KVM: Add kvm_read_guest_atomic()
[safe/jmp/linux-2.6] / drivers / scsi / scsi_lib.c
index 9a05076..7c4c889 100644 (file)
 #include <linux/pci.h>
 #include <linux/delay.h>
 #include <linux/hardirq.h>
+#include <linux/scatterlist.h>
 
 #include <scsi/scsi.h>
+#include <scsi/scsi_cmnd.h>
 #include <scsi/scsi_dbg.h>
 #include <scsi/scsi_device.h>
 #include <scsi/scsi_driver.h>
 #include <scsi/scsi_eh.h>
 #include <scsi/scsi_host.h>
-#include <scsi/scsi_request.h>
 
 #include "scsi_priv.h"
 #include "scsi_logging.h"
 
 
-#define SG_MEMPOOL_NR          (sizeof(scsi_sg_pools)/sizeof(struct scsi_host_sg_pool))
-#define SG_MEMPOOL_SIZE                32
+#define SG_MEMPOOL_NR          ARRAY_SIZE(scsi_sg_pools)
+#define SG_MEMPOOL_SIZE                2
+
+/*
+ * The maximum number of SG segments that we will put inside a scatterlist
+ * (unless chaining is used). Should ideally fit inside a single page, to
+ * avoid a higher order allocation.
+ */
+#define SCSI_MAX_SG_SEGMENTS   128
 
 struct scsi_host_sg_pool {
        size_t          size;
-       char            *name; 
-       kmem_cache_t    *slab;
+       char            *name;
+       struct kmem_cache       *slab;
        mempool_t       *pool;
 };
 
-#if (SCSI_MAX_PHYS_SEGMENTS < 32)
-#error SCSI_MAX_PHYS_SEGMENTS is too small
-#endif
-
-#define SP(x) { x, "sgpool-" #x } 
+#define SP(x) { x, "sgpool-" #x }
 static struct scsi_host_sg_pool scsi_sg_pools[] = {
        SP(8),
        SP(16),
+#if (SCSI_MAX_SG_SEGMENTS > 16)
        SP(32),
-#if (SCSI_MAX_PHYS_SEGMENTS > 32)
+#if (SCSI_MAX_SG_SEGMENTS > 32)
        SP(64),
-#if (SCSI_MAX_PHYS_SEGMENTS > 64)
+#if (SCSI_MAX_SG_SEGMENTS > 64)
        SP(128),
-#if (SCSI_MAX_PHYS_SEGMENTS > 128)
-       SP(256),
-#if (SCSI_MAX_PHYS_SEGMENTS > 256)
-#error SCSI_MAX_PHYS_SEGMENTS is too large
-#endif
 #endif
 #endif
 #endif
-};     
+};
 #undef SP
 
 static void scsi_run_queue(struct request_queue *q);
@@ -82,8 +82,8 @@ static void scsi_unprep_request(struct request *req)
 {
        struct scsi_cmnd *cmd = req->special;
 
-       req->flags &= ~REQ_DONTPREP;
-       req->special = (req->flags & REQ_SPECIAL) ? cmd->sc_request : NULL;
+       req->cmd_flags &= ~REQ_DONTPREP;
+       req->special = NULL;
 
        scsi_put_command(cmd);
 }
@@ -161,72 +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
-        *
-        * 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,
-                          1, sreq);
-}
-EXPORT_SYMBOL(scsi_do_req);
-
 /**
  * scsi_execute - insert request and wait for the result
  * @sdev:      scsi device
@@ -239,9 +173,9 @@ EXPORT_SYMBOL(scsi_do_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,
                 int data_direction, void *buffer, unsigned bufflen,
                 unsigned char *sense, int timeout, int retries, int flags)
@@ -262,7 +196,8 @@ int scsi_execute(struct scsi_device *sdev, const unsigned char *cmd,
        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
@@ -306,7 +241,7 @@ struct scsi_io_context {
        char sense[SCSI_SENSE_BUFFERSIZE];
 };
 
-static kmem_cache_t *scsi_io_context_cache;
+static struct kmem_cache *scsi_io_context_cache;
 
 static void scsi_end_async(struct request *req, int uptodate)
 {
@@ -328,33 +263,18 @@ static int scsi_merge_bio(struct request *rq, struct bio *bio)
                bio->bi_rw |= (1 << BIO_RW);
        blk_queue_bounce(q, &bio);
 
-       if (!rq->bio)
-               blk_rq_bio_prep(q, rq, bio);
-       else if (!q->back_merge_fn(q, rq, bio))
-               return -EINVAL;
-       else {
-               rq->biotail->bi_next = bio;
-               rq->biotail = bio;
-               rq->hard_nr_sectors += bio_sectors(bio);
-               rq->nr_sectors = rq->hard_nr_sectors;
-       }
-
-       return 0;
+       return blk_rq_append_bio(q, rq, bio);
 }
 
-static int scsi_bi_endio(struct bio *bio, unsigned int bytes_done, int error)
+static void scsi_bi_endio(struct bio *bio, 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
+ * @sgl:       scatterlist
  * @nsegs:     number of elements
  * @bufflen:   len of buffer
  * @gfp:       memory allocation flags
@@ -367,20 +287,27 @@ 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 + PAGE_SIZE - 1) >> PAGE_SHIFT;
-       unsigned int data_len = 0, len, bytes, off;
+       int nr_pages = (bufflen + sgl[0].offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
+       unsigned int data_len = bufflen, len, bytes, off;
+       struct scatterlist *sg;
        struct page *page;
        struct bio *bio = NULL;
        int i, err, nr_vecs = 0;
 
-       for (i = 0; i < nsegs; i++) {
-               page = sgl[i].page;
-               off = sgl[i].offset;
-               len = sgl[i].length;
-               data_len += len;
+       for_each_sg(sgl, sg, nsegs, i) {
+               page = sg_page(sg);
+               off = sg->offset;
+               len = sg->length;
+               data_len += len;
 
