scsi_device_quiesce(struct scsi_device *sdev)
2417 {
2418         int err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2419         if (err)
2420                 return err;
2421
2422         scsi_run_queue(sdev->request_queue);
2423         while (sdev->device_busy) {
2424                 msleep_interruptible(200);
2425                 scsi_run_queue(sdev->request_queue);
2426         }
2427         return 0;
2428 }
2429 EXPORT_SYMBOL(scsi_device_quiesce);
2430
2431 /**
2432  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2433  *      @sdev:  scsi device to resume.
2434  *
2435  *      Moves the device from quiesced back to running and restarts the
2436  *      queues.
2437  *
2438  *      Must be called with user context, may sleep.
2439  */
2440 void
2441 scsi_device_resume(struct scsi_device *sdev)
2442 {
2443         if(scsi_device_set_state(sdev, SDEV_RUNNING))
2444                 return;
2445         scsi_run_queue(sdev->request_queue);
2446 }
2447 EXPORT_SYMBOL(scsi_device_resume);
2448
2449 static void
2450 device_quiesce_fn(struct scsi_device *sdev, void *data)
2451 {
2452         scsi_device_quiesce(sdev);
2453 }
2454
2455 void
2456 scsi_target_quiesce(struct scsi_target *starget)
2457 {
2458         starget_for_each_device(starget, NULL, device_quiesce_fn);
2459 }
2460 EXPORT_SYMBOL(scsi_target_quiesce);
2461
2462 static void
2463 device_resume_fn(struct scsi_device *sdev, void *data)
2464 {
2465         scsi_device_resume(sdev);
2466 }
2467
2468 void
2469 scsi_target_resume(struct scsi_target *starget)
2470 {
2471         starget_for_each_device(starget, NULL, device_resume_fn);
2472 }
2473 EXPORT_SYMBOL(scsi_target_resume);
2474
2475 /**
2476  * scsi_internal_device_block - internal function to put a device temporarily into the SDEV_BLOCK state
2477  * @sdev:       device to block
2478  *
2479  * Block request made by scsi lld's to temporarily stop all
2480  * scsi commands on the specified device.  Called from interrupt
2481  * or normal process context.
2482  *
2483  * Returns zero if successful or error if not
2484  *
2485  * Notes:       
2486  *      This routine transitions the device to the SDEV_BLOCK state
2487  *      (which must be a legal transition).  When the device is in this
2488  *      state, all commands are deferred until the scsi lld reenables
2489  *      the device with scsi_device_unblock or device_block_tmo fires.
2490  *      This routine assumes the host_lock is held on entry.
2491  */
2492 int
2493 scsi_internal_device_block(struct scsi_device *sdev)
2494 {
2495         struct request_queue *q = sdev->request_queue;
2496         unsigned long flags;
2497         int err = 0;
2498
2499         err = scsi_device_set_state(sdev, SDEV_BLOCK);
2500         if (err) {
2501                 err = scsi_device_set_state(sdev, SDEV_CREATED_BLOCK);
2502
2503                 if (err)
2504                         return err;
2505         }
2506
2507         /* 
2508          * The device has transitioned to SDEV_BLOCK.  Stop the
2509          * block layer from calling the midlayer with this device's
2510          * request queue. 
2511          */
2512         spin_lock_irqsave(q->queue_lock, flags);
2513         blk_stop_queue(q);
2514         spin_unlock_irqrestore(q->queue_lock, flags);
2515
2516         return 0;
2517 }
2518 EXPORT_SYMBOL_GPL(scsi_internal_device_block);
2519  
2520 /**
2521  * scsi_internal_device_unblock - resume a device after a block request
2522  * @sdev:       device to resume
2523  *
2524  * Called by scsi lld's or the midlayer to restart the device queue
2525  * for the previously suspended scsi device.  Called from interrupt or
2526  * normal process context.
2527  *
2528  * Returns zero if successful or error if not.
2529  *
2530  * Notes:       
2531  *      This routine transitions the device to the SDEV_RUNNING state
2532  *      (which must be a legal transition) allowing the midlayer to
2533  *      goose the queue for this device.  This routine assumes the 
2534  *      host_lock is held upon entry.
2535  */
2536 int
2537 scsi_internal_device_unblock(struct scsi_device *sdev)
2538 {
2539         struct request_queue *q = sdev->request_queue; 
2540         int err;
2541         unsigned long flags;
2542         
2543         /* 
2544          * Try to transition the scsi device to SDEV_RUNNING
2545          * and goose the device queue if successful.  
