USB: xhci: Fix TRB physical to virtual address translation.
[safe/jmp/linux-2.6] / drivers / scsi / scsi_lib.c
index 91c74c5..5987da8 100644 (file)
@@ -91,26 +91,19 @@ static void scsi_unprep_request(struct request *req)
        scsi_put_command(cmd);
 }
 
-/*
- * Function:    scsi_queue_insert()
- *
- * Purpose:     Insert a command in the midlevel queue.
- *
- * Arguments:   cmd    - command that we are adding to queue.
- *              reason - why we are inserting command to queue.
- *
- * Lock status: Assumed that lock is not held upon entry.
- *
- * Returns:     Nothing.
- *
- * Notes:       We do this for one of two cases.  Either the host is busy
- *              and it cannot accept any more commands for the time being,
- *              or the device returned QUEUE_FULL and can accept no more
- *              commands.
- * Notes:       This could be called either from an interrupt context or a
- *              normal process context.
+/**
+ * __scsi_queue_insert - private queue insertion
+ * @cmd: The SCSI command being requeued
+ * @reason:  The reason for the requeue
+ * @unbusy: Whether the queue should be unbusied
+ *
+ * This is a private queue insertion.  The public interface
+ * scsi_queue_insert() always assumes the queue should be unbusied
+ * because it's always called before the completion.  This function is
+ * for a requeue after completion, which should only occur in this
+ * file.
  */
-int scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
+static int __scsi_queue_insert(struct scsi_cmnd *cmd, int reason, int unbusy)
 {
        struct Scsi_Host *host = cmd->device->host;
        struct scsi_device *device = cmd->device;
@@ -150,7 +143,8 @@ int scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
         * Decrement the counters, since these commands are no longer
         * active on the host/device.
         */
-       scsi_device_unbusy(device);
+       if (unbusy)
+               scsi_device_unbusy(device);
 
        /*
         * Requeue this command.  It will go before all other commands
@@ -172,6 +166,29 @@ int scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
        return 0;
 }
 
+/*
+ * Function:    scsi_queue_insert()
+ *
+ * Purpose:     Insert a command in the midlevel queue.
+ *
+ * Arguments:   cmd    - command that we are adding to queue.
+ *              reason - why we are inserting command to queue.
+ *
+ * Lock status: Assumed that lock is not held upon entry.
+ *
+ * Returns:     Nothing.
+ *
+ * Notes:       We do this for one of two cases.  Either the host is busy
+ *              and it cannot accept any more commands for the time being,
+ *              or the device returned QUEUE_FULL and can accept no more
+ *              commands.
+ * Notes:       This could be called either from an interrupt context or a
+ *              normal process context.
+ */
+int scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
+{
+       return __scsi_queue_insert(cmd, reason, 1);
+}
 /**
  * scsi_execute - insert request and wait for the result
  * @sdev:      scsi device
@@ -183,13 +200,15 @@ int scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
  * @timeout:   request timeout in seconds
  * @retries:   number of times to retry request
  * @flags:     or into request flags;
+ * @resid:     optional residual length
  *
  * returns the req->errors value which is the scsi_cmnd result
  * field.
  */
 int scsi_execute(struct scsi_device *sdev, const unsigned char *cmd,
                 int data_direction, void *buffer, unsigned bufflen,
-                unsigned char *sense, int timeout, int retries, int flags)
+                unsigned char *sense, int timeout, int retries, int flags,
+                int *resid)
 {
        struct request *req;
        int write = (data_direction == DMA_TO_DEVICE);
@@ -221,9 +240,11 @@ int scsi_execute(struct scsi_device *sdev, const unsigned char *cmd,
         * is invalid.  Prevent the garbage from being misinterpreted
         * and prevent security leaks by zeroing out the excess data.
         */
-       if (unlikely(req->data_len > 0 && req->data_len <= bufflen))
-               memset(buffer + (bufflen - req->data_len), 0, req->data_len);
+       if (unlikely(req->resid_len > 0 && req->resid_len <= bufflen))
+               memset(buffer + (bufflen - req->resid_len), 0, req->resid_len);
 
+       if (resid)
+               *resid = req->resid_len;
        ret = req->errors;
  out:
        blk_put_request(req);
@@ -235,7 +256,8 @@ EXPORT_SYMBOL(scsi_execute);
 
 int scsi_execute_req(struct scsi_device *sdev, const unsigned char *cmd,
                     int data_direction, void *buffer, unsigned bufflen,
-                    struct scsi_sense_hdr *sshdr, int timeout, int retries)
+                    struct scsi_sense_hdr *sshdr, int timeout, int retries,
+                    int *resid)
 {
        char *sense = NULL;
        int result;
@@ -246,7 +268,7 @@ int scsi_execute_req(struct scsi_device *sdev, const unsigned char *cmd,
                        return DRIVER_ERROR << 24;
        }
        result = scsi_execute(sdev, cmd, data_direction, buffer, bufflen,
-                             sense, timeout, retries, 0);
+                             sense, timeout, retries, 0, resid);
        if (sshdr)
                scsi_normalize_sense(sense, SCSI_SENSE_BUFFERSIZE, sshdr);
 
@@ -255,196 +277,6 @@ int scsi_execute_req(struct scsi_device *sdev, const unsigned char *cmd,
 }
 EXPORT_SYMBOL(scsi_execute_req);
 
