[SCSI] libfc: call ddp setup for only FCP reads to avoid accessing junk fsp pointer
[safe/jmp/linux-2.6] / drivers / scsi / scsi_lib.c
index dc9c772..c664242 100644 (file)
@@ -8,6 +8,7 @@
  */
 
 #include <linux/bio.h>
+#include <linux/bitops.h>
 #include <linux/blkdev.h>
 #include <linux/completion.h>
 #include <linux/kernel.h>
 #include <linux/init.h>
 #include <linux/pci.h>
 #include <linux/delay.h>
+#include <linux/hardirq.h>
+#include <linux/scatterlist.h>
 
 #include <scsi/scsi.h>
+#include <scsi/scsi_cmnd.h>
 #include <scsi/scsi_dbg.h>
 #include <scsi/scsi_device.h>
 #include <scsi/scsi_driver.h>
 #include <scsi/scsi_eh.h>
 #include <scsi/scsi_host.h>
-#include <scsi/scsi_request.h>
 
 #include "scsi_priv.h"
 #include "scsi_logging.h"
 
 
-#define SG_MEMPOOL_NR          (sizeof(scsi_sg_pools)/sizeof(struct scsi_host_sg_pool))
-#define SG_MEMPOOL_SIZE                32
+#define SG_MEMPOOL_NR          ARRAY_SIZE(scsi_sg_pools)
+#define SG_MEMPOOL_SIZE                2
 
 struct scsi_host_sg_pool {
        size_t          size;
-       char            *name; 
-       kmem_cache_t    *slab;
+       char            *name;
+       struct kmem_cache       *slab;
        mempool_t       *pool;
 };
 
-#if (SCSI_MAX_PHYS_SEGMENTS < 32)
-#error SCSI_MAX_PHYS_SEGMENTS is too small
+#define SP(x) { x, "sgpool-" __stringify(x) }
+#if (SCSI_MAX_SG_SEGMENTS < 32)
+#error SCSI_MAX_SG_SEGMENTS is too small (must be 32 or greater)
 #endif
-
-#define SP(x) { x, "sgpool-" #x } 
 static struct scsi_host_sg_pool scsi_sg_pools[] = {
        SP(8),
        SP(16),
+#if (SCSI_MAX_SG_SEGMENTS > 32)
        SP(32),
-#if (SCSI_MAX_PHYS_SEGMENTS > 32)
+#if (SCSI_MAX_SG_SEGMENTS > 64)
        SP(64),
-#if (SCSI_MAX_PHYS_SEGMENTS > 64)
+#if (SCSI_MAX_SG_SEGMENTS > 128)
        SP(128),
-#if (SCSI_MAX_PHYS_SEGMENTS > 128)
-       SP(256),
-#if (SCSI_MAX_PHYS_SEGMENTS > 256)
-#error SCSI_MAX_PHYS_SEGMENTS is too large
+#if (SCSI_MAX_SG_SEGMENTS > 256)
+#error SCSI_MAX_SG_SEGMENTS is too large (256 MAX)
 #endif
 #endif
 #endif
 #endif
-};     
+       SP(SCSI_MAX_SG_SEGMENTS)
+};
 #undef SP
 
-
-/*
- * Function:    scsi_insert_special_req()
- *
- * Purpose:     Insert pre-formed request into request queue.
- *
- * Arguments:   sreq   - request that is ready to be queued.
- *              at_head        - boolean.  True if we should insert at head
- *                        of queue, false if we should insert at tail.
- *
- * Lock status: Assumed that lock is not held upon entry.
- *
- * Returns:     Nothing
- *
- * Notes:       This function is called from character device and from
- *              ioctl types of functions where the caller knows exactly
- *              what SCSI command needs to be issued.   The idea is that
- *              we merely inject the command into the queue (at the head
- *              for now), and then call the queue request function to actually
- *              process it.
- */
-int scsi_insert_special_req(struct scsi_request *sreq, int at_head)
-{
-       /*
-        * Because users of this function are apt to reuse requests with no
-        * modification, we have to sanitise the request flags here
-        */
-       sreq->sr_request->flags &= ~REQ_DONTPREP;
-       blk_insert_request(sreq->sr_device->request_queue, sreq->sr_request,
-                          at_head, sreq);
-       return 0;
-}
+struct kmem_cache *scsi_sdb_cache;
 
 static void scsi_run_queue(struct request_queue *q);
-static void scsi_release_buffers(struct scsi_cmnd *cmd);
 
 /*
  * Function:   scsi_unprep_request()
@@ -115,35 +85,29 @@ static void scsi_unprep_request(struct request *req)
 {
        struct scsi_cmnd *cmd = req->special;
 
-       req->flags &= ~REQ_DONTPREP;
-       req->special = (req->flags & REQ_SPECIAL) ? cmd->sc_request : NULL;
+       req->cmd_flags &= ~REQ_DONTPREP;
+       req->special = NULL;
 
        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;
+       struct scsi_target *starget = scsi_target(device);
        struct request_queue *q = device->request_queue;
        unsigned long flags;
 
@@ -163,16 +127,24 @@ int scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
         * if a command is requeued with no other commands outstanding
         * either for the device or for the host.
         */
-       if (reason == SCSI_MLQUEUE_HOST_BUSY)
+       switch (reason) {
+       case SCSI_MLQUEUE_HOST_BUSY:
                host->host_blocked = host->max_host_blocked;
-       else if (reason == SCSI_MLQUEUE_DEVICE_BUSY)
+               break;
+       case SCSI_MLQUEUE_DEVICE_BUSY:
                device->device_blocked = device->max_device_blocked;
+               break;
+       case SCSI_MLQUEUE_TARGET_BUSY:
+               starget->target_blocked = starget->max_target_blocked;
+               break;
+       }
 
        /*
         * 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
@@ -195,115 +167,28 @@ int scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
 }
 
 /*
- * Function:    scsi_do_req
+ * Function:    scsi_queue_insert()
  *
- * Purpose:     Queue a SCSI request
+ * Purpose:     Insert a command in the midlevel queue.
  *
- * Arguments:   sreq     - command descriptor.
- *              cmnd      - actual SCSI command to be performed.
- *              buffer    - data buffer.
- *              bufflen   - size of data buffer.
- *              done      - completion function to be run.
- *              timeout   - how long to let it run before timeout.
- *              retries   - number of retries we allow.
+ * Arguments:   cmd    - command that we are adding to queue.
+ *              reason - why we are inserting command to queue.
  *
- * Lock status: No locks held upon entry.
+ * Lock status: Assumed that lock is not held upon entry.
  *
  * Returns:     Nothing.
  *
- * Notes:      This function is only used for queueing requests for things
- *             like ioctls and character device requests - this is because
- *             we essentially just inject a request into the queue for the
- *             device.
- *
- *             In order to support the scsi_device_quiesce function, we
- *             now inject requests on the *head* of the device queue
- *             rather than the tail.
+ * 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.
  */
-void scsi_do_req(struct scsi_request *sreq, const void *cmnd,
-                void *buffer, unsigned bufflen,
-                void (*done)(struct scsi_cmnd *),
-                int timeout, int retries)
-{
-       /*
-        * If the upper level driver is reusing these things, then
-        * we should release the low-level block now.  Another one will
-        * be allocated later when this request is getting queued.
-        */
-       __scsi_release_request(sreq);
-
-       /*
-        * Our own function scsi_done (which marks the host as not busy,
-        * disables the timeout counter, etc) will be called by us or by the
-        * scsi_hosts[host].queuecommand() function needs to also call
-        * the completion function for the high level driver.
-        */
-       memcpy(sreq->sr_cmnd, cmnd, sizeof(sreq->sr_cmnd));
-       sreq->sr_bufflen = bufflen;
-       sreq->sr_buffer = buffer;
-       sreq->sr_allowed = retries;
-       sreq->sr_done = done;
-       sreq->sr_timeout_per_command = timeout;
-
-       if (sreq->sr_cmd_len == 0)
-               sreq->sr_cmd_len = COMMAND_SIZE(sreq->sr_cmnd[0]);
-
-       /*
-        * head injection *required* here otherwise quiesce won't work
-        */
-       scsi_insert_special_req(sreq, 1);
-}
-EXPORT_SYMBOL(scsi_do_req);
-
-/* This is the end routine we get to if a command was never attached
- * to the request.  Simply complete the request without changing
- * rq_status; this will cause a DRIVER_ERROR. */
-static void scsi_wait_req_end_io(struct request *req)
-{
-       BUG_ON(!req->waiting);
-
-       complete(req->waiting);
-}
-
-void scsi_wait_req(struct scsi_request *sreq, const void *cmnd, void *buffer,
-                  unsigned bufflen, int timeout, int retries)
+int scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
 {
-       DECLARE_COMPLETION(wait);
-       int write = (sreq->sr_data_direction == DMA_TO_DEVICE);
-       struct request *req;
-
-       req = blk_get_request(sreq->sr_device->request_queue, write,
-                             __GFP_WAIT);
-       if (bufflen && blk_rq_map_kern(sreq->sr_device->request_queue, req,
-                                      buffer, bufflen, __GFP_WAIT)) {
-               sreq->sr_result = DRIVER_ERROR << 24;
-               blk_put_request(req);
-               return;
-       }
-
-       req->flags |= REQ_NOMERGE;
-       req->waiting = &wait;
-       req->end_io = scsi_wait_req_end_io;
-       req->cmd_len = COMMAND_SIZE(((u8 *)cmnd)[0]);
-       req->sense = sreq->sr_sense_buffer;
-       req->sense_len = 0;
-       memcpy(req->cmd, cmnd, req->cmd_len);
-       req->timeout = timeout;
-       req->flags |= REQ_BLOCK_PC;
-       req->rq_disk = NULL;
-       blk_insert_request(sreq->sr_device->request_queue, req,
-                          sreq->sr_data_direction == DMA_TO_DEVICE, NULL);
-       wait_for_completion(&wait);
-       sreq->sr_request->waiting = NULL;
-       sreq->sr_result = req->errors;
-       if (req->errors)
-               sreq->sr_result |= (DRIVER_ERROR << 24);
-
-       blk_put_request(req);
+       return __scsi_queue_insert(cmd, reason, 1);
 }
-
-EXPORT_SYMBOL(scsi_wait_req);
-
 /**
  * scsi_execute - insert request and wait for the result
  * @sdev:      scsi device
@@ -315,13 +200,15 @@ EXPORT_SYMBOL(scsi_wait_req);
  * @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 the scsi_cmnd result
+ * 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);
@@ -337,14 +224,27 @@ int scsi_execute(struct scsi_device *sdev, const unsigned char *cmd,
        memcpy(req->cmd, cmd, req->cmd_len);
        req->sense = sense;
        req->sense_len = 0;
+       req->retries = retries;
        req->timeout = timeout;
-       req->flags |= flags | REQ_BLOCK_PC | REQ_SPECIAL | REQ_QUIET;
+       req->cmd_type = REQ_TYPE_BLOCK_PC;
+       req->cmd_flags |= flags | REQ_QUIET | REQ_PREEMPT;
 
