]> nv-tegra.nvidia Code Review - linux-2.6.git/blobdiff - drivers/scsi/scsi_lib.c
[SCSI] don't build scsi_dma_{map,unmap} for !HAS_DMA
[linux-2.6.git] / drivers / scsi / scsi_lib.c
index 863bb6495daaeb6bd6fcdb1a00c7721507e6fb8d..1f5a07bf2a7535e439300b4daa51f898983973ef 100644 (file)
 #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_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;
+       struct kmem_cache       *slab;
        mempool_t       *pool;
 };
 
@@ -63,41 +64,7 @@ static struct scsi_host_sg_pool scsi_sg_pools[] = {
 };     
 #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;
-}
-
 static void scsi_run_queue(struct request_queue *q);
-static void scsi_release_buffers(struct scsi_cmnd *cmd);
 
 /*
  * Function:   scsi_unprep_request()
@@ -115,10 +82,9 @@ 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_release_buffers(cmd);
        scsi_put_command(cmd);
 }
 
@@ -140,14 +106,12 @@ static void scsi_unprep_request(struct request *req)
  *              commands.
  * Notes:       This could be called either from an interrupt context or a
  *              normal process context.
- * Notes:      Upon return, cmd is a stale pointer.
  */
 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;
-       struct request *req = cmd->request;
        unsigned long flags;
 
        SCSI_LOG_MLQUEUE(1,
@@ -188,9 +152,8 @@ int scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
         * function.  The SCSI request function detects the blocked condition
         * and plugs the queue appropriately.
          */
-       scsi_unprep_request(req);
        spin_lock_irqsave(q->queue_lock, flags);
-       blk_requeue_request(q, req);
+       blk_requeue_request(q, cmd->request);
        spin_unlock_irqrestore(q->queue_lock, flags);
 
        scsi_run_queue(q);
@@ -198,116 +161,6 @@ int scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
        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)
-{
-       /*
-        * 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)
-{
-       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);
-}
-
-EXPORT_SYMBOL(scsi_wait_req);
-
 /**
  * scsi_execute - insert request and wait for the result
  * @sdev:      scsi device
@@ -320,7 +173,7 @@ EXPORT_SYMBOL(scsi_wait_req);
  * @retries:   number of times to retry request
  * @flags:     or into request flags;
  *
- * 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,
@@ -341,8 +194,10 @@ 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
@@ -366,13 +221,12 @@ int scsi_execute_req(struct scsi_device *sdev, const unsigned char *cmd,
        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);
        if (sshdr)
                scsi_normalize_sense(sense, SCSI_SENSE_BUFFERSIZE, sshdr);
 
@@ -381,6 +235,200 @@ int scsi_execute_req(struct scsi_device *sdev, const unsigned char *cmd,
 }
 EXPORT_SYMBOL(scsi_execute_req);
 
