rbd: fix I/O error propagation for reads
authorJosh Durgin <josh.durgin@inktank.com>
Tue, 27 Aug 2013 00:55:38 +0000 (17:55 -0700)
committerSage Weil <sage@inktank.com>
Wed, 4 Sep 2013 05:06:10 +0000 (22:06 -0700)
When a request returns an error, the driver needs to report the entire
extent of the request as completed.  Writes already did this, since
they always set xferred = length, but reads were skipping that step if
an error other than -ENOENT occurred.  Instead, rbd would end up
passing 0 xferred to blk_end_request(), which would always report
needing more data.  This resulted in an assert failing when more data
was required by the block layer, but all the object requests were
done:

[ 1868.719077] rbd: obj_request read result -108 xferred 0
[ 1868.719077]
[ 1868.719518] end_request: I/O error, dev rbd1, sector 0
[ 1868.719739]
[ 1868.719739] Assertion failure in rbd_img_obj_callback() at line 1736:
[ 1868.719739]
[ 1868.719739]   rbd_assert(more ^ (which == img_request->obj_request_count));

Without this assert, reads that hit errors would hang forever, since
the block layer considered them incomplete.

Fixes: http://tracker.ceph.com/issues/5647
CC: stable@vger.kernel.org # v3.10
Signed-off-by: Josh Durgin <josh.durgin@inktank.com>
Reviewed-by: Alex Elder <alex.elder@linaro.org>
drivers/block/rbd.c

index 0d669ae80d61d4e2a2b33733054bf13b047f22f6..f8fd7d3c13ba52bc414660f4d063f57b04472f19 100644 (file)
@@ -1557,11 +1557,12 @@ rbd_img_obj_request_read_callback(struct rbd_obj_request *obj_request)
                obj_request, obj_request->img_request, obj_request->result,
                xferred, length);
        /*
-        * ENOENT means a hole in the image.  We zero-fill the
-        * entire length of the request.  A short read also implies
-        * zero-fill to the end of the request.  Either way we
-        * update the xferred count to indicate the whole request
-        * was satisfied.
+        * ENOENT means a hole in the image.  We zero-fill the entire
+        * length of the request.  A short read also implies zero-fill
+        * to the end of the request.  An error requires the whole
+        * length of the request to be reported finished with an error
+        * to the block layer.  In each case we update the xferred
+        * count to indicate the whole request was satisfied.
         */
        rbd_assert(obj_request->type != OBJ_REQUEST_NODATA);
        if (obj_request->result == -ENOENT) {
@@ -1570,14 +1571,13 @@ rbd_img_obj_request_read_callback(struct rbd_obj_request *obj_request)
                else
                        zero_pages(obj_request->pages, 0, length);
                obj_request->result = 0;
-               obj_request->xferred = length;
        } else if (xferred < length && !obj_request->result) {
                if (obj_request->type == OBJ_REQUEST_BIO)
                        zero_bio_chain(obj_request->bio_list, xferred);
                else
                        zero_pages(obj_request->pages, xferred, length);
-               obj_request->xferred = length;
        }
+       obj_request->xferred = length;
        obj_request_done_set(obj_request);
 }