        /*
         * head injection *required* here otherwise quiesce won't work
         */
        blk_execute_rq(req->q, NULL, req, 1);
 
+       /*
+        * Some devices (USB mass-storage in particular) may transfer
+        * garbage data together with a residue indicating that the data
+        * is invalid.  Prevent the garbage from being misinterpreted
+        * and prevent security leaks by zeroing out the excess data.
+        */
+       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);
@@ -356,19 +256,19 @@ 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;
        
        if (sshdr) {
-               sense = kmalloc(SCSI_SENSE_BUFFERSIZE, GFP_NOIO);
+               sense = kzalloc(SCSI_SENSE_BUFFERSIZE, GFP_NOIO);
                if (!sense)
                        return DRIVER_ERROR << 24;
-               memset(sense, 0, SCSI_SENSE_BUFFERSIZE);
        }
        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);
 
@@ -384,71 +284,30 @@ EXPORT_SYMBOL(scsi_execute_req);
  *
  * Arguments:   cmd    - command that is ready to be queued.
  *
- * Returns:     Nothing
- *
  * Notes:       This function has the job of initializing a number of
  *              fields related to error handling.   Typically this will
  *              be called once for each command, as required.
  */
-static int scsi_init_cmd_errh(struct scsi_cmnd *cmd)
+static void scsi_init_cmd_errh(struct scsi_cmnd *cmd)
 {
        cmd->serial_number = 0;
-
-       memset(cmd->sense_buffer, 0, sizeof cmd->sense_buffer);
-
+       scsi_set_resid(cmd, 0);
+       memset(cmd->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
        if (cmd->cmd_len == 0)
-               cmd->cmd_len = COMMAND_SIZE(cmd->cmnd[0]);
-
-       /*
-        * We need saved copies of a number of fields - this is because
-        * error handling may need to overwrite these with different values
-        * to run different commands, and once error handling is complete,
-        * we will need to restore these values prior to running the actual
-        * command.
-        */
-       cmd->old_use_sg = cmd->use_sg;
-       cmd->old_cmd_len = cmd->cmd_len;
-       cmd->sc_old_data_direction = cmd->sc_data_direction;
-       cmd->old_underflow = cmd->underflow;
-       memcpy(cmd->data_cmnd, cmd->cmnd, sizeof(cmd->cmnd));
-       cmd->buffer = cmd->request_buffer;
-       cmd->bufflen = cmd->request_bufflen;
-
-       return 1;
-}
-
-/*
- * Function:   scsi_setup_cmd_retry()
- *
- * Purpose:    Restore the command state for a retry
- *
- * Arguments:  cmd     - command to be restored
- *
- * Returns:    Nothing
- *
- * Notes:      Immediately prior to retrying a command, we need
- *             to restore certain fields that we saved above.
- */
-void scsi_setup_cmd_retry(struct scsi_cmnd *cmd)
-{
-       memcpy(cmd->cmnd, cmd->data_cmnd, sizeof(cmd->data_cmnd));
-       cmd->request_buffer = cmd->buffer;
-       cmd->request_bufflen = cmd->bufflen;
-       cmd->use_sg = cmd->old_use_sg;
-       cmd->cmd_len = cmd->old_cmd_len;
-       cmd->sc_data_direction = cmd->sc_old_data_direction;
-       cmd->underflow = cmd->old_underflow;
+               cmd->cmd_len = scsi_command_size(cmd->cmnd);
 }
 
 void scsi_device_unbusy(struct scsi_device *sdev)
 {
        struct Scsi_Host *shost = sdev->host;
+       struct scsi_target *starget = scsi_target(sdev);
        unsigned long flags;
 
        spin_lock_irqsave(shost->host_lock, flags);
        shost->host_busy--;
+       starget->target_busy--;
        if (unlikely(scsi_host_in_recovery(shost) &&
-                    shost->host_failed))
+                    (shost->host_failed || shost->host_eh_scheduled)))
                scsi_eh_wakeup(shost);
        spin_unlock(shost->host_lock);
        spin_lock(sdev->request_queue->queue_lock);
@@ -502,6 +361,30 @@ 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 &&
+                starget->target_busy >= starget->can_queue) ||
+                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()
  *
@@ -518,16 +401,18 @@ static void scsi_run_queue(struct request_queue *q)
 {
        struct scsi_device *sdev = q->queuedata;
        struct Scsi_Host *shost = sdev->host;
+       LIST_HEAD(starved_list);
        unsigned long flags;
 
-       if (sdev->single_lun)
+       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;
+
                /*
                 * As long as shost is accepting commands and we have
                 * starved queues, call blk_run_queue. scsi_request_fn
@@ -538,22 +423,35 @@ 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);
+               if (scsi_host_is_busy(shost))
+                       break;
+
+               sdev = list_entry(starved_list.next,
+                                 struct scsi_device, starved_entry);
                list_del_init(&sdev->starved_entry);
-               spin_unlock_irqrestore(shost->host_lock, flags);
+               if (scsi_target_is_busy(scsi_target(sdev))) {
+                       list_move_tail(&sdev->starved_entry,
+                                      &shost->starved_list);
+                       continue;
+               }
 
-               blk_run_queue(sdev->request_queue);
+               spin_unlock(shost->host_lock);
 
-               spin_lock_irqsave(shost->host_lock, flags);
-               if (unlikely(!list_empty(&sdev->starved_entry)))
-                       /*
-                        * sdev lost a race, and was put back on the
-                        * starved list. This is unlikely but without this
-                        * in theory we could loop forever.
-                        */
-                       break;
+               spin_lock(sdev->request_queue->queue_lock);
+               flagset = test_bit(QUEUE_FLAG_REENTER, &q->queue_flags) &&
+                               !test_bit(QUEUE_FLAG_REENTER,
+                                       &sdev->request_queue->queue_flags);
+               if (flagset)
+                       queue_flag_set(QUEUE_FLAG_REENTER, sdev->request_queue);
+               __blk_run_queue(sdev->request_queue);
+               if (flagset)
+                       queue_flag_clear(QUEUE_FLAG_REENTER, sdev->request_queue);
+               spin_unlock(sdev->request_queue->queue_lock);
+
+               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);
@@ -582,8 +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);
 
@@ -592,10 +490,17 @@ static void scsi_requeue_command(struct request_queue *q, struct scsi_cmnd *cmd)
 
 void scsi_next_command(struct scsi_cmnd *cmd)
 {
-       struct request_queue *q = cmd->device->request_queue;
+       struct scsi_device *sdev = cmd->device;
+       struct request_queue *q = sdev->request_queue;
+
+       /* need to hold a reference on the device before we let go of the cmd */
+       get_device(&sdev->sdev_gendev);
 
        scsi_put_command(cmd);
        scsi_run_queue(q);
+
+       /* ok to remove device now */
+       put_device(&sdev->sdev_gendev);
 }
 
 void scsi_run_host_queues(struct Scsi_Host *shost)
@@ -606,6 +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()
  *
@@ -613,7 +520,7 @@ void scsi_run_host_queues(struct Scsi_Host *shost)
  *             of upper level post-processing and scsi_io_completion).
  *
  * Arguments:   cmd     - command that is complete.
- *              uptodate - 1 if I/O indicates success, <= 0 for I/O error.
+ *              error    - 0 if I/O indicates success, < 0 for I/O error.
  *              bytes    - number of bytes of completed I/O
  *             requeue  - indicates whether we should requeue leftovers.
  *
@@ -628,26 +535,20 @@ void scsi_run_host_queues(struct Scsi_Host *shost)
  *             at some point during this call.
  * Notes:      If cmd was requeued, upon return it will be a stale pointer.
  */
-static struct scsi_cmnd *scsi_end_request(struct scsi_cmnd *cmd, int uptodate,
+static struct scsi_cmnd *scsi_end_request(struct scsi_cmnd *cmd, int error,
                                          int bytes, int requeue)
 {
-       request_queue_t *q = cmd->device->request_queue;
+       struct request_queue *q = cmd->device->request_queue;
        struct request *req = cmd->request;
-       unsigned long flags;
 
        /*
         * If there are blocks left over at the end, set up the command
         * to queue the remainder of them.
         */
-       if (end_that_request_chunk(req, uptodate, bytes)) {
-               int leftover = (req->hard_nr_sectors << 9);
-
-               if (blk_pc_request(req))
-                       leftover = req->data_len;
-
+       if (blk_end_request(req, error, bytes)) {
                /* kill remainder if no retrys */
-               if (!uptodate && blk_noretry_request(req))
-                       end_that_request_chunk(req, 0, leftover);
+               if (error && scsi_noretry_cmd(cmd))
+                       blk_end_request_all(req, error);
                else {
                        if (requeue) {
                                /*
@@ -655,6 +556,7 @@ static struct scsi_cmnd *scsi_end_request(struct scsi_cmnd *cmd, int uptodate,
                                 * leftovers in the front of the
                                 * queue, and goose the queue again.
                                 */
+                               scsi_release_buffers(cmd);
                                scsi_requeue_command(q, cmd);
                                cmd = NULL;
                        }
@@ -662,73 +564,86 @@ static struct scsi_cmnd *scsi_end_request(struct scsi_cmnd *cmd, int uptodate,
                }
        }
 
-       add_disk_randomness(req->rq_disk);
-
-       spin_lock_irqsave(q->queue_lock, flags);
-       if (blk_rq_tagged(req))
-               blk_queue_end_tag(q, req);
-       end_that_request_last(req);
-       spin_unlock_irqrestore(q->queue_lock, flags);
-
        /*
         * 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;
 }
 
-static struct scatterlist *scsi_alloc_sgtable(struct scsi_cmnd *cmd, int gfp_mask)
+static inline unsigned int scsi_sgtable_index(unsigned short nents)
 {
-       struct scsi_host_sg_pool *sgp;
-       struct scatterlist *sgl;
+       unsigned int index;
 