-struct scsi_io_context {
-       void *data;
-       void (*done)(void *data, char *sense, int result, int resid);
-       char sense[SCSI_SENSE_BUFFERSIZE];
-};
-
-static struct kmem_cache *scsi_io_context_cache;
-
-static void scsi_end_async(struct request *req, int uptodate)
-{
-       struct scsi_io_context *sioc = req->end_io_data;
-
-       if (sioc->done)
-               sioc->done(sioc->data, sioc->sense, req->errors, req->data_len);
-
-       kmem_cache_free(scsi_io_context_cache, sioc);
-       __blk_put_request(req->q, req);
-}
-
-static int scsi_merge_bio(struct request *rq, struct bio *bio)
-{
-       struct request_queue *q = rq->q;
-
-       bio->bi_flags &= ~(1 << BIO_SEG_VALID);
-       if (rq_data_dir(rq) == WRITE)
-               bio->bi_rw |= (1 << BIO_RW);
-       blk_queue_bounce(q, &bio);
-
-       return blk_rq_append_bio(q, rq, bio);
-}
-
-static void scsi_bi_endio(struct bio *bio, int error)
-{
-       bio_put(bio);
-}
-
-/**
- * scsi_req_map_sg - map a scatterlist into a request
- * @rq:                request to fill
- * @sgl:       scatterlist
- * @nsegs:     number of elements
- * @bufflen:   len of buffer
- * @gfp:       memory allocation flags
- *
- * scsi_req_map_sg maps a scatterlist into a request so that the
- * request can be sent to the block layer. We do not trust the scatterlist
- * sent to use, as some ULDs use that struct to only organize the pages.
- */
-static int scsi_req_map_sg(struct request *rq, struct scatterlist *sgl,
-                          int nsegs, unsigned bufflen, gfp_t gfp)
-{
-       struct request_queue *q = rq->q;
-       int nr_pages = (bufflen + sgl[0].offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
-       unsigned int data_len = bufflen, len, bytes, off;
-       struct scatterlist *sg;
-       struct page *page;
-       struct bio *bio = NULL;
-       int i, err, nr_vecs = 0;
-
-       for_each_sg(sgl, sg, nsegs, i) {
-               page = sg_page(sg);
-               off = sg->offset;
-               len = sg->length;
-
-               while (len > 0 && data_len > 0) {
-                       /*
-                        * sg sends a scatterlist that is larger than
-                        * the data_len it wants transferred for certain
-                        * IO sizes
-                        */
-                       bytes = min_t(unsigned int, len, PAGE_SIZE - off);
-                       bytes = min(bytes, data_len);
-
-                       if (!bio) {
-                               nr_vecs = min_t(int, BIO_MAX_PAGES, nr_pages);
-                               nr_pages -= nr_vecs;
-
-                               bio = bio_alloc(gfp, nr_vecs);
-                               if (!bio) {
-                                       err = -ENOMEM;
-                                       goto free_bios;
-                               }
-                               bio->bi_end_io = scsi_bi_endio;
-                       }
-
-                       if (bio_add_pc_page(q, bio, page, bytes, off) !=
-                           bytes) {
-                               bio_put(bio);
-                               err = -EINVAL;
-                               goto free_bios;
-                       }
-
-                       if (bio->bi_vcnt >= nr_vecs) {
-                               err = scsi_merge_bio(rq, bio);
-                               if (err) {
-                                       bio_endio(bio, 0);
-                                       goto free_bios;
-                               }
-                               bio = NULL;
-                       }
-
-                       page++;
-                       len -= bytes;
-                       data_len -=bytes;
-                       off = 0;
-               }
-       }
-
-       rq->buffer = rq->data = NULL;
-       rq->data_len = bufflen;
-       return 0;
-
-free_bios:
-       while ((bio = rq->bio) != NULL) {
-               rq->bio = bio->bi_next;
-               /*
-                * call endio instead of bio_put incase it was bounced
-                */
-               bio_endio(bio, 0);
-       }
-
-       return err;
-}
-
-/**
- * scsi_execute_async - insert request
- * @sdev:      scsi device
- * @cmd:       scsi command
- * @cmd_len:   length of scsi cdb
- * @data_direction: DMA_TO_DEVICE, DMA_FROM_DEVICE, or DMA_NONE
- * @buffer:    data buffer (this can be a kernel buffer or scatterlist)
- * @bufflen:   len of buffer
- * @use_sg:    if buffer is a scatterlist this is the number of elements
- * @timeout:   request timeout in seconds
- * @retries:   number of times to retry request
- * @privdata:  data passed to done()
- * @done:      callback function when done
- * @gfp:       memory allocation flags
- */
-int scsi_execute_async(struct scsi_device *sdev, const unsigned char *cmd,
-                      int cmd_len, int data_direction, void *buffer, unsigned bufflen,
-                      int use_sg, int timeout, int retries, void *privdata,
-                      void (*done)(void *, char *, int, int), gfp_t gfp)
-{
-       struct request *req;
-       struct scsi_io_context *sioc;
-       int err = 0;
-       int write = (data_direction == DMA_TO_DEVICE);
-
-       sioc = kmem_cache_zalloc(scsi_io_context_cache, gfp);
-       if (!sioc)
-               return DRIVER_ERROR << 24;
-
-       req = blk_get_request(sdev->request_queue, write, gfp);
-       if (!req)
-               goto free_sense;
-       req->cmd_type = REQ_TYPE_BLOCK_PC;
-       req->cmd_flags |= REQ_QUIET;
-
-       if (use_sg)
-               err = scsi_req_map_sg(req, buffer, use_sg, bufflen, gfp);
-       else if (bufflen)
-               err = blk_rq_map_kern(req->q, req, buffer, bufflen, gfp);
-
-       if (err)
-               goto free_req;
-
-       req->cmd_len = cmd_len;
-       memset(req->cmd, 0, BLK_MAX_CDB); /* ATAPI hates garbage after CDB */
-       memcpy(req->cmd, cmd, req->cmd_len);
-       req->sense = sioc->sense;
-       req->sense_len = 0;
-       req->timeout = timeout;
-       req->retries = retries;
-       req->end_io_data = sioc;
-
-       sioc->data = privdata;
-       sioc->done = done;
-
-       blk_execute_rq_nowait(req->q, NULL, req, 1, scsi_end_async);
-       return 0;
-
-free_req:
-       blk_put_request(req);
-free_sense:
-       kmem_cache_free(scsi_io_context_cache, sioc);
-       return DRIVER_ERROR << 24;
-}
-EXPORT_SYMBOL_GPL(scsi_execute_async);
-
 /*
  * Function:    scsi_init_cmd_errh()
  *
@@ -529,6 +361,14 @@ static void scsi_single_lun_run(struct scsi_device *current_sdev)
        spin_unlock_irqrestore(shost->host_lock, flags);
 }
 
