On 15.02.22 23:22, Eric Blake wrote:
On Tue, Feb 15, 2022 at 02:57:27PM +0100, Hanna Reitz wrote:
Test the following scenario:
1. Some block node (null-co) attached to a user (here: NBD server) that
performs I/O and keeps the node in an I/O thread
2. Repeatedly run blockdev-add/blockdev-del to add/remove an overlay
to/from that node
Each blockdev-add triggers bdrv_refresh_limits(), and because
blockdev-add runs in the main thread, it does not stop the I/O requests.
I/O can thus happen while the limits are refreshed, and when such a
request sees a temporarily invalid block limit (e.g. alignment is 0),
this may easily crash qemu (or the storage daemon in this case).
The block layer needs to ensure that I/O requests to a node are paused
while that node's BlockLimits are refreshed.
Signed-off-by: Hanna Reitz <hre...@redhat.com>
---
.../qemu-iotests/tests/graph-changes-while-io | 91 +++++++++++++++++++
.../tests/graph-changes-while-io.out | 5 +
2 files changed, 96 insertions(+)
create mode 100755 tests/qemu-iotests/tests/graph-changes-while-io
create mode 100644 tests/qemu-iotests/tests/graph-changes-while-io.out
Reviewed-by: Eric Blake <ebl...@redhat.com>
Since we found this with the help of NBD, should I be considering this
series for my NBD queue, or is there a better block-related maintainer
queue that it should go through?
Well, we found it by using a guest, it’s just that using a guest in the
iotests is not quite that great, so we need some other way to induce I/O
(concurrently to monitor commands). I could’ve used FUSE, too, but NBD
is always compiled in, so. :)
In any case, of course I don’t mind who takes this series. If you want
to take it, go ahead (and thanks!) – I’ll be sending a v2 to split the
`try` block in patch 2, though.
Hanna