-       BUG_ON(!cmd->use_sg);
+       BUG_ON(nents > SCSI_MAX_SG_SEGMENTS);
 
-       switch (cmd->use_sg) {
-       case 1 ... 8:
-               cmd->sglist_len = 0;
-               break;
-       case 9 ... 16:
-               cmd->sglist_len = 1;
-               break;
-       case 17 ... 32:
-               cmd->sglist_len = 2;
-               break;
-#if (SCSI_MAX_PHYS_SEGMENTS > 32)
-       case 33 ... 64:
-               cmd->sglist_len = 3;
-               break;
-#if (SCSI_MAX_PHYS_SEGMENTS > 64)
-       case 65 ... 128:
-               cmd->sglist_len = 4;
-               break;
-#if (SCSI_MAX_PHYS_SEGMENTS  > 128)
-       case 129 ... 256:
-               cmd->sglist_len = 5;
-               break;
-#endif
-#endif
-#endif
-       default:
-               return NULL;
-       }
+       if (nents <= 8)
+               index = 0;
+       else
+               index = get_count_order(nents) - 3;
 
-       sgp = scsi_sg_pools + cmd->sglist_len;
-       sgl = mempool_alloc(sgp->pool, gfp_mask);
-       return sgl;
+       return index;
 }
 
-static void scsi_free_sgtable(struct scatterlist *sgl, int index)
+static void scsi_sg_free(struct scatterlist *sgl, unsigned int nents)
 {
        struct scsi_host_sg_pool *sgp;
 
-       BUG_ON(index >= SG_MEMPOOL_NR);
-
-       sgp = scsi_sg_pools + index;
+       sgp = scsi_sg_pools + scsi_sgtable_index(nents);
        mempool_free(sgl, sgp->pool);
 }
 
+static struct scatterlist *scsi_sg_alloc(unsigned int nents, gfp_t gfp_mask)
+{
+       struct scsi_host_sg_pool *sgp;
+
+       sgp = scsi_sg_pools + scsi_sgtable_index(nents);
+       return mempool_alloc(sgp->pool, gfp_mask);
+}
+
+static int scsi_alloc_sgtable(struct scsi_data_buffer *sdb, int nents,
+                             gfp_t gfp_mask)
+{
+       int ret;
+
+       BUG_ON(!nents);
+
+       ret = __sg_alloc_table(&sdb->table, nents, SCSI_MAX_SG_SEGMENTS,
+                              gfp_mask, scsi_sg_alloc);
+       if (unlikely(ret))
+               __sg_free_table(&sdb->table, SCSI_MAX_SG_SEGMENTS,
+                               scsi_sg_free);
+
+       return ret;
+}
+
+static void scsi_free_sgtable(struct scsi_data_buffer *sdb)
+{
+       __sg_free_table(&sdb->table, SCSI_MAX_SG_SEGMENTS, scsi_sg_free);
+}
+
+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 (do_bidi_check && scsi_bidi_cmnd(cmd)) {
+               struct scsi_data_buffer *bidi_sdb =
+                       cmd->request->next_rq->special;
+               scsi_free_sgtable(bidi_sdb);
+               kmem_cache_free(scsi_sdb_cache, bidi_sdb);
+               cmd->request->next_rq->special = NULL;
+       }
+
+       if (scsi_prot_sg_count(cmd))
+               scsi_free_sgtable(cmd->prot_sdb);
+}
+
 /*
  * Function:    scsi_release_buffers()
  *
@@ -746,27 +661,11 @@ static void scsi_free_sgtable(struct scatterlist *sgl, int index)
  *             the scatter-gather table, and potentially any bounce
  *             buffers.
  */
-static void scsi_release_buffers(struct scsi_cmnd *cmd)
+void scsi_release_buffers(struct scsi_cmnd *cmd)
 {
-       struct request *req = cmd->request;
-
-       /*
-        * Free up any indirection buffers we allocated for DMA purposes. 
-        */
-       if (cmd->use_sg)
-               scsi_free_sgtable(cmd->request_buffer, cmd->sglist_len);
-       else if (cmd->request_buffer != req->buffer)
-               kfree(cmd->request_buffer);
-
-       /*
-        * Zero these out.  They now point to freed memory, and it is
-        * dangerous to hang onto the pointers.
-        */
-       cmd->buffer  = NULL;
-       cmd->bufflen = 0;
-       cmd->request_buffer = NULL;
-       cmd->request_bufflen = 0;
+       __scsi_release_buffers(cmd, 1);
 }
+EXPORT_SYMBOL(scsi_release_buffers);
 
 /*
  * Function:    scsi_io_completion()
@@ -785,58 +684,47 @@ static void scsi_release_buffers(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) 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.
  *
- *             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.
+ *             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.
  *
- *             b) We can just use scsi_requeue_command() here.  This would
- *                be used if we just wanted to retry, for example.
+ *             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,
-                       unsigned int block_bytes)
+void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
 {
        int result = cmd->result;
-       int this_count = cmd->bufflen;
-       request_queue_t *q = cmd->device->request_queue;
+       struct request_queue *q = cmd->device->request_queue;
        struct request *req = cmd->request;
-       int clear_errors = 1;
+       int error = 0;
        struct scsi_sense_hdr sshdr;
        int sense_valid = 0;
        int sense_deferred = 0;
-
-       if (blk_complete_barrier_rq(q, req, good_bytes >> 9))
-               return;
-
-       /*
-        * Free up any indirection buffers we allocated for DMA purposes. 
-        * For the case of a READ, we need to copy the data out of the
-        * bounce buffer and into the real buffer.
-        */
-       if (cmd->use_sg)
-               scsi_free_sgtable(cmd->buffer, cmd->sglist_len);
-       else if (cmd->buffer != req->buffer) {
-               if (rq_data_dir(req) == READ) {
-                       unsigned long flags;
-                       char *to = bio_kmap_irq(req->bio, &flags);
-                       memcpy(to, cmd->buffer, cmd->bufflen);
-                       bio_kunmap_irq(to, &flags);
-               }
-               kfree(cmd->buffer);
-       }
+       enum {ACTION_FAIL, ACTION_REPREP, ACTION_RETRY,
+             ACTION_DELAYED_RETRY} action;
+       char *description = NULL;
 
        if (result) {
                sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
                if (sense_valid)
                        sense_deferred = scsi_sense_is_deferred(&sshdr);
        }
+
        if (blk_pc_request(req)) { /* SG_IO ioctl from block level */
                req->errors = result;
                if (result) {
-                       clear_errors = 0;
                        if (sense_valid && req->sense) {
                                /*
                                 * SG_IO wants current and deferred errors
@@ -848,158 +736,217 @@ void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes,
                                memcpy(req->sense, cmd->sense_buffer,  len);
                                req->sense_len = len;
                        }
-               } else
-                       req->data_len = cmd->resid;
+                       if (!sense_deferred)
+                               error = -EIO;
+               }
+
+               req->resid_len = scsi_get_resid(cmd);
+
+               if (scsi_bidi_cmnd(cmd)) {
+                       /*
+                        * Bidi commands Must be complete as a whole,
+                        * both sides at once.
+                        */
+                       req->next_rq->resid_len = scsi_in(cmd)->resid;
+
+                       scsi_release_buffers(cmd);
+                       blk_end_request_all(req, 0);
+
+                       scsi_next_command(cmd);
+                       return;
+               }
        }
 
-       /*
-        * Zero these out.  They now point to freed memory, and it is
-        * dangerous to hang onto the pointers.
-        */
-       cmd->buffer  = NULL;
-       cmd->bufflen = 0;
-       cmd->request_buffer = NULL;
-       cmd->request_bufflen = 0;
+       BUG_ON(blk_bidi_rq(req)); /* bidi not support for !blk_pc_request yet */
 
        /*
         * Next deal with any sectors which we were able to correctly
         * handle.
         */
-       if (good_bytes >= 0) {
-               SCSI_LOG_HLCOMPLETE(1, printk("%ld sectors total, %d bytes done.\n",
-                                             req->nr_sectors, good_bytes));
-               SCSI_LOG_HLCOMPLETE(1, printk("use_sg is %d\n", cmd->use_sg));
-
-               if (clear_errors)
-                       req->errors = 0;
-               /*
-                * If multiple sectors are requested in one buffer, then
-                * they will have been finished off by the first command.
-                * If not, then we have a multi-buffer command.
-                *
-                * If block_bytes != 0, it means we had a medium error
-                * of some sort, and that we want to mark some number of
-                * sectors as not uptodate.  Thus we want to inhibit
-                * requeueing right here - we will requeue down below
-                * when we handle the bad sectors.
-                */
+       SCSI_LOG_HLCOMPLETE(1, printk("%u sectors total, "
+                                     "%d bytes done.\n",
+                                     blk_rq_sectors(req), good_bytes));
 
