fedora core 6 1.2949 + vserver 2.2.0
[linux-2.6.git] / drivers / scsi / scsi_lib.c
index e30d3f2..f02f48a 100644 (file)
 #include <linux/slab.h>
 #include <linux/init.h>
 #include <linux/pci.h>
+#include <linux/delay.h>
+#include <linux/hardirq.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_NR          ARRAY_SIZE(scsi_sg_pools)
 #define SG_MEMPOOL_SIZE                32
 
 struct scsi_host_sg_pool {
        size_t          size;
        char            *name; 
-       kmem_cache_t    *slab;
+       struct kmem_cache       *slab;
        mempool_t       *pool;
 };
 
@@ -43,7 +45,7 @@ struct scsi_host_sg_pool {
 #endif
 
 #define SP(x) { x, "sgpool-" #x } 
-struct scsi_host_sg_pool scsi_sg_pools[] = { 
+static struct scsi_host_sg_pool scsi_sg_pools[] = {
        SP(8),
        SP(16),
        SP(32),
@@ -62,37 +64,28 @@ struct scsi_host_sg_pool scsi_sg_pools[] = {
 };     
 #undef SP
 
+static void scsi_run_queue(struct request_queue *q);
 
 /*
- * Function:    scsi_insert_special_req()
+ * Function:   scsi_unprep_request()
  *
- * Purpose:     Insert pre-formed request into request queue.
+ * Purpose:    Remove all preparation done for a request, including its
+ *             associated scsi_cmnd, so that it can be requeued.
  *
- * 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.
+ * Arguments:  req     - request to unprepare
  *
- * Lock status: Assumed that lock is not held upon entry.
- *
- * Returns:     Nothing
+ * Lock status:        Assumed that no locks are held upon entry.
  *
- * 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.
+ * Returns:    Nothing.
  */
-int scsi_insert_special_req(struct scsi_request *sreq, int at_head)
+static void scsi_unprep_request(struct request *req)
 {
-       /*
-        * 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, 0);
-       return 0;
+       struct scsi_cmnd *cmd = req->special;
+
+       req->cmd_flags &= ~REQ_DONTPREP;
+       req->special = NULL;
+
+       scsi_put_command(cmd);
 }
 
 /*
@@ -118,18 +111,14 @@ int scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
 {
        struct Scsi_Host *host = cmd->device->host;
        struct scsi_device *device = cmd->device;
+       struct request_queue *q = device->request_queue;
+       unsigned long flags;
 
        SCSI_LOG_MLQUEUE(1,
                 printk("Inserting command %p into mlqueue\n", cmd));
 
        /*
-        * We are inserting the command into the ml queue.  First, we
-        * cancel the timer, so it doesn't time out.
-        */
-       scsi_delete_timer(cmd);
-
-       /*
-        * Next, set the appropriate busy bit for the device/host.
+        * Set the appropriate busy bit for the device/host.
         *
         * If the host/device isn't busy, assume that something actually
         * completed, and that we should be able to queue a command now.
@@ -146,12 +135,6 @@ int scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
        else if (reason == SCSI_MLQUEUE_DEVICE_BUSY)
                device->device_blocked = device->max_device_blocked;
 
-       /*
-        * Register the fact that we own the thing for now.
-        */
-       cmd->state = SCSI_STATE_MLQUEUE;
-       cmd->owner = SCSI_OWNER_MIDLEVEL;
-
        /*
         * Decrement the counters, since these commands are no longer
         * active on the host/device.
@@ -159,113 +142,293 @@ int scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
        scsi_device_unbusy(device);
 
        /*
-        * Insert this command at the head of the queue for it's device.
-        * It will go before all other commands that are already in the queue.
+        * Requeue this command.  It will go before all other commands
+        * that are already in the queue.
         *
         * NOTE: there is magic here about the way the queue is plugged if
         * we have no outstanding commands.
         * 
-        * Although this *doesn't* plug the queue, it does call the request
+        * Although we *don't* plug the queue, we call the request
         * function.  The SCSI request function detects the blocked condition
         * and plugs the queue appropriately.
-        */
-       blk_insert_request(device->request_queue, cmd->request, 1, cmd, 1);
+         */
+       spin_lock_irqsave(q->queue_lock, flags);
+       blk_requeue_request(q, cmd->request);
+       spin_unlock_irqrestore(q->queue_lock, flags);
+
+       scsi_run_queue(q);
+
        return 0;
 }
 
-/*
- * Function:    scsi_do_req
- *
- * Purpose:     Queue a SCSI request
- *
- * 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.
- *
- * Lock status: No locks 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.
- */
-void scsi_do_req(struct scsi_request *sreq, const void *cmnd,
-                void *buffer, unsigned bufflen,
-                void (*done)(struct scsi_cmnd *),
-                int timeout, int retries)
+/**
+ * scsi_execute - insert request and wait for the result
+ * @sdev:      scsi device
+ * @cmd:       scsi command
+ * @data_direction: data direction
+ * @buffer:    data buffer
+ * @bufflen:   len of buffer
+ * @sense:     optional sense buffer
+ * @timeout:   request timeout in seconds
+ * @retries:   number of times to retry request
+ * @flags:     or into request flags;
+ *
+ * returns the req->errors value which is the 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)
 {
-       /*
-        * 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);
+       struct request *req;
+       int write = (data_direction == DMA_TO_DEVICE);
+       int ret = DRIVER_ERROR << 24;
 
-       /*
-        * 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;
+       req = blk_get_request(sdev->request_queue, write, __GFP_WAIT);
+
+       if (bufflen &&  blk_rq_map_kern(sdev->request_queue, req,
+                                       buffer, bufflen, __GFP_WAIT))
+               goto out;
 
-       if (sreq->sr_cmd_len == 0)
-               sreq->sr_cmd_len = COMMAND_SIZE(sreq->sr_cmnd[0]);
+       req->cmd_len = COMMAND_SIZE(cmd[0]);
+       memcpy(req->cmd, cmd, req->cmd_len);
+       req->sense = sense;
+       req->sense_len = 0;
+       req->retries = retries;
+       req->timeout = timeout;
+       req->cmd_type = REQ_TYPE_BLOCK_PC;
+       req->cmd_flags |= flags | REQ_QUIET | REQ_PREEMPT;
 
        /*
         * head injection *required* here otherwise quiesce won't work
         */
-       scsi_insert_special_req(sreq, 1);
+       blk_execute_rq(req->q, NULL, req, 1);
+
+       ret = req->errors;
+ out:
+       blk_put_request(req);
+
+       return ret;
 }
-static void scsi_wait_done(struct scsi_cmnd *cmd)
+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 request *req = cmd->request;
-       struct request_queue *q = cmd->device->request_queue;
-       unsigned long flags;
+       char *sense = NULL;
+       int result;
+       
+       if (sshdr) {
+               sense = kzalloc(SCSI_SENSE_BUFFERSIZE, GFP_NOIO);
+               if (!sense)
+                       return DRIVER_ERROR << 24;
+       }
+       result = scsi_execute(sdev, cmd, data_direction, buffer, bufflen,
+                             sense, timeout, retries, 0);
+       if (sshdr)
+               scsi_normalize_sense(sense, SCSI_SENSE_BUFFERSIZE, sshdr);
 