+static inline int scsi_device_is_busy(struct scsi_device *sdev)
+{
+       if (sdev->device_busy >= sdev->queue_depth || sdev->device_blocked)
+               return 1;
+
+       return 0;
+}
+
 static inline int scsi_target_is_busy(struct scsi_target *starget)
 {
        return ((starget->can_queue > 0 &&
@@ -536,6 +376,15 @@ static inline int scsi_target_is_busy(struct scsi_target *starget)
                 starget->target_blocked);
 }
 
+static inline int scsi_host_is_busy(struct Scsi_Host *shost)
+{
+       if ((shost->can_queue > 0 && shost->host_busy >= shost->can_queue) ||
+           shost->host_blocked || shost->host_self_blocked)
+               return 1;
+
+       return 0;
+}
+
 /*
  * Function:   scsi_run_queue()
  *
@@ -550,19 +399,18 @@ static inline int scsi_target_is_busy(struct scsi_target *starget)
  */
 static void scsi_run_queue(struct request_queue *q)
 {
-       struct scsi_device *starved_head = NULL, *sdev = q->queuedata;
+       struct scsi_device *sdev = q->queuedata;
        struct Scsi_Host *shost = sdev->host;
+       LIST_HEAD(starved_list);
        unsigned long flags;
 
        if (scsi_target(sdev)->single_lun)
                scsi_single_lun_run(sdev);
 
        spin_lock_irqsave(shost->host_lock, flags);
-       while (!list_empty(&shost->starved_list) &&
-              !shost->host_blocked && !shost->host_self_blocked &&
-               !((shost->can_queue > 0) &&
-                 (shost->host_busy >= shost->can_queue))) {
+       list_splice_init(&shost->starved_list, &starved_list);
 
+       while (!list_empty(&starved_list)) {
                int flagset;
 
                /*
@@ -575,24 +423,18 @@ static void scsi_run_queue(struct request_queue *q)
                 * scsi_request_fn must get the host_lock before checking
                 * or modifying starved_list or starved_entry.
                 */
-               sdev = list_entry(shost->starved_list.next,
-                                         struct scsi_device, starved_entry);
-               /*
-                * The *queue_ready functions can add a device back onto the
-                * starved list's tail, so we must check for a infinite loop.
-                */
-               if (sdev == starved_head)
+               if (scsi_host_is_busy(shost))
                        break;
-               if (!starved_head)
-                       starved_head = sdev;
 
+               sdev = list_entry(starved_list.next,
+                                 struct scsi_device, starved_entry);
+               list_del_init(&sdev->starved_entry);
                if (scsi_target_is_busy(scsi_target(sdev))) {
                        list_move_tail(&sdev->starved_entry,
                                       &shost->starved_list);
                        continue;
                }
 
-               list_del_init(&sdev->starved_entry);
                spin_unlock(shost->host_lock);
 
                spin_lock(sdev->request_queue->queue_lock);
@@ -608,6 +450,8 @@ static void scsi_run_queue(struct request_queue *q)
 
                spin_lock(shost->host_lock);
        }
+       /* put any unprocessed entries back */
+       list_splice(&starved_list, &shost->starved_list);
        spin_unlock_irqrestore(shost->host_lock, flags);
 
        blk_run_queue(q);
@@ -636,8 +480,8 @@ static void scsi_requeue_command(struct request_queue *q, struct scsi_cmnd *cmd)
        struct request *req = cmd->request;
        unsigned long flags;
 
-       scsi_unprep_request(req);
        spin_lock_irqsave(q->queue_lock, flags);
+       scsi_unprep_request(req);
        blk_requeue_request(q, req);
        spin_unlock_irqrestore(q->queue_lock, flags);
 