-               /*
-                * If the command completed without error, then either
-                * finish off the rest of the command, or start a new one.
-                */
-               if (scsi_end_request(cmd, 1, good_bytes, result == 0) == NULL)
-                       return;
+       /*
+        * 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;
        }
+
        /*
-        * Now, if we were good little boys and girls, Santa left us a request
-        * sense buffer.  We can extract information from this, so we
-        * can choose a block to remap, etc.
+        * A number of bytes were successfully read.  If there
+        * are leftovers and there is some kind of error
+        * (result != 0), retry the rest.
         */
-       if (sense_valid && !sense_deferred) {
+       if (scsi_end_request(cmd, error, good_bytes, result == 0) == NULL)
+               return;
+
+       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) {
-                               /* detected disc change.  set a bit 
+                               /* Detected disc change.  Set a bit
                                 * and quietly refuse further access.
                                 */
                                cmd->device->changed = 1;
-                               scsi_end_request(cmd, 0,
-                                               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.  
-                               */
-                               scsi_requeue_command(q, cmd);
-                               return;
+                               /* Must have been a power glitch, or a
+                                * bus reset.  Could not have been a
+                                * media change, so we just retry the
+                                * command and see what happens.
+                                */
+                               action = ACTION_RETRY;
                        }
                        break;
                case ILLEGAL_REQUEST:
-                       /*
-                       * If we had an ILLEGAL REQUEST returned, then we may
-                       * have performed an unsupported command.  The only
-                       * thing this should be would be a ten byte read where
-                       * only a six byte read was supported.  Also, on a
-                       * system where READ CAPACITY failed, we may have read
-                       * past the end of the disk.
-                       */
-                       if (cmd->device->use_10_for_rw &&
+                       /* If we had an ILLEGAL REQUEST returned, then
+                        * we may have performed an unsupported
+                        * command.  The only thing this should be
+                        * would be a ten byte read where only a six
+                        * byte read was supported.  Also, on a system
+                        * where READ CAPACITY failed, we may have
+                        * read past the end of the disk.
+                        */
+                       if ((cmd->device->use_10_for_rw &&
+                           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);
-                               result = 0;
-                       } else {
-                               scsi_end_request(cmd, 0, 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 */
+                               description = "Target Data Integrity Failure";
+                               error = -EILSEQ;
                        }
                        break;
                case NOT_READY:
-                       /*
-                        * If the device is in the process of becoming ready,
-                        * retry.
+                       /* If the device is in the process of becoming
+                        * ready, or has a temporary blockage, retry.
                         */
-                       if (sshdr.asc == 0x04 && sshdr.ascq == 0x01) {
-                               scsi_requeue_command(q, cmd);
-                               return;
+                       if (sshdr.asc == 0x04) {
+                               switch (sshdr.ascq) {
+                               case 0x01: /* becoming ready */
+                               case 0x04: /* format in progress */
+                               case 0x05: /* rebuild in progress */
+                               case 0x06: /* recalculation in progress */
+                               case 0x07: /* operation in progress */
+                               case 0x08: /* Long write in progress */
+                               case 0x09: /* self test in progress */
+                               case 0x14: /* space allocation in progress */
+                                       action = ACTION_DELAYED_RETRY;
+                                       break;
+                               default:
+                                       description = "Device not ready";
+                                       action = ACTION_FAIL;
+                                       break;
+                               }
+                       } else {
+                               description = "Device not ready";
+                               action = ACTION_FAIL;
                        }
-                       if (!(req->flags & REQ_QUIET))
-                               dev_printk(KERN_INFO,
-                                          &cmd->device->sdev_gendev,
-                                          "Device not ready.\n");
-                       scsi_end_request(cmd, 0, this_count, 1);
-                       return;
+                       break;
                case VOLUME_OVERFLOW:
-                       if (!(req->flags & REQ_QUIET)) {
-                               dev_printk(KERN_INFO,
-                                          &cmd->device->sdev_gendev,
-                                          "Volume overflow, CDB: ");
-                               __scsi_print_command(cmd->data_cmnd);
-                               scsi_print_sense("", cmd);
-                       }
-                       scsi_end_request(cmd, 0, block_bytes, 1);
-                       return;
+                       /* See SSC3rXX or current. */
+                       action = ACTION_FAIL;
+                       break;
                default:
+                       description = "Unhandled sense code";
+                       action = ACTION_FAIL;
                        break;
                }
-       }                       /* driver byte != 0 */
-       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;
+       } else {
+               description = "Unhandled error code";
+               action = ACTION_FAIL;
        }
-       if (result) {
-               if (!(req->flags & REQ_QUIET)) {
-                       dev_printk(KERN_INFO, &cmd->device->sdev_gendev,
-                                  "SCSI error: return code = 0x%x\n", 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);
                }
-               /*
-                * Mark a single buffer as not uptodate.  Queue the remainder.
-                * We sometimes get this cruft in the event that a medium error
-                * isn't properly reported.
+               if (blk_end_request_err(req, error))
+                       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.
                 */
-               block_bytes = req->hard_cur_sectors << 9;
-               if (!block_bytes)
-                       block_bytes = req->data_len;
-               scsi_end_request(cmd, 0, block_bytes, 1);
+               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;
        }
 }
-EXPORT_SYMBOL(scsi_io_completion);
+
+static int scsi_init_sgtable(struct request *req, struct scsi_data_buffer *sdb,
+                            gfp_t gfp_mask)
+{
+       int count;
+
+       /*
+        * If sg table allocation fails, requeue request later.
+        */
+       if (unlikely(scsi_alloc_sgtable(sdb, req->nr_phys_segments,
+                                       gfp_mask))) {
+               return BLKPREP_DEFER;
+       }
+
+       req->buffer = NULL;
+
+       /* 
+        * Next, walk the list, and fill in the addresses and sizes of
+        * each segment.
+        */
+       count = blk_rq_map_sg(req->q, req, sdb->table.sgl);
+       BUG_ON(count > sdb->table.nents);
+       sdb->table.nents = count;
+       sdb->length = blk_rq_bytes(req);
+       return BLKPREP_OK;
+}
 
 /*
  * Function:    scsi_init_io()
@@ -1012,290 +959,259 @@ EXPORT_SYMBOL(scsi_io_completion);
  *             BLKPREP_DEFER if the failure is retryable
  *             BLKPREP_KILL if the failure is fatal
  */
-static int scsi_init_io(struct scsi_cmnd *cmd)
+int scsi_init_io(struct scsi_cmnd *cmd, gfp_t gfp_mask)
 {
-       struct request     *req = cmd->request;
-       struct scatterlist *sgpnt;
-       int                count;
+       int error = scsi_init_sgtable(cmd->request, &cmd->sdb, gfp_mask);
+       if (error)
+               goto err_exit;
+
+       if (blk_bidi_rq(cmd->request)) {
+               struct scsi_data_buffer *bidi_sdb = kmem_cache_zalloc(
+                       scsi_sdb_cache, GFP_ATOMIC);
+               if (!bidi_sdb) {
+                       error = BLKPREP_DEFER;
+                       goto err_exit;
+               }
 
-       /*
-        * if this is a rq->data based REQ_BLOCK_PC, setup for a non-sg xfer
-        */
-       if ((req->flags & REQ_BLOCK_PC) && !req->bio) {
-               cmd->request_bufflen = req->data_len;
-               cmd->request_buffer = req->data;
-               req->buffer = req->data;
-               cmd->use_sg = 0;
-               return 0;
+               cmd->request->next_rq->special = bidi_sdb;
+               error = scsi_init_sgtable(cmd->request->next_rq, bidi_sdb,
+                                                                   GFP_ATOMIC);
+               if (error)
+                       goto err_exit;
        }
 
-       /*
-        * we used to not use scatter-gather for single segment request,
-        * but now we do (it makes highmem I/O easier to support without
-        * kmapping pages)
-        */
-       cmd->use_sg = req->nr_phys_segments;
+       if (blk_integrity_rq(cmd->request)) {
+               struct scsi_data_buffer *prot_sdb = cmd->prot_sdb;
+               int ivecs, count;
 
-       /*
-        * if sg table allocation fails, requeue request later.
-        */
-       sgpnt = scsi_alloc_sgtable(cmd, GFP_ATOMIC);
-       if (unlikely(!sgpnt))
-               return BLKPREP_DEFER;
+               BUG_ON(prot_sdb == NULL);
+               ivecs = blk_rq_count_integrity_sg(cmd->request);
 
-       cmd->request_buffer = (char *) sgpnt;
-       cmd->request_bufflen = req->nr_sectors << 9;
-       if (blk_pc_request(req))
-               cmd->request_bufflen = req->data_len;
-       req->buffer = NULL;
+               if (scsi_alloc_sgtable(prot_sdb, ivecs, gfp_mask)) {
+                       error = BLKPREP_DEFER;
+                       goto err_exit;
+               }
 
-       /* 
-        * Next, walk the list, and fill in the addresses and sizes of
-        * each segment.
-        */
-       count = blk_rq_map_sg(req->q, req, cmd->request_buffer);
+               count = blk_rq_map_integrity_sg(cmd->request,
+                                               prot_sdb->table.sgl);
+               BUG_ON(unlikely(count > ivecs));
 
-       /*
-        * mapped well, send it off
-        */
-       if (likely(count <= cmd->use_sg)) {
-               cmd->use_sg = count;
-               return 0;
+               cmd->prot_sdb = prot_sdb;
+               cmd->prot_sdb->table.nents = count;
        }
 
-       printk(KERN_ERR "Incorrect number of segments after building list\n");
-       printk(KERN_ERR "counted %d, received %d\n", count, cmd->use_sg);
-       printk(KERN_ERR "req nr_sec %lu, cur_nr_sec %u\n", req->nr_sectors,
-                       req->current_nr_sectors);
+       return BLKPREP_OK ;
 
-       /* release the command and kill it */
+err_exit:
        scsi_release_buffers(cmd);
-       scsi_put_command(cmd);
-       return BLKPREP_KILL;
+       if (error == BLKPREP_KILL)
+               scsi_put_command(cmd);
+       else /* BLKPREP_DEFER */
+               scsi_unprep_request(cmd->request);
+
+       return error;
 }
+EXPORT_SYMBOL(scsi_init_io);
 
-static int scsi_prepare_flush_fn(request_queue_t *q, struct request *rq)
+static struct scsi_cmnd *scsi_get_cmd_from_req(struct scsi_device *sdev,
+               struct request *req)
 {
-       struct scsi_device *sdev = q->queuedata;
-       struct scsi_driver *drv;
-
-       if (sdev->sdev_state == SDEV_RUNNING) {
-               drv = *(struct scsi_driver **) rq->rq_disk->private_data;
+       struct scsi_cmnd *cmd;
 
-               if (drv->prepare_flush)
-                       return drv->prepare_flush(q, rq);
+       if (!req->special) {
+               cmd = scsi_get_command(sdev, GFP_ATOMIC);
+               if (unlikely(!cmd))
+                       return NULL;
+               req->special = cmd;
+       } else {
+               cmd = req->special;
        }
 
-       return 0;
+       /* pull a tag out of the request if we have one */
+       cmd->tag = req->tag;
+       cmd->request = req;
+
+       cmd->cmnd = req->cmd;
+
+       return cmd;
 }
 
-static void scsi_end_flush_fn(request_queue_t *q, struct request *rq)
+int scsi_setup_blk_pc_cmnd(struct scsi_device *sdev, struct request *req)
 {
-       struct scsi_device *sdev = q->queuedata;
-       struct request *flush_rq = rq->end_io_data;
-       struct scsi_driver *drv;
+       struct scsi_cmnd *cmd;
+       int ret = scsi_prep_state_check(sdev, req);
 