-       req->rq_status = RQ_SCSI_DONE;  /* Busy, but indicate request done */
+       kfree(sense);
+       return result;
+}
+EXPORT_SYMBOL(scsi_execute_req);
 
-       spin_lock_irqsave(q->queue_lock, flags);
-       if (blk_rq_tagged(req))
-               blk_queue_end_tag(q, req);
-       spin_unlock_irqrestore(q->queue_lock, flags);
+struct scsi_io_context {
+       void *data;
+       void (*done)(void *data, char *sense, int result, int resid);
+       char sense[SCSI_SENSE_BUFFERSIZE];
+};
+
+static struct kmem_cache *scsi_io_context_cache;
+
+static void scsi_end_async(struct request *req, int uptodate)
+{
+       struct scsi_io_context *sioc = req->end_io_data;
 
-       if (req->waiting)
-               complete(req->waiting);
+       if (sioc->done)
+               sioc->done(sioc->data, sioc->sense, req->errors, req->data_len);
+
+       kmem_cache_free(scsi_io_context_cache, sioc);
+       __blk_put_request(req->q, req);
 }
 
-void scsi_wait_req(struct scsi_request *sreq, const void *cmnd, void *buffer,
-                  unsigned bufflen, int timeout, int retries)
+static int scsi_merge_bio(struct request *rq, struct bio *bio)
 {
-       DECLARE_COMPLETION(wait);
-       
-       sreq->sr_request->waiting = &wait;
-       sreq->sr_request->rq_status = RQ_SCSI_BUSY;
-       scsi_do_req(sreq, cmnd, buffer, bufflen, scsi_wait_done,
-                       timeout, retries);
-       wait_for_completion(&wait);
-       sreq->sr_request->waiting = NULL;
-       if (sreq->sr_request->rq_status != RQ_SCSI_DONE)
-               sreq->sr_result |= (DRIVER_ERROR << 24);
-
-       __scsi_release_request(sreq);
+       struct request_queue *q = rq->q;
+
+       bio->bi_flags &= ~(1 << BIO_SEG_VALID);
+       if (rq_data_dir(rq) == WRITE)
+               bio->bi_rw |= (1 << BIO_RW);
+       blk_queue_bounce(q, &bio);
+
+       if (!rq->bio)
+               blk_rq_bio_prep(q, rq, bio);
+       else if (!ll_back_merge_fn(q, rq, bio))
+               return -EINVAL;
+       else {
+               rq->biotail->bi_next = bio;
+               rq->biotail = bio;
+       }
+
+       return 0;
+}
+
+static int scsi_bi_endio(struct bio *bio, unsigned int bytes_done, int error)
+{
+       if (bio->bi_size)
+               return 1;
+
+       bio_put(bio);
+       return 0;
+}
+
+/**
+ * scsi_req_map_sg - map a scatterlist into a request
+ * @rq:                request to fill
+ * @sg:                scatterlist
+ * @nsegs:     number of elements
+ * @bufflen:   len of buffer
+ * @gfp:       memory allocation flags
+ *
+ * scsi_req_map_sg maps a scatterlist into a request so that the
+ * request can be sent to the block layer. We do not trust the scatterlist
+ * sent to use, as some ULDs use that struct to only organize the pages.
+ */
+static int scsi_req_map_sg(struct request *rq, struct scatterlist *sgl,
+                          int nsegs, unsigned bufflen, gfp_t gfp)
+{
+       struct request_queue *q = rq->q;
+       int nr_pages = (bufflen + sgl[0].offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
+       unsigned int data_len = 0, len, bytes, off;
+       struct page *page;
+       struct bio *bio = NULL;
+       int i, err, nr_vecs = 0;
+
+       for (i = 0; i < nsegs; i++) {
+               page = sgl[i].page;
+               off = sgl[i].offset;
+               len = sgl[i].length;
+               data_len += len;
+
+               while (len > 0) {
+                       bytes = min_t(unsigned int, len, PAGE_SIZE - off);
+
+                       if (!bio) {
+                               nr_vecs = min_t(int, BIO_MAX_PAGES, nr_pages);
+                               nr_pages -= nr_vecs;
+
+                               bio = bio_alloc(gfp, nr_vecs);
+                               if (!bio) {
+                                       err = -ENOMEM;
+                                       goto free_bios;
+                               }
+                               bio->bi_end_io = scsi_bi_endio;
+                       }
+
+                       if (bio_add_pc_page(q, bio, page, bytes, off) !=
+                           bytes) {
+                               bio_put(bio);
+                               err = -EINVAL;
+                               goto free_bios;
+                       }
+
+                       if (bio->bi_vcnt >= nr_vecs) {
+                               err = scsi_merge_bio(rq, bio);
+                               if (err) {
+                                       bio_endio(bio, bio->bi_size, 0);
+                                       goto free_bios;
+                               }
+                               bio = NULL;
+                       }
+
+                       page++;
+                       len -= bytes;
+                       off = 0;
+               }
+       }
+
+       rq->buffer = rq->data = NULL;
+       rq->data_len = data_len;
+       return 0;
+
+free_bios:
+       while ((bio = rq->bio) != NULL) {
+               rq->bio = bio->bi_next;
+               /*
+                * call endio instead of bio_put incase it was bounced
+                */
+               bio_endio(bio, bio->bi_size, 0);
+       }
+
+       return err;
+}
+
+/**
+ * scsi_execute_async - insert request
+ * @sdev:      scsi device
+ * @cmd:       scsi command
+ * @cmd_len:   length of scsi cdb
+ * @data_direction: data direction
+ * @buffer:    data buffer (this can be a kernel buffer or scatterlist)
+ * @bufflen:   len of buffer
+ * @use_sg:    if buffer is a scatterlist this is the number of elements
+ * @timeout:   request timeout in seconds
+ * @retries:   number of times to retry request
+ * @flags:     or into request flags
+ **/
+int scsi_execute_async(struct scsi_device *sdev, const unsigned char *cmd,
+                      int cmd_len, int data_direction, void *buffer, unsigned bufflen,
+                      int use_sg, int timeout, int retries, void *privdata,
+                      void (*done)(void *, char *, int, int), gfp_t gfp)
+{
+       struct request *req;
+       struct scsi_io_context *sioc;
+       int err = 0;
+       int write = (data_direction == DMA_TO_DEVICE);
+
+       sioc = kmem_cache_alloc(scsi_io_context_cache, gfp);
+       if (!sioc)
+               return DRIVER_ERROR << 24;
+       memset(sioc, 0, sizeof(*sioc));
+
+       req = blk_get_request(sdev->request_queue, write, gfp);
+       if (!req)
+               goto free_sense;
+       req->cmd_type = REQ_TYPE_BLOCK_PC;
+       req->cmd_flags |= REQ_QUIET;
+
+       if (use_sg)
+               err = scsi_req_map_sg(req, buffer, use_sg, bufflen, gfp);
+       else if (bufflen)
+               err = blk_rq_map_kern(req->q, req, buffer, bufflen, gfp);
+
+       if (err)
+               goto free_req;
+
+       req->cmd_len = cmd_len;
+       memset(req->cmd, 0, BLK_MAX_CDB); /* ATAPI hates garbage after CDB */
+       memcpy(req->cmd, cmd, req->cmd_len);
+       req->sense = sioc->sense;
+       req->sense_len = 0;
+       req->timeout = timeout;
+       req->retries = retries;
+       req->end_io_data = sioc;
+
+       sioc->data = privdata;
+       sioc->done = done;
+
+       blk_execute_rq_nowait(req->q, NULL, req, 1, scsi_end_async);
+       return 0;
+
+free_req:
+       blk_put_request(req);
+free_sense:
+       kmem_cache_free(scsi_io_context_cache, sioc);
+       return DRIVER_ERROR << 24;
 }
+EXPORT_SYMBOL_GPL(scsi_execute_async);
 