-               while (len > 0) {
+               while (len > 0 && data_len > 0) {
+                       /*
+                        * sg sends a scatterlist that is larger than
+                        * the data_len it wants transferred for certain
+                        * IO sizes
+                        */
                        bytes = min_t(unsigned int, len, PAGE_SIZE - off);
+                       bytes = min(bytes, data_len);
 
                        if (!bio) {
                                nr_vecs = min_t(int, BIO_MAX_PAGES, nr_pages);
@@ -404,7 +331,7 @@ static int scsi_req_map_sg(struct request *rq, struct scatterlist *sgl,
                        if (bio->bi_vcnt >= nr_vecs) {
                                err = scsi_merge_bio(rq, bio);
                                if (err) {
-                                       bio_endio(bio, bio->bi_size, 0);
+                                       bio_endio(bio, 0);
                                        goto free_bios;
                                }
                                bio = NULL;
@@ -412,12 +339,13 @@ static int scsi_req_map_sg(struct request *rq, struct scatterlist *sgl,
 
                        page++;
                        len -= bytes;
+                       data_len -=bytes;
                        off = 0;
                }
        }
 
        rq->buffer = rq->data = NULL;
-       rq->data_len = data_len;
+       rq->data_len = bufflen;
        return 0;
 
 free_bios:
@@ -426,7 +354,7 @@ free_bios:
                /*
                 * call endio instead of bio_put incase it was bounced
                 */
-               bio_endio(bio, bio->bi_size, 0);
+               bio_endio(bio, 0);
        }
 
        return err;
@@ -437,14 +365,16 @@ free_bios:
  * @sdev:      scsi device
  * @cmd:       scsi command
  * @cmd_len:   length of scsi cdb
- * @data_direction: data direction
+ * @data_direction: DMA_TO_DEVICE, DMA_FROM_DEVICE, or DMA_NONE
  * @buffer:    data buffer (this can be a kernel buffer or scatterlist)
  * @bufflen:   len of buffer
  * @use_sg:    if buffer is a scatterlist this is the number of elements
  * @timeout:   request timeout in seconds
  * @retries:   number of times to retry request
- * @flags:     or into request flags
- **/
+ * @privdata:  data passed to done()
+ * @done:      callback function when done
+ * @gfp:       memory allocation flags
+ */
 int scsi_execute_async(struct scsi_device *sdev, const unsigned char *cmd,
                       int cmd_len, int data_direction, void *buffer, unsigned bufflen,
                       int use_sg, int timeout, int retries, void *privdata,
@@ -455,15 +385,15 @@ int scsi_execute_async(struct scsi_device *sdev, const unsigned char *cmd,
        int err = 0;
        int write = (data_direction == DMA_TO_DEVICE);
 
-       sioc = kmem_cache_alloc(scsi_io_context_cache, gfp);
+       sioc = kmem_cache_zalloc(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->flags |= REQ_BLOCK_PC | REQ_QUIET;
+       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);
@@ -474,6 +404,7 @@ int scsi_execute_async(struct scsi_device *sdev, const unsigned char *cmd,
                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;
@@ -490,7 +421,7 @@ int scsi_execute_async(struct scsi_device *sdev, const unsigned char *cmd,
 free_req:
        blk_put_request(req);
 free_sense:
-       kfree(sioc);
+       kmem_cache_free(scsi_io_context_cache, sioc);
        return DRIVER_ERROR << 24;
 }
 EXPORT_SYMBOL_GPL(scsi_execute_async);
@@ -502,60 +433,17 @@ EXPORT_SYMBOL_GPL(scsi_execute_async);
  *
  * 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);
-
+       cmd->resid = 0;
+       memset(cmd->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
        if (cmd->cmd_len == 0)
                cmd->cmd_len = COMMAND_SIZE(cmd->cmnd[0]);
-
-       /*
-        * We need saved copies of a number of fields - this is because
-        * error handling may need to overwrite these with different values
-        * to run different commands, and once error handling is complete,
-        * we will need to restore these values prior to running the actual
-        * command.
-        */
-       cmd->old_use_sg = cmd->use_sg;
-       cmd->old_cmd_len = cmd->cmd_len;
-       cmd->sc_old_data_direction = cmd->sc_data_direction;
-       cmd->old_underflow = cmd->underflow;
-       memcpy(cmd->data_cmnd, cmd->cmnd, sizeof(cmd->cmnd));
-       cmd->buffer = cmd->request_buffer;
-       cmd->bufflen = cmd->request_bufflen;
-
-       return 1;
-}
-
-/*
- * Function:   scsi_setup_cmd_retry()
- *
- * Purpose:    Restore the command state for a retry
- *
- * Arguments:  cmd     - command to be restored
- *
- * Returns:    Nothing
- *
- * Notes:      Immediately prior to retrying a command, we need
- *             to restore certain fields that we saved above.
- */
-void scsi_setup_cmd_retry(struct scsi_cmnd *cmd)
-{
-       memcpy(cmd->cmnd, cmd->data_cmnd, sizeof(cmd->data_cmnd));
-       cmd->request_buffer = cmd->buffer;
-       cmd->request_bufflen = cmd->bufflen;
-       cmd->use_sg = cmd->old_use_sg;
-       cmd->cmd_len = cmd->old_cmd_len;
-       cmd->sc_data_direction = cmd->sc_old_data_direction;
-       cmd->underflow = cmd->old_underflow;
 }
 
 void scsi_device_unbusy(struct scsi_device *sdev)
@@ -566,7 +454,7 @@ void scsi_device_unbusy(struct scsi_device *sdev)
        spin_lock_irqsave(shost->host_lock, flags);
        shost->host_busy--;
        if (unlikely(scsi_host_in_recovery(shost) &&
-                    shost->host_failed))
+                    (shost->host_failed || shost->host_eh_scheduled)))
                scsi_eh_wakeup(shost);
        spin_unlock(shost->host_lock);
        spin_lock(sdev->request_queue->queue_lock);
@@ -638,7 +526,7 @@ static void scsi_run_queue(struct request_queue *q)
        struct Scsi_Host *shost = sdev->host;
        unsigned long flags;
 
