xen/blkfront: use blk_rq_map_sg to generate ring entries
authorJens Axboe <jens.axboe@oracle.com>
Tue, 24 Feb 2009 07:10:09 +0000 (08:10 +0100)
committerJens Axboe <jens.axboe@oracle.com>
Thu, 26 Feb 2009 09:45:48 +0000 (10:45 +0100)
On occasion, the request will apparently have more segments than we
fit into the ring. Jens says:

> The second problem is that the block layer then appears to create one
> too many segments, but from the dump it has rq->nr_phys_segments ==
> BLKIF_MAX_SEGMENTS_PER_REQUEST. I suspect the latter is due to
> xen-blkfront not handling the merging on its own. It should check that
> the new page doesn't form part of the previous page. The
> rq_for_each_segment() iterates all single bits in the request, not dma
> segments. The "easiest" way to do this is to call blk_rq_map_sg() and
> then iterate the mapped sg list. That will give you what you are
> looking for.

> Here's a test patch, compiles but otherwise untested. I spent more
> time figuring out how to enable XEN than to code it up, so YMMV!
> Probably the sg list wants to be put inside the ring and only
> initialized on allocation, then you can get rid of the sg on stack and
> sg_init_table() loop call in the function. I'll leave that, and the
> testing, to you.

[Moved sg array into info structure, and initialize once. -J]

Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
drivers/block/xen-blkfront.c

index 918ef725de41c849f3f78b7e30f80d7709e01a39..b6c8ce25435994558885e9e4b9c4ac6aafaa9357 100644 (file)
@@ -40,6 +40,7 @@
 #include <linux/hdreg.h>
 #include <linux/cdrom.h>
 #include <linux/module.h>
+#include <linux/scatterlist.h>
 
 #include <xen/xenbus.h>
 #include <xen/grant_table.h>
@@ -82,6 +83,7 @@ struct blkfront_info
        enum blkif_state connected;
        int ring_ref;
        struct blkif_front_ring ring;
+       struct scatterlist sg[BLKIF_MAX_SEGMENTS_PER_REQUEST];
        unsigned int evtchn, irq;
        struct request_queue *rq;
        struct work_struct work;
@@ -204,12 +206,11 @@ static int blkif_queue_request(struct request *req)
        struct blkfront_info *info = req->rq_disk->private_data;
        unsigned long buffer_mfn;
        struct blkif_request *ring_req;
-       struct req_iterator iter;
-       struct bio_vec *bvec;
        unsigned long id;
        unsigned int fsect, lsect;
-       int ref;
+       int i, ref;
        grant_ref_t gref_head;
+       struct scatterlist *sg;
 
        if (unlikely(info->connected != BLKIF_STATE_CONNECTED))
                return 1;
@@ -238,12 +239,13 @@ static int blkif_queue_request(struct request *req)
        if (blk_barrier_rq(req))
                ring_req->operation = BLKIF_OP_WRITE_BARRIER;
 
-       ring_req->nr_segments = 0;
-       rq_for_each_segment(bvec, req, iter) {
-               BUG_ON(ring_req->nr_segments == BLKIF_MAX_SEGMENTS_PER_REQUEST);
-               buffer_mfn = pfn_to_mfn(page_to_pfn(bvec->bv_page));
-               fsect = bvec->bv_offset >> 9;
-               lsect = fsect + (bvec->bv_len >> 9) - 1;
+       ring_req->nr_segments = blk_rq_map_sg(req->q, req, info->sg);
+       BUG_ON(ring_req->nr_segments > BLKIF_MAX_SEGMENTS_PER_REQUEST);
+
+       for_each_sg(info->sg, sg, ring_req->nr_segments, i) {
+               buffer_mfn = pfn_to_mfn(page_to_pfn(sg_page(sg)));
+               fsect = sg->offset >> 9;
+               lsect = fsect + (sg->length >> 9) - 1;
                /* install a grant reference. */
                ref = gnttab_claim_grant_reference(&gref_head);
                BUG_ON(ref == -ENOSPC);
@@ -254,16 +256,12 @@ static int blkif_queue_request(struct request *req)
                                buffer_mfn,
                                rq_data_dir(req) );
 
-               info->shadow[id].frame[ring_req->nr_segments] =
-                               mfn_to_pfn(buffer_mfn);
-
-               ring_req->seg[ring_req->nr_segments] =
+               info->shadow[id].frame[i] = mfn_to_pfn(buffer_mfn);
+               ring_req->seg[i] =
                                (struct blkif_request_segment) {
                                        .gref       = ref,
                                        .first_sect = fsect,
                                        .last_sect  = lsect };
-
-               ring_req->nr_segments++;
        }
 
        info->ring.req_prod_pvt++;
@@ -622,6 +620,8 @@ static int setup_blkring(struct xenbus_device *dev,
        SHARED_RING_INIT(sring);
        FRONT_RING_INIT(&info->ring, sring, PAGE_SIZE);
 
+       sg_init_table(info->sg, BLKIF_MAX_SEGMENTS_PER_REQUEST);
+
        err = xenbus_grant_ring(dev, virt_to_mfn(info->ring.sring));
        if (err < 0) {
                free_page((unsigned long)sring);