 /*
  * Function:    scsi_init_cmd_errh()
@@ -274,65 +437,16 @@ void scsi_wait_req(struct scsi_request *sreq, const void *cmnd, void *buffer,
  *
  * 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->owner = SCSI_OWNER_MIDLEVEL;
        cmd->serial_number = 0;
-       cmd->serial_number_at_timeout = 0;
-       cmd->abort_reason = 0;
-
        memset(cmd->sense_buffer, 0, sizeof cmd->sense_buffer);
-
        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;
-       cmd->internal_timeout = NORMAL_TIMEOUT;
-       cmd->abort_reason = 0;
-
-       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;
 }
 
 void scsi_device_unbusy(struct scsi_device *sdev)
@@ -342,13 +456,13 @@ void scsi_device_unbusy(struct scsi_device *sdev)
 
        spin_lock_irqsave(shost->host_lock, flags);
        shost->host_busy--;
-       if (unlikely(test_bit(SHOST_RECOVERY, &shost->shost_state) &&
-                    shost->host_failed))
+       if (unlikely(scsi_host_in_recovery(shost) &&
+                    (shost->host_failed || shost->host_eh_scheduled)))
                scsi_eh_wakeup(shost);
        spin_unlock(shost->host_lock);
-       spin_lock(&sdev->sdev_lock);
+       spin_lock(sdev->request_queue->queue_lock);
        sdev->device_busy--;
-       spin_unlock_irqrestore(&sdev->sdev_lock, flags);
+       spin_unlock_irqrestore(sdev->request_queue->queue_lock, flags);
 }
 
 /*
@@ -362,10 +476,11 @@ static void scsi_single_lun_run(struct scsi_device *current_sdev)
 {
        struct Scsi_Host *shost = current_sdev->host;
        struct scsi_device *sdev, *tmp;
+       struct scsi_target *starget = scsi_target(current_sdev);
        unsigned long flags;
 
        spin_lock_irqsave(shost->host_lock, flags);
-       current_sdev->sdev_target->starget_sdev_user = NULL;
+       starget->starget_sdev_user = NULL;
        spin_unlock_irqrestore(shost->host_lock, flags);
 
        /*
@@ -377,10 +492,12 @@ static void scsi_single_lun_run(struct scsi_device *current_sdev)
        blk_run_queue(current_sdev->request_queue);
 
        spin_lock_irqsave(shost->host_lock, flags);
-       if (current_sdev->sdev_target->starget_sdev_user)
+       if (starget->starget_sdev_user)
                goto out;
-       list_for_each_entry_safe(sdev, tmp, &current_sdev->same_target_siblings,
+       list_for_each_entry_safe(sdev, tmp, &starget->devices,
                        same_target_siblings) {
+               if (sdev == current_sdev)
+                       continue;
                if (scsi_device_get(sdev))
                        continue;
 
@@ -435,7 +552,15 @@ static void scsi_run_queue(struct request_queue *q)
                list_del_init(&sdev->starved_entry);
                spin_unlock_irqrestore(shost->host_lock, flags);
 
-               blk_run_queue(sdev->request_queue);
+
+               if (test_bit(QUEUE_FLAG_REENTER, &q->queue_flags) &&
+                   !test_and_set_bit(QUEUE_FLAG_REENTER,
+                                     &sdev->request_queue->queue_flags)) {
+                       blk_run_queue(sdev->request_queue);
+                       clear_bit(QUEUE_FLAG_REENTER,
+                                 &sdev->request_queue->queue_flags);
+               } else
+                       blk_run_queue(sdev->request_queue);
 
                spin_lock_irqsave(shost->host_lock, flags);
                if (unlikely(!list_empty(&sdev->starved_entry)))
@@ -467,21 +592,34 @@ static void scsi_run_queue(struct request_queue *q)
  *             I/O errors in the middle of the request, in which case
  *             we need to request the blocks that come after the bad
  *             sector.
+ * Notes:      Upon return, cmd is a stale pointer.
  */
 static void scsi_requeue_command(struct request_queue *q, struct scsi_cmnd *cmd)
 {
-       cmd->request->flags &= ~REQ_DONTPREP;
-       blk_insert_request(q, cmd->request, 1, cmd, 1);
+       struct request *req = cmd->request;
+       unsigned long flags;
+
+       scsi_unprep_request(req);
+       spin_lock_irqsave(q->queue_lock, flags);
+       blk_requeue_request(q, req);
+       spin_unlock_irqrestore(q->queue_lock, flags);
 
        scsi_run_queue(q);
 }
 
 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)
@@ -495,25 +633,24 @@ void scsi_run_host_queues(struct Scsi_Host *shost)
 /*
  * Function:    scsi_end_request()
  *
- * Purpose:     Post-processing of completed commands called from interrupt
- *              handler or a bottom-half handler.
+ * Purpose:     Post-processing of completed commands (usually invoked at end
+ *             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.
- *              sectors  - number of sectors we want to mark.
+ *              uptodate - 1 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.
- *             frequeue - indicates that if we release the command block
- *                        that the queue request function should be called.
  *
  * Lock status: Assumed that lock is not held upon entry.
  *
- * Returns:     Nothing
+ * Returns:     cmd if requeue required, NULL otherwise.
  *
  * Notes:       This is called for block device requests in order to
  *              mark some number of sectors as complete.
  * 
  *             We are guaranteeing that the request queue will be goosed
  *             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,
                                          int bytes, int requeue)
@@ -536,14 +673,15 @@ static struct scsi_cmnd *scsi_end_request(struct scsi_cmnd *cmd, int uptodate,
                if (!uptodate && blk_noretry_request(req))
                        end_that_request_chunk(req, 0, leftover);
                else {
-                       if (requeue)
+                       if (requeue) {
                                /*
                                 * Bleah.  Leftovers again.  Stick the
                                 * leftovers in the front of the
                                 * queue, and goose the queue again.
                                 */
                                scsi_requeue_command(q, cmd);
-
+                               cmd = NULL;
+                       }
                        return cmd;
                }
        }
@@ -553,7 +691,7 @@ static struct scsi_cmnd *scsi_end_request(struct scsi_cmnd *cmd, int uptodate,
        spin_lock_irqsave(q->queue_lock, flags);
        if (blk_rq_tagged(req))
                blk_queue_end_tag(q, req);
-       end_that_request_last(req);
+       end_that_request_last(req, uptodate);
        spin_unlock_irqrestore(q->queue_lock, flags);
 
        /*
@@ -564,7 +702,7 @@ static struct scsi_cmnd *scsi_end_request(struct scsi_cmnd *cmd, int uptodate,
        return NULL;
 }
 
