[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <0c4e532ed8b58f8253a14f8ed59d93523a096f16.camel@kernel.org>
Date: Wed, 30 Apr 2025 21:07:20 -0700
From: Jeff Layton <jlayton@...nel.org>
To: Kuniyuki Iwashima <kuniyu@...zon.com>
Cc: airlied@...il.com, akpm@...ux-foundation.org, andrew@...n.ch,
davem@...emloft.net, dri-devel@...ts.freedesktop.org, edumazet@...gle.com,
horms@...nel.org, intel-gfx@...ts.freedesktop.org,
jani.nikula@...ux.intel.com, joonas.lahtinen@...ux.intel.com,
kuba@...nel.org, linux-kernel@...r.kernel.org,
maarten.lankhorst@...ux.intel.com, mripard@...nel.org, nathan@...nel.org,
netdev@...r.kernel.org, pabeni@...hat.com, qasdev00@...il.com,
rodrigo.vivi@...el.com, simona@...ll.ch, tursulin@...ulin.net,
tzimmermann@...e.de
Subject: Re: [PATCH v6 08/10] net: add symlinks to ref_tracker_dir for netns
On Wed, 2025-04-30 at 20:50 -0700, Kuniyuki Iwashima wrote:
> From: Jeff Layton <jlayton@...nel.org>
> Date: Wed, 30 Apr 2025 20:42:40 -0700
> > On Wed, 2025-04-30 at 20:07 -0700, Kuniyuki Iwashima wrote:
> > > From: Jeff Layton <jlayton@...nel.org>
> > > Date: Wed, 30 Apr 2025 19:59:23 -0700
> > > > On Wed, 2025-04-30 at 14:29 -0700, Kuniyuki Iwashima wrote:
> > > > > From: Jeff Layton <jlayton@...nel.org>
> > > > > Date: Wed, 30 Apr 2025 08:06:54 -0700
> > > > > > After assigning the inode number to the namespace, use it to create a
> > > > > > unique name for each netns refcount tracker with the ns.inum value in
> > > > > > it, and register a symlink to the debugfs file for it.
> > > > > >
> > > > > > init_net is registered before the ref_tracker dir is created, so add a
> > > > > > late_initcall() to register its files and symlinks.
> > > > > >
> > > > > > Signed-off-by: Jeff Layton <jlayton@...nel.org>
> > > > > > ---
> > > > > > net/core/net_namespace.c | 28 +++++++++++++++++++++++++++-
> > > > > > 1 file changed, 27 insertions(+), 1 deletion(-)
> > > > > >
> > > > > > diff --git a/net/core/net_namespace.c b/net/core/net_namespace.c
> > > > > > index 008de9675ea98fa8c18628b2f1c3aee7f3ebc9c6..6cbc8eabb8e56c847fc34fa8ec9994e8b275b0af 100644
> > > > > > --- a/net/core/net_namespace.c
> > > > > > +++ b/net/core/net_namespace.c
> > > > > > @@ -763,12 +763,38 @@ struct net *get_net_ns_by_pid(pid_t pid)
> > > > > > }
> > > > > > EXPORT_SYMBOL_GPL(get_net_ns_by_pid);
> > > > > >
> > > > > > +#ifdef CONFIG_NET_NS_REFCNT_TRACKER
> > > > > > +static void net_ns_net_debugfs(struct net *net)
> > > > > > +{
> > > > > > + ref_tracker_dir_symlink(&net->refcnt_tracker, "netns-%u-refcnt", net->ns.inum);
> > > > > > + ref_tracker_dir_symlink(&net->notrefcnt_tracker, "netns-%u-notrefcnt", net->ns.inum);
> > > > >
> > > > > Could you use net->net_cookie ?
> > > > >
> > > > > net->ns.inum is always 1 when CONFIG_PROC_FS=n.
> > > >
> > > > My main use-case for this is to be able to match the inode number in
> > > > the /proc/<pid>/ns/net symlink with the correct ref_tracker debugfs
> > > > file. Is there a way to use the net_cookie to make that association?
> > >
> > > It's roundabout, but net_cookie can be retrieved by creating a
> > > random socket in the netns and calling setsockopt(SO_NETNS_COOKIE).
> > >
> > > Ido proposed a handy ip-netns subcommand here, and I guess it will
> > > be implemented soon(?)
> > > https://lore.kernel.org/netdev/1d99d7ccfc3a7a18840948ab6ba1c0b5fad90901.camel@fejes.dev/
> >
> > For the cases where I was looking at netns leaks, there were no more
> > processes in the container, so there was no way to enter the container
> > and spawn a socket at that point.
>
> Then how do you get net->ns.inum ?
>
In my case, I was looking at /sys/kernel/debug/sunrpc/rpc_xprt/*/info.
That also displays net->ns.inum in the same format. When I was
originally working on this, the problem I was chasing was due to stuck
RPC transports (rpc_xprt).
>
> >
> > The point of the symlinks is to have a way to easily identify what
> > you're tracking. NAME_MAX is 255. We could do something like this
> > instead:
> >
> > snprintf(..., "netns-%u-%llx-refcnt", net->ns.inum, net->net_cookie);
> >
> > Obviously the inums would all be 1 when PROC_FS=n, but the cookies
> > would be unique. Would that work?
>
> This works, but depending on the question above, there's no point in
> using inum ?
Having the inum is definitely useful to me, particularly since the
net_cookie would be pretty useless for the problems I've been chasing.
I'll plan to respin this to include the net_cookie though, since that
would disambiguate the names when PROC_FS=n. It might also turn out to
be useful for someone, once there is a way to fetch the net_cookie from
userland.
--
Jeff Layton <jlayton@...nel.org>
Powered by blists - more mailing lists