@@ -667,6 +511,8 @@ void scsi_run_host_queues(struct Scsi_Host *shost)
                scsi_run_queue(sdev->request_queue);
 }
 
+static void __scsi_release_buffers(struct scsi_cmnd *, int);
+
 /*
  * Function:    scsi_end_request()
  *
@@ -700,14 +546,9 @@ static struct scsi_cmnd *scsi_end_request(struct scsi_cmnd *cmd, int error,
         * to queue the remainder of them.
         */
        if (blk_end_request(req, error, bytes)) {
-               int leftover = (req->hard_nr_sectors << 9);
-
-               if (blk_pc_request(req))
-                       leftover = req->data_len;
-
                /* kill remainder if no retrys */
-               if (error && blk_noretry_request(req))
-                       blk_end_request(req, error, leftover);
+               if (error && scsi_noretry_cmd(cmd))
+                       blk_end_request_all(req, error);
                else {
                        if (requeue) {
                                /*
@@ -715,6 +556,7 @@ static struct scsi_cmnd *scsi_end_request(struct scsi_cmnd *cmd, int error,
                                 * leftovers in the front of the
                                 * queue, and goose the queue again.
                                 */
+                               scsi_release_buffers(cmd);
                                scsi_requeue_command(q, cmd);
                                cmd = NULL;
                        }
@@ -726,6 +568,7 @@ static struct scsi_cmnd *scsi_end_request(struct scsi_cmnd *cmd, int error,
         * This will goose the queue request function at the end, so we don't
         * need to worry about launching another command.
         */
+       __scsi_release_buffers(cmd, 0);
        scsi_next_command(cmd);
        return NULL;
 }
@@ -781,31 +624,15 @@ static void scsi_free_sgtable(struct scsi_data_buffer *sdb)
        __sg_free_table(&sdb->table, SCSI_MAX_SG_SEGMENTS, scsi_sg_free);
 }
 
-/*
- * Function:    scsi_release_buffers()
- *
- * Purpose:     Completion processing for block device I/O requests.
- *
- * Arguments:   cmd    - command that we are bailing.
- *
- * Lock status: Assumed that no lock is held upon entry.
- *
- * Returns:     Nothing
- *
- * Notes:       In the event that an upper level driver rejects a
- *             command, we must release resources allocated during
- *             the __init_io() function.  Primarily this would involve
- *             the scatter-gather table, and potentially any bounce
- *             buffers.
- */
-void scsi_release_buffers(struct scsi_cmnd *cmd)
+static void __scsi_release_buffers(struct scsi_cmnd *cmd, int do_bidi_check)
 {
+
        if (cmd->sdb.table.nents)
                scsi_free_sgtable(&cmd->sdb);
 
        memset(&cmd->sdb, 0, sizeof(cmd->sdb));
 
-       if (scsi_bidi_cmnd(cmd)) {
+       if (do_bidi_check && scsi_bidi_cmnd(cmd)) {
                struct scsi_data_buffer *bidi_sdb =
                        cmd->request->next_rq->special;
                scsi_free_sgtable(bidi_sdb);
@@ -816,35 +643,29 @@ void scsi_release_buffers(struct scsi_cmnd *cmd)
        if (scsi_prot_sg_count(cmd))
                scsi_free_sgtable(cmd->prot_sdb);
 }
-EXPORT_SYMBOL(scsi_release_buffers);
 
 /*
- * Bidi commands Must be complete as a whole, both sides at once.
- * If part of the bytes were written and lld returned
- * scsi_in()->resid and/or scsi_out()->resid this information will be left
- * in req->data_len and req->next_rq->data_len. The upper-layer driver can
- * decide what to do with this information.
+ * Function:    scsi_release_buffers()
+ *
+ * Purpose:     Completion processing for block device I/O requests.
+ *
+ * Arguments:   cmd    - command that we are bailing.
+ *
+ * Lock status: Assumed that no lock is held upon entry.
+ *
+ * Returns:     Nothing
+ *
+ * Notes:       In the event that an upper level driver rejects a
+ *             command, we must release resources allocated during
+ *             the __init_io() function.  Primarily this would involve
+ *             the scatter-gather table, and potentially any bounce
+ *             buffers.
  */
-static void scsi_end_bidi_request(struct scsi_cmnd *cmd)
+void scsi_release_buffers(struct scsi_cmnd *cmd)
 {
-       struct request *req = cmd->request;
-       unsigned int dlen = req->data_len;
-       unsigned int next_dlen = req->next_rq->data_len;
-
-       req->data_len = scsi_out(cmd)->resid;
-       req->next_rq->data_len = scsi_in(cmd)->resid;
-
-       /* The req and req->next_rq have not been completed */
-       BUG_ON(blk_end_bidi_request(req, 0, dlen, next_dlen));
-
-       scsi_release_buffers(cmd);
-
-       /*
-        * This will goose the queue request function at the end, so we don't
-        * need to worry about launching another command.
-        */
-       scsi_next_command(cmd);
+       __scsi_release_buffers(cmd, 1);
 }
+EXPORT_SYMBOL(scsi_release_buffers);
 