-static struct scatterlist *scsi_alloc_sgtable(struct scsi_cmnd *cmd, int gfp_mask)
+struct scatterlist *scsi_alloc_sgtable(struct scsi_cmnd *cmd, gfp_t gfp_mask)
 {
        struct scsi_host_sg_pool *sgp;
        struct scatterlist *sgl;
@@ -602,21 +740,23 @@ static struct scatterlist *scsi_alloc_sgtable(struct scsi_cmnd *cmd, int gfp_mas
 
        sgp = scsi_sg_pools + cmd->sglist_len;
        sgl = mempool_alloc(sgp->pool, gfp_mask);
-       if (sgl)
-               memset(sgl, 0, sgp->size);
        return sgl;
 }
 
-static void scsi_free_sgtable(struct scatterlist *sgl, int index)
+EXPORT_SYMBOL(scsi_alloc_sgtable);
+
+void scsi_free_sgtable(struct scatterlist *sgl, int index)
 {
        struct scsi_host_sg_pool *sgp;
 
-       BUG_ON(index > SG_MEMPOOL_NR);
+       BUG_ON(index >= SG_MEMPOOL_NR);
 
        sgp = scsi_sg_pools + index;
        mempool_free(sgl, sgp->pool);
 }
 
+EXPORT_SYMBOL(scsi_free_sgtable);
+
 /*
  * Function:    scsi_release_buffers()
  *
@@ -636,22 +776,13 @@ static void scsi_free_sgtable(struct scatterlist *sgl, int index)
  */
 static 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;
 }
@@ -684,38 +815,33 @@ static void scsi_release_buffers(struct scsi_cmnd *cmd)
  *             b) We can just use scsi_requeue_command() here.  This would
  *                be used if we just wanted to retry, for example.
  */
-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;
+       int this_count = cmd->request_bufflen;
        request_queue_t *q = cmd->device->request_queue;
        struct request *req = cmd->request;
        int clear_errors = 1;
+       struct scsi_sense_hdr sshdr;
+       int sense_valid = 0;
+       int sense_deferred = 0;
 
-       /*
-        * 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);
+       scsi_release_buffers(cmd);
+
+       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 = (driver_byte(result) & DRIVER_SENSE) ?
-                             (CHECK_CONDITION << 1) : (result & 0xff);
+               req->errors = result;
                if (result) {
                        clear_errors = 0;
-                       if (cmd->sense_buffer[0] & 0x70) {
+                       if (sense_valid && req->sense) {
+                               /*
+                                * SG_IO wants current and deferred errors
+                                */
                                int len = 8 + cmd->sense_buffer[7];
 
                                if (len > SCSI_SENSE_BUFFERSIZE)
@@ -724,160 +850,135 @@ void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes,
                                req->sense_len = len;
                        }
                } else
-                       req->data_len -= cmd->bufflen;
+                       req->data_len = cmd->resid;
        }
 
-       /*
-        * 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;
-
        /*
         * 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));
+       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.
-                */
-               cmd = scsi_end_request(cmd, 1, good_bytes, result == 0);
+       if (clear_errors)
+               req->errors = 0;
 
-               /*
-                * If the command completed without error, then either finish off the
-                * rest of the command, or start a new one.
-                */
-               if (result == 0 || cmd == NULL ) {
-                       return;
-               }
-       }
-       /*
-        * 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 (driver_byte(result) != 0) {
-               if ((cmd->sense_buffer[0] & 0x7f) == 0x70) {
-                       /*
-                        * If the device is in the process of becoming ready,
-                        * retry.
-                        */
-                       if (cmd->sense_buffer[12] == 0x04 &&
-                           cmd->sense_buffer[13] == 0x01) {
+       if (scsi_end_request(cmd, 1, good_bytes, result == 0) == NULL)
+               return;
+
+       /* good_bytes = 0, or (inclusive) there were leftovers and
+        * result = 0, so scsi_end_request couldn't retry.
+        */
+       if (sense_valid && !sense_deferred) {
+               switch (sshdr.sense_key) {
+               case UNIT_ATTENTION:
+                       if (cmd->device->removable) {
+                               /* Detected disc change.  Set a bit
+                                * and quietly refuse further access.
+                                */
+                               cmd->device->changed = 1;
+                               scsi_end_request(cmd, 0, this_count, 1);
+                               return;
+                       } 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;
                        }
-                       if ((cmd->sense_buffer[2] & 0xf) == UNIT_ATTENTION) {
-                               if (cmd->device->removable) {
-                                       /* detected disc change.  set a bit 
-                                        * and quietly refuse further access.
-                                        */
-                                       cmd->device->changed = 1;
-                                       cmd = scsi_end_request(cmd, 0,
-                                                       this_count, 1);
-                                       return;
-                               } 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;
-                               }
-                       }
-               }
-               /*
-                * 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.
-                */
-
-               switch (cmd->sense_buffer[2]) {
+                       break;
                case ILLEGAL_REQUEST:
-                       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)) {
                                cmd->device->use_10_for_rw = 0;
-                               /*
-                                * This will cause a retry with a 6-byte
-                                * command.
+                               /* This will cause a retry with a
+                                * 6-byte command.
                                 */
                                scsi_requeue_command(q, cmd);
-                               result = 0;
+                               return;
                        } else {
-                               cmd = scsi_end_request(cmd, 0, this_count, 1);
+                               scsi_end_request(cmd, 0, this_count, 1);
                                return;
                        }
                        break;
                case NOT_READY:
-                       printk(KERN_INFO "Device %s not ready.\n",
-                              req->rq_disk ? req->rq_disk->disk_name : "");
-                       cmd = scsi_end_request(cmd, 0, this_count, 1);
+                       /* If the device is in the process of becoming
+                        * ready, or has a temporary blockage, retry.
+                        */
+                       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 */
+                                       scsi_requeue_command(q, cmd);
+                                       return;
+                               default:
+                                       break;
+                               }
+                       }
+                       if (!(req->cmd_flags & REQ_QUIET)) {
+                               scmd_printk(KERN_INFO, cmd,
+                                           "Device not ready: ");
+                               scsi_print_sense_hdr("", &sshdr);
+                       }
+                       scsi_end_request(cmd, 0, this_count, 1);
                        return;
-                       break;
-               case MEDIUM_ERROR:
                case VOLUME_OVERFLOW:
-                       printk("scsi%d: ERROR on channel %d, id %d, lun %d, CDB: ",
-                              cmd->device->host->host_no, (int) cmd->device->channel,
-                              (int) cmd->device->id, (int) cmd->device->lun);
-                       __scsi_print_command(cmd->data_cmnd);
-                       scsi_print_sense("", cmd);
-                       cmd = scsi_end_request(cmd, 0, block_bytes, 1);
+                       if (!(req->cmd_flags & REQ_QUIET)) {
+                               scmd_printk(KERN_INFO, cmd,
+                                           "Volume overflow, CDB: ");
+                               __scsi_print_command(cmd->cmnd);
+                               scsi_print_sense("", cmd);
+                       }
+                       /* See SSC3rXX or current. */
+                       scsi_end_request(cmd, 0, this_count, 1);
                        return;
                default:
                        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.  
+               /* Third party bus reset or reset for error recovery
+                * reasons.  Just retry the request and see what
+                * happens.
                 */
                scsi_requeue_command(q, cmd);
                return;
        }
        if (result) {
-               printk("SCSI error : <%d %d %d %d> return code = 0x%x\n",
-                      cmd->device->host->host_no,
-                      cmd->device->channel,
-                      cmd->device->id,
-                      cmd->device->lun, result);
-
-               if (driver_byte(result) & DRIVER_SENSE)
-                       scsi_print_sense("", 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.
-                */
-               block_bytes = req->hard_cur_sectors << 9;
-               if (!block_bytes)
-                       block_bytes = req->data_len;
-               cmd = scsi_end_request(cmd, 0, block_bytes, 1);
+               if (!(req->cmd_flags & REQ_QUIET)) {
+                       scmd_printk(KERN_INFO, cmd,
+                                   "SCSI error: return code = 0x%08x\n",
+                                   result);
+                       if (driver_byte(result) & DRIVER_SENSE)
+                               scsi_print_sense("", cmd);
+               }
        }
+       scsi_end_request(cmd, 0, this_count, !result);
 }
+EXPORT_SYMBOL(scsi_io_completion);
 
