[SCSI] export busy state via q->lld_busy_fn()
[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  * Busy state exporting function for request stacking drivers.
1469  *
1470  * For efficiency, no lock is taken to check the busy state of
1471  * shost/starget/sdev, since the returned value is not guaranteed and
1472  * may be changed after request stacking drivers call the function,
1473  * regardless of taking lock or not.
1474  *
1475  * When scsi can't dispatch I/Os anymore and needs to kill I/Os
1476  * (e.g. !sdev), scsi needs to return 'not busy'.
1477  * Otherwise, request stacking drivers may hold requests forever.
1478  */
1479 static int scsi_lld_busy(struct request_queue *q)
1480 {
1481         struct scsi_device *sdev = q->queuedata;
1482         struct Scsi_Host *shost;
1483         struct scsi_target *starget;
1484
1485         if (!sdev)
1486                 return 0;
1487
1488         shost = sdev->host;
1489         starget = scsi_target(sdev);
1490
1491         if (scsi_host_in_recovery(shost) || scsi_host_is_busy(shost) ||
1492             scsi_target_is_busy(starget) || scsi_device_is_busy(sdev))
1493                 return 1;
1494
1495         return 0;
1496 }
1497
1498 /*
1499  * Kill a request for a dead device
1500  */
1501 static void scsi_kill_request(struct request *req, struct request_queue *q)
1502 {
1503         struct scsi_cmnd *cmd = req->special;
1504         struct scsi_device *sdev = cmd->device;
1505         struct scsi_target *starget = scsi_target(sdev);
1506         struct Scsi_Host *shost = sdev->host;
1507
1508         blkdev_dequeue_request(req);
1509
1510         if (unlikely(cmd == NULL)) {
1511                 printk(KERN_CRIT "impossible request in %s.\n",
1512                                  __func__);
1513                 BUG();
1514         }
1515
1516         scsi_init_cmd_errh(cmd);
1517         cmd->result = DID_NO_CONNECT << 16;
1518         atomic_inc(&cmd->device->iorequest_cnt);
1519
1520         /*
1521          * SCSI request completion path will do scsi_device_unbusy(),
1522          * bump busy counts.  To bump the counters, we need to dance
1523          * with the locks as normal issue path does.
1524          */
1525         sdev->device_busy++;
1526         spin_unlock(sdev->request_queue->queue_lock);
1527         spin_lock(shost->host_lock);
1528         shost->host_busy++;
1529         starget->target_busy++;
1530         spin_unlock(shost->host_lock);
1531         spin_lock(sdev->request_queue->queue_lock);
1532
1533         blk_complete_request(req);
1534 }
1535
1536 static void scsi_softirq_done(struct request *rq)
1537 {
1538         struct scsi_cmnd *cmd = rq->special;
1539         unsigned long wait_for = (cmd->allowed + 1) * rq->timeout;
1540         int disposition;
1541
1542         INIT_LIST_HEAD(&cmd->eh_entry);
1543
1544         /*
1545          * Set the serial numbers back to zero
1546          */
1547         cmd->serial_number = 0;
1548
1549         atomic_inc(&cmd->device->iodone_cnt);
1550         if (cmd->result)
1551                 atomic_inc(&cmd->device->ioerr_cnt);
1552
1553         disposition = scsi_decide_disposition(cmd);
1554         if (disposition != SUCCESS &&
1555             time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1556                 sdev_printk(KERN_ERR, cmd->device,
1557                             "timing out command, waited %lus\n",
1558                             wait_for/HZ);
1559                 disposition = SUCCESS;
1560         }
1561                         
1562         scsi_log_completion(cmd, disposition);
1563
1564         switch (disposition) {
1565                 case SUCCESS:
1566                         scsi_finish_command(cmd);
1567                         break;
1568                 case NEEDS_RETRY:
1569                         scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1570                         break;
1571                 case ADD_TO_MLQUEUE:
1572                         scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1573                         break;
1574                 default:
1575                         if (!scsi_eh_scmd_add(cmd, 0))
1576                                 scsi_finish_command(cmd);
1577         }
1578 }
1579
1580 /*
1581  * Function:    scsi_request_fn()
1582  *
1583  * Purpose:     Main strategy routine for SCSI.
1584  *
1585  * Arguments:   q       - Pointer to actual queue.
1586  *
1587  * Returns:     Nothing
1588  *
1589  * Lock status: IO request lock assumed to be held when called.
1590  */
1591 static void scsi_request_fn(struct request_queue *q)
1592 {
1593         struct scsi_device *sdev = q->queuedata;
1594         struct Scsi_Host *shost;
1595         struct scsi_cmnd *cmd;
1596         struct request *req;
1597
1598         if (!sdev) {
1599                 printk("scsi: killing requests for dead queue\n");
1600                 while ((req = elv_next_request(q)) != NULL)
1601                         scsi_kill_request(req, q);
1602                 return;
1603         }
1604
1605         if(!get_device(&sdev->sdev_gendev))
1606                 /* We must be tearing the block queue down already */
1607                 return;
1608
1609         /*
1610          * To start with, we keep looping until the queue is empty, or until
1611          * the host is no longer able to accept any more requests.
1612          */
1613         shost = sdev->host;
1614         while (!blk_queue_plugged(q)) {
1615                 int rtn;
1616                 /*
1617                  * get next queueable request.  We do this early to make sure
1618                  * that the request is fully prepared even if we cannot 
1619                  * accept it.