 /*
  * Function:    scsi_io_completion()
@@ -863,27 +684,37 @@ static void scsi_end_bidi_request(struct scsi_cmnd *cmd)
  *              (the normal case for most drivers), we don't need
  *              the logic to deal with cleaning up afterwards.
  *
- *             We must do one of several things here:
+ *             We must call scsi_end_request().  This will finish off
+ *             the specified number of sectors.  If we are done, the
+ *             command block will be released and the queue function
+ *             will be goosed.  If we are not done then we have to
+ *             figure out what to do next:
  *
- *             a) Call scsi_end_request.  This will finish off the
- *                specified number of sectors.  If we are done, the
- *                command block will be released, and the queue
- *                function will be goosed.  If we are not done, then
- *                scsi_end_request will directly goose the queue.
+ *             a) We can call scsi_requeue_command().  The request
+ *                will be unprepared and put back on the queue.  Then
+ *                a new command will be created for it.  This should
+ *                be used if we made forward progress, or if we want
+ *                to switch from READ(10) to READ(6) for example.
  *
- *             b) We can just use scsi_requeue_command() here.  This would
- *                be used if we just wanted to retry, for example.
+ *             b) We can call scsi_queue_insert().  The request will
+ *                be put back on the queue and retried using the same
+ *                command as before, possibly after a delay.
+ *
+ *             c) We can call blk_end_request() with -EIO to fail
+ *                the remainder of the request.
  */
 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
 {
        int result = cmd->result;
-       int this_count;
        struct request_queue *q = cmd->device->request_queue;
        struct request *req = cmd->request;
        int error = 0;
        struct scsi_sense_hdr sshdr;
        int sense_valid = 0;
        int sense_deferred = 0;
+       enum {ACTION_FAIL, ACTION_REPREP, ACTION_RETRY,
+             ACTION_DELAYED_RETRY} action;
+       char *description = NULL;
 
        if (result) {
                sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
@@ -908,37 +739,65 @@ void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
                        if (!sense_deferred)
                                error = -EIO;
                }
+
+               req->resid_len = scsi_get_resid(cmd);
+
                if (scsi_bidi_cmnd(cmd)) {
-                       /* will also release_buffers */
-                       scsi_end_bidi_request(cmd);
+                       /*
+                        * Bidi commands Must be complete as a whole,
+                        * both sides at once.
+                        */
+                       req->next_rq->resid_len = scsi_in(cmd)->resid;
+
+                       blk_end_request_all(req, 0);
+
+                       scsi_release_buffers(cmd);
+                       scsi_next_command(cmd);
                        return;
                }
-               req->data_len = scsi_get_resid(cmd);
        }
 
        BUG_ON(blk_bidi_rq(req)); /* bidi not support for !blk_pc_request yet */
-       scsi_release_buffers(cmd);
 
        /*
         * Next deal with any sectors which we were able to correctly
         * handle.
         */
-       SCSI_LOG_HLCOMPLETE(1, printk("%ld sectors total, "
+       SCSI_LOG_HLCOMPLETE(1, printk("%u sectors total, "
                                      "%d bytes done.\n",
-                                     req->nr_sectors, good_bytes));
+                                     blk_rq_sectors(req), good_bytes));
 
-       /* A number of bytes were successfully read.  If there
+       /*
+        * Recovered errors need reporting, but they're always treated
+        * as success, so fiddle the result code here.  For BLOCK_PC
+        * we already took a copy of the original into rq->errors which
+        * is what gets returned to the user
+        */
+       if (sense_valid && sshdr.sense_key == RECOVERED_ERROR) {
+               if (!(req->cmd_flags & REQ_QUIET))
+                       scsi_print_sense("", cmd);
+               result = 0;
+               /* BLOCK_PC may have set error */
+               error = 0;
+       }
+
+       /*
+        * A number of bytes were successfully read.  If there
         * are leftovers and there is some kind of error
         * (result != 0), retry the rest.
         */
        if (scsi_end_request(cmd, error, good_bytes, result == 0) == NULL)
                return;
-       this_count = blk_rq_bytes(req);
 
