I see, I will test on my system.
Alexey Dobriyan <adobriyan@xxxxxxxxx> 于2022年3月21日周一 18:40写道：
>
> On Mon, Mar 21, 2022 at 05:15:02PM +0800, hui li wrote:
> > Alexey Dobriyan <adobriyan@xxxxxxxxx> 于2022年