1620                  */
1621                 req = elv_next_request(q);
1622                 if (!req || !scsi_dev_queue_ready(q, sdev))
1623                         break;
1624
1625                 if (unlikely(!scsi_device_online(sdev))) {
1626                         sdev_printk(KERN_ERR, sdev,
1627                                     "rejecting I/O to offline device\n");
1628                         scsi_kill_request(req, q);
1629                         continue;
1630                 }
1631
1632
1633                 /*
1634                  * Remove the request from the request list.
1635                  */
1636                 if (!(blk_queue_tagged(q) && !blk_queue_start_tag(q, req)))
1637                         blkdev_dequeue_request(req);
1638                 sdev->device_busy++;
1639
1640                 spin_unlock(q->queue_lock);
1641                 cmd = req->special;
1642                 if (unlikely(cmd == NULL)) {
1643                         printk(KERN_CRIT "impossible request in %s.\n"
1644                                          "please mail a stack trace to "
1645                                          "linux-scsi@vger.kernel.org\n",
1646                                          __func__);
1647                         blk_dump_rq_flags(req, "foo");
1648                         BUG();
1649                 }
1650                 spin_lock(shost->host_lock);
1651
1652                 /*
1653                  * We hit this when the driver is using a host wide
1654                  * tag map. For device level tag maps the queue_depth check
1655                  * in the device ready fn would prevent us from trying
1656                  * to allocate a tag. Since the map is a shared host resource
1657                  * we add the dev to the starved list so it eventually gets
1658                  * a run when a tag is freed.
1659                  */
1660                 if (blk_queue_tagged(q) && !blk_rq_tagged(req)) {
1661                         if (list_empty(&sdev->starved_entry))
1662                                 list_add_tail(&sdev->starved_entry,
1663                                               &shost->starved_list);
1664                         goto not_ready;
1665                 }
1666
1667                 if (!scsi_target_queue_ready(shost, sdev))
1668                         goto not_ready;
1669
1670                 if (!scsi_host_queue_ready(q, shost, sdev))
1671                         goto not_ready;
1672
1673                 scsi_target(sdev)->target_busy++;
1674                 shost->host_busy++;
1675
1676                 /*
1677                  * XXX(hch): This is rather suboptimal, scsi_dispatch_cmd will
1678                  *              take the lock again.
1679                  */
1680                 spin_unlock_irq(shost->host_lock);
1681
1682                 /*
1683                  * Finally, initialize any error handling parameters, and set up
1684                  * the timers for timeouts.
1685                  */
1686                 scsi_init_cmd_errh(cmd);
1687
1688                 /*
1689                  * Dispatch the command to the low-level driver.
1690                  */
1691                 rtn = scsi_dispatch_cmd(cmd);
1692                 spin_lock_irq(q->queue_lock);
1693                 if(rtn) {
1694                         /* we're refusing the command; because of
1695                          * the way locks get dropped, we need to 
1696                          * check here if plugging is required */
1697                         if(sdev->device_busy == 0)
1698                                 blk_plug_device(q);
1699
1700                         break;
1701                 }
1702         }
1703
1704         goto out;
1705
1706  not_ready:
1707         spin_unlock_irq(shost->host_lock);
1708
1709         /*
1710          * lock q, handle tag, requeue req, and decrement device_busy. We
1711          * must return with queue_lock held.
1712          *
1713          * Decrementing device_busy without checking it is OK, as all such
1714          * cases (host limits or settings) should run the queue at some
1715          * later time.
1716          */
1717         spin_lock_irq(q->queue_lock);
1718         blk_requeue_request(q, req);
1719         sdev->device_busy--;
1720         if(sdev->device_busy == 0)
1721                 blk_plug_device(q);
1722  out:
1723         /* must be careful here...if we trigger the ->remove() function
1724          * we cannot be holding the q lock */
1725         spin_unlock_irq(q->queue_lock);
1726         put_device(&sdev->sdev_gendev);
1727         spin_lock_irq(q->queue_lock);
1728 }
1729
1730 u64 scsi_calculate_bounce_limit(struct Scsi_Host *shost)
1731 {
1732         struct device *host_dev;
1733         u64 bounce_limit = 0xffffffff;
1734
1735         if (shost->unchecked_isa_dma)
1736                 return BLK_BOUNCE_ISA;
1737         /*
1738          * Platforms with virtual-DMA translation
1739          * hardware have no practical limit.