-       /* good_bytes = 0, or (inclusive) there were leftovers and
-        * result = 0, so scsi_end_request couldn't retry.
-        */
-       if (sense_valid && !sense_deferred) {
+       error = -EIO;
+
+       if (host_byte(result) == DID_RESET) {
+               /* Third party bus reset or reset for error recovery
+                * reasons.  Just retry the command and see what
+                * happens.
+                */
+               action = ACTION_RETRY;
+       } else if (sense_valid && !sense_deferred) {
                switch (sshdr.sense_key) {
                case UNIT_ATTENTION:
                        if (cmd->device->removable) {
@@ -946,16 +805,15 @@ void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
                                 * and quietly refuse further access.
                                 */
                                cmd->device->changed = 1;
-                               scsi_end_request(cmd, -EIO, this_count, 1);
-                               return;
+                               description = "Media Changed";
+                               action = ACTION_FAIL;
                        } else {
                                /* Must have been a power glitch, or a
                                 * bus reset.  Could not have been a
                                 * media change, so we just retry the
-                                * request and see what happens.
+                                * command and see what happens.
                                 */
-                               scsi_requeue_command(q, cmd);
-                               return;
+                               action = ACTION_RETRY;
                        }
                        break;
                case ILLEGAL_REQUEST:
@@ -971,20 +829,21 @@ void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
                            sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
                            (cmd->cmnd[0] == READ_10 ||
                             cmd->cmnd[0] == WRITE_10)) {
+                               /* This will issue a new 6-byte command. */
                                cmd->device->use_10_for_rw = 0;
-                               /* This will cause a retry with a
-                                * 6-byte command.
-                                */
-                               scsi_requeue_command(q, cmd);
-                       } else if (sshdr.asc == 0x10) /* DIX */
-                               scsi_end_request(cmd, -EIO, this_count, 0);
-                       else
-                               scsi_end_request(cmd, -EIO, this_count, 1);
-                       return;
+                               action = ACTION_REPREP;
+                       } else if (sshdr.asc == 0x10) /* DIX */ {
+                               description = "Host Data Integrity Failure";
+                               action = ACTION_FAIL;
+                               error = -EILSEQ;
+                       } else
+                               action = ACTION_FAIL;
+                       break;
                case ABORTED_COMMAND:
+                       action = ACTION_FAIL;
                        if (sshdr.asc == 0x10) { /* DIF */
-                               scsi_end_request(cmd, -EIO, this_count, 0);
-                               return;
+                               description = "Target Data Integrity Failure";
+                               error = -EILSEQ;
                        }
                        break;
                case NOT_READY:
@@ -1000,49 +859,66 @@ void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
                                case 0x07: /* operation in progress */
                                case 0x08: /* Long write in progress */
                                case 0x09: /* self test in progress */
-                                       scsi_requeue_command(q, cmd);
-                                       return;
+                                       action = ACTION_DELAYED_RETRY;
+                                       break;
                                default:
+                                       description = "Device not ready";
+                                       action = ACTION_FAIL;
                                        break;
                                }
+                       } else {
+                               description = "Device not ready";
+                               action = ACTION_FAIL;
                        }
-                       if (!(req->cmd_flags & REQ_QUIET))
-                               scsi_cmd_print_sense_hdr(cmd,
-                                                        "Device not ready",
-                                                        &sshdr);
-
-                       scsi_end_request(cmd, -EIO, this_count, 1);
-                       return;
+                       break;
                case VOLUME_OVERFLOW:
-                       if (!(req->cmd_flags & REQ_QUIET)) {
-                               scmd_printk(KERN_INFO, cmd,
-                                           "Volume overflow, CDB: ");
-                               __scsi_print_command(cmd->cmnd);
-                               scsi_print_sense("", cmd);
-                       }
                        /* See SSC3rXX or current. */
-                       scsi_end_request(cmd, -EIO, this_count, 1);
-                       return;
+                       action = ACTION_FAIL;
+                       break;
                default:
+                       description = "Unhandled sense code";
+                       action = ACTION_FAIL;
                        break;
                }
+       } else {
+               description = "Unhandled error code";
+               action = ACTION_FAIL;
        }
-       if (host_byte(result) == DID_RESET) {
-               /* Third party bus reset or reset for error recovery
-                * reasons.  Just retry the request and see what
-                * happens.
-                */
-               scsi_requeue_command(q, cmd);
-               return;
-       }
-       if (result) {
+
+       switch (action) {
+       case ACTION_FAIL:
+               /* Give up and fail the remainder of the request */
+               scsi_release_buffers(cmd);
                if (!(req->cmd_flags & REQ_QUIET)) {
+                       if (description)
+                               scmd_printk(KERN_INFO, cmd, "%s\n",
+                                           description);
                        scsi_print_result(cmd);
                        if (driver_byte(result) & DRIVER_SENSE)
                                scsi_print_sense("", cmd);
+                       scsi_print_command(cmd);
                }
+               if (blk_end_request_err(req, -EIO))
+                       scsi_requeue_command(q, cmd);
+               else
+                       scsi_next_command(cmd);
+               break;
+       case ACTION_REPREP:
+               /* Unprep the request and put it back at the head of the queue.
+                * A new command will be prepared and issued.
+                */
+               scsi_release_buffers(cmd);
+               scsi_requeue_command(q, cmd);
+               break;
+       case ACTION_RETRY:
+               /* Retry the same command immediately */
+               __scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY, 0);
+               break;
+       case ACTION_DELAYED_RETRY:
+               /* Retry the same command after a delay */
+               __scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY, 0);
+               break;
        }
-       scsi_end_request(cmd, -EIO, this_count, !result);
 }
 
 static int scsi_init_sgtable(struct request *req, struct scsi_data_buffer *sdb,
@@ -1067,10 +943,7 @@ static int scsi_init_sgtable(struct request *req, struct scsi_data_buffer *sdb,
        count = blk_rq_map_sg(req->q, req, sdb->table.sgl);
        BUG_ON(count > sdb->table.nents);
        sdb->table.nents = count;
-       if (blk_pc_request(req))
-               sdb->length = req->data_len;
-       else
-               sdb->length = req->nr_sectors << 9;
+       sdb->length = blk_rq_bytes(req);
        return BLKPREP_OK;
 }
 
@@ -1189,22 +1062,21 @@ int scsi_setup_blk_pc_cmnd(struct scsi_device *sdev, struct request *req)
                if (unlikely(ret))
                        return ret;
        } else {
-               BUG_ON(req->data_len);
-               BUG_ON(req->data);
+               BUG_ON(blk_rq_bytes(req));
 
                memset(&cmd->sdb, 0, sizeof(cmd->sdb));
                req->buffer = NULL;
        }
 
        cmd->cmd_len = req->cmd_len;
-       if (!req->data_len)
+       if (!blk_rq_bytes(req))
                cmd->sc_data_direction = DMA_NONE;
        else if (rq_data_dir(req) == WRITE)
                cmd->sc_data_direction = DMA_TO_DEVICE;
        else
                cmd->sc_data_direction = DMA_FROM_DEVICE;
        
