[SCSI] untangle scsi_prep_fn
authorChristoph Hellwig <hch@lst.de>
Sat, 4 Nov 2006 19:10:55 +0000 (20:10 +0100)
committerJames Bottomley <jejb@mulgrave.il.steeleye.com>
Wed, 15 Nov 2006 20:14:20 +0000 (14:14 -0600)
I wanted to add some BUG checks to scsi_prep_fn to make sure no one
sends us a non-sg command, but this function is a horrible mess.

So I decided to detangle the function and document what the valid
cases are.  While doing that I found that REQ_TYPE_SPECIAL commands
aren't used by the SCSI layer anymore and we can get rid of the code
handling them.

The new structure of scsi_prep_fn is:

 (1) check if we're allowed to send this command
 (2) big switch on cmd_type.  For the two valid types call into
     a function to set the command up, else error
 (3) code to handle error cases

Because FS and BLOCK_PC commands are handled entirely separate after
the patch this introduces a tiny amount of code duplication.  This
improves readabiulity though and will help to avoid the bidi command
overhead for FS commands so it's a good thing.

I've tested this on both sata and mptsas.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
drivers/scsi/scsi_lib.c

index 743f67e..ee35a62 100644 (file)
@@ -995,25 +995,14 @@ 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 (blk_pc_request(req) && !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)) {
@@ -1021,24 +1010,21 @@ static int scsi_init_io(struct scsi_cmnd *cmd)
                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");
@@ -1068,6 +1054,27 @@ static int scsi_issue_flush_fn(request_queue_t *q, struct gendisk *disk,
        return -EOPNOTSUPP;
 }
 
+static struct scsi_cmnd *scsi_get_cmd_from_req(struct scsi_device *sdev,
+               struct request *req)
+{
+       struct scsi_cmnd *cmd;
+
+       if (!req->special) {
+               cmd = scsi_get_command(sdev, GFP_ATOMIC);
+               if (unlikely(!cmd))
+                       return NULL;
+               req->special = cmd;
+       } else {
+               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));
@@ -1080,9 +1087,37 @@ static void scsi_blk_pc_done(struct scsi_cmnd *cmd)
        scsi_io_completion(cmd, cmd->request_bufflen);
 }
 
-static void scsi_setup_blk_pc_cmnd(struct scsi_cmnd *cmd)
+static int scsi_setup_blk_pc_cmnd(struct scsi_device *sdev, struct request *req)
 {
-       struct request *req = cmd->request;
+       struct scsi_cmnd *cmd;
+
+       cmd = scsi_get_cmd_from_req(sdev, req);
+       if (unlikely(!cmd))
+               return BLKPREP_DEFER;
+
+       /*
+        * BLOCK_PC requests may transfer data, in which case they must
+        * a bio attached to them.  Or they might contain a SCSI command
+        * that does not transfer data, in which case they may optionally
+        * submit a request without an attached bio.
+        */
+       if (req->bio) {
+               int ret;
+
+               BUG_ON(!req->nr_phys_segments);
+
+               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;
+       }
 
        BUG_ON(sizeof(req->cmd) > sizeof(cmd->cmnd));
        memcpy(cmd->cmnd, req->cmd, sizeof(cmd->cmnd));
@@ -1098,154 +1133,138 @@ static void scsi_setup_blk_pc_cmnd(struct scsi_cmnd *cmd)
        cmd->allowed = req->retries;
        cmd->timeout_per_command = req->timeout;
        cmd->done = scsi_blk_pc_done;
+       return BLKPREP_OK;
 }
 
-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.
+ */
+static 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;
+       struct scsi_driver *drv;
+       int 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;
+
+       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;
        }
 
+       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;
+
        /*
-        * 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 (blk_special_request(req) && req->special)
-               cmd = req->special;
-       else if (blk_pc_request(req) || blk_fs_request(req)) {
-               if (unlikely(specials_only) && !(req->cmd_flags & REQ_PREEMPT)){
-                       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;
+
+               if (ret != BLKPREP_OK)
+                       goto out;
        }
-       
-       /* 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 (blk_fs_request(req) || blk_pc_request(req)) {
-               int ret;
 
+       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) {
-                       /* For BLKPREP_KILL/DEFER the cmd was released */
-               case BLKPREP_KILL:
-                       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 (blk_pc_request(req)) {
-                       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->cmd_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;
 }
 
 /*