 /*
  * Function:    scsi_init_io()
@@ -897,50 +998,36 @@ static int scsi_init_io(struct scsi_cmnd *cmd)
        int                count;
 
        /*
-        * 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;
-       }
-
-       /*
-        * we used to not use scatter-gather for single segment request,
+        * 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 sg table allocation fails, requeue request later.
+        * If sg table allocation fails, requeue request later.
         */
        sgpnt = scsi_alloc_sgtable(cmd, GFP_ATOMIC);
        if (unlikely(!sgpnt)) {
-               req->flags |= REQ_SPECIAL;
+               scsi_unprep_request(req);
                return BLKPREP_DEFER;
        }
 
+       req->buffer = NULL;
        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;
+       else
+               cmd->request_bufflen = req->nr_sectors << 9;
 
        /* 
         * 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);
-
-       /*
-        * mapped well, send it off
-        */
        if (likely(count <= cmd->use_sg)) {
                cmd->use_sg = count;
-               return 0;
+               return BLKPREP_OK;
        }
 
        printk(KERN_ERR "Incorrect number of segments after building list\n");
@@ -954,148 +1041,233 @@ static int scsi_init_io(struct scsi_cmnd *cmd)
        return BLKPREP_KILL;
 }
 
-static int scsi_prep_fn(struct request_queue *q, struct request *req)
+static int scsi_issue_flush_fn(request_queue_t *q, struct gendisk *disk,
+                              sector_t *error_sector)
 {
        struct scsi_device *sdev = q->queuedata;
-       struct scsi_cmnd *cmd;
-       int specials_only = 0;
+       struct scsi_driver *drv;
 
-       /*
-        * 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 (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);
-               return BLKPREP_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);
-                       return BLKPREP_KILL;
-               }
-               /* OK, we only allow special commands (i.e. not
-                * user initiated ones */
-               specials_only = sdev->sdev_state;
-       }
+       if (sdev->sdev_state != SDEV_RUNNING)
+               return -ENXIO;
 
-       /*
-        * 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) {
-               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)) {
+       drv = *(struct scsi_driver **) disk->private_data;
+       if (drv->issue_flush)
+               return drv->issue_flush(&sdev->sdev_gendev, error_sector);
 
-               if(unlikely(specials_only)) {
-                       if(specials_only == SDEV_QUIESCE)
-                               return BLKPREP_DEFER;
-                       
-                       printk(KERN_ERR "scsi%d (%d:%d): rejecting I/O to device being removed\n",
-                              sdev->host->host_no, sdev->id, sdev->lun);
-                       return BLKPREP_KILL;
-               }
-                       
-                       
-               /*
-                * Now try and find a command block that we can use.
-                */
-               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;
+       return -EOPNOTSUPP;
+}
+
+static struct scsi_cmnd *scsi_get_cmd_from_req(struct scsi_device *sdev,
+               struct request *req)
+{
+       struct scsi_cmnd *cmd;
+
+       if (!req->special) {
+               cmd = scsi_get_command(sdev, GFP_ATOMIC);
+               if (unlikely(!cmd))
+                       return NULL;
+               req->special = cmd;
        } else {
-               blk_dump_rq_flags(req, "SCSI bad req");
-               return BLKPREP_KILL;
+               cmd = req->special;
        }
-       
-       /* 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;
+
+       /* pull a tag out of the request if we have one */
+       cmd->tag = req->tag;
        cmd->request = req;
-       
+
+       return cmd;
+}
+
+static void scsi_blk_pc_done(struct scsi_cmnd *cmd)
+{
+       BUG_ON(!blk_pc_request(cmd->request));
        /*
-        * 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.
+        * This will complete the whole command with uptodate=1 so
+        * as far as the block layer is concerned the command completed
+        * successfully. Since this is a REQ_BLOCK_PC command the
+        * caller should check the request's errors value
         */
-       if (req->flags & (REQ_CMD | REQ_BLOCK_PC)) {
-               struct scsi_driver *drv;
-               int ret;
+       scsi_io_completion(cmd, cmd->request_bufflen);
+}
 
-               /*
-                * 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.
-                */
+static int scsi_setup_blk_pc_cmnd(struct scsi_device *sdev, struct request *req)
+{
+       struct scsi_cmnd *cmd;
+
+       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);
 
-               /* 
-                * This sets up the scatter-gather table (allocating if
-                * required).
-                */
                ret = scsi_init_io(cmd);
-               if (ret)        /* BLKPREP_KILL return also releases the command */
+               if (unlikely(ret))
                        return ret;
-               
-               /*
-                * Initialize the actual SCSI command for this request.
-                */
-               drv = *(struct scsi_driver **)req->rq_disk->private_data;
-               if (unlikely(!drv->init_command(cmd))) {
-                       scsi_release_buffers(cmd);
-                       scsi_put_command(cmd);
-                       return BLKPREP_KILL;
-               }
+       } else {
+               BUG_ON(req->data_len);
+               BUG_ON(req->data);
+
+               cmd->request_bufflen = 0;
+               cmd->request_buffer = NULL;
+               cmd->use_sg = 0;
+               req->buffer = NULL;
        }
 
+       BUILD_BUG_ON(sizeof(req->cmd) > sizeof(cmd->cmnd));
+       memcpy(cmd->cmnd, req->cmd, sizeof(cmd->cmnd));
+       cmd->cmd_len = req->cmd_len;
+       if (!req->data_len)
+               cmd->sc_data_direction = DMA_NONE;
+       else if (rq_data_dir(req) == WRITE)
+               cmd->sc_data_direction = DMA_TO_DEVICE;
+       else
+               cmd->sc_data_direction = DMA_FROM_DEVICE;
+       
+       cmd->transfersize = req->data_len;
+       cmd->allowed = req->retries;
+       cmd->timeout_per_command = req->timeout;
+       cmd->done = scsi_blk_pc_done;
+       return BLKPREP_OK;
+}
+
+/*
+ * 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.
+ */
+static int scsi_setup_fs_cmnd(struct scsi_device *sdev, struct request *req)
+{
+       struct scsi_cmnd *cmd;
+       struct scsi_driver *drv;
+       int ret;
+
+       /*
+        * Filesystem requests must transfer data.
+        */
+       BUG_ON(!req->nr_phys_segments);
+
+       cmd = scsi_get_cmd_from_req(sdev, req);
+       if (unlikely(!cmd))
+               return BLKPREP_DEFER;
+
+       ret = scsi_init_io(cmd);
+       if (unlikely(ret))
+               return ret;
+
        /*
-        * The request is now prepped, no need to come back here
+        * Initialize the actual SCSI command for this request.
         */
-       req->flags |= REQ_DONTPREP;
+       drv = *(struct scsi_driver **)req->rq_disk->private_data;
+       if (unlikely(!drv->init_command(cmd))) {
+               scsi_release_buffers(cmd);
+               scsi_put_command(cmd);
+               return BLKPREP_KILL;
+       }
+
        return BLKPREP_OK;
+}
 