-       if (sdev->single_lun)
+       if (scsi_target(sdev)->single_lun)
                scsi_single_lun_run(sdev);
 
        spin_lock_irqsave(shost->host_lock, flags);
@@ -661,7 +549,15 @@ 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)))
@@ -738,7 +634,7 @@ void scsi_run_host_queues(struct Scsi_Host *shost)
  *             of upper level post-processing and scsi_io_completion).
  *
  * Arguments:   cmd     - command that is complete.
- *              uptodate - 1 if I/O indicates success, <= 0 for I/O error.
+ *              error    - 0 if I/O indicates success, < 0 for I/O error.
  *              bytes    - number of bytes of completed I/O
  *             requeue  - indicates whether we should requeue leftovers.
  *
@@ -753,26 +649,25 @@ void scsi_run_host_queues(struct Scsi_Host *shost)
  *             at some point during this call.
  * Notes:      If cmd was requeued, upon return it will be a stale pointer.
  */
-static struct scsi_cmnd *scsi_end_request(struct scsi_cmnd *cmd, int uptodate,
+static struct scsi_cmnd *scsi_end_request(struct scsi_cmnd *cmd, int error,
                                          int bytes, int requeue)
 {
-       request_queue_t *q = cmd->device->request_queue;
+       struct request_queue *q = cmd->device->request_queue;
        struct request *req = cmd->request;
-       unsigned long flags;
 
        /*
         * If there are blocks left over at the end, set up the command
         * to queue the remainder of them.
         */
-       if (end_that_request_chunk(req, uptodate, bytes)) {
+       if (blk_end_request(req, error, bytes)) {
                int leftover = (req->hard_nr_sectors << 9);
 
                if (blk_pc_request(req))
                        leftover = req->data_len;
 
                /* kill remainder if no retrys */
-               if (!uptodate && blk_noretry_request(req))
-                       end_that_request_chunk(req, 0, leftover);
+               if (error && blk_noretry_request(req))
+                       blk_end_request(req, error, leftover);
                else {
                        if (requeue) {
                                /*
@@ -787,14 +682,6 @@ static struct scsi_cmnd *scsi_end_request(struct scsi_cmnd *cmd, int uptodate,
                }
        }
 
-       add_disk_randomness(req->rq_disk);
-
-       spin_lock_irqsave(q->queue_lock, flags);
-       if (blk_rq_tagged(req))
-               blk_queue_end_tag(q, req);
-       end_that_request_last(req, uptodate);
-       spin_unlock_irqrestore(q->queue_lock, flags);
-
        /*
         * This will goose the queue request function at the end, so we don't
         * need to worry about launching another command.
@@ -803,57 +690,87 @@ 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, gfp_t gfp_mask)
-{
-       struct scsi_host_sg_pool *sgp;
-       struct scatterlist *sgl;
+/*
+ * Like SCSI_MAX_SG_SEGMENTS, but for archs that have sg chaining. This limit
+ * is totally arbitrary, a setting of 2048 will get you at least 8mb ios.
+ */
+#define SCSI_MAX_SG_CHAIN_SEGMENTS     2048
 
-       BUG_ON(!cmd->use_sg);
+static inline unsigned int scsi_sgtable_index(unsigned short nents)
+{
+       unsigned int index;
 
-       switch (cmd->use_sg) {
+       switch (nents) {
        case 1 ... 8:
-               cmd->sglist_len = 0;
+               index = 0;
                break;
        case 9 ... 16:
-               cmd->sglist_len = 1;
+               index = 1;
                break;
+#if (SCSI_MAX_SG_SEGMENTS > 16)
        case 17 ... 32:
-               cmd->sglist_len = 2;
+               index = 2;
                break;
-#if (SCSI_MAX_PHYS_SEGMENTS > 32)
+#if (SCSI_MAX_SG_SEGMENTS > 32)
        case 33 ... 64:
-               cmd->sglist_len = 3;
+               index = 3;
                break;
-#if (SCSI_MAX_PHYS_SEGMENTS > 64)
+#if (SCSI_MAX_SG_SEGMENTS > 64)
        case 65 ... 128:
-               cmd->sglist_len = 4;
-               break;
-#if (SCSI_MAX_PHYS_SEGMENTS  > 128)
-       case 129 ... 256:
-               cmd->sglist_len = 5;
+               index = 4;
                break;
 #endif
 #endif
 #endif
        default:
-               return NULL;
+               printk(KERN_ERR "scsi: bad segment count=%d\n", nents);
+               BUG();
        }
 
-       sgp = scsi_sg_pools + cmd->sglist_len;
-       sgl = mempool_alloc(sgp->pool, gfp_mask);
-       return sgl;
+       return index;
 }
 
-static void scsi_free_sgtable(struct scatterlist *sgl, int index)
+static void scsi_sg_free(struct scatterlist *sgl, unsigned int nents)
 {
        struct scsi_host_sg_pool *sgp;
 
-       BUG_ON(index >= SG_MEMPOOL_NR);
-
-       sgp = scsi_sg_pools + index;
+       sgp = scsi_sg_pools + scsi_sgtable_index(nents);
        mempool_free(sgl, sgp->pool);
 }
 