1740          */
1741         if (!PCI_DMA_BUS_IS_PHYS)
1742                 return BLK_BOUNCE_ANY;
1743
1744         host_dev = scsi_get_device(shost);
1745         if (host_dev && host_dev->dma_mask)
1746                 bounce_limit = *host_dev->dma_mask;
1747
1748         return bounce_limit;
1749 }
1750 EXPORT_SYMBOL(scsi_calculate_bounce_limit);
1751
1752 struct request_queue *__scsi_alloc_queue(struct Scsi_Host *shost,
1753                                          request_fn_proc *request_fn)
1754 {
1755         struct request_queue *q;
1756         struct device *dev = shost->shost_gendev.parent;
1757
1758         q = blk_init_queue(request_fn, NULL);
1759         if (!q)
1760                 return NULL;
1761
1762         /*
1763          * this limit is imposed by hardware restrictions
1764          */
1765         blk_queue_max_hw_segments(q, shost->sg_tablesize);
1766         blk_queue_max_phys_segments(q, SCSI_MAX_SG_CHAIN_SEGMENTS);
1767
1768         blk_queue_max_sectors(q, shost->max_sectors);
1769         blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
1770         blk_queue_segment_boundary(q, shost->dma_boundary);
1771         dma_set_seg_boundary(dev, shost->dma_boundary);
1772
1773         blk_queue_max_segment_size(q, dma_get_max_seg_size(dev));
1774
1775         /* New queue, no concurrency on queue_flags */
1776         if (!shost->use_clustering)
1777                 queue_flag_clear_unlocked(QUEUE_FLAG_CLUSTER, q);
1778
1779         /*
1780          * set a reasonable default alignment on word boundaries: the
1781          * host and device may alter it using
1782          * blk_queue_update_dma_alignment() later.
1783          */
1784         blk_queue_dma_alignment(q, 0x03);
1785
1786         return q;
1787 }
1788 EXPORT_SYMBOL(__scsi_alloc_queue);
1789
1790 struct request_queue *scsi_alloc_queue(struct scsi_device *sdev)
1791 {
1792         struct request_queue *q;
1793
1794         q = __scsi_alloc_queue(sdev->host, scsi_request_fn);
1795         if (!q)
1796                 return NULL;
1797
1798         blk_queue_prep_rq(q, scsi_prep_fn);
1799         blk_queue_softirq_done(q, scsi_softirq_done);
1800         blk_queue_rq_timed_out(q, scsi_times_out);
1801         blk_queue_lld_busy(q, scsi_lld_busy);
1802         return q;
1803 }
1804
1805 void scsi_free_queue(struct request_queue *q)
1806 {
1807         blk_cleanup_queue(q);
1808 }
1809
1810 /*
1811  * Function:    scsi_block_requests()
1812  *
1813  * Purpose:     Utility function used by low-level drivers to prevent further
1814  *              commands from being queued to the device.
1815  *
1816  * Arguments:   shost       - Host in question
1817  *
1818  * Returns:     Nothing
1819  *
1820  * Lock status: No locks are assumed held.
1821  *
1822  * Notes:       There is no timer nor any other means by which the requests
1823  *              get unblocked other than the low-level driver calling
1824  *              scsi_unblock_requests().
1825  */
1826 void scsi_block_requests(struct Scsi_Host *shost)
1827 {
1828         shost->host_self_blocked = 1;
1829 }
1830 EXPORT_SYMBOL(scsi_block_requests);
1831
1832 /*
1833  * Function:    scsi_unblock_requests()
1834  *
1835  * Purpose:     Utility function used by low-level drivers to allow further
1836  *              commands from being queued to the device.
1837  *
1838  * Arguments:   shost       - Host in question
1839  *
1840  * Returns:     Nothing
1841  *
1842  * Lock status: No locks are assumed held.
1843  *
1844  * Notes:       There is no timer nor any other means by which the requests
1845  *              get unblocked other than the low-level driver calling
1846  *              scsi_unblock_requests().
1847  *
1848  *              This is done as an API function so that changes to the
1849  *              internals of the scsi mid-layer won't require wholesale
1850  *              changes to drivers that use this feature.
1851  */
1852 void scsi_unblock_requests(struct Scsi_Host *shost)
1853 {
1854         shost->host_self_blocked = 0;
1855         scsi_run_host_queues(shost);
1856 }
1857 EXPORT_SYMBOL(scsi_unblock_requests);
1858
1859 int __init scsi_init_queue(void)
1860 {
1861         int i;
1862
1863         scsi_io_context_cache = kmem_cache_create("scsi_io_context",
1864                                         sizeof(struct scsi_io_context),
1865                                         0, 0, NULL);
1866         if (!scsi_io_context_cache) {
1867                 printk(KERN_ERR "SCSI: can't init scsi io context cache\n");
1868                 return -ENOMEM;
1869         }
1870
1871         scsi_sdb_cache = kmem_cache_create("scsi_data_buffer",
1872                                            sizeof(struct scsi_data_buffer),
1873                                            0, 0, NULL);
1874         if (!scsi_sdb_cache) {
1875                 printk(KERN_ERR "SCSI: can't init scsi sdb cache\n");
1876                 goto cleanup_io_context;
1877         }
1878
1879         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1880                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1881                 int size = sgp->size * sizeof(struct scatterlist);
1882
1883                 sgp->slab = kmem_cache_create(sgp->name, size, 0,
1884                                 SLAB_HWCACHE_ALIGN, NULL);
1885                 if (!sgp->slab) {
1886                         printk(KERN_ERR "SCSI: can't init sg slab %s\n",
1887                                         sgp->name);
1888                         goto cleanup_sdb;
1889                 }
1890
1891                 sgp->pool = mempool_create_slab_pool(SG_MEMPOOL_SIZE,
1892                                                      sgp->slab);
1893                 if (!sgp->pool) {
1894                         printk(KERN_ERR "SCSI: can't init sg mempool %s\n",
1895                                         sgp->name);
1896                         goto cleanup_sdb;
1897                 }
1898         }
1899
1900         return 0;
1901
1902 cleanup_sdb:
1903         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1904                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1905                 if (sgp->pool)
1906                         mempool_destroy(sgp->pool);
1907                 if (sgp->slab)
1908                         kmem_cache_destroy(sgp->slab);
1909         }
1910         kmem_cache_destroy(scsi_sdb_cache);
1911 cleanup_io_context:
1912         kmem_cache_destroy(scsi_io_context_cache);
1913
1914         return -ENOMEM;
1915 }
1916
1917 void scsi_exit_queue(void)
1918 {
1919         int i;
1920
1921         kmem_cache_destroy(scsi_io_context_cache);
1922         kmem_cache_destroy(scsi_sdb_cache);
1923
1924         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1925                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1926                 mempool_destroy(sgp->pool);
1927                 kmem_cache_destroy(sgp->slab);
1928         }
1929 }
1930
1931 /**
1932  *      scsi_mode_select - issue a mode select
1933  *      @sdev:  SCSI device to be queried
1934  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
1935  *      @sp:    Save page bit (0 == don't save, 1 == save)
1936  *      @modepage: mode page being requested
1937  *      @buffer: request buffer (may not be smaller than eight bytes)
1938  *      @len:   length of request buffer.