-       if (flush_rq->errors) {
-               printk("scsi: barrier error, disabling flush support\n");
-               blk_queue_ordered(q, QUEUE_ORDERED_NONE);
-       }
+       if (ret != BLKPREP_OK)
+               return ret;
+
+       cmd = scsi_get_cmd_from_req(sdev, req);
+       if (unlikely(!cmd))
+               return BLKPREP_DEFER;
+
+       /*
+        * BLOCK_PC requests may transfer data, in which case they must
+        * a bio attached to them.  Or they might contain a SCSI command
+        * that does not transfer data, in which case they may optionally
+        * submit a request without an attached bio.
+        */
+       if (req->bio) {
+               int ret;
+
+               BUG_ON(!req->nr_phys_segments);
+
+               ret = scsi_init_io(cmd, GFP_ATOMIC);
+               if (unlikely(ret))
+                       return ret;
+       } else {
+               BUG_ON(blk_rq_bytes(req));
 
-       if (sdev->sdev_state == SDEV_RUNNING) {
-               drv = *(struct scsi_driver **) rq->rq_disk->private_data;
-               drv->end_flush(q, rq);
+               memset(&cmd->sdb, 0, sizeof(cmd->sdb));
+               req->buffer = NULL;
        }
+
+       cmd->cmd_len = req->cmd_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 = blk_rq_bytes(req);
+       cmd->allowed = req->retries;
+       return BLKPREP_OK;
 }
+EXPORT_SYMBOL(scsi_setup_blk_pc_cmnd);
 
-static int scsi_issue_flush_fn(request_queue_t *q, struct gendisk *disk,
-                              sector_t *error_sector)
+/*
+ * Setup a REQ_TYPE_FS command.  These are simple read/write request
+ * from filesystems that still need to be translated to SCSI CDBs from
+ * the ULD.
+ */
+int scsi_setup_fs_cmnd(struct scsi_device *sdev, struct request *req)
 {
-       struct scsi_device *sdev = q->queuedata;
-       struct scsi_driver *drv;
+       struct scsi_cmnd *cmd;
+       int ret = scsi_prep_state_check(sdev, req);
 
-       if (sdev->sdev_state != SDEV_RUNNING)
-               return -ENXIO;
+       if (ret != BLKPREP_OK)
+               return ret;
 
-       drv = *(struct scsi_driver **) disk->private_data;
-       if (drv->issue_flush)
-               return drv->issue_flush(&sdev->sdev_gendev, error_sector);
+       if (unlikely(sdev->scsi_dh_data && sdev->scsi_dh_data->scsi_dh
+                        && sdev->scsi_dh_data->scsi_dh->prep_fn)) {
+               ret = sdev->scsi_dh_data->scsi_dh->prep_fn(sdev, req);
+               if (ret != BLKPREP_OK)
+                       return ret;
+       }
 
-       return -EOPNOTSUPP;
-}
+       /*
+        * Filesystem requests must transfer data.
+        */
+       BUG_ON(!req->nr_phys_segments);
 
-static void scsi_generic_done(struct scsi_cmnd *cmd)
-{
-       BUG_ON(!blk_pc_request(cmd->request));
-       scsi_io_completion(cmd, cmd->result == 0 ? cmd->bufflen : 0, 0);
+       cmd = scsi_get_cmd_from_req(sdev, req);
+       if (unlikely(!cmd))
+               return BLKPREP_DEFER;
+
+       memset(cmd->cmnd, 0, BLK_MAX_CDB);
+       return scsi_init_io(cmd, GFP_ATOMIC);
 }
+EXPORT_SYMBOL(scsi_setup_fs_cmnd);
 
-static int scsi_prep_fn(struct request_queue *q, struct request *req)
+int scsi_prep_state_check(struct scsi_device *sdev, struct request *req)
 {
-       struct scsi_device *sdev = q->queuedata;
-       struct scsi_cmnd *cmd;
-       int specials_only = 0;
+       int ret = BLKPREP_OK;
 
        /*
-        * Just check to see if the device is online.  If it isn't, we
-        * refuse to process any commands.  The device must be brought
-        * online before trying any recovery commands
+        * If the device is not in running state we will reject some
+        * or all commands.
         */
-       if (unlikely(!scsi_device_online(sdev))) {
-               printk(KERN_ERR "scsi%d (%d:%d): rejecting I/O to offline device\n",
-                      sdev->host->host_no, sdev->id, sdev->lun);
-               goto kill;
-       }
        if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
-               /* OK, we're not in a running state don't prep
-                * user commands */
-               if (sdev->sdev_state == SDEV_DEL) {
-                       /* Device is fully deleted, no commands
-                        * at all allowed down */
-                       printk(KERN_ERR "scsi%d (%d:%d): rejecting I/O to dead device\n",
-                              sdev->host->host_no, sdev->id, sdev->lun);
-                       goto kill;
+               switch (sdev->sdev_state) {
+               case SDEV_OFFLINE:
+                       /*
+                        * If the device is offline we refuse to process any
+                        * commands.  The device must be brought online
+                        * before trying any recovery commands.
+                        */
+                       sdev_printk(KERN_ERR, sdev,
+                                   "rejecting I/O to offline device\n");
+                       ret = BLKPREP_KILL;
+                       break;
+               case SDEV_DEL:
+                       /*
+                        * If the device is fully deleted, we refuse to
+                        * process any commands as well.
+                        */
+                       sdev_printk(KERN_ERR, sdev,
+                                   "rejecting I/O to dead device\n");
+                       ret = BLKPREP_KILL;
+                       break;
+               case SDEV_QUIESCE:
+               case SDEV_BLOCK:
+               case SDEV_CREATED_BLOCK:
+                       /*
+                        * If the devices is blocked we defer normal commands.
+                        */
+                       if (!(req->cmd_flags & REQ_PREEMPT))
+                               ret = BLKPREP_DEFER;
+                       break;
+               default:
+                       /*
+                        * For any other not fully online state we only allow
+                        * special commands.  In particular any user initiated
+                        * command is not allowed.
+                        */
+                       if (!(req->cmd_flags & REQ_PREEMPT))
+                               ret = BLKPREP_KILL;
+                       break;
                }
-               /* OK, we only allow special commands (i.e. not
-                * user initiated ones */
-               specials_only = sdev->sdev_state;
        }
+       return ret;
+}
+EXPORT_SYMBOL(scsi_prep_state_check);
 
-       /*
-        * Find the actual device driver associated with this command.
-        * The SPECIAL requests are things like character device or
-        * ioctls, which did not originate from ll_rw_blk.  Note that
-        * the special field is also used to indicate the cmd for
-        * the remainder of a partially fulfilled request that can 
-        * come up when there is a medium error.  We have to treat
-        * these two cases differently.  We differentiate by looking
-        * at request->cmd, as this tells us the real story.
-        */
-       if (req->flags & REQ_SPECIAL && req->special) {
-               struct scsi_request *sreq = req->special;
-
-               if (sreq->sr_magic == SCSI_REQ_MAGIC) {
-                       cmd = scsi_get_command(sreq->sr_device, GFP_ATOMIC);
-                       if (unlikely(!cmd))
-                               goto defer;
-                       scsi_init_cmd_from_req(cmd, sreq);
-               } else
-                       cmd = req->special;
-       } else if (req->flags & (REQ_CMD | REQ_BLOCK_PC)) {
+int scsi_prep_return(struct request_queue *q, struct request *req, int ret)
+{
+       struct scsi_device *sdev = q->queuedata;
 
-               if(unlikely(specials_only) && !(req->flags & REQ_SPECIAL)) {
-                       if(specials_only == SDEV_QUIESCE ||
-                                       specials_only == SDEV_BLOCK)
-                               goto defer;
-                       
-                       printk(KERN_ERR "scsi%d (%d:%d): rejecting I/O to device being removed\n",
-                              sdev->host->host_no, sdev->id, sdev->lun);
-                       goto kill;
+       switch (ret) {
+       case BLKPREP_KILL:
+               req->errors = DID_NO_CONNECT << 16;
+               /* release the command and kill it */
+               if (req->special) {
+                       struct scsi_cmnd *cmd = req->special;
+                       scsi_release_buffers(cmd);
+                       scsi_put_command(cmd);
+                       req->special = NULL;
                }
-                       
-                       
+               break;
+       case BLKPREP_DEFER:
                /*
-                * Now try and find a command block that we can use.
+                * 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.
                 */
-               if (!req->special) {
-                       cmd = scsi_get_command(sdev, GFP_ATOMIC);
-                       if (unlikely(!cmd))
-                               goto defer;
-               } else
-                       cmd = req->special;
-               
-               /* pull a tag out of the request if we have one */
-               cmd->tag = req->tag;
-       } else {
-               blk_dump_rq_flags(req, "SCSI bad req");
-               goto kill;
+               if (sdev->device_busy == 0)
+                       blk_plug_device(q);
+               break;
+       default:
+               req->cmd_flags |= REQ_DONTPREP;
        }
-       
-       /* note the overloading of req->special.  When the tag
-        * is active it always means cmd.  If the tag goes
-        * back for re-queueing, it may be reset */
-       req->special = cmd;
-       cmd->request = req;
-       
-       /*
-        * FIXME: drop the lock here because the functions below
-        * expect to be called without the queue lock held.  Also,
-        * previously, we dequeued the request before dropping the
-        * lock.  We hope REQ_STARTED prevents anything untoward from
-        * happening now.
-        */
-       if (req->flags & (REQ_CMD | REQ_BLOCK_PC)) {
-               struct scsi_driver *drv;
-               int ret;
-
-               /*
-                * This will do a couple of things:
-                *  1) Fill in the actual SCSI command.
-                *  2) Fill in any other upper-level specific fields
-                * (timeout).
-                *
-                * If this returns 0, it means that the request failed
-                * (reading past end of disk, reading offline device,
-                * etc).   This won't actually talk to the device, but
-                * some kinds of consistency checking may cause the     
-                * request to be rejected immediately.
-                */
 
-               /* 
-                * This sets up the scatter-gather table (allocating if
-                * required).
-                */
-               ret = scsi_init_io(cmd);
-               switch(ret) {
-               case BLKPREP_KILL:
-                       /* BLKPREP_KILL return also releases the command */
-                       goto kill;
-               case BLKPREP_DEFER:
-                       goto defer;
-               }
-               
-               /*
-                * Initialize the actual SCSI command for this request.
-                */
-               if (req->rq_disk) {
-                       drv = *(struct scsi_driver **)req->rq_disk->private_data;
-                       if (unlikely(!drv->init_command(cmd))) {
-                               scsi_release_buffers(cmd);
-                               scsi_put_command(cmd);
-                               goto kill;
-                       }
-               } else {
-                       memcpy(cmd->cmnd, req->cmd, sizeof(cmd->cmnd));
-                       cmd->cmd_len = req->cmd_len;
-                       if (rq_data_dir(req) == WRITE)
-                               cmd->sc_data_direction = DMA_TO_DEVICE;
-                       else if (req->data_len)
-                               cmd->sc_data_direction = DMA_FROM_DEVICE;
-                       else
-                               cmd->sc_data_direction = DMA_NONE;
-                       
-                       cmd->transfersize = req->data_len;
-                       cmd->allowed = 3;
-                       cmd->timeout_per_command = req->timeout;
-                       cmd->done = scsi_generic_done;
-               }
-       }
+       return ret;
+}
+EXPORT_SYMBOL(scsi_prep_return);
 