+struct scsi_io_context {
+       void *data;
+       void (*done)(void *data, char *sense, int result, int resid);
+       char sense[SCSI_SENSE_BUFFERSIZE];
+};
+
+static struct kmem_cache *scsi_io_context_cache;
+
+static void scsi_end_async(struct request *req, int uptodate)
+{
+       struct scsi_io_context *sioc = req->end_io_data;
+
+       if (sioc->done)
+               sioc->done(sioc->data, sioc->sense, req->errors, req->data_len);
+
+       kmem_cache_free(scsi_io_context_cache, sioc);
+       __blk_put_request(req->q, req);
+}
+
+static int scsi_merge_bio(struct request *rq, struct bio *bio)
+{
+       struct request_queue *q = rq->q;
+
+       bio->bi_flags &= ~(1 << BIO_SEG_VALID);
+       if (rq_data_dir(rq) == WRITE)
+               bio->bi_rw |= (1 << BIO_RW);
+       blk_queue_bounce(q, &bio);
+
+       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_zalloc(scsi_io_context_cache, gfp);
+       if (!sioc)
+               return DRIVER_ERROR << 24;
+
+       req = blk_get_request(sdev->request_queue, write, gfp);
+       if (!req)
+               goto free_sense;
+       req->cmd_type = REQ_TYPE_BLOCK_PC;
+       req->cmd_flags |= REQ_QUIET;
+
+       if (use_sg)
+               err = scsi_req_map_sg(req, buffer, use_sg, bufflen, gfp);
+       else if (bufflen)
+               err = blk_rq_map_kern(req->q, req, buffer, bufflen, gfp);
+
+       if (err)
+               goto free_req;
+
+       req->cmd_len = cmd_len;
+       memset(req->cmd, 0, BLK_MAX_CDB); /* ATAPI hates garbage after CDB */
+       memcpy(req->cmd, cmd, req->cmd_len);
+       req->sense = sioc->sense;
+       req->sense_len = 0;
+       req->timeout = timeout;
+       req->retries = retries;
+       req->end_io_data = sioc;
+
+       sioc->data = privdata;
+       sioc->done = done;
+
+       blk_execute_rq_nowait(req->q, NULL, req, 1, scsi_end_async);
+       return 0;
+
+free_req:
+       blk_put_request(req);
+free_sense:
+       kmem_cache_free(scsi_io_context_cache, sioc);
+       return DRIVER_ERROR << 24;
+}
+EXPORT_SYMBOL_GPL(scsi_execute_async);
+
 /*
  * Function:    scsi_init_cmd_errh()
  *
@@ -388,60 +436,16 @@ 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);
-
        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;
 }
 
 void scsi_device_unbusy(struct scsi_device *sdev)
@@ -451,8 +455,8 @@ void scsi_device_unbusy(struct scsi_device *sdev)
 
        spin_lock_irqsave(shost->host_lock, flags);
        shost->host_busy--;
-       if (unlikely((shost->shost_state == SHOST_RECOVERY) &&
-                    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->request_queue->queue_lock);
@@ -547,7 +551,15 @@ 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)))
@@ -596,10 +608,17 @@ 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)
@@ -671,7 +690,7 @@ 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);
 
        /*
@@ -682,7 +701,7 @@ 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;
@@ -723,7 +742,9 @@ static struct scatterlist *scsi_alloc_sgtable(struct scsi_cmnd *cmd, int gfp_mas
        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;
 
@@ -733,6 +754,8 @@ static void scsi_free_sgtable(struct scatterlist *sgl, int index)
        mempool_free(sgl, sgp->pool);
 }
 
+EXPORT_SYMBOL(scsi_free_sgtable);
+
 /*
  * Function:    scsi_release_buffers()
  *
@@ -752,22 +775,13 @@ 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;
 }
@@ -800,11 +814,10 @@ 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;
@@ -812,31 +825,14 @@ void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes,
        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);
-       }
+       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 = result;
                if (result) {
@@ -852,156 +848,132 @@ 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;
+               }
+               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.
-                */
+       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 (scsi_end_request(cmd, 1, good_bytes, result == 0) == 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 (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 
+                               /* Detected disc change.  Set a bit
                                 * and quietly refuse further access.
                                 */
                                cmd->device->changed = 1;
-                               scsi_end_request(cmd, 0,
-                                               this_count, 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.  
-                               */
+                               /* 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;
                        }
                        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)) {
                                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 {
                                scsi_end_request(cmd, 0, this_count, 1);
                                return;
                        }
                        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 */
+                                       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);
                        }
-                       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;
                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);
+                       if (!(req->cmd_flags & REQ_QUIET)) {
+                               scmd_printk(KERN_INFO, cmd,
+                                           "Volume overflow, CDB: ");
+                               __scsi_print_command(cmd->cmnd);
                                scsi_print_sense("", cmd);
                        }
-                       scsi_end_request(cmd, 0, block_bytes, 1);
+                       /* 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) {
-               if (!(req->flags & REQ_QUIET)) {
-                       dev_printk(KERN_INFO, &cmd->device->sdev_gendev,
-                                  "SCSI error: return code = 0x%x\n", result);
-
+               if (!(req->cmd_flags & REQ_QUIET)) {
+                       scsi_print_result(cmd);
                        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;
-               scsi_end_request(cmd, 0, block_bytes, 1);
        }
+       scsi_end_request(cmd, 0, this_count, !result);
 }
 EXPORT_SYMBOL(scsi_io_completion);
 