1939  *      @timeout: command timeout
1940  *      @retries: number of retries before failing
1941  *      @data: returns a structure abstracting the mode header data
1942  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
1943  *              must be SCSI_SENSE_BUFFERSIZE big.
1944  *
1945  *      Returns zero if successful; negative error number or scsi
1946  *      status on error
1947  *
1948  */
1949 int
1950 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
1951                  unsigned char *buffer, int len, int timeout, int retries,
1952                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1953 {
1954         unsigned char cmd[10];
1955         unsigned char *real_buffer;
1956         int ret;
1957
1958         memset(cmd, 0, sizeof(cmd));
1959         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
1960
1961         if (sdev->use_10_for_ms) {
1962                 if (len > 65535)
1963                         return -EINVAL;
1964                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
1965                 if (!real_buffer)
1966                         return -ENOMEM;
1967                 memcpy(real_buffer + 8, buffer, len);
1968                 len += 8;
1969                 real_buffer[0] = 0;
1970                 real_buffer[1] = 0;
1971                 real_buffer[2] = data->medium_type;
1972                 real_buffer[3] = data->device_specific;
1973                 real_buffer[4] = data->longlba ? 0x01 : 0;
1974                 real_buffer[5] = 0;
1975                 real_buffer[6] = data->block_descriptor_length >> 8;
1976                 real_buffer[7] = data->block_descriptor_length;
1977
1978                 cmd[0] = MODE_SELECT_10;
1979                 cmd[7] = len >> 8;
1980                 cmd[8] = len;
1981         } else {
1982                 if (len > 255 || data->block_descriptor_length > 255 ||
1983                     data->longlba)
1984                         return -EINVAL;
1985
1986                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
1987                 if (!real_buffer)
1988                         return -ENOMEM;
1989                 memcpy(real_buffer + 4, buffer, len);
1990                 len += 4;
1991                 real_buffer[0] = 0;
1992                 real_buffer[1] = data->medium_type;
1993                 real_buffer[2] = data->device_specific;
1994                 real_buffer[3] = data->block_descriptor_length;
1995                 
1996
1997                 cmd[0] = MODE_SELECT;
1998                 cmd[4] = len;
1999         }
2000
2001         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
2002                                sshdr, timeout, retries);
2003         kfree(real_buffer);
2004         return ret;
2005 }
2006 EXPORT_SYMBOL_GPL(scsi_mode_select);
2007
2008 /**
2009  *      scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
2010  *      @sdev:  SCSI device to be queried
2011  *      @dbd:   set if mode sense will allow block descriptors to be returned
2012  *      @modepage: mode page being requested
2013  *      @buffer: request buffer (may not be smaller than eight bytes)
2014  *      @len:   length of request buffer.
2015  *      @timeout: command timeout
2016  *      @retries: number of retries before failing
2017  *      @data: returns a structure abstracting the mode header data
2018  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
2019  *              must be SCSI_SENSE_BUFFERSIZE big.
2020  *
2021  *      Returns zero if unsuccessful, or the header offset (either 4
2022  *      or 8 depending on whether a six or ten byte command was
2023  *      issued) if successful.
