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)
commit17c1cc1d9293a568a00545469078e29555cc7f39
treef20e923740a3b20fb449bff813ca7438ce00b1e5
parent7d6e1f5461d0c16eb6aa8d226976995856d85e4e
rbd: fix I/O error propagation for reads

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