+static struct scatterlist *scsi_sg_alloc(unsigned int nents, gfp_t gfp_mask)
+{
+       struct scsi_host_sg_pool *sgp;
+
+       sgp = scsi_sg_pools + scsi_sgtable_index(nents);
+       return mempool_alloc(sgp->pool, gfp_mask);
+}
+
+int scsi_alloc_sgtable(struct scsi_cmnd *cmd, gfp_t gfp_mask)
+{
+       int ret;
+
+       BUG_ON(!cmd->use_sg);
+
+       ret = __sg_alloc_table(&cmd->sg_table, cmd->use_sg,
+                              SCSI_MAX_SG_SEGMENTS, gfp_mask, scsi_sg_alloc);
+       if (unlikely(ret))
+               __sg_free_table(&cmd->sg_table, SCSI_MAX_SG_SEGMENTS,
+                               scsi_sg_free);
+
+       cmd->request_buffer = cmd->sg_table.sgl;
+       return ret;
+}
+
+EXPORT_SYMBOL(scsi_alloc_sgtable);
+
+void scsi_free_sgtable(struct scsi_cmnd *cmd)
+{
+       __sg_free_table(&cmd->sg_table, SCSI_MAX_SG_SEGMENTS, scsi_sg_free);
+}
+
+EXPORT_SYMBOL(scsi_free_sgtable);
+
 /*
  * Function:    scsi_release_buffers()
  *
@@ -873,22 +790,13 @@ 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);
+               scsi_free_sgtable(cmd);
 
        /*
         * 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;
 }
@@ -921,40 +829,25 @@ 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;
-       request_queue_t *q = cmd->device->request_queue;
+       int this_count = cmd->request_bufflen;
+       struct request_queue *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 = result;
                if (result) {
@@ -970,157 +863,133 @@ 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, 0, 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, -EIO, 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 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);
+                               scsi_end_request(cmd, -EIO, 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->flags & REQ_QUIET))
-                               scmd_printk(KERN_INFO, cmd,
-                                          "Device not ready.\n");
-                       scsi_end_request(cmd, 0, this_count, 1);
+                       if (!(req->cmd_flags & REQ_QUIET))
+                               scsi_cmd_print_sense_hdr(cmd,
+                                                        "Device not ready",
+                                                        &sshdr);
+
+                       scsi_end_request(cmd, -EIO, this_count, 1);
                        return;
                case VOLUME_OVERFLOW:
-                       if (!(req->flags & REQ_QUIET)) {
+                       if (!(req->cmd_flags & REQ_QUIET)) {
                                scmd_printk(KERN_INFO, cmd,
-                                          "Volume overflow, CDB: ");
-                               __scsi_print_command(cmd->data_cmnd);
+                                           "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, -EIO, 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)) {
-                       scmd_printk(KERN_INFO, cmd,
-                                  "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, -EIO, this_count, !result);
 }
-EXPORT_SYMBOL(scsi_io_completion);
 
 /*
  * Function:    scsi_init_io()
@@ -1131,105 +1000,101 @@ EXPORT_SYMBOL(scsi_io_completion);
  *
  * Returns:     0 on success
  *             BLKPREP_DEFER if the failure is retryable
- *             BLKPREP_KILL if the failure is fatal
  */
 static int scsi_init_io(struct scsi_cmnd *cmd)
 {
        struct request     *req = cmd->request;
-       struct scatterlist *sgpnt;
        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(scsi_alloc_sgtable(cmd, GFP_ATOMIC))) {
                scsi_unprep_request(req);
                return BLKPREP_DEFER;
        }
 
-       cmd->request_buffer = (char *) sgpnt;
-       cmd->request_bufflen = req->nr_sectors << 9;
+       req->buffer = NULL;
        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);
+       BUG_ON(count > cmd->use_sg);
+       cmd->use_sg = count;
+       return BLKPREP_OK;
+}
 
-       /*
-        * mapped well, send it off
-        */
-       if (likely(count <= cmd->use_sg)) {
-               cmd->use_sg = count;
-               return 0;
+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;
        }
 
-       printk(KERN_ERR "Incorrect number of segments after building list\n");
-       printk(KERN_ERR "counted %d, received %d\n", count, cmd->use_sg);
-       printk(KERN_ERR "req nr_sec %lu, cur_nr_sec %u\n", req->nr_sectors,
-                       req->current_nr_sectors);
+       /* pull a tag out of the request if we have one */
+       cmd->tag = req->tag;
+       cmd->request = req;
 
-       /* release the command and kill it */
-       scsi_release_buffers(cmd);
-       scsi_put_command(cmd);
-       return BLKPREP_KILL;
+       return cmd;
 }
 
-static int scsi_issue_flush_fn(request_queue_t *q, struct gendisk *disk,
-                              sector_t *error_sector)
+int scsi_setup_blk_pc_cmnd(struct scsi_device *sdev, struct request *req)
 {
-       struct scsi_device *sdev = q->queuedata;
-       struct scsi_driver *drv;
-
-       if (sdev->sdev_state != SDEV_RUNNING)
-               return -ENXIO;
+       struct scsi_cmnd *cmd;
+       int ret = scsi_prep_state_check(sdev, req);
 
-       drv = *(struct scsi_driver **) disk->private_data;
-       if (drv->issue_flush)
-               return drv->issue_flush(&sdev->sdev_gendev, error_sector);
+       if (ret != BLKPREP_OK)
+               return ret;
 
-       return -EOPNOTSUPP;
-}
+       cmd = scsi_get_cmd_from_req(sdev, req);
+       if (unlikely(!cmd))
+               return BLKPREP_DEFER;
 
-static void scsi_blk_pc_done(struct scsi_cmnd *cmd)
-{
-       BUG_ON(!blk_pc_request(cmd->request));
        /*
-        * 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
+        * 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.
         */
-       scsi_io_completion(cmd, cmd->bufflen, 0);
-}
+       if (req->bio) {
+               int ret;
 
-static void scsi_setup_blk_pc_cmnd(struct scsi_cmnd *cmd)
-{
-       struct request *req = cmd->request;
+               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);
 
-       BUG_ON(sizeof(req->cmd) > sizeof(cmd->cmnd));
+               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)
@@ -1242,165 +1107,127 @@ static void scsi_setup_blk_pc_cmnd(struct scsi_cmnd *cmd)
        cmd->transfersize = req->data_len;
        cmd->allowed = req->retries;
        cmd->timeout_per_command = req->timeout;
-       cmd->done = scsi_blk_pc_done;
+       return BLKPREP_OK;
 }
+EXPORT_SYMBOL(scsi_setup_blk_pc_cmnd);
 
