09.07.2020 20:14, Vladimir Sementsov-Ogievskiy wrote:
09.07.2020 18:34, Eric Blake wrote:
On 7/9/20 3:35 AM, Daniel P. Berrangé wrote:
On Wed, Jul 08, 2020 at 10:15:37PM +0300, Vladimir Sementsov-Ogievskiy wrote:
Hi all!
We understood, that keepalive is almost superfluous with default 2 hours
in /proc/tcp_keepalive_time. Forcing user to setup keepalive for the
whole system doesn't seem right, better setup it per-socket.
Adding the ability to explicitly configure the keepalive settings makes
sense for QEMU. Completely ignoring system defaults when no explicit
settings are given though is not valid IMHO.
We already have the ability to add per-socket keepalive (see commit aec21d3175,
in 4.2). I guess what you are trying to further do is determine whether the
default value for that field, when not explicitly specified by the user, can
have saner semantics (default off for chardev sockets, default on for nbd
clients where retry was enabled). But since you already have to explicitly
opt-in to nbd retry, what's so hard about opting in to keepalive at the same
time, other than more typing? Given that the easiest way to do this is via a
machine-coded generation of the command line or QMP command, it doesn't seem
that hard to just keep things as they are with explicit opt-in to per-socket
keepalive.
Hmm. Looking at the code, I remember that reconnect is not optional, it works by default
now. The option we have is "reconnect-delay" which only specify, after how much
seconds we'll automatically fail the requests, not retrying them (0 seconds by default).
Still, NBD tries to reconnect in background anyway.
In our downstream we have now old version of nbd-reconnect interface and enabled non-zero
"delay" by default.
And now we need to migrate to upstream code. Hmm. So I have several options:
1. Set a downstream default for reconnect-delay to be something like 5min.
[most simple thing to do]
2. Set an upstream default to 5min. [needs a discussion, but may be useful]
3. Force all users (not customers I mean, but other teams (and even one another
company)) to set reconnect-delay option. [just for completeness, I will not go
this way]
So, what do you think about [2]? This includes:
- non-zero reconnect-delay by default, so requests will wait some time for
reconnect and retry
- enabled keep-alive and some keep-alive default parameters, to not hang in
recvmsg for a long time
Some other related ideas:
- non-blocking nbd_open
- move open to the coroutine
- use non-blocking socket connect (for example use
qio_channel_socket_connect_async, which runs connect in thread). Currently, if
you try to blockdev-add some unavailable nbd host, vm hangs during connect()
call which may be about a minute
- make blockdev-add to be async qmp command (Kevin's series)
- allow reconnect on open
also, recently I noted, that bdrv_close may hang due to reconnect: it does
bdrv_flush, which waits for NBD to reconnect.. This seems not convenient,
probably we should disable reconnect before bdrv_drained_begin() in
bdrv_close().
--
Best regards,
Vladimir