-       /*
-        * The request is now prepped, no need to come back here
-        */
-       req->flags |= REQ_DONTPREP;
-       return BLKPREP_OK;
+int scsi_prep_fn(struct request_queue *q, struct request *req)
+{
+       struct scsi_device *sdev = q->queuedata;
+       int ret = BLKPREP_KILL;
 
- defer:
-       /* If we defer, the elv_next_request() returns NULL, but the
-        * queue must be restarted, so we plug here if no returning
-        * command will automatically do that. */
-       if (sdev->device_busy == 0)
-               blk_plug_device(q);
-       return BLKPREP_DEFER;
- kill:
-       req->errors = DID_NO_CONNECT << 16;
-       return BLKPREP_KILL;
+       if (req->cmd_type == REQ_TYPE_BLOCK_PC)
+               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
@@ -1306,25 +1222,66 @@ static 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
                 */
                if (--sdev->device_blocked == 0) {
                        SCSI_LOG_MLQUEUE(3,
-                               printk("scsi%d (%d:%d) unblocking device at"
-                                      " zero depth\n", sdev->host->host_no,
-                                      sdev->id, sdev->lun));
+                                  sdev_printk(KERN_INFO, sdev,
+                                  "unblocking device at zero depth\n"));
                } else {
                        blk_plug_device(q);
                        return 0;
                }
        }
-       if (sdev->device_blocked)
-               return 0;
+       if (scsi_device_is_busy(sdev))
+               return 0;
+
+       return 1;
+}
+
+
+/*
+ * scsi_target_queue_ready: checks if there we can send commands to target
+ * @sdev: scsi device on starget to check.
+ *
+ * Called with the host lock held.
+ */
+static inline int scsi_target_queue_ready(struct Scsi_Host *shost,
+                                          struct scsi_device *sdev)
+{
+       struct scsi_target *starget = scsi_target(sdev);
+
+       if (starget->single_lun) {
+               if (starget->starget_sdev_user &&
+                   starget->starget_sdev_user != sdev)
+                       return 0;
+               starget->starget_sdev_user = sdev;
+       }
+
+       if (starget->target_busy == 0 && starget->target_blocked) {
+               /*
+                * unblock after target_blocked iterates to zero
+                */
+               if (--starget->target_blocked == 0) {
+                       SCSI_LOG_MLQUEUE(3, starget_printk(KERN_INFO, starget,
+                                        "unblocking target at zero depth\n"));
+               } else
+                       return 0;
+       }
+
+       if (scsi_target_is_busy(starget)) {
+               if (list_empty(&sdev->starved_entry)) {
+                       list_add_tail(&sdev->starved_entry,
+                                     &shost->starved_list);
+                       return 0;
+               }
+       }
 
+       /* We're OK to process the command, so we can't be starved */
+       if (!list_empty(&sdev->starved_entry))
+               list_del_init(&sdev->starved_entry);
        return 1;
 }
 
@@ -1350,12 +1307,10 @@ static inline int scsi_host_queue_ready(struct request_queue *q,
                                printk("scsi%d unblocking host at zero depth\n",
                                        shost->host_no));
                } else {
-                       blk_plug_device(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;
@@ -1369,24 +1324,119 @@ 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, request_queue_t *q)
+static void scsi_kill_request(struct request *req, struct request_queue *q)
 {
        struct scsi_cmnd *cmd = req->special;
+       struct scsi_device *sdev;
+       struct scsi_target *starget;
+       struct Scsi_Host *shost;
 
-       blkdev_dequeue_request(req);
+       blk_start_request(req);
 
        if (unlikely(cmd == NULL)) {
                printk(KERN_CRIT "impossible request in %s.\n",
-                                __FUNCTION__);
+                                __func__);
                BUG();
        }
 
+       sdev = cmd->device;
+       starget = scsi_target(sdev);
+       shost = sdev->host;
        scsi_init_cmd_errh(cmd);
        cmd->result = DID_NO_CONNECT << 16;
        atomic_inc(&cmd->device->iorequest_cnt);
-       __scsi_done(cmd);
+
+       /*
+        * SCSI request completion path will do scsi_device_unbusy(),
+        * bump busy counts.  To bump the counters, we need to dance
+        * with the locks as normal issue path does.
+        */
+       sdev->device_busy++;
+       spin_unlock(sdev->request_queue->queue_lock);
+       spin_lock(shost->host_lock);
+       shost->host_busy++;
+       starget->target_busy++;
+       spin_unlock(shost->host_lock);
+       spin_lock(sdev->request_queue->queue_lock);
+
+       blk_complete_request(req);
+}
+
+static void scsi_softirq_done(struct request *rq)
+{
+       struct scsi_cmnd *cmd = rq->special;
+       unsigned long wait_for = (cmd->allowed + 1) * rq->timeout;
+       int disposition;
+
+       INIT_LIST_HEAD(&cmd->eh_entry);
+
+       /*
+        * Set the serial numbers back to zero
+        */
+       cmd->serial_number = 0;
+
+       atomic_inc(&cmd->device->iodone_cnt);
+       if (cmd->result)
+               atomic_inc(&cmd->device->ioerr_cnt);
+
+       disposition = scsi_decide_disposition(cmd);
+       if (disposition != SUCCESS &&
+           time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
+               sdev_printk(KERN_ERR, cmd->device,
+                           "timing out command, waited %lus\n",
+                           wait_for/HZ);
+               disposition = SUCCESS;
+       }
+                       
+       scsi_log_completion(cmd, disposition);
+
+       switch (disposition) {
+               case SUCCESS:
+                       scsi_finish_command(cmd);
+                       break;
+               case NEEDS_RETRY:
+                       scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
+                       break;
+               case ADD_TO_MLQUEUE:
+                       scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
+                       break;
+               default:
+                       if (!scsi_eh_scmd_add(cmd, 0))
+                               scsi_finish_command(cmd);
+       }
 }
 
 /*
@@ -1409,7 +1459,7 @@ 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;
        }
@@ -1430,13 +1480,13 @@ 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;
 
                if (unlikely(!scsi_device_online(sdev))) {
-                       printk(KERN_ERR "scsi%d (%d:%d): rejecting I/O to offline device\n",
-                              sdev->host->host_no, sdev->id, sdev->lun);
+                       sdev_printk(KERN_ERR, sdev,
+                                   "rejecting I/O to offline device\n");
                        scsi_kill_request(req, q);
                        continue;
                }
@@ -1446,7 +1496,7 @@ 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);
@@ -1454,20 +1504,35 @@ static void scsi_request_fn(struct request_queue *q)
                if (unlikely(cmd == NULL)) {
                        printk(KERN_CRIT "impossible request in %s.\n"
                                         "please mail a stack trace to "
-                                        "linux-scsi@vger.kernel.org",
-                                        __FUNCTION__);
+                                        "linux-scsi@vger.kernel.org\n",
+                                        __func__);
+                       blk_dump_rq_flags(req, "foo");
                        BUG();
                }
                spin_lock(shost->host_lock);
 
-               if (!scsi_host_queue_ready(q, shost, sdev))
+               /*
+                * We hit this when the driver is using a host wide
+                * tag map. For device level tag maps the queue_depth check
+                * in the device ready fn would prevent us from trying
+                * to allocate a tag. Since the map is a shared host resource
+                * we add the dev to the starved list so it eventually gets
+                * a run when a tag is freed.
+                */
+               if (blk_queue_tagged(q) && !blk_rq_tagged(req)) {
+                       if (list_empty(&sdev->starved_entry))
+                               list_add_tail(&sdev->starved_entry,
+                                             &shost->starved_list);
                        goto not_ready;
-               if (sdev->single_lun) {
-                       if (scsi_target(sdev)->starget_sdev_user &&
-                           scsi_target(sdev)->starget_sdev_user != sdev)
-                               goto not_ready;
-                       scsi_target(sdev)->starget_sdev_user = sdev;
                }
+
+               if (!scsi_target_queue_ready(shost, sdev))
+                       goto not_ready;
+
+               if (!scsi_host_queue_ready(q, shost, sdev))
+                       goto not_ready;
+
+               scsi_target(sdev)->target_busy++;
                shost->host_busy++;
 
                /*
@@ -1546,37 +1611,56 @@ u64 scsi_calculate_bounce_limit(struct Scsi_Host *shost)
 }
 EXPORT_SYMBOL(scsi_calculate_bounce_limit);
 
-struct request_queue *scsi_alloc_queue(struct scsi_device *sdev)
+struct request_queue *__scsi_alloc_queue(struct Scsi_Host *shost,
+                                        request_fn_proc *request_fn)
 {
-       struct Scsi_Host *shost = sdev->host;
        struct request_queue *q;
+       struct device *dev = shost->shost_gendev.parent;
 
-       q = blk_init_queue(scsi_request_fn, NULL);
+       q = blk_init_queue(request_fn, NULL);
        if (!q)
                return NULL;
 
-       blk_queue_prep_rq(q, scsi_prep_fn);
-
+       /*
+        * this limit is imposed by hardware restrictions
+        */
        blk_queue_max_hw_segments(q, shost->sg_tablesize);
-       blk_queue_max_phys_segments(q, SCSI_MAX_PHYS_SEGMENTS);
+       blk_queue_max_phys_segments(q, SCSI_MAX_SG_CHAIN_SEGMENTS);
+
        blk_queue_max_sectors(q, shost->max_sectors);
        blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
        blk_queue_segment_boundary(q, shost->dma_boundary);
-       blk_queue_issue_flush_fn(q, scsi_issue_flush_fn);
+       dma_set_seg_boundary(dev, shost->dma_boundary);
+
+       blk_queue_max_segment_size(q, dma_get_max_seg_size(dev));
+
+       /* New queue, no concurrency on queue_flags */
+       if (!shost->use_clustering)
+               queue_flag_clear_unlocked(QUEUE_FLAG_CLUSTER, q);
 