- 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;
+static int scsi_prep_fn(struct request_queue *q, struct request *req)
+{
+       struct scsi_device *sdev = q->queuedata;
+       int ret = BLKPREP_OK;
+
+       /*
+        * If the device is not in running state we will reject some
+        * or all commands.
+        */
+       if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
+               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:
+                       /*
+                        * 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;
+               }
+
+               if (ret != BLKPREP_OK)
+                       goto out;
+       }
+
+       switch (req->cmd_type) {
+       case REQ_TYPE_BLOCK_PC:
+               ret = scsi_setup_blk_pc_cmnd(sdev, req);
+               break;
+       case REQ_TYPE_FS:
+               ret = scsi_setup_fs_cmnd(sdev, req);
+               break;
+       default:
+               /*
+                * All other command types are not supported.
+                *
+                * Note that these days the SCSI subsystem does not use
+                * REQ_TYPE_SPECIAL requests anymore.  These are only used
+                * (directly or via blk_insert_request) by non-SCSI drivers.
+                */
+               blk_dump_rq_flags(req, "SCSI bad req");
+               ret = BLKPREP_KILL;
+               break;
+       }
+
+ out:
+       switch (ret) {
+       case BLKPREP_KILL:
+               req->errors = DID_NO_CONNECT << 16;
+               break;
+       case BLKPREP_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);
+               break;
+       default:
+               req->cmd_flags |= REQ_DONTPREP;
+       }
+
+       return ret;
 }
 
 /*
@@ -1115,9 +1287,8 @@ static inline int scsi_dev_queue_ready(struct request_queue *q,
                 */
                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;
@@ -1140,7 +1311,7 @@ static inline int scsi_host_queue_ready(struct request_queue *q,
                                   struct Scsi_Host *shost,
                                   struct scsi_device *sdev)
 {
-       if (test_bit(SHOST_RECOVERY, &shost->shost_state))
+       if (scsi_host_in_recovery(shost))
                return 0;
        if (shost->host_busy == 0 && shost->host_blocked) {
                /*
@@ -1169,6 +1340,77 @@ static inline int scsi_host_queue_ready(struct request_queue *q,
        return 1;
 }
 
+/*
+ * Kill a request for a dead device
+ */
+static void scsi_kill_request(struct request *req, request_queue_t *q)
+{
+       struct scsi_cmnd *cmd = req->special;
+       struct scsi_device *sdev = cmd->device;
+       struct Scsi_Host *shost = sdev->host;
+
+       blkdev_dequeue_request(req);
+
+       if (unlikely(cmd == NULL)) {
+               printk(KERN_CRIT "impossible request in %s.\n",
+                                __FUNCTION__);
+               BUG();
+       }
+
+       scsi_init_cmd_errh(cmd);
+       cmd->result = DID_NO_CONNECT << 16;
+       atomic_inc(&cmd->device->iorequest_cnt);
+
+       /*
+        * 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++;
+       spin_unlock(shost->host_lock);
+       spin_lock(sdev->request_queue->queue_lock);
+
+       __scsi_done(cmd);
+}
+
+static void scsi_softirq_done(struct request *rq)
+{
+       struct scsi_cmnd *cmd = rq->completion_data;
+       unsigned long wait_for = (cmd->allowed + 1) * cmd->timeout_per_command;
+       int disposition;
+
+       INIT_LIST_HEAD(&cmd->eh_entry);
+
+       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_retry_command(cmd);
+                       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);
+       }
+}
+
 /*
  * Function:    scsi_request_fn()
  *
@@ -1183,10 +1425,17 @@ static inline int scsi_host_queue_ready(struct request_queue *q,
 static void scsi_request_fn(struct request_queue *q)
 {
        struct scsi_device *sdev = q->queuedata;
-       struct Scsi_Host *shost = sdev->host;
+       struct Scsi_Host *shost;
        struct scsi_cmnd *cmd;
        struct request *req;
 
+       if (!sdev) {
+               printk("scsi: killing requests for dead queue\n");
+               while ((req = elv_next_request(q)) != NULL)
+                       scsi_kill_request(req, q);
+               return;
+       }
+
        if(!get_device(&sdev->sdev_gendev))
                /* We must be tearing the block queue down already */
                return;
@@ -1195,6 +1444,7 @@ static void scsi_request_fn(struct request_queue *q)
         * To start with, we keep looping until the queue is empty, or until
         * the host is no longer able to accept any more requests.
         */
+       shost = sdev->host;
        while (!blk_queue_plugged(q)) {
                int rtn;
                /*
@@ -1207,13 +1457,9 @@ static void scsi_request_fn(struct request_queue *q)
                        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);
-                       blkdev_dequeue_request(req);
-                       req->flags |= REQ_QUIET;
-                       while (end_that_request_first(req, 0, req->nr_sectors))
-                               ;
-                       end_that_request_last(req);
+                       sdev_printk(KERN_ERR, sdev,
+                                   "rejecting I/O to offline device\n");
+                       scsi_kill_request(req, q);
                        continue;
                }
 
@@ -1226,15 +1472,24 @@ static void scsi_request_fn(struct request_queue *q)
                sdev->device_busy++;
 
                spin_unlock(q->queue_lock);
+               cmd = req->special;
+               if (unlikely(cmd == NULL)) {
+                       printk(KERN_CRIT "impossible request in %s.\n"
+                                        "please mail a stack trace to "
+                                        "linux-scsi@vger.kernel.org\n",
+                                        __FUNCTION__);
+                       blk_dump_rq_flags(req, "foo");
+                       BUG();
+               }
                spin_lock(shost->host_lock);
 
                if (!scsi_host_queue_ready(q, shost, sdev))
                        goto not_ready;
                if (sdev->single_lun) {
-                       if (sdev->sdev_target->starget_sdev_user &&
-                           sdev->sdev_target->starget_sdev_user != sdev)
+                       if (scsi_target(sdev)->starget_sdev_user &&
+                           scsi_target(sdev)->starget_sdev_user != sdev)
                                goto not_ready;
-                       sdev->sdev_target->starget_sdev_user = sdev;
+                       scsi_target(sdev)->starget_sdev_user = sdev;
                }
                shost->host_busy++;
 
@@ -1244,15 +1499,6 @@ static void scsi_request_fn(struct request_queue *q)
                 */
                spin_unlock_irq(shost->host_lock);
 