@@ -1023,48 +995,36 @@ 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))
+       if (unlikely(!sgpnt)) {
+               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");
@@ -1078,38 +1038,6 @@ static int scsi_init_io(struct scsi_cmnd *cmd)
        return BLKPREP_KILL;
 }
 
-static int scsi_prepare_flush_fn(request_queue_t *q, struct request *rq)
-{
-       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;
-
-               if (drv->prepare_flush)
-                       return drv->prepare_flush(q, rq);
-       }
-
-       return 0;
-}
-
-static void scsi_end_flush_fn(request_queue_t *q, struct request *rq)
-{
-       struct scsi_device *sdev = q->queuedata;
-       struct request *flush_rq = rq->end_io_data;
-       struct scsi_driver *drv;
-
-       if (flush_rq->errors) {
-               printk("scsi: barrier error, disabling flush support\n");
-               blk_queue_ordered(q, QUEUE_ORDERED_NONE);
-       }
-
-       if (sdev->sdev_state == SDEV_RUNNING) {
-               drv = *(struct scsi_driver **) rq->rq_disk->private_data;
-               drv->end_flush(q, rq);
-       }
-}
-
 static int scsi_issue_flush_fn(request_queue_t *q, struct gendisk *disk,
                               sector_t *error_sector)
 {
@@ -1126,178 +1054,217 @@ static int scsi_issue_flush_fn(request_queue_t *q, struct gendisk *disk,
        return -EOPNOTSUPP;
 }
 
-static void scsi_generic_done(struct scsi_cmnd *cmd)
+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 {
+               cmd = req->special;
+       }
+
+       /* 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));
-       scsi_io_completion(cmd, cmd->result == 0 ? cmd->bufflen : 0, 0);
+       /*
+        * 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
+        */
+       scsi_io_completion(cmd, cmd->request_bufflen);
 }
 
-static int scsi_prep_fn(struct request_queue *q, struct request *req)
+static int scsi_setup_blk_pc_cmnd(struct scsi_device *sdev, struct request *req)
 {
-       struct scsi_device *sdev = q->queuedata;
        struct scsi_cmnd *cmd;
-       int specials_only = 0;
+
+       cmd = scsi_get_cmd_from_req(sdev, req);
+       if (unlikely(!cmd))
+               return BLKPREP_DEFER;
 
        /*
-        * 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
+        * 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 (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;
-               }
-               /* OK, we only allow special commands (i.e. not
-                * user initiated ones */
-               specials_only = sdev->sdev_state;
+       if (req->bio) {
+               int ret;
+
+               BUG_ON(!req->nr_phys_segments);
+
+               ret = scsi_init_io(cmd);
+               if (unlikely(ret))
+                       return ret;
+       } 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;
+
        /*
-        * 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.
+        * Filesystem requests must transfer data.
         */
-       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)) {
+       BUG_ON(!req->nr_phys_segments);
 
-               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;
-               }
-                       
-                       
-               /*
-                * 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;
-       } else {
-               blk_dump_rq_flags(req, "SCSI bad req");
-               goto kill;
+       cmd = scsi_get_cmd_from_req(sdev, req);
+       if (unlikely(!cmd))
+               return BLKPREP_DEFER;
+
+       ret = scsi_init_io(cmd);
+       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;
        }
-       
-       /* 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;
-       
+
+       return BLKPREP_OK;
+}
+
+static int scsi_prep_fn(struct request_queue *q, struct request *req)
+{
+       struct scsi_device *sdev = q->queuedata;
+       int ret = BLKPREP_OK;
+
        /*
-        * 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 the device is not in running state we will reject some
+        * or all commands.
         */
-       if (req->flags & (REQ_CMD | REQ_BLOCK_PC)) {
-               struct scsi_driver *drv;
-               int ret;
+       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:
                /*
-                * This will do a couple of things:
-                *  1) Fill in the actual SCSI command.
-                *  2) Fill in any other upper-level specific fields
-                * (timeout).
+                * All other command types are not supported.
                 *
-                * 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.
+                * 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;
+       }
 