2024  */
2025 int
2026 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
2027                   unsigned char *buffer, int len, int timeout, int retries,
2028                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2029 {
2030         unsigned char cmd[12];
2031         int use_10_for_ms;
2032         int header_length;
2033         int result;
2034         struct scsi_sense_hdr my_sshdr;
2035
2036         memset(data, 0, sizeof(*data));
2037         memset(&cmd[0], 0, 12);
2038         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
2039         cmd[2] = modepage;
2040
2041         /* caller might not be interested in sense, but we need it */
2042         if (!sshdr)
2043                 sshdr = &my_sshdr;
2044
2045  retry:
2046         use_10_for_ms = sdev->use_10_for_ms;
2047
2048         if (use_10_for_ms) {
2049                 if (len < 8)
2050                         len = 8;
2051
2052                 cmd[0] = MODE_SENSE_10;
2053                 cmd[8] = len;
2054                 header_length = 8;
2055         } else {
2056                 if (len < 4)
2057                         len = 4;
2058
2059                 cmd[0] = MODE_SENSE;
2060                 cmd[4] = len;
2061                 header_length = 4;
2062         }
2063
2064         memset(buffer, 0, len);
2065
2066         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
2067                                   sshdr, timeout, retries);
2068
2069         /* This code looks awful: what it's doing is making sure an
2070          * ILLEGAL REQUEST sense return identifies the actual command
2071          * byte as the problem.  MODE_SENSE commands can return
2072          * ILLEGAL REQUEST if the code page isn't supported */
2073
2074         if (use_10_for_ms && !scsi_status_is_good(result) &&
2075             (driver_byte(result) & DRIVER_SENSE)) {
2076                 if (scsi_sense_valid(sshdr)) {
2077                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
2078                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
2079                                 /* 
2080                                  * Invalid command operation code
2081                                  */
2082                                 sdev->use_10_for_ms = 0;
2083                                 goto retry;
2084                         }
2085                 }
2086         }
2087
2088         if(scsi_status_is_good(result)) {
2089                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
2090                              (modepage == 6 || modepage == 8))) {
2091                         /* Initio breakage? */
2092                         header_length = 0;
2093                         data->length = 13;
2094                         data->medium_type = 0;
2095                         data->device_specific = 0;
2096                         data->longlba = 0;
2097                         data->block_descriptor_length = 0;
2098                 } else if(use_10_for_ms) {
2099                         data->length = buffer[0]*256 + buffer[1] + 2;
2100                         data->medium_type = buffer[2];
2101                         data->device_specific = buffer[3];
2102                         data->longlba = buffer[4] & 0x01;
2103                         data->block_descriptor_length = buffer[6]*256
2104                                 + buffer[7];
2105                 } else {
2106                         data->length = buffer[0] + 1;
2107                         data->medium_type = buffer[1];
2108                         data->device_specific = buffer[2];
2109                         data->block_descriptor_length = buffer[3];
2110                 }
2111                 data->header_length = header_length;
2112         }
2113
2114         return result;
2115 }
2116 EXPORT_SYMBOL(scsi_mode_sense);
2117
2118 /**
2119  *      scsi_test_unit_ready - test if unit is ready
2120  *      @sdev:  scsi device to change the state of.
2121  *      @timeout: command timeout
2122  *      @retries: number of retries before failing
2123  *      @sshdr_external: Optional pointer to struct scsi_sense_hdr for
2124  *              returning sense. Make sure that this is cleared before passing
2125  *              in.
2126  *
2127  *      Returns zero if unsuccessful or an error if TUR failed.  For
2128  *      removable media, a return of NOT_READY or UNIT_ATTENTION is
2129  *      translated to success, with the ->changed flag updated.
2130  **/
2131 int
2132 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
2133                      struct scsi_sense_hdr *sshdr_external)
2134 {
2135         char cmd[] = {
2136                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
2137         };
2138         struct scsi_sense_hdr *sshdr;
2139         int result;
2140
2141         if (!sshdr_external)
2142                 sshdr = kzalloc(sizeof(*sshdr), GFP_KERNEL);
2143         else
2144                 sshdr = sshdr_external;
2145
2146         /* try to eat the UNIT_ATTENTION if there are enough retries */
2147         do {
2148                 result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
2149                                           timeout, retries);
2150                 if (sdev->removable && scsi_sense_valid(sshdr) &&
2151                     sshdr->sense_key == UNIT_ATTENTION)
2152                         sdev->changed = 1;
2153         } while (scsi_sense_valid(sshdr) &&
2154                  sshdr->sense_key == UNIT_ATTENTION && --retries);
2155
2156         if (!sshdr)
2157                 /* could not allocate sense buffer, so can't process it */
2158                 return result;
2159
2160         if (sdev->removable && scsi_sense_valid(sshdr) &&
2161             (sshdr->sense_key == UNIT_ATTENTION ||
2162              sshdr->sense_key == NOT_READY)) {
2163                 sdev->changed = 1;
2164                 result = 0;
2165         }
2166         if (!sshdr_external)
2167                 kfree(sshdr);
2168         return result;
2169 }
2170 EXPORT_SYMBOL(scsi_test_unit_ready);
2171
2172 /**
2173  *      scsi_device_set_state - Take the given device through the device state model.
2174  *      @sdev:  scsi device to change the state of.
2175  *      @state: state to change to.
2176  *
2177  *      Returns zero if unsuccessful or an error if the requested 
2178  *      transition is illegal.