-static int scsi_prep_fn(struct request_queue *q, struct request *req)
+/*
+ * Setup a REQ_TYPE_FS command.  These are simple read/write request
+ * from filesystems that still need to be translated to SCSI CDBs from
+ * the ULD.
+ */
+int scsi_setup_fs_cmnd(struct scsi_device *sdev, struct request *req)
 {
-       struct scsi_device *sdev = q->queuedata;
        struct scsi_cmnd *cmd;
-       int specials_only = 0;
+       int ret = scsi_prep_state_check(sdev, req);
 
+       if (ret != BLKPREP_OK)
+               return ret;
        /*
-        * 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
+        * Filesystem requests must transfer data.
         */
-       if (unlikely(!scsi_device_online(sdev))) {
-               sdev_printk(KERN_ERR, sdev,
-                           "rejecting I/O to offline device\n");
-               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 */
-                       sdev_printk(KERN_ERR, sdev,
-                                   "rejecting I/O to dead device\n");
-                       goto kill;
-               }
-               /* OK, we only allow special commands (i.e. not
-                * user initiated ones */
-               specials_only = sdev->sdev_state;
-       }
+       BUG_ON(!req->nr_phys_segments);
+
+       cmd = scsi_get_cmd_from_req(sdev, req);
+       if (unlikely(!cmd))
+               return BLKPREP_DEFER;
+
+       return scsi_init_io(cmd);
+}
+EXPORT_SYMBOL(scsi_setup_fs_cmnd);
+
+int scsi_prep_state_check(struct scsi_device *sdev, struct request *req)
+{
+       int ret = BLKPREP_OK;
 
        /*
-        * 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 the device is not in running state we will reject some
+        * or all commands.
         */
-       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)) {
-
-               if(unlikely(specials_only) && !(req->flags & REQ_SPECIAL)) {
-                       if(specials_only == SDEV_QUIESCE ||
-                                       specials_only == SDEV_BLOCK)
-                               goto defer;
-                       
+       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 device being removed\n");
-                       goto kill;
+                                   "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;
                }
-                       
-                       
-               /*
-                * 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;
        }
-       
-       /* note the overloading of req->special.  When the tag
-        * is active it always means cmd.  If the tag goes
-        * back for re-queueing, it may be reset */
-       req->special = cmd;
-       cmd->request = req;
-       
-       /*
-        * FIXME: drop the lock here because the functions below
-        * expect to be called without the queue lock held.  Also,
-        * previously, we dequeued the request before dropping the
-        * lock.  We hope REQ_STARTED prevents anything untoward from
-        * happening now.
-        */
-       if (req->flags & (REQ_CMD | REQ_BLOCK_PC)) {
-               int ret;
+       return ret;
+}
+EXPORT_SYMBOL(scsi_prep_state_check);
 
-               /*
-                * 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.
-                */
+int scsi_prep_return(struct request_queue *q, struct request *req, int ret)
+{
+       struct scsi_device *sdev = q->queuedata;
 
-               /* 
-                * This sets up the scatter-gather table (allocating if
-                * required).
-                */
-               ret = scsi_init_io(cmd);
-               switch(ret) {
-                       /* For BLKPREP_KILL/DEFER the cmd was released */
-               case BLKPREP_KILL:
-                       goto kill;
-               case BLKPREP_DEFER:
-                       goto defer;
+       switch (ret) {
+       case BLKPREP_KILL:
+               req->errors = DID_NO_CONNECT << 16;
+               /* release the command and kill it */
+               if (req->special) {
+                       struct scsi_cmnd *cmd = req->special;
+                       scsi_release_buffers(cmd);
+                       scsi_put_command(cmd);
+                       req->special = NULL;
                }
-               
+               break;
+       case BLKPREP_DEFER:
                /*
-                * 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->flags & REQ_BLOCK_PC) {
-                       scsi_setup_blk_pc_cmnd(cmd);
-               } else if (req->rq_disk) {
-                       struct scsi_driver *drv;
-
-                       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;
-                       }
-               }
+               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;
+       return ret;
+}
+EXPORT_SYMBOL(scsi_prep_return);
 
- 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;
+int scsi_prep_fn(struct request_queue *q, struct request *req)
+{
+       struct scsi_device *sdev = q->queuedata;
+       int ret = BLKPREP_KILL;
+
+       if (req->cmd_type == REQ_TYPE_BLOCK_PC)
+               ret = scsi_setup_blk_pc_cmnd(sdev, req);
+       return scsi_prep_return(q, req, ret);
 }
 
 /*
@@ -1476,9 +1303,11 @@ static inline int scsi_host_queue_ready(struct request_queue *q,
 /*
  * Kill a request for a dead device
  */
-static void scsi_kill_request(struct request *req, request_queue_t *q)
+static void scsi_kill_request(struct request *req, struct request_queue *q)
 {
        struct scsi_cmnd *cmd = req->special;
+       struct scsi_device *sdev = cmd->device;
+       struct Scsi_Host *shost = sdev->host;
 
        blkdev_dequeue_request(req);
 
@@ -1491,6 +1320,19 @@ 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);
 }
 