-               /* 
-                * 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;
-               }
-               
+ out:
+       switch (ret) {
+       case BLKPREP_KILL:
+               req->errors = DID_NO_CONNECT << 16;
+               break;
+       case BLKPREP_DEFER:
                /*
-                * Initialize the actual SCSI command for this request.
+                * 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 (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));
-                       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;
-               }
+               if (sdev->device_busy == 0)
+                       blk_plug_device(q);
+               break;
+       default:
+               req->cmd_flags |= REQ_DONTPREP;
        }
 
-       /*
-        * The request is now prepped, no need to come back here
-        */
-       req->flags |= REQ_DONTPREP;
-       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;
- kill:
-       req->errors = DID_NO_CONNECT << 16;
-       return BLKPREP_KILL;
+       return ret;
 }
 
 /*
@@ -1317,9 +1284,8 @@ 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;
@@ -1342,7 +1308,7 @@ static inline int scsi_host_queue_ready(struct request_queue *q,
                                   struct Scsi_Host *shost,
                                   struct scsi_device *sdev)
 {
-       if (shost->shost_state == SHOST_RECOVERY)
+       if (scsi_host_in_recovery(shost))
                return 0;
        if (shost->host_busy == 0 && shost->host_blocked) {
                /*
@@ -1377,6 +1343,8 @@ static inline int scsi_host_queue_ready(struct request_queue *q,
 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);
 
@@ -1389,9 +1357,57 @@ static void scsi_kill_request(struct request *req, request_queue_t *q)
        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_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);
+       }
+}
+
 /*
  * Function:    scsi_request_fn()
  *
@@ -1438,8 +1454,8 @@ 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);
+                       sdev_printk(KERN_ERR, sdev,
+                                   "rejecting I/O to offline device\n");
                        scsi_kill_request(req, q);
                        continue;
                }
@@ -1457,8 +1473,9 @@ 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",
+                                        "linux-scsi@vger.kernel.org\n",
                                         __FUNCTION__);
+                       blk_dump_rq_flags(req, "foo");
                        BUG();
                }
                spin_lock(shost->host_lock);
@@ -1514,7 +1531,6 @@ static void scsi_request_fn(struct request_queue *q)
         * cases (host limits or settings) should run the queue at some
         * later time.
         */
-       scsi_unprep_request(req);
        spin_lock_irq(q->queue_lock);
        blk_requeue_request(q, req);
        sdev->device_busy--;
@@ -1550,39 +1566,40 @@ 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;
 
-       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);
-
        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);
-       blk_queue_issue_flush_fn(q, scsi_issue_flush_fn);
-
-       /*
-        * ordered tags are superior to flush ordering
-        */
-       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;
-       }
 
        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)
 {
@@ -1642,6 +1659,14 @@ 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);
@@ -1653,9 +1678,8 @@ 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);
@@ -1669,12 +1693,92 @@ 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 
  *             six bytes if necessary.
@@ -1696,7 +1800,8 @@ void scsi_exit_queue(void)
 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;
@@ -1756,8 +1861,16 @@ 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];
@@ -1770,6 +1883,7 @@ 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;
@@ -1873,6 +1987,7 @@ 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;
@@ -1883,6 +1998,9 @@ 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:
@@ -1896,10 +2014,10 @@ 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;
 }
@@ -2112,3 +2230,63 @@ 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
+ * @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;
+
+       WARN_ON(!irqs_disabled());
+
+       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);