2546          */
2547         err = scsi_device_set_state(sdev, SDEV_RUNNING);
2548         if (err) {
2549                 err = scsi_device_set_state(sdev, SDEV_CREATED);
2550
2551                 if (err)
2552                         return err;
2553         }
2554
2555         spin_lock_irqsave(q->queue_lock, flags);
2556         blk_start_queue(q);
2557         spin_unlock_irqrestore(q->queue_lock, flags);
2558
2559         return 0;
2560 }
2561 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock);
2562
2563 static void
2564 device_block(struct scsi_device *sdev, void *data)
2565 {
2566         scsi_internal_device_block(sdev);
2567 }
2568
2569 static int
2570 target_block(struct device *dev, void *data)
2571 {
2572         if (scsi_is_target_device(dev))
2573                 starget_for_each_device(to_scsi_target(dev), NULL,
2574                                         device_block);
2575         return 0;
2576 }
2577
2578 void
2579 scsi_target_block(struct device *dev)
2580 {
2581         if (scsi_is_target_device(dev))
2582                 starget_for_each_device(to_scsi_target(dev), NULL,
2583                                         device_block);
2584         else
2585                 device_for_each_child(dev, NULL, target_block);
2586 }
2587 EXPORT_SYMBOL_GPL(scsi_target_block);
2588
2589 static void
2590 device_unblock(struct scsi_device *sdev, void *data)
2591 {
2592         scsi_internal_device_unblock(sdev);
2593 }
2594
2595 static int
2596 target_unblock(struct device *dev, void *data)
2597 {
2598         if (scsi_is_target_device(dev))
2599                 starget_for_each_device(to_scsi_target(dev), NULL,
2600                                         device_unblock);
2601         return 0;
2602 }
2603
2604 void
2605 scsi_target_unblock(struct device *dev)
2606 {
2607         if (scsi_is_target_device(dev))
2608                 starget_for_each_device(to_scsi_target(dev), NULL,
2609                                         device_unblock);
2610         else
2611                 device_for_each_child(dev, NULL, target_unblock);
2612 }
2613 EXPORT_SYMBOL_GPL(scsi_target_unblock);
2614
2615 /**
2616  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2617  * @sgl:        scatter-gather list
2618  * @sg_count:   number of segments in sg
2619  * @offset:     offset in bytes into sg, on return offset into the mapped area
2620  * @len:        bytes to map, on return number of bytes mapped
2621  *
2622  * Returns virtual address of the start of the mapped page
2623  */
2624 void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
2625                           size_t *offset, size_t *len)
2626 {
2627         int i;
2628         size_t sg_len = 0, len_complete = 0;
2629         struct scatterlist *sg;
2630         struct page *page;
2631
2632         WARN_ON(!irqs_disabled());
2633
2634         for_each_sg(sgl, sg, sg_count, i) {
2635                 len_complete = sg_len; /* Complete sg-entries */
2636                 sg_len += sg->length;
2637                 if (sg_len > *offset)
2638                         break;
2639         }
2640
2641         if (unlikely(i == sg_count)) {
2642                 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2643                         "elements %d\n",
2644                        __func__, sg_len, *offset, sg_count);
2645                 WARN_ON(1);
2646                 return NULL;
2647         }
2648
2649         /* Offset starting from the beginning of first page in this sg-entry */
2650         *offset = *offset - len_complete + sg->offset;
2651
2652         /* Assumption: contiguous pages can be accessed as "page + i" */
2653         page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
2654         *offset &= ~PAGE_MASK;
2655
2656         /* Bytes in this sg-entry from *offset to the end of the page */
2657         sg_len = PAGE_SIZE - *offset;
2658         if (*len > sg_len)
2659                 *len = sg_len;
2660
2661         return kmap_atomic(page, KM_BIO_SRC_IRQ);
2662 }
2663 EXPORT_SYMBOL(scsi_kmap_atomic_sg);
2664
2665 /**
2666  * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
2667  * @virt:       virtual address to be unmapped
2668  */
2669 void scsi_kunmap_atomic_sg(void *virt)
2670 {
2671         kunmap_atomic(virt, KM_BIO_SRC_IRQ);
2672 }
2673 EXPORT_SYMBOL(scsi_kunmap_atomic_sg);