@@ -1518,7 +1360,7 @@ static void scsi_softirq_done(struct request *rq)
                        scsi_finish_command(cmd);
                        break;
                case NEEDS_RETRY:
-                       scsi_retry_command(cmd);
+                       scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
                        break;
                case ADD_TO_MLQUEUE:
                        scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
@@ -1594,15 +1436,16 @@ 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);
 
                if (!scsi_host_queue_ready(q, shost, sdev))
                        goto not_ready;
-               if (sdev->single_lun) {
+               if (scsi_target(sdev)->single_lun) {
                        if (scsi_target(sdev)->starget_sdev_user &&
                            scsi_target(sdev)->starget_sdev_user != sdev)
                                goto not_ready;
@@ -1686,27 +1529,62 @@ 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);
-
+       /*
+        * this limit is imposed by hardware restrictions
+        */
        blk_queue_max_hw_segments(q, shost->sg_tablesize);
-       blk_queue_max_phys_segments(q, SCSI_MAX_PHYS_SEGMENTS);
+
+       /*
+        * In the future, sg chaining support will be mandatory and this
+        * ifdef can then go away. Right now we don't have all archs
+        * converted, so better keep it safe.
+        */
+#ifdef ARCH_HAS_SG_CHAIN
+       if (shost->use_sg_chaining)
+               blk_queue_max_phys_segments(q, SCSI_MAX_SG_CHAIN_SEGMENTS);
+       else
+               blk_queue_max_phys_segments(q, SCSI_MAX_SG_SEGMENTS);
+#else
+       blk_queue_max_phys_segments(q, SCSI_MAX_SG_SEGMENTS);
+#endif
+
        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);
-       blk_queue_softirq_done(q, scsi_softirq_done);
 
        if (!shost->use_clustering)
                clear_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags);
+
+       /*
+        * set a reasonable default alignment on word boundaries: the
+        * host and device may alter it using
+        * blk_queue_update_dma_alignment() later.
+        */
+       blk_queue_dma_alignment(q, 0x03);
+
+       return q;
+}
+EXPORT_SYMBOL(__scsi_alloc_queue);
+
+struct request_queue *scsi_alloc_queue(struct scsi_device *sdev)
+{
+       struct request_queue *q;
+
+       q = __scsi_alloc_queue(sdev->host, scsi_request_fn);
+       if (!q)
+               return NULL;
+
+       blk_queue_prep_rq(q, scsi_prep_fn);
+       blk_queue_softirq_done(q, scsi_softirq_done);
        return q;
 }
 
@@ -1770,7 +1648,7 @@ int __init scsi_init_queue(void)
 
        scsi_io_context_cache = kmem_cache_create("scsi_io_context",
                                        sizeof(struct scsi_io_context),
-                                       0, 0, NULL, NULL);
+                                       0, 0, NULL);
        if (!scsi_io_context_cache) {
                printk(KERN_ERR "SCSI: can't init scsi io context cache\n");
                return -ENOMEM;
@@ -1781,15 +1659,14 @@ int __init scsi_init_queue(void)
                int size = sgp->size * sizeof(struct scatterlist);
 
                sgp->slab = kmem_cache_create(sgp->name, size, 0,
-                               SLAB_HWCACHE_ALIGN, NULL, NULL);
+                               SLAB_HWCACHE_ALIGN, NULL);
                if (!sgp->slab) {
                        printk(KERN_ERR "SCSI: can't init sg slab %s\n",
                                        sgp->name);
                }
 
-               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);
@@ -1811,9 +1688,86 @@ void scsi_exit_queue(void)
                kmem_cache_destroy(sgp->slab);
        }
 }
+
 /**
- *     scsi_mode_sense - issue a mode sense, falling back from 10 to 
- *             six bytes if necessary.
+ *     scsi_mode_select - issue a mode select
+ *     @sdev:  SCSI device to be queried
+ *     @pf:    Page format bit (1 == standard, 0 == vendor specific)
+ *     @sp:    Save page bit (0 == don't save, 1 == save)
+ *     @modepage: mode page being requested
+ *     @buffer: request buffer (may not be smaller than eight bytes)
+ *     @len:   length of request buffer.
+ *     @timeout: command timeout
+ *     @retries: number of retries before failing
+ *     @data: returns a structure abstracting the mode header data
+ *     @sshdr: place to put sense data (or NULL if no sense to be collected).
+ *             must be SCSI_SENSE_BUFFERSIZE big.
+ *
+ *     Returns zero if successful; negative error number or scsi
+ *     status on error
+ *
+ */
+int
+scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
+                unsigned char *buffer, int len, int timeout, int retries,
+                struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
+{
+       unsigned char cmd[10];
+       unsigned char *real_buffer;
+       int ret;
+
+       memset(cmd, 0, sizeof(cmd));
+       cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
+
+       if (sdev->use_10_for_ms) {
+               if (len > 65535)
+                       return -EINVAL;
+               real_buffer = kmalloc(8 + len, GFP_KERNEL);
+               if (!real_buffer)
+                       return -ENOMEM;
+               memcpy(real_buffer + 8, buffer, len);
+               len += 8;
+               real_buffer[0] = 0;
+               real_buffer[1] = 0;
+               real_buffer[2] = data->medium_type;
+               real_buffer[3] = data->device_specific;
+               real_buffer[4] = data->longlba ? 0x01 : 0;
+               real_buffer[5] = 0;
+               real_buffer[6] = data->block_descriptor_length >> 8;
+               real_buffer[7] = data->block_descriptor_length;
+
+               cmd[0] = MODE_SELECT_10;
+               cmd[7] = len >> 8;
+               cmd[8] = len;
+       } else {
+               if (len > 255 || data->block_descriptor_length > 255 ||
+                   data->longlba)
+                       return -EINVAL;
+
+               real_buffer = kmalloc(4 + len, GFP_KERNEL);
+               if (!real_buffer)
+                       return -ENOMEM;
+               memcpy(real_buffer + 4, buffer, len);
+               len += 4;
+               real_buffer[0] = 0;
+               real_buffer[1] = data->medium_type;
+               real_buffer[2] = data->device_specific;
+               real_buffer[3] = data->block_descriptor_length;
+               
+
+               cmd[0] = MODE_SELECT;
+               cmd[4] = len;
+       }
+
+       ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
+                              sshdr, timeout, retries);
+       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.
  *     @sdev:  SCSI device to be queried
  *     @dbd:   set if mode sense will allow block descriptors to be returned
  *     @modepage: mode page being requested
@@ -1822,17 +1776,18 @@ void scsi_exit_queue(void)
  *     @timeout: command timeout
  *     @retries: number of retries before failing
  *     @data: returns a structure abstracting the mode header data
- *     @sense: place to put sense data (or NULL if no sense to be collected).
+ *     @sshdr: place to put sense data (or NULL if no sense to be collected).
  *             must be SCSI_SENSE_BUFFERSIZE big.
  *
  *     Returns zero if unsuccessful, or the header offset (either 4
  *     or 8 depending on whether a six or ten byte command was
  *     issued) if successful.
- **/
+ */
 int
 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
                  unsigned char *buffer, int len, int timeout, int retries,
-                 struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr) {
+                 struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
+{
        unsigned char cmd[12];
        int use_10_for_ms;
        int header_length;
@@ -1921,40 +1876,69 @@ scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
 }
 EXPORT_SYMBOL(scsi_mode_sense);
 
