[libvirt] [PATCH] logging: ensure pending I/O is drained before reading position

[Michal Privoznik]

On 12/14/18 2:04 PM, Daniel P. Berrangé wrote:
> The virtualization driver has two connections to the virtlogd daemon,
> one pipe fd for writing to the log file, and one socket fd for making
> RPC calls. The typical sequence is to write some data to the pipe fd and
> then make an RPC call to determine the current log file offset.
> 
> Unfortunately these two operations are not guaranteed to be handling in
> order by virtlogd. The event loop for virtlogd may identify an incoming
> event on both the pipe fd and socket fd in the same iteration of the
> event loop. It is then entirely possible that it will process the socket
> fd RPC call before reading the pending log data from the pipe fd.
> 
> As a result the virtualization driver will get an outdated log file
> offset reported back.
> 
> This can be seen with the QEMU driver where, when a guest fails to
> start, it will randomly include too much data in the error message it
> has fetched from the log file.
> 
> The solution is to ensure we have drained all pending data from the pipe
> fd before reporting the log file offset. The pipe fd is always in
> blocking mode, so cares needs to be taken to avoid blocking. When
> draining this is taken care of by using poll(). The extra complication
> is that they might already be an event loop dispatch pending on the pipe
> fd. If we have just drained the pipe this pending event will be invalid
> so must be discarded.
> 
> Signed-off-by: Daniel P. Berrangé <berrange at redhat.com>
> ---
>  src/logging/log_handler.c | 48 +++++++++++++++++++++++++++++++++++++++
>  1 file changed, 48 insertions(+)

ACK

Michal