-               cmd = req->special;
-               if (unlikely(cmd == NULL)) {
-                       printk(KERN_CRIT "impossible request in %s.\n"
-                                        "please mail a stack trace to "
-                                        "linux-scsi@vger.kernel.org",
-                                        __FUNCTION__);
-                       BUG();
-               }
-
                /*
                 * Finally, initialize any error handling parameters, and set up
                 * the timers for timeouts.
@@ -1304,42 +1550,59 @@ static void scsi_request_fn(struct request_queue *q)
 u64 scsi_calculate_bounce_limit(struct Scsi_Host *shost)
 {
        struct device *host_dev;
+       u64 bounce_limit = 0xffffffff;
 
        if (shost->unchecked_isa_dma)
                return BLK_BOUNCE_ISA;
-
-       host_dev = scsi_get_device(shost);
-       if (PCI_DMA_BUS_IS_PHYS && host_dev && host_dev->dma_mask)
-               return *host_dev->dma_mask;
-
        /*
         * Platforms with virtual-DMA translation
         * hardware have no practical limit.
         */
-       return BLK_BOUNCE_ANY;
+       if (!PCI_DMA_BUS_IS_PHYS)
+               return BLK_BOUNCE_ANY;
+
+       host_dev = scsi_get_device(shost);
+       if (host_dev && host_dev->dma_mask)
+               bounce_limit = *host_dev->dma_mask;
+
+       return bounce_limit;
 }
+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;
 
-       q = blk_init_queue(scsi_request_fn, &sdev->sdev_lock);
+       q = blk_init_queue(request_fn, NULL);
        if (!q)
                return NULL;
 
-       blk_queue_prep_rq(q, scsi_prep_fn);
-
        blk_queue_max_hw_segments(q, shost->sg_tablesize);
        blk_queue_max_phys_segments(q, SCSI_MAX_PHYS_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);
+
        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_issue_flush_fn(q, scsi_issue_flush_fn);
+       blk_queue_softirq_done(q, scsi_softirq_done);
+       return q;
+}
 
 void scsi_free_queue(struct request_queue *q)
 {
@@ -1366,6 +1629,7 @@ void scsi_block_requests(struct Scsi_Host *shost)
 {
        shost->host_self_blocked = 1;
 }
+EXPORT_SYMBOL(scsi_block_requests);
 
 /*
  * Function:    scsi_unblock_requests()
@@ -1392,11 +1656,20 @@ void scsi_unblock_requests(struct Scsi_Host *shost)
        shost->host_self_blocked = 0;
        scsi_run_host_queues(shost);
 }
+EXPORT_SYMBOL(scsi_unblock_requests);
 
 int __init scsi_init_queue(void)
 {
        int i;
 
+       scsi_io_context_cache = kmem_cache_create("scsi_io_context",
+                                       sizeof(struct scsi_io_context),
+                                       0, 0, NULL, NULL);
+       if (!scsi_io_context_cache) {
+               printk(KERN_ERR "SCSI: can't init scsi io context 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);
@@ -1408,9 +1681,8 @@ int __init scsi_init_queue(void)
                                        sgp->name);
                }
 
-               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);
@@ -1424,16 +1696,96 @@ void scsi_exit_queue(void)
 {
        int i;
 
+       kmem_cache_destroy(scsi_io_context_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
+ *     @sense: 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);
+       kfree(real_buffer);
+       return ret;
+}
+EXPORT_SYMBOL_GPL(scsi_mode_select);
+
 /**
- *     __scsi_mode_sense - issue a mode sense, falling back from 10 to 
+ *     scsi_mode_sense - issue a mode sense, falling back from 10 to 
  *             six bytes if necessary.
- *     @sreq:  SCSI request to fill in with the MODE_SENSE
+ *     @sdev:  SCSI device to be queried
  *     @dbd:   set if mode sense will allow block descriptors to be returned
  *     @modepage: mode page being requested
  *     @buffer: request buffer (may not be smaller than eight bytes)
@@ -1441,26 +1793,35 @@ 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).
+ *             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_request *sreq, int dbd, int modepage,
+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_mode_data *data, struct scsi_sense_hdr *sshdr)
+{
        unsigned char cmd[12];
        int use_10_for_ms;
        int header_length;
+       int result;
+       struct scsi_sense_hdr my_sshdr;
 
        memset(data, 0, sizeof(*data));
        memset(&cmd[0], 0, 12);
        cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
        cmd[2] = modepage;
 
+       /* caller might not be interested in sense, but we need it */
+       if (!sshdr)
+               sshdr = &my_sshdr;
+
  retry:
-       use_10_for_ms = sreq->sr_device->use_10_for_ms;
+       use_10_for_ms = sdev->use_10_for_ms;
 
        if (use_10_for_ms) {
                if (len < 8)
@@ -1478,32 +1839,41 @@ __scsi_mode_sense(struct scsi_request *sreq, int dbd, int modepage,
                header_length = 4;
        }
 
-       sreq->sr_cmd_len = 0;
-       sreq->sr_sense_buffer[0] = 0;
-       sreq->sr_sense_buffer[2] = 0;
-       sreq->sr_data_direction = DMA_FROM_DEVICE;
-
        memset(buffer, 0, len);
 
-       scsi_wait_req(sreq, cmd, buffer, len, timeout, retries);
+       result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
+                                 sshdr, timeout, retries);
 
        /* This code looks awful: what it's doing is making sure an
         * ILLEGAL REQUEST sense return identifies the actual command
         * byte as the problem.  MODE_SENSE commands can return
         * ILLEGAL REQUEST if the code page isn't supported */
-       if (use_10_for_ms && ! scsi_status_is_good(sreq->sr_result) &&
-           (driver_byte(sreq->sr_result) & DRIVER_SENSE) &&
-           sreq->sr_sense_buffer[2] == ILLEGAL_REQUEST &&
-           (sreq->sr_sense_buffer[4] & 0x40) == 0x40 &&
-           sreq->sr_sense_buffer[5] == 0 &&
-           sreq->sr_sense_buffer[6] == 0 ) {
-               sreq->sr_device->use_10_for_ms = 0;
-               goto retry;
+
+       if (use_10_for_ms && !scsi_status_is_good(result) &&
+           (driver_byte(result) & DRIVER_SENSE)) {
+               if (scsi_sense_valid(sshdr)) {
+                       if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
+                           (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
+                               /* 
+                                * Invalid command operation code
+                                */
+                               sdev->use_10_for_ms = 0;
+                               goto retry;
+                       }
+               }
        }
 
-       if(scsi_status_is_good(sreq->sr_result)) {
-               data->header_length = header_length;
-               if(use_10_for_ms) {
+       if(scsi_status_is_good(result)) {
+               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];
@@ -1516,44 +1886,37 @@ __scsi_mode_sense(struct scsi_request *sreq, int dbd, int modepage,
                        data->device_specific = buffer[2];
                        data->block_descriptor_length = buffer[3];
                }
+               data->header_length = header_length;
        }
 
-       return sreq->sr_result;
+       return result;
 }
+EXPORT_SYMBOL(scsi_mode_sense);
 