        /*
-        * ordered tags are superior to flush ordering
+        * set a reasonable default alignment on word boundaries: the
+        * host and device may alter it using
+        * blk_queue_update_dma_alignment() later.
         */
-       if (shost->ordered_tag)
-               blk_queue_ordered(q, QUEUE_ORDERED_TAG);
-       else if (shost->ordered_flush) {
-               blk_queue_ordered(q, QUEUE_ORDERED_FLUSH);
-               q->prepare_flush_fn = scsi_prepare_flush_fn;
-               q->end_flush_fn = scsi_end_flush_fn;
-       }
+       blk_queue_dma_alignment(q, 0x03);
 
-       if (!shost->use_clustering)
-               clear_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags);
+       return q;
+}
+EXPORT_SYMBOL(__scsi_alloc_queue);
+
+struct request_queue *scsi_alloc_queue(struct scsi_device *sdev)
+{
+       struct request_queue *q;
+
+       q = __scsi_alloc_queue(sdev->host, scsi_request_fn);
+       if (!q)
+               return NULL;
+
+       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;
 }
 
@@ -1638,42 +1722,142 @@ int __init scsi_init_queue(void)
 {
        int i;
 
+       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");
+               return -ENOMEM;
+       }
+
        for (i = 0; i < SG_MEMPOOL_NR; i++) {
                struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
                int size = sgp->size * sizeof(struct scatterlist);
 
                sgp->slab = kmem_cache_create(sgp->name, size, 0,
-                               SLAB_HWCACHE_ALIGN, NULL, NULL);
+                               SLAB_HWCACHE_ALIGN, NULL);
                if (!sgp->slab) {
                        printk(KERN_ERR "SCSI: can't init sg slab %s\n",
                                        sgp->name);
+                       goto cleanup_sdb;
                }
 
-               sgp->pool = mempool_create(SG_MEMPOOL_SIZE,
-                               mempool_alloc_slab, mempool_free_slab,
-                               sgp->slab);
+               sgp->pool = mempool_create_slab_pool(SG_MEMPOOL_SIZE,
+                                                    sgp->slab);
                if (!sgp->pool) {
                        printk(KERN_ERR "SCSI: can't init sg mempool %s\n",
                                        sgp->name);
+                       goto cleanup_sdb;
                }
        }
 
        return 0;
+
+cleanup_sdb:
+       for (i = 0; i < SG_MEMPOOL_NR; i++) {
+               struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
+               if (sgp->pool)
+                       mempool_destroy(sgp->pool);
+               if (sgp->slab)
+                       kmem_cache_destroy(sgp->slab);
+       }
+       kmem_cache_destroy(scsi_sdb_cache);
+
+       return -ENOMEM;
 }
 
 void scsi_exit_queue(void)
 {
        int i;
 
+       kmem_cache_destroy(scsi_sdb_cache);
+
        for (i = 0; i < SG_MEMPOOL_NR; i++) {
                struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
                mempool_destroy(sgp->pool);
                kmem_cache_destroy(sgp->slab);
        }
 }
+
+/**
+ *     scsi_mode_select - issue a mode select
+ *     @sdev:  SCSI device to be queried
+ *     @pf:    Page format bit (1 == standard, 0 == vendor specific)
+ *     @sp:    Save page bit (0 == don't save, 1 == save)
+ *     @modepage: mode page being requested
+ *     @buffer: request buffer (may not be smaller than eight bytes)
+ *     @len:   length of request buffer.
+ *     @timeout: command timeout
+ *     @retries: number of retries before failing
+ *     @data: returns a structure abstracting the mode header data
+ *     @sshdr: place to put sense data (or NULL if no sense to be collected).
+ *             must be SCSI_SENSE_BUFFERSIZE big.
+ *
+ *     Returns zero if successful; negative error number or scsi
+ *     status on error
+ *
+ */
+int
+scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
+                unsigned char *buffer, int len, int timeout, int retries,
+                struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
+{
+       unsigned char cmd[10];
+       unsigned char *real_buffer;
+       int ret;
+
+       memset(cmd, 0, sizeof(cmd));
+       cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
+
+       if (sdev->use_10_for_ms) {
+               if (len > 65535)
+                       return -EINVAL;
+               real_buffer = kmalloc(8 + len, GFP_KERNEL);
+               if (!real_buffer)
+                       return -ENOMEM;
+               memcpy(real_buffer + 8, buffer, len);
+               len += 8;
+               real_buffer[0] = 0;
+               real_buffer[1] = 0;
+               real_buffer[2] = data->medium_type;
+               real_buffer[3] = data->device_specific;
+               real_buffer[4] = data->longlba ? 0x01 : 0;
+               real_buffer[5] = 0;
+               real_buffer[6] = data->block_descriptor_length >> 8;
+               real_buffer[7] = data->block_descriptor_length;
+
+               cmd[0] = MODE_SELECT_10;
+               cmd[7] = len >> 8;
+               cmd[8] = len;
+       } else {
+               if (len > 255 || data->block_descriptor_length > 255 ||
+                   data->longlba)
+                       return -EINVAL;
+
+               real_buffer = kmalloc(4 + len, GFP_KERNEL);
+               if (!real_buffer)
+                       return -ENOMEM;
+               memcpy(real_buffer + 4, buffer, len);
+               len += 4;
+               real_buffer[0] = 0;
+               real_buffer[1] = data->medium_type;
+               real_buffer[2] = data->device_specific;
+               real_buffer[3] = data->block_descriptor_length;
+               
+
+               cmd[0] = MODE_SELECT;
+               cmd[4] = len;
+       }
+
+       ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
+                              sshdr, timeout, retries, NULL);
+       kfree(real_buffer);
+       return ret;
+}
+EXPORT_SYMBOL_GPL(scsi_mode_select);
+
 /**
- *     scsi_mode_sense - issue a mode sense, falling back from 10 to 
- *             six bytes if necessary.
+ *     scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
  *     @sdev:  SCSI device to be queried
  *     @dbd:   set if mode sense will allow block descriptors to be returned
  *     @modepage: mode page being requested
@@ -1682,17 +1866,18 @@ void scsi_exit_queue(void)
  *     @timeout: command timeout
  *     @retries: number of retries before failing
  *     @data: returns a structure abstracting the mode header data
- *     @sense: place to put sense data (or NULL if no sense to be collected).
+ *     @sshdr: place to put sense data (or NULL if no sense to be collected).
  *             must be SCSI_SENSE_BUFFERSIZE big.
  *
  *     Returns zero if unsuccessful, or the header offset (either 4
  *     or 8 depending on whether a six or ten byte command was
  *     issued) if successful.
- **/
+ */
 int
 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
                  unsigned char *buffer, int len, int timeout, int retries,
-                 struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr) {
+                 struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
+{
        unsigned char cmd[12];
        int use_10_for_ms;
        int header_length;
@@ -1730,7 +1915,7 @@ 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
@@ -1752,8 +1937,16 @@ scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
        }
 
        if(scsi_status_is_good(result)) {
-               data->header_length = header_length;
-               if(use_10_for_ms) {
+               if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
+                            (modepage == 6 || modepage == 8))) {
+                       /* Initio breakage? */
+                       header_length = 0;
+                       data->length = 13;
+                       data->medium_type = 0;
+                       data->device_specific = 0;
+                       data->longlba = 0;
+                       data->block_descriptor_length = 0;
+               } else if(use_10_for_ms) {
                        data->length = buffer[0]*256 + buffer[1] + 2;
                        data->medium_type = buffer[2];
                        data->device_specific = buffer[3];
@@ -1766,46 +1959,75 @@ scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
                        data->device_specific = buffer[2];
                        data->block_descriptor_length = buffer[3];
                }
+               data->header_length = header_length;
        }
 
        return result;
 }
 EXPORT_SYMBOL(scsi_mode_sense);
 
+/**
+ *     scsi_test_unit_ready - test if unit is ready
+ *     @sdev:  scsi device to change the state of.
+ *     @timeout: command timeout
+ *     @retries: number of retries before failing
+ *     @sshdr_external: Optional pointer to struct scsi_sense_hdr for
+ *             returning sense. Make sure that this is cleared before passing
+ *             in.
+ *
+ *     Returns zero if unsuccessful or an error if TUR failed.  For
+ *     removable media, a return of NOT_READY or UNIT_ATTENTION is
+ *     translated to success, with the ->changed flag updated.
+ **/
 int
-scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries)
+scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
+                    struct scsi_sense_hdr *sshdr_external)
 {
        char cmd[] = {
                TEST_UNIT_READY, 0, 0, 0, 0, 0,
        };
-       struct scsi_sense_hdr sshdr;
+       struct scsi_sense_hdr *sshdr;
        int result;
-       
-       result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, &sshdr,
-                                 timeout, retries);
-
-       if ((driver_byte(result) & DRIVER_SENSE) && sdev->removable) {
 
-               if ((scsi_sense_valid(&sshdr)) &&
-                   ((sshdr.sense_key == UNIT_ATTENTION) ||
-                    (sshdr.sense_key == NOT_READY))) {
+       if (!sshdr_external)
+               sshdr = kzalloc(sizeof(*sshdr), GFP_KERNEL);
+       else
+               sshdr = sshdr_external;
+
+       /* 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, NULL);
+               if (sdev->removable && scsi_sense_valid(sshdr) &&
+                   sshdr->sense_key == UNIT_ATTENTION)
                        sdev->changed = 1;
-                       result = 0;
-               }
+       } 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 (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);
        return result;
 }
 EXPORT_SYMBOL(scsi_test_unit_ready);
 
