Il 28/08/2012 15:04, Stefan Hajnoczi ha scritto:
> Zero writes have cluster granularity in QED. Therefore they can only be
> used to zero entire clusters.
>
> If the zero write request leaves sectors untouched, zeroing the entire
> cluster would obscure the backing file. Instead return -ENOTSUP, which
> is handled by block.c:bdrv_co_do_write_zeroes() and falls back to a
> regular write.
>
> The qemu-iotests 034 test cases covers this scenario.
Reviewed-by: Paolo Bonzini <pbonz...@redhat.com>
Makes sense since both streaming and copy-on-read will do cluster-aligned
writes.
The "right fix" would not be much more complex though, something like this,
right?
(untested).
diff --git a/block/qed.c b/block/qed.c
index a02dbfd..a885671 100644
--- a/block/qed.c
+++ b/block/qed.c
@@ -1133,8 +1133,14 @@ static void qed_aio_write_alloc(QEDAIOCB *acb, size_t
len)
return;
}
+ if (qed_offset_into_cluster(s, acb->cur_pos) != 0 ||
+ qed_offset_into_cluster(s, acb->cur_pos + acb->cur_qiov.size) !=
0) {
+ goto copy;
+ }
+
cb = qed_aio_write_zero_cluster;
} else {
+copy:
cb = qed_aio_write_prefill;
acb->cur_cluster = qed_alloc_clusters(s, acb->cur_nclusters);
}
Paolo