+/**
+ *     scsi_test_unit_ready - test if unit is ready
+ *     @sdev:  scsi device to change the state of.
+ *     @timeout: command timeout
+ *     @retries: number of retries before failing
+ *     @sshdr_external: Optional pointer to struct scsi_sense_hdr for
+ *             returning sense. Make sure that this is cleared before passing
+ *             in.
+ *
+ *     Returns zero if unsuccessful or an error if TUR failed.  For
+ *     removable media, a return of NOT_READY or UNIT_ATTENTION is
+ *     translated to success, with the ->changed flag updated.
+ **/
 int
-scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries)
+scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
+                    struct scsi_sense_hdr *sshdr_external)
 {
        char cmd[] = {
                TEST_UNIT_READY, 0, 0, 0, 0, 0,
        };
-       struct scsi_sense_hdr sshdr;
+       struct scsi_sense_hdr *sshdr;
        int result;
-       
-       result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, &sshdr,
-                                 timeout, retries);
+
+       if (!sshdr_external)
+               sshdr = kzalloc(sizeof(*sshdr), GFP_KERNEL);
+       else
+               sshdr = sshdr_external;
+
+       /* try to eat the UNIT_ATTENTION if there are enough retries */
+       do {
+               result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
+                                         timeout, retries);
+       } while ((driver_byte(result) & DRIVER_SENSE) &&
+                sshdr && sshdr->sense_key == UNIT_ATTENTION &&
+                --retries);
+
+       if (!sshdr)
+               /* could not allocate sense buffer, so can't process it */
+               return result;
 
        if ((driver_byte(result) & DRIVER_SENSE) && sdev->removable) {
 
-               if ((scsi_sense_valid(&sshdr)) &&
-                   ((sshdr.sense_key == UNIT_ATTENTION) ||
-                    (sshdr.sense_key == NOT_READY))) {
+               if ((scsi_sense_valid(sshdr)) &&
+                   ((sshdr->sense_key == UNIT_ATTENTION) ||
+                    (sshdr->sense_key == NOT_READY))) {
                        sdev->changed = 1;
                        result = 0;
                }
        }
+       if (!sshdr_external)
+               kfree(sshdr);
        return result;
 }
 EXPORT_SYMBOL(scsi_test_unit_ready);
 
 /**
- *     scsi_device_set_state - Take the given device through the device
- *             state model.
+ *     scsi_device_set_state - Take the given device through the device state model.
  *     @sdev:  scsi device to change the state of.
  *     @state: state to change to.
  *
  *     Returns zero if unsuccessful or an error if the requested 
  *     transition is illegal.
- **/
+ */
 int
 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
 {
@@ -2018,6 +2002,7 @@ 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;
@@ -2028,6 +2013,9 @@ 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:
@@ -2051,6 +2039,142 @@ scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
 EXPORT_SYMBOL(scsi_device_set_state);
 