-       cmd->transfersize = req->data_len;
+       cmd->transfersize = blk_rq_bytes(req);
        cmd->allowed = req->retries;
        return BLKPREP_OK;
 }
@@ -1314,7 +1186,7 @@ int scsi_prep_return(struct request_queue *q, struct request *req, int ret)
                break;
        case BLKPREP_DEFER:
                /*
-                * If we defer, the elv_next_request() returns NULL, but the
+                * If we defer, the blk_peek_request() returns NULL, but the
                 * queue must be restarted, so we plug here if no returning
                 * command will automatically do that.
                 */
@@ -1338,6 +1210,7 @@ int scsi_prep_fn(struct request_queue *q, struct request *req)
                ret = scsi_setup_blk_pc_cmnd(sdev, req);
        return scsi_prep_return(q, req, ret);
 }
+EXPORT_SYMBOL(scsi_prep_fn);
 
 /*
  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
@@ -1348,8 +1221,6 @@ int scsi_prep_fn(struct request_queue *q, struct request *req)
 static inline int scsi_dev_queue_ready(struct request_queue *q,
                                  struct scsi_device *sdev)
 {
-       if (sdev->device_busy >= sdev->queue_depth)
-               return 0;
        if (sdev->device_busy == 0 && sdev->device_blocked) {
                /*
                 * unblock after device_blocked iterates to zero
@@ -1363,7 +1234,7 @@ static inline int scsi_dev_queue_ready(struct request_queue *q,
                        return 0;
                }
        }
-       if (sdev->device_blocked)
+       if (scsi_device_is_busy(sdev))
                return 0;
 
        return 1;
@@ -1395,10 +1266,8 @@ static inline int scsi_target_queue_ready(struct Scsi_Host *shost,
                if (--starget->target_blocked == 0) {
                        SCSI_LOG_MLQUEUE(3, starget_printk(KERN_INFO, starget,
                                         "unblocking target at zero depth\n"));
-               } else {
-                       blk_plug_device(sdev->request_queue);
+               } else
                        return 0;
-               }
        }
 
        if (scsi_target_is_busy(starget)) {
@@ -1440,8 +1309,7 @@ static inline int scsi_host_queue_ready(struct request_queue *q,
                        return 0;
                }
        }
-       if ((shost->can_queue > 0 && shost->host_busy >= shost->can_queue) ||
-           shost->host_blocked || shost->host_self_blocked) {
+       if (scsi_host_is_busy(shost)) {
                if (list_empty(&sdev->starved_entry))
                        list_add_tail(&sdev->starved_entry, &shost->starved_list);
                return 0;
@@ -1455,6 +1323,37 @@ static inline int scsi_host_queue_ready(struct request_queue *q,
 }
 
 /*
+ * Busy state exporting function for request stacking drivers.
+ *
+ * For efficiency, no lock is taken to check the busy state of
+ * shost/starget/sdev, since the returned value is not guaranteed and
+ * may be changed after request stacking drivers call the function,
+ * regardless of taking lock or not.
+ *
+ * When scsi can't dispatch I/Os anymore and needs to kill I/Os
+ * (e.g. !sdev), scsi needs to return 'not busy'.
+ * Otherwise, request stacking drivers may hold requests forever.
+ */
+static int scsi_lld_busy(struct request_queue *q)
+{
+       struct scsi_device *sdev = q->queuedata;
+       struct Scsi_Host *shost;
+       struct scsi_target *starget;
+
+       if (!sdev)
+               return 0;
+
+       shost = sdev->host;
+       starget = scsi_target(sdev);
+
+       if (scsi_host_in_recovery(shost) || scsi_host_is_busy(shost) ||
+           scsi_target_is_busy(starget) || scsi_device_is_busy(sdev))
+               return 1;
+
+       return 0;
+}
+
+/*
  * Kill a request for a dead device
  */
 static void scsi_kill_request(struct request *req, struct request_queue *q)
@@ -1464,7 +1363,7 @@ static void scsi_kill_request(struct request *req, struct request_queue *q)
        struct scsi_target *starget = scsi_target(sdev);
        struct Scsi_Host *shost = sdev->host;
 
-       blkdev_dequeue_request(req);
+       blk_start_request(req);
 
        if (unlikely(cmd == NULL)) {
                printk(KERN_CRIT "impossible request in %s.\n",
@@ -1556,7 +1455,7 @@ static void scsi_request_fn(struct request_queue *q)
 
        if (!sdev) {
                printk("scsi: killing requests for dead queue\n");
-               while ((req = elv_next_request(q)) != NULL)
+               while ((req = blk_peek_request(q)) != NULL)
                        scsi_kill_request(req, q);
                return;
        }
@@ -1577,7 +1476,7 @@ static void scsi_request_fn(struct request_queue *q)
                 * that the request is fully prepared even if we cannot 
                 * accept it.
                 */
-               req = elv_next_request(q);
+               req = blk_peek_request(q);
                if (!req || !scsi_dev_queue_ready(q, sdev))
                        break;
 