2179  */
2180 int
2181 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2182 {
2183         enum scsi_device_state oldstate = sdev->sdev_state;
2184
2185         if (state == oldstate)
2186                 return 0;
2187
2188         switch (state) {
2189         case SDEV_CREATED:
2190                 switch (oldstate) {
2191                 case SDEV_CREATED_BLOCK:
2192                         break;
2193                 default:
2194                         goto illegal;
2195                 }
2196                 break;
2197                         
2198         case SDEV_RUNNING:
2199                 switch (oldstate) {
2200                 case SDEV_CREATED:
2201                 case SDEV_OFFLINE:
2202                 case SDEV_QUIESCE:
2203                 case SDEV_BLOCK:
2204                         break;
2205                 default:
2206                         goto illegal;
2207                 }
2208                 break;
2209
2210         case SDEV_QUIESCE:
2211                 switch (oldstate) {
2212                 case SDEV_RUNNING:
2213                 case SDEV_OFFLINE:
2214                         break;
2215                 default:
2216                         goto illegal;
2217                 }
2218                 break;
2219
2220         case SDEV_OFFLINE:
2221                 switch (oldstate) {
2222                 case SDEV_CREATED:
2223                 case SDEV_RUNNING:
2224                 case SDEV_QUIESCE:
2225                 case SDEV_BLOCK:
2226                         break;
2227                 default:
2228                         goto illegal;
2229                 }
2230                 break;
2231
2232         case SDEV_BLOCK:
2233                 switch (oldstate) {
2234                 case SDEV_RUNNING:
2235                 case SDEV_CREATED_BLOCK:
2236                         break;
2237                 default:
2238                         goto illegal;
2239                 }
2240                 break;
2241
2242         case SDEV_CREATED_BLOCK:
2243                 switch (oldstate) {
2244                 case SDEV_CREATED:
2245                         break;
2246                 default:
2247                         goto illegal;
2248                 }
2249                 break;
2250
2251         case SDEV_CANCEL:
2252                 switch (oldstate) {
2253                 case SDEV_CREATED:
2254                 case SDEV_RUNNING:
2255                 case SDEV_QUIESCE:
2256                 case SDEV_OFFLINE:
2257                 case SDEV_BLOCK:
2258                         break;
2259                 default:
2260                         goto illegal;
2261                 }
2262                 break;
2263
2264         case SDEV_DEL:
2265                 switch (oldstate) {
2266                 case SDEV_CREATED:
2267                 case SDEV_RUNNING:
2268                 case SDEV_OFFLINE:
2269                 case SDEV_CANCEL:
2270                         break;
2271                 default:
2272                         goto illegal;
2273                 }
2274                 break;
2275
2276         }
2277         sdev->sdev_state = state;
2278         return 0;
2279
2280  illegal:
2281         SCSI_LOG_ERROR_RECOVERY(1, 
2282                                 sdev_printk(KERN_ERR, sdev,
2283                                             "Illegal state transition %s->%s\n",
2284                                             scsi_device_state_name(oldstate),
2285                                             scsi_device_state_name(state))
2286                                 );
2287         return -EINVAL;
2288 }
2289 EXPORT_SYMBOL(scsi_device_set_state);
2290
2291 /**
2292  *      sdev_evt_emit - emit a single SCSI device uevent
2293  *      @sdev: associated SCSI device
2294  *      @evt: event to emit
2295  *
2296  *      Send a single uevent (scsi_event) to the associated scsi_device.
2297  */
2298 static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2299 {
2300         int idx = 0;
2301         char *envp[3];
2302
2303         switch (evt->evt_type) {
2304         case SDEV_EVT_MEDIA_CHANGE:
2305                 envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2306                 break;
2307
2308         default:
2309                 /* do nothing */
2310                 break;
2311         }
2312
2313         envp[idx++] = NULL;
2314
2315         kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2316 }
2317
2318 /**
2319  *      sdev_evt_thread - send a uevent for each scsi event
2320  *      @work: work struct for scsi_device
2321  *
2322  *      Dispatch queued events to their associated scsi_device kobjects
2323  *      as uevents.
2324  */
2325 void scsi_evt_thread(struct work_struct *work)
2326 {
2327         struct scsi_device *sdev;
2328         LIST_HEAD(event_list);
2329
2330         sdev = container_of(work, struct scsi_device, event_work);
2331
2332         while (1) {
2333                 struct scsi_event *evt;
2334                 struct list_head *this, *tmp;
2335                 unsigned long flags;
2336
2337                 spin_lock_irqsave(&sdev->list_lock, flags);
2338                 list_splice_init(&sdev->event_list, &event_list);
2339                 spin_unlock_irqrestore(&sdev->list_lock, flags);
2340
2341                 if (list_empty(&event_list))
2342                         break;
2343
2344                 list_for_each_safe(this, tmp, &event_list) {
2345                         evt = list_entry(this, struct scsi_event, node);
2346                         list_del(&evt->node);
2347                         scsi_evt_emit(sdev, evt);
2348                         kfree(evt);
2349                 }
2350         }
2351 }
2352
2353 /**
2354  *      sdev_evt_send - send asserted event to uevent thread
2355  *      @sdev: scsi_device event occurred on
2356  *      @evt: event to send
2357  *
2358  *      Assert scsi device event asynchronously.
2359  */
2360 void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2361 {
2362         unsigned long flags;
2363
2364 #if 0
2365         /* FIXME: currently this check eliminates all media change events
2366          * for polled devices.  Need to update to discriminate between AN
2367          * and polled events */
2368         if (!test_bit(evt->evt_type, sdev->supported_events)) {
2369                 kfree(evt);
2370                 return;
2371         }
2372 #endif
2373
2374         spin_lock_irqsave(&sdev->list_lock, flags);
2375         list_add_tail(&evt->node, &sdev->event_list);
2376         schedule_work(&sdev->event_work);
2377         spin_unlock_irqrestore(&sdev->list_lock, flags);
2378 }
2379 EXPORT_SYMBOL_GPL(sdev_evt_send);
2380
2381 /**
2382  *      sdev_evt_alloc - allocate a new scsi event
2383  *      @evt_type: type of event to allocate
2384  *      @gfpflags: GFP flags for allocation
2385  *
2386  *      Allocates and returns a new scsi_event.