 /**
+ *     sdev_evt_emit - emit a single SCSI device uevent
+ *     @sdev: associated SCSI device
+ *     @evt: event to emit
+ *
+ *     Send a single uevent (scsi_event) to the associated scsi_device.
+ */
+static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
+{
+       int idx = 0;
+       char *envp[3];
+
+       switch (evt->evt_type) {
+       case SDEV_EVT_MEDIA_CHANGE:
+               envp[idx++] = "SDEV_MEDIA_CHANGE=1";
+               break;
+
+       default:
+               /* do nothing */
+               break;
+       }
+
+       envp[idx++] = NULL;
+
+       kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
+}
+
+/**
+ *     sdev_evt_thread - send a uevent for each scsi event
+ *     @work: work struct for scsi_device
+ *
+ *     Dispatch queued events to their associated scsi_device kobjects
+ *     as uevents.
+ */
+void scsi_evt_thread(struct work_struct *work)
+{
+       struct scsi_device *sdev;
+       LIST_HEAD(event_list);
+
+       sdev = container_of(work, struct scsi_device, event_work);
+
+       while (1) {
+               struct scsi_event *evt;
+               struct list_head *this, *tmp;
+               unsigned long flags;
+
+               spin_lock_irqsave(&sdev->list_lock, flags);
+               list_splice_init(&sdev->event_list, &event_list);
+               spin_unlock_irqrestore(&sdev->list_lock, flags);
+
+               if (list_empty(&event_list))
+                       break;
+
+               list_for_each_safe(this, tmp, &event_list) {
+                       evt = list_entry(this, struct scsi_event, node);
+                       list_del(&evt->node);
+                       scsi_evt_emit(sdev, evt);
+                       kfree(evt);
+               }
+       }
+}
+
+/**
+ *     sdev_evt_send - send asserted event to uevent thread
+ *     @sdev: scsi_device event occurred on
+ *     @evt: event to send
+ *
+ *     Assert scsi device event asynchronously.
+ */
+void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
+{
+       unsigned long flags;
+
+       if (!test_bit(evt->evt_type, sdev->supported_events)) {
+               kfree(evt);
+               return;
+       }
+
+       spin_lock_irqsave(&sdev->list_lock, flags);
+       list_add_tail(&evt->node, &sdev->event_list);
+       schedule_work(&sdev->event_work);
+       spin_unlock_irqrestore(&sdev->list_lock, flags);
+}
+EXPORT_SYMBOL_GPL(sdev_evt_send);
+
+/**
+ *     sdev_evt_alloc - allocate a new scsi event
+ *     @evt_type: type of event to allocate
+ *     @gfpflags: GFP flags for allocation
+ *
+ *     Allocates and returns a new scsi_event.
+ */
+struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
+                                 gfp_t gfpflags)
+{
+       struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
+       if (!evt)
+               return NULL;
+
+       evt->evt_type = evt_type;
+       INIT_LIST_HEAD(&evt->node);
+
+       /* evt_type-specific initialization, if any */
+       switch (evt_type) {
+       case SDEV_EVT_MEDIA_CHANGE:
+       default:
+               /* do nothing */
+               break;
+       }
+
+       return evt;
+}
+EXPORT_SYMBOL_GPL(sdev_evt_alloc);
+
+/**
+ *     sdev_evt_send_simple - send asserted event to uevent thread
+ *     @sdev: scsi_device event occurred on
+ *     @evt_type: type of event to send
+ *     @gfpflags: GFP flags for allocation
+ *
+ *     Assert scsi device event asynchronously, given an event type.
+ */
+void sdev_evt_send_simple(struct scsi_device *sdev,
+                         enum scsi_device_event evt_type, gfp_t gfpflags)
+{
+       struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
+       if (!evt) {
+               sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
+                           evt_type);
+               return;
+       }
+
+       sdev_evt_send(sdev, evt);
+}
+EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
+
+/**
  *     scsi_device_quiesce - Block user issued commands.
  *     @sdev:  scsi device to quiesce.
  *
@@ -2064,7 +2188,7 @@ EXPORT_SYMBOL(scsi_device_set_state);
  *     Must be called with user context, may sleep.
  *
  *     Returns zero if unsuccessful or an error if not.
- **/
+ */
 int
 scsi_device_quiesce(struct scsi_device *sdev)
 {
@@ -2089,7 +2213,7 @@ EXPORT_SYMBOL(scsi_device_quiesce);
  *     queues.
  *
  *     Must be called with user context, may sleep.
- **/
+ */
 void
 scsi_device_resume(struct scsi_device *sdev)
 {
@@ -2126,8 +2250,7 @@ scsi_target_resume(struct scsi_target *starget)
 EXPORT_SYMBOL(scsi_target_resume);
 
 /**
- * scsi_internal_device_block - internal function to put a device
- *                             temporarily into the SDEV_BLOCK state
+ * scsi_internal_device_block - internal function to put a device temporarily into the SDEV_BLOCK state
  * @sdev:      device to block
  *
  * Block request made by scsi lld's to temporarily stop all
@@ -2142,11 +2265,11 @@ EXPORT_SYMBOL(scsi_target_resume);
  *     state, all commands are deferred until the scsi lld reenables
  *     the device with scsi_device_unblock or device_block_tmo fires.
  *     This routine assumes the host_lock is held on entry.
- **/
+ */
 int
 scsi_internal_device_block(struct scsi_device *sdev)
 {
-       request_queue_t *q = sdev->request_queue;
+       struct request_queue *q = sdev->request_queue;
        unsigned long flags;
        int err = 0;
 
@@ -2182,11 +2305,11 @@ EXPORT_SYMBOL_GPL(scsi_internal_device_block);
  *     (which must be a legal transition) allowing the midlayer to
  *     goose the queue for this device.  This routine assumes the 
  *     host_lock is held upon entry.
- **/
+ */
 int
 scsi_internal_device_unblock(struct scsi_device *sdev)
 {
-       request_queue_t *q = sdev->request_queue; 
+       struct request_queue *q = sdev->request_queue; 
        int err;
        unsigned long flags;
        
@@ -2257,3 +2380,63 @@ scsi_target_unblock(struct device *dev)
                device_for_each_child(dev, NULL, target_unblock);
 }
 EXPORT_SYMBOL_GPL(scsi_target_unblock);
+
+/**
+ * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
+ * @sgl:       scatter-gather list
+ * @sg_count:  number of segments in sg
+ * @offset:    offset in bytes into sg, on return offset into the mapped area
+ * @len:       bytes to map, on return number of bytes mapped
+ *
+ * Returns virtual address of the start of the mapped page
+ */
+void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
+                         size_t *offset, size_t *len)
+{
+       int i;
+       size_t sg_len = 0, len_complete = 0;
+       struct scatterlist *sg;
+       struct page *page;
+
+       WARN_ON(!irqs_disabled());
+
+       for_each_sg(sgl, sg, sg_count, i) {
+               len_complete = sg_len; /* Complete sg-entries */
+               sg_len += sg->length;
+               if (sg_len > *offset)
+                       break;
+       }
+
+       if (unlikely(i == sg_count)) {
+               printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
+                       "elements %d\n",
+                      __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->offset;
+
+       /* Assumption: contiguous pages can be accessed as "page + i" */
+       page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
+       *offset &= ~PAGE_MASK;
+
+       /* Bytes in this sg-entry from *offset to the end of the page */
+       sg_len = PAGE_SIZE - *offset;
+       if (*len > sg_len)
+               *len = sg_len;
+
+       return kmap_atomic(page, KM_BIO_SRC_IRQ);
+}
+EXPORT_SYMBOL(scsi_kmap_atomic_sg);
+
+/**
+ * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
+ * @virt:      virtual address to be unmapped
+ */
+void scsi_kunmap_atomic_sg(void *virt)
+{
+       kunmap_atomic(virt, KM_BIO_SRC_IRQ);
+}
+EXPORT_SYMBOL(scsi_kunmap_atomic_sg);