@@ -1593,7 +1492,7 @@ static void scsi_request_fn(struct request_queue *q)
                 * Remove the request from the request list.
                 */
                if (!(blk_queue_tagged(q) && !blk_queue_start_tag(q, req)))
-                       blkdev_dequeue_request(req);
+                       blk_start_request(req);
                sdev->device_busy++;
 
                spin_unlock(q->queue_lock);
@@ -1757,6 +1656,7 @@ struct request_queue *scsi_alloc_queue(struct scsi_device *sdev)
        blk_queue_prep_rq(q, scsi_prep_fn);
        blk_queue_softirq_done(q, scsi_softirq_done);
        blk_queue_rq_timed_out(q, scsi_times_out);
+       blk_queue_lld_busy(q, scsi_lld_busy);
        return q;
 }
 
@@ -1818,20 +1718,12 @@ int __init scsi_init_queue(void)
 {
        int i;
 
-       scsi_io_context_cache = kmem_cache_create("scsi_io_context",
-                                       sizeof(struct scsi_io_context),
-                                       0, 0, NULL);
-       if (!scsi_io_context_cache) {
-               printk(KERN_ERR "SCSI: can't init scsi io context cache\n");
-               return -ENOMEM;
-       }
-
        scsi_sdb_cache = kmem_cache_create("scsi_data_buffer",
                                           sizeof(struct scsi_data_buffer),
                                           0, 0, NULL);
        if (!scsi_sdb_cache) {
                printk(KERN_ERR "SCSI: can't init scsi sdb cache\n");
-               goto cleanup_io_context;
+               return -ENOMEM;
        }
 
        for (i = 0; i < SG_MEMPOOL_NR; i++) {
@@ -1866,8 +1758,6 @@ cleanup_sdb:
                        kmem_cache_destroy(sgp->slab);
        }
        kmem_cache_destroy(scsi_sdb_cache);
-cleanup_io_context:
-       kmem_cache_destroy(scsi_io_context_cache);
 
        return -ENOMEM;
 }
@@ -1876,7 +1766,6 @@ void scsi_exit_queue(void)
 {
        int i;
 
-       kmem_cache_destroy(scsi_io_context_cache);
        kmem_cache_destroy(scsi_sdb_cache);
 
        for (i = 0; i < SG_MEMPOOL_NR; i++) {
@@ -1957,7 +1846,7 @@ scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
        }
 
        ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
-                              sshdr, timeout, retries);
+                              sshdr, timeout, retries, NULL);
        kfree(real_buffer);
        return ret;
 }
@@ -2022,7 +1911,7 @@ scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
        memset(buffer, 0, len);
 
        result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
-                                 sshdr, timeout, retries);
+                                 sshdr, timeout, retries, NULL);
 
        /* This code looks awful: what it's doing is making sure an
         * ILLEGAL REQUEST sense return identifies the actual command
@@ -2104,23 +1993,22 @@ scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
        /* try to eat the UNIT_ATTENTION if there are enough retries */
        do {
                result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
-                                         timeout, retries);
-       } while ((driver_byte(result) & DRIVER_SENSE) &&
-                sshdr && sshdr->sense_key == UNIT_ATTENTION &&
-                --retries);
+                                         timeout, retries, NULL);
+               if (sdev->removable && scsi_sense_valid(sshdr) &&
+                   sshdr->sense_key == UNIT_ATTENTION)
+                       sdev->changed = 1;
+       } while (scsi_sense_valid(sshdr) &&
+                sshdr->sense_key == UNIT_ATTENTION && --retries);
 
        if (!sshdr)
                /* could not allocate sense buffer, so can't process it */
                return result;
 
-       if ((driver_byte(result) & DRIVER_SENSE) && sdev->removable) {
-
-               if ((scsi_sense_valid(sshdr)) &&
-                   ((sshdr->sense_key == UNIT_ATTENTION) ||
-                    (sshdr->sense_key == NOT_READY))) {
-                       sdev->changed = 1;
-                       result = 0;
-               }
+       if (sdev->removable && scsi_sense_valid(sshdr) &&
+           (sshdr->sense_key == UNIT_ATTENTION ||
+            sshdr->sense_key == NOT_READY)) {
+               sdev->changed = 1;
+               result = 0;
        }
        if (!sshdr_external)
                kfree(sshdr);
@@ -2528,20 +2416,18 @@ int
 scsi_internal_device_unblock(struct scsi_device *sdev)
 {
        struct request_queue *q = sdev->request_queue; 
-       int err;
        unsigned long flags;
        
        /* 
         * Try to transition the scsi device to SDEV_RUNNING
         * and goose the device queue if successful.  
         */
-       err = scsi_device_set_state(sdev, SDEV_RUNNING);
-       if (err) {
-               err = scsi_device_set_state(sdev, SDEV_CREATED);
-
-               if (err)
-                       return err;
-       }
+       if (sdev->sdev_state == SDEV_BLOCK)
+               sdev->sdev_state = SDEV_RUNNING;
+       else if (sdev->sdev_state == SDEV_CREATED_BLOCK)
+               sdev->sdev_state = SDEV_CREATED;
+       else
+               return -EINVAL;
 
        spin_lock_irqsave(q->queue_lock, flags);
        blk_start_queue(q);