2387  */
2388 struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2389                                   gfp_t gfpflags)
2390 {
2391         struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2392         if (!evt)
2393                 return NULL;
2394
2395         evt->evt_type = evt_type;
2396         INIT_LIST_HEAD(&evt->node);
2397
2398         /* evt_type-specific initialization, if any */
2399         switch (evt_type) {
2400         case SDEV_EVT_MEDIA_CHANGE:
2401         default:
2402                 /* do nothing */
2403                 break;
2404         }
2405
2406         return evt;
2407 }
2408 EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2409
2410 /**
2411  *      sdev_evt_send_simple - send asserted event to uevent thread
2412  *      @sdev: scsi_device event occurred on
2413  *      @evt_type: type of event to send
2414  *      @gfpflags: GFP flags for allocation
2415  *
2416  *      Assert scsi device event asynchronously, given an event type.
2417  */
2418 void sdev_evt_send_simple(struct scsi_device *sdev,
2419                           enum scsi_device_event evt_type, gfp_t gfpflags)
2420 {
2421         struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2422         if (!evt) {
2423                 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2424                             evt_type);
2425                 return;
2426         }
2427
2428         sdev_evt_send(sdev, evt);
2429 }
2430 EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2431
2432 /**
2433  *      scsi_device_quiesce - Block user issued commands.
2434  *      @sdev:  scsi device to quiesce.
2435  *
2436  *      This works by trying to transition to the SDEV_QUIESCE state
2437  *      (which must be a legal transition).  When the device is in this
2438  *      state, only special requests will be accepted, all others will
2439  *      be deferred.  Since special requests may also be requeued requests,
2440  *      a successful return doesn't guarantee the device will be 
2441  *      totally quiescent.
2442  *
2443  *      Must be called with user context, may sleep.
2444  *
2445  *      Returns zero if unsuccessful or an error if not.
2446  */
2447 int
2448 scsi_device_quiesce(struct scsi_device *sdev)
2449 {
2450         int err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2451         if (err)
2452                 return err;
2453
2454         scsi_run_queue(sdev->request_queue);
2455         while (sdev->device_busy) {
2456                 msleep_interruptible(200);
2457                 scsi_run_queue(sdev->request_queue);
2458         }
2459         return 0;
2460 }
2461 EXPORT_SYMBOL(scsi_device_quiesce);
2462
2463 /**
2464  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2465  *      @sdev:  scsi device to resume.
2466  *
2467  *      Moves the device from quiesced back to running and restarts the
2468  *      queues.
2469  *
2470  *      Must be called with user context, may sleep.
2471  */
2472 void
2473 scsi_device_resume(struct scsi_device *sdev)
2474 {
2475         if(scsi_device_set_state(sdev, SDEV_RUNNING))
2476                 return;
2477         scsi_run_queue(sdev->request_queue);
2478 }
2479 EXPORT_SYMBOL(scsi_device_resume);
2480
2481 static void
2482 device_quiesce_fn(struct scsi_device *sdev, void *data)
2483 {
2484         scsi_device_quiesce(sdev);
2485 }
2486
2487 void
2488 scsi_target_quiesce(struct scsi_target *starget)
2489 {
2490         starget_for_each_device(starget, NULL, device_quiesce_fn);
2491 }
2492 EXPORT_SYMBOL(scsi_target_quiesce);
2493
2494 static void
2495 device_resume_fn(struct scsi_device *sdev, void *data)
2496 {
2497         scsi_device_resume(sdev);
2498 }
2499
2500 void
2501 scsi_target_resume(struct scsi_target *starget)
2502 {
2503         starget_for_each_device(starget, NULL, device_resume_fn);
2504 }
2505 EXPORT_SYMBOL(scsi_target_resume);
2506
2507 /**
2508  * scsi_internal_device_block - internal function to put a device temporarily into the SDEV_BLOCK state
2509  * @sdev:       device to block
2510  *
2511  * Block request made by scsi lld's to temporarily stop all
2512  * scsi commands on the specified device.  Called from interrupt
2513  * or normal process context.
2514  *
2515  * Returns zero if successful or error if not
2516  *
2517  * Notes:       
2518  *      This routine transitions the device to the SDEV_BLOCK state
2519  *      (which must be a legal transition).  When the device is in this
2520  *      state, all commands are deferred until the scsi lld reenables
2521  *      the device with scsi_device_unblock or device_block_tmo fires.
2522  *      This routine assumes the host_lock is held on entry.
2523  */
2524 int
2525 scsi_internal_device_block(struct scsi_device *sdev)
2526 {
2527         struct request_queue *q = sdev->request_queue;
2528         unsigned long flags;
2529         int err = 0;
2530
2531         err = scsi_device_set_state(sdev, SDEV_BLOCK);
2532         if (err) {
2533                 err = scsi_device_set_state(sdev, SDEV_CREATED_BLOCK);
2534
2535                 if (err)
2536                         return err;
2537         }
2538
2539         /* 
2540          * The device has transitioned to SDEV_BLOCK.  Stop the
2541          * block layer from calling the midlayer with this device's
2542          * request queue. 