-/**
- *     scsi_mode_sense - issue a mode sense, falling back from 10 to 
- *             six bytes if necessary.
- *     @sdev:  scsi device to send command to.
- *     @dbd:   set if mode sense will disable block descriptors in the return
- *     @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
- *
- *     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)
+scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries)
 {
-       struct scsi_request *sreq = scsi_allocate_request(sdev, GFP_KERNEL);
-       int ret;
-
-       if (!sreq)
-               return -1;
-
-       ret = __scsi_mode_sense(sreq, dbd, modepage, buffer, len,
-                               timeout, retries, data);
+       char cmd[] = {
+               TEST_UNIT_READY, 0, 0, 0, 0, 0,
+       };
+       struct scsi_sense_hdr sshdr;
+       int result;
+       
+       result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, &sshdr,
+                                 timeout, retries);
 
-       scsi_release_request(sreq);
+       if ((driver_byte(result) & DRIVER_SENSE) && sdev->removable) {
 
-       return ret;
+               if ((scsi_sense_valid(&sshdr)) &&
+                   ((sshdr.sense_key == UNIT_ATTENTION) ||
+                    (sshdr.sense_key == NOT_READY))) {
+                       sdev->changed = 1;
+                       result = 0;
+               }
+       }
+       return result;
 }
+EXPORT_SYMBOL(scsi_test_unit_ready);
 
 /**
  *     scsi_device_set_state - Take the given device through the device
@@ -1584,6 +1947,7 @@ scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
                case SDEV_CREATED:
                case SDEV_OFFLINE:
                case SDEV_QUIESCE:
+               case SDEV_BLOCK:
                        break;
                default:
                        goto illegal;
@@ -1605,6 +1969,17 @@ scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
                case SDEV_CREATED:
                case SDEV_RUNNING:
                case SDEV_QUIESCE:
+               case SDEV_BLOCK:
+                       break;
+               default:
+                       goto illegal;
+               }
+               break;
+
+       case SDEV_BLOCK:
+               switch (oldstate) {
+               case SDEV_CREATED:
+               case SDEV_RUNNING:
                        break;
                default:
                        goto illegal;
@@ -1615,7 +1990,9 @@ 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;
                default:
                        goto illegal;
@@ -1624,6 +2001,9 @@ 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:
@@ -1636,11 +2016,12 @@ scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
        return 0;
 
  illegal:
-       dev_printk(KERN_ERR, &sdev->sdev_gendev,
-                  "Illegal state transition %s->%s\n",
-                  scsi_device_state_name(oldstate),
-                  scsi_device_state_name(state));
-       WARN_ON(1);
+       SCSI_LOG_ERROR_RECOVERY(1, 
+                               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);
@@ -1669,7 +2050,7 @@ scsi_device_quiesce(struct scsi_device *sdev)
 
        scsi_run_queue(sdev->request_queue);
        while (sdev->device_busy) {
-               schedule_timeout(HZ/5);
+               msleep_interruptible(200);
                scsi_run_queue(sdev->request_queue);
        }
        return 0;
@@ -1694,3 +2075,219 @@ scsi_device_resume(struct scsi_device *sdev)
 }
 EXPORT_SYMBOL(scsi_device_resume);
 
+static void
+device_quiesce_fn(struct scsi_device *sdev, void *data)
+{
+       scsi_device_quiesce(sdev);
+}
+
+void
+scsi_target_quiesce(struct scsi_target *starget)
+{
+       starget_for_each_device(starget, NULL, device_quiesce_fn);
+}
+EXPORT_SYMBOL(scsi_target_quiesce);
+
+static void
+device_resume_fn(struct scsi_device *sdev, void *data)
+{
+       scsi_device_resume(sdev);
+}
+
+void
+scsi_target_resume(struct scsi_target *starget)
+{
+       starget_for_each_device(starget, NULL, device_resume_fn);
+}
+EXPORT_SYMBOL(scsi_target_resume);
+
+/**
+ * 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
+ * scsi commands on the specified device.  Called from interrupt
+ * or normal process context.
+ *
+ * Returns zero if successful or error if not
+ *
+ * Notes:       
+ *     This routine transitions the device to the SDEV_BLOCK state
+ *     (which must be a legal transition).  When the device is in this
+ *     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;
+       unsigned long flags;
+       int err = 0;
+
+       err = scsi_device_set_state(sdev, SDEV_BLOCK);
+       if (err)
+               return err;
+
+       /* 
+        * The device has transitioned to SDEV_BLOCK.  Stop the
+        * block layer from calling the midlayer with this device's
+        * request queue. 
+        */
+       spin_lock_irqsave(q->queue_lock, flags);
+       blk_stop_queue(q);
+       spin_unlock_irqrestore(q->queue_lock, flags);
+
+       return 0;
+}
+EXPORT_SYMBOL_GPL(scsi_internal_device_block);
+/**
+ * scsi_internal_device_unblock - resume a device after a block request
+ * @sdev:      device to resume
+ *
+ * Called by scsi lld's or the midlayer to restart the device queue
+ * for the previously suspended scsi device.  Called from interrupt or
+ * normal process context.
+ *
+ * Returns zero if successful or error if not.
+ *
+ * Notes:       
+ *     This routine transitions the device to the SDEV_RUNNING state
+ *     (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;
+       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;
+
+       spin_lock_irqsave(q->queue_lock, flags);
+       blk_start_queue(q);
+       spin_unlock_irqrestore(q->queue_lock, flags);
+
+       return 0;
+}
+EXPORT_SYMBOL_GPL(scsi_internal_device_unblock);
+
+static void
+device_block(struct scsi_device *sdev, void *data)
+{
+       scsi_internal_device_block(sdev);
+}
+
+static int
+target_block(struct device *dev, void *data)
+{
+       if (scsi_is_target_device(dev))
+               starget_for_each_device(to_scsi_target(dev), NULL,
+                                       device_block);
+       return 0;
+}
+
+void
+scsi_target_block(struct device *dev)
+{
+       if (scsi_is_target_device(dev))
+               starget_for_each_device(to_scsi_target(dev), NULL,
+                                       device_block);
+       else
+               device_for_each_child(dev, NULL, target_block);
+}
+EXPORT_SYMBOL_GPL(scsi_target_block);
+
+static void
+device_unblock(struct scsi_device *sdev, void *data)
+{
+       scsi_internal_device_unblock(sdev);
+}
+
+static int
+target_unblock(struct device *dev, void *data)
+{
+       if (scsi_is_target_device(dev))
+               starget_for_each_device(to_scsi_target(dev), NULL,
+                                       device_unblock);
+       return 0;
+}
+
+void
+scsi_target_unblock(struct device *dev)
+{
+       if (scsi_is_target_device(dev))
+               starget_for_each_device(to_scsi_target(dev), NULL,
+                                       device_unblock);
+       else
+               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
+ * @sg:                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 *sg, int sg_count,
+                         size_t *offset, size_t *len)
+{
+       int i;
+       size_t sg_len = 0, len_complete = 0;
+       struct page *page;
+
+       for (i = 0; i < sg_count; i++) {
+               len_complete = sg_len; /* Complete sg-entries */
+               sg_len += sg[i].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",
+                      __FUNCTION__, 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[i].offset;
+
+       /* Assumption: contiguous pages can be accessed as "page + i" */
+       page = nth_page(sg[i].page, (*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);