 /**
- *     scsi_device_set_state - Take the given device through the device
- *             state model.
+ *     scsi_device_set_state - Take the given device through the device state model.
  *     @sdev:  scsi device to change the state of.
  *     @state: state to change to.
  *
  *     Returns zero if unsuccessful or an error if the requested 
  *     transition is illegal.
- **/
+ */
 int
 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
 {
@@ -1816,10 +2038,13 @@ scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
 
        switch (state) {
        case SDEV_CREATED:
-               /* There are no legal states that come back to
-                * created.  This is the manually initialised start
-                * state */
-               goto illegal;
+               switch (oldstate) {
+               case SDEV_CREATED_BLOCK:
+                       break;
+               default:
+                       goto illegal;
+               }
+               break;
                        
        case SDEV_RUNNING:
                switch (oldstate) {
@@ -1857,8 +2082,17 @@ scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
 
        case SDEV_BLOCK:
                switch (oldstate) {
-               case SDEV_CREATED:
                case SDEV_RUNNING:
+               case SDEV_CREATED_BLOCK:
+                       break;
+               default:
+                       goto illegal;
+               }
+               break;
+
+       case SDEV_CREATED_BLOCK:
+               switch (oldstate) {
+               case SDEV_CREATED:
                        break;
                default:
                        goto illegal;
@@ -1869,6 +2103,7 @@ scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
                switch (oldstate) {
                case SDEV_CREATED:
                case SDEV_RUNNING:
+               case SDEV_QUIESCE:
                case SDEV_OFFLINE:
                case SDEV_BLOCK:
                        break;
@@ -1879,6 +2114,9 @@ scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
 
        case SDEV_DEL:
                switch (oldstate) {
+               case SDEV_CREATED:
+               case SDEV_RUNNING:
+               case SDEV_OFFLINE:
                case SDEV_CANCEL:
                        break;
                default:
@@ -1892,16 +2130,157 @@ scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
 
  illegal:
        SCSI_LOG_ERROR_RECOVERY(1, 
-                               dev_printk(KERN_ERR, &sdev->sdev_gendev,
-                                          "Illegal state transition %s->%s\n",
-                                          scsi_device_state_name(oldstate),
-                                          scsi_device_state_name(state))
+                               sdev_printk(KERN_ERR, sdev,
+                                           "Illegal state transition %s->%s\n",
+                                           scsi_device_state_name(oldstate),
+                                           scsi_device_state_name(state))
                                );
        return -EINVAL;
 }
 EXPORT_SYMBOL(scsi_device_set_state);
 
 /**
+ *     sdev_evt_emit - emit a single SCSI device uevent
+ *     @sdev: associated SCSI device
+ *     @evt: event to emit
+ *
+ *     Send a single uevent (scsi_event) to the associated scsi_device.
+ */
+static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
+{
+       int idx = 0;
+       char *envp[3];
+
+       switch (evt->evt_type) {
+       case SDEV_EVT_MEDIA_CHANGE:
+               envp[idx++] = "SDEV_MEDIA_CHANGE=1";
+               break;
+
+       default:
+               /* do nothing */
+               break;
+       }
+
+       envp[idx++] = NULL;
+
+       kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
+}
+
+/**
+ *     sdev_evt_thread - send a uevent for each scsi event
+ *     @work: work struct for scsi_device
+ *
+ *     Dispatch queued events to their associated scsi_device kobjects
+ *     as uevents.
+ */
+void scsi_evt_thread(struct work_struct *work)
+{
+       struct scsi_device *sdev;
+       LIST_HEAD(event_list);
+
+       sdev = container_of(work, struct scsi_device, event_work);
+
+       while (1) {
+               struct scsi_event *evt;
+               struct list_head *this, *tmp;
+               unsigned long flags;
+
+               spin_lock_irqsave(&sdev->list_lock, flags);
+               list_splice_init(&sdev->event_list, &event_list);
+               spin_unlock_irqrestore(&sdev->list_lock, flags);
+
+               if (list_empty(&event_list))
+                       break;
+
+               list_for_each_safe(this, tmp, &event_list) {
+                       evt = list_entry(this, struct scsi_event, node);
+                       list_del(&evt->node);
+                       scsi_evt_emit(sdev, evt);
+                       kfree(evt);
+               }
+       }
+}
+
+/**
+ *     sdev_evt_send - send asserted event to uevent thread
+ *     @sdev: scsi_device event occurred on
+ *     @evt: event to send
+ *
+ *     Assert scsi device event asynchronously.
+ */
+void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
+{
+       unsigned long flags;
+
+#if 0
+       /* FIXME: currently this check eliminates all media change events
+        * for polled devices.  Need to update to discriminate between AN
+        * and polled events */
+       if (!test_bit(evt->evt_type, sdev->supported_events)) {
+               kfree(evt);
+               return;
+       }
+#endif
+
+       spin_lock_irqsave(&sdev->list_lock, flags);
+       list_add_tail(&evt->node, &sdev->event_list);
+       schedule_work(&sdev->event_work);
+       spin_unlock_irqrestore(&sdev->list_lock, flags);
+}
+EXPORT_SYMBOL_GPL(sdev_evt_send);
+
+/**
+ *     sdev_evt_alloc - allocate a new scsi event
+ *     @evt_type: type of event to allocate
+ *     @gfpflags: GFP flags for allocation
+ *
+ *     Allocates and returns a new scsi_event.
+ */
+struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
+                                 gfp_t gfpflags)
+{
+       struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
+       if (!evt)
+               return NULL;
+
+       evt->evt_type = evt_type;
+       INIT_LIST_HEAD(&evt->node);
+
+       /* evt_type-specific initialization, if any */
+       switch (evt_type) {
+       case SDEV_EVT_MEDIA_CHANGE:
+       default:
+               /* do nothing */
+               break;
+       }
+
+       return evt;
+}
+EXPORT_SYMBOL_GPL(sdev_evt_alloc);
+
+/**
+ *     sdev_evt_send_simple - send asserted event to uevent thread
+ *     @sdev: scsi_device event occurred on
+ *     @evt_type: type of event to send
+ *     @gfpflags: GFP flags for allocation
+ *
+ *     Assert scsi device event asynchronously, given an event type.
+ */
+void sdev_evt_send_simple(struct scsi_device *sdev,
+                         enum scsi_device_event evt_type, gfp_t gfpflags)
+{
+       struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
+       if (!evt) {
+               sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
+                           evt_type);
+               return;
+       }
+
+       sdev_evt_send(sdev, evt);
+}
+EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
+
+/**
  *     scsi_device_quiesce - Block user issued commands.
  *     @sdev:  scsi device to quiesce.
  *
@@ -1915,7 +2294,7 @@ EXPORT_SYMBOL(scsi_device_set_state);
  *     Must be called with user context, may sleep.
  *
  *     Returns zero if unsuccessful or an error if not.
- **/
+ */
 int
 scsi_device_quiesce(struct scsi_device *sdev)
 {
@@ -1940,7 +2319,7 @@ EXPORT_SYMBOL(scsi_device_quiesce);
  *     queues.
  *
  *     Must be called with user context, may sleep.
- **/
+ */
 void
 scsi_device_resume(struct scsi_device *sdev)
 {
@@ -1977,8 +2356,7 @@ scsi_target_resume(struct scsi_target *starget)
 EXPORT_SYMBOL(scsi_target_resume);
 
 /**
- * scsi_internal_device_block - internal function to put a device
- *                             temporarily into the SDEV_BLOCK state
+ * scsi_internal_device_block - internal function to put a device temporarily into the SDEV_BLOCK state
  * @sdev:      device to block
  *
  * Block request made by scsi lld's to temporarily stop all
@@ -1993,17 +2371,21 @@ EXPORT_SYMBOL(scsi_target_resume);
  *     state, all commands are deferred until the scsi lld reenables
  *     the device with scsi_device_unblock or device_block_tmo fires.
  *     This routine assumes the host_lock is held on entry.
- **/
+ */
 int
 scsi_internal_device_block(struct scsi_device *sdev)
 {
-       request_queue_t *q = sdev->request_queue;
+       struct request_queue *q = sdev->request_queue;
        unsigned long flags;
        int err = 0;
 
        err = scsi_device_set_state(sdev, SDEV_BLOCK);
-       if (err)
-               return err;
+       if (err) {
+               err = scsi_device_set_state(sdev, SDEV_CREATED_BLOCK);
+
+               if (err)
+                       return err;
+       }
 
        /* 
         * The device has transitioned to SDEV_BLOCK.  Stop the
@@ -2033,21 +2415,23 @@ EXPORT_SYMBOL_GPL(scsi_internal_device_block);
  *     (which must be a legal transition) allowing the midlayer to
  *     goose the queue for this device.  This routine assumes the 
  *     host_lock is held upon entry.
- **/
+ */
 int
 scsi_internal_device_unblock(struct scsi_device *sdev)
 {
-       request_queue_t *q = sdev->request_queue; 
-       int err;
+       struct request_queue *q = sdev->request_queue; 
        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)
-               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);
@@ -2108,3 +2492,63 @@ scsi_target_unblock(struct device *dev)
                device_for_each_child(dev, NULL, target_unblock);
 }
 EXPORT_SYMBOL_GPL(scsi_target_unblock);
+
+/**
+ * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
+ * @sgl:       scatter-gather list
+ * @sg_count:  number of segments in sg
+ * @offset:    offset in bytes into sg, on return offset into the mapped area
+ * @len:       bytes to map, on return number of bytes mapped
+ *
+ * Returns virtual address of the start of the mapped page
+ */
+void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
+                         size_t *offset, size_t *len)
+{
+       int i;
+       size_t sg_len = 0, len_complete = 0;
+       struct scatterlist *sg;
+       struct page *page;
+
+       WARN_ON(!irqs_disabled());
+
+       for_each_sg(sgl, sg, sg_count, i) {
+               len_complete = sg_len; /* Complete sg-entries */
+               sg_len += sg->length;
+               if (sg_len > *offset)
+                       break;
+       }
+
+       if (unlikely(i == sg_count)) {
+               printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
+                       "elements %d\n",
+                      __func__, sg_len, *offset, sg_count);
+               WARN_ON(1);
+               return NULL;
+       }
+
+       /* Offset starting from the beginning of first page in this sg-entry */
+       *offset = *offset - len_complete + sg->offset;
+
+       /* Assumption: contiguous pages can be accessed as "page + i" */
+       page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
+       *offset &= ~PAGE_MASK;
+
+       /* Bytes in this sg-entry from *offset to the end of the page */
+       sg_len = PAGE_SIZE - *offset;
+       if (*len > sg_len)
+               *len = sg_len;
+
+       return kmap_atomic(page, KM_BIO_SRC_IRQ);
+}
+EXPORT_SYMBOL(scsi_kmap_atomic_sg);
+
+/**
+ * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
+ * @virt:      virtual address to be unmapped
+ */
+void scsi_kunmap_atomic_sg(void *virt)
+{
+       kunmap_atomic(virt, KM_BIO_SRC_IRQ);
+}
+EXPORT_SYMBOL(scsi_kunmap_atomic_sg);