[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20231018085717.454931c3@hermes.local>
Date: Wed, 18 Oct 2023 08:57:17 -0700
From: Stephen Hemminger <stephen@...workplumber.org>
To: Antoine Tenart <atenart@...nel.org>
Cc: davem@...emloft.net, kuba@...nel.org, pabeni@...hat.com,
edumazet@...gle.com, netdev@...r.kernel.org, gregkh@...uxfoundation.org,
mhocko@...e.com
Subject: Re: [RFC PATCH net-next 0/4] net-sysfs: remove
rtnl_trylock/restart_syscall use
On Wed, 18 Oct 2023 17:47:42 +0200
Antoine Tenart <atenart@...nel.org> wrote:
> Hi,
>
> This is sent as an RFC because I believe this should be discussed (and
> some might want to do additional testing), but the code itself is ready.
>
> Some time ago we tried to improve the rtnl_trylock/restart_syscall
> situation[1]. What happens is when there is rtnl contention, userspace
> accessing net sysfs attributes will spin and experience delays. This can
> happen in different situations, when sysfs attributes are accessed
> (networking daemon, configuration, monitoring) while operations under
> rtnl are performed (veth creation, driver configuration, etc). A few
> improvements can be done in userspace to ease things, like using the
> netlink interface instead, or polling less (or more selectively) the
> attributes; but in the end the root cause is always there and this keeps
> happening from time to time.
>
> That initial effort however wasn't successful, although I think there
> was an interest, mostly because we found technical flaws and didn't find
> a working solution at the time. Some time later, we gave it a new try
> and found something more promising, but the patches fell off my radar. I
> recently had another look at this series, made more tests and cleaned it
> up.
>
> The technical aspect is described in patch 1 directly in the code
> comments, with an additional important comment in patch 3. This was
> mostly tested by stress-testing net sysfs attributes (read/write ops)
> while adding/removing queues and adding/removing veths, all in parallel.
>
> All comments are welcomed.
The trylock was introduced to deal with lock inversion.
It is not clear how this more complex solution prevents that.
Powered by blists - more mailing lists