2543          */
2544         spin_lock_irqsave(q->queue_lock, flags);
2545         blk_stop_queue(q);
2546         spin_unlock_irqrestore(q->queue_lock, flags);
2547
2548         return 0;
2549 }
2550 EXPORT_SYMBOL_GPL(scsi_internal_device_block);
2551  
2552 /**
2553  * scsi_internal_device_unblock - resume a device after a block request
2554  * @sdev:       device to resume
2555  *
2556  * Called by scsi lld's or the midlayer to restart the device queue
2557  * for the previously suspended scsi device.  Called from interrupt or
2558  * normal process context.
2559  *
2560  * Returns zero if successful or error if not.
2561  *
2562  * Notes:       
2563  *      This routine transitions the device to the SDEV_RUNNING state
2564  *      (which must be a legal transition) allowing the midlayer to
2565  *      goose the queue for this device.  This routine assumes the 
2566  *      host_lock is held upon entry.
2567  */
2568 int
2569 scsi_internal_device_unblock(struct scsi_device *sdev)
2570 {
2571         struct request_queue *q = sdev->request_queue; 
2572         int err;
2573         unsigned long flags;
2574         
2575         /* 
2576          * Try to transition the scsi device to SDEV_RUNNING
2577          * and goose the device queue if successful.  
2578          */
2579         err = scsi_device_set_state(sdev, SDEV_RUNNING);
2580         if (err) {
2581                 err = scsi_device_set_state(sdev, SDEV_CREATED);
2582
2583                 if (err)
2584                         return err;
2585         }
2586
2587         spin_lock_irqsave(q->queue_lock, flags);
2588         blk_start_queue(q);
2589         spin_unlock_irqrestore(q->queue_lock, flags);
2590
2591         return 0;
2592 }
2593 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock);
2594
2595 static void
2596 device_block(struct scsi_device *sdev, void *data)
2597 {
2598         scsi_internal_device_block(sdev);
2599 }
2600
2601 static int
2602 target_block(struct device *dev, void *data)
2603 {
2604         if (scsi_is_target_device(dev))
2605                 starget_for_each_device(to_scsi_target(dev), NULL,
2606                                         device_block);
2607         return 0;
2608 }
2609
2610 void
2611 scsi_target_block(struct device *dev)
2612 {
2613         if (scsi_is_target_device(dev))
2614                 starget_for_each_device(to_scsi_target(dev), NULL,
2615                                         device_block);
2616         else
2617                 device_for_each_child(dev, NULL, target_block);
2618 }
2619 EXPORT_SYMBOL_GPL(scsi_target_block);
2620
2621 static void
2622 device_unblock(struct scsi_device *sdev, void *data)
2623 {
2624         scsi_internal_device_unblock(sdev);
2625 }
2626
2627 static int
2628 target_unblock(struct device *dev, void *data)
2629 {
2630         if (scsi_is_target_device(dev))
2631                 starget_for_each_device(to_scsi_target(dev), NULL,
2632                                         device_unblock);
2633         return 0;
2634 }
2635
2636 void
2637 scsi_target_unblock(struct device *dev)
2638 {
2639         if (scsi_is_target_device(dev))
2640                 starget_for_each_device(to_scsi_target(dev), NULL,
2641                                         device_unblock);
2642         else
2643                 device_for_each_child(dev, NULL, target_unblock);
2644 }
2645 EXPORT_SYMBOL_GPL(scsi_target_unblock);
2646
2647 /**
2648  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2649  * @sgl:        scatter-gather list
2650  * @sg_count:   number of segments in sg
2651  * @offset:     offset in bytes into sg, on return offset into the mapped area
2652  * @len:        bytes to map, on return number of bytes mapped
2653  *
2654  * Returns virtual address of the start of the mapped page
2655  */
2656 void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
2657                           size_t *offset, size_t *len)
2658 {
2659         int i;
2660         size_t sg_len = 0, len_complete = 0;
2661         struct scatterlist *sg;
2662         struct page *page;
2663
2664         WARN_ON(!irqs_disabled());
2665
2666         for_each_sg(sgl, sg, sg_count, i) {
2667                 len_complete = sg_len; /* Complete sg-entries */
2668                 sg_len += sg->length;
2669                 if (sg_len > *offset)
2670                         break;
2671         }
2672
2673         if (unlikely(i == sg_count)) {
2674                 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2675                         "elements %d\n",
2676                        __func__, sg_len, *offset, sg_count);
2677                 WARN_ON(1);
2678                 return NULL;
2679         }
2680
2681         /* Offset starting from the beginning of first page in this sg-entry */
2682         *offset = *offset - len_complete + sg->offset;
2683
2684         /* Assumption: contiguous pages can be accessed as "page + i" */
2685         page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
2686         *offset &= ~PAGE_MASK;
2687
2688         /* Bytes in this sg-entry from *offset to the end of the page */
2689         sg_len = PAGE_SIZE - *offset;
2690         if (*len > sg_len)
2691                 *len = sg_len;
2692
2693         return kmap_atomic(page, KM_BIO_SRC_IRQ);
2694 }
2695 EXPORT_SYMBOL(scsi_kmap_atomic_sg);
2696
2697 /**
2698  * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
2699  * @virt:       virtual address to be unmapped
2700  */
2701 void scsi_kunmap_atomic_sg(void *virt)
2702 {
2703         kunmap_atomic(virt, KM_BIO_SRC_IRQ);
2704 }
2705 EXPORT_SYMBOL(scsi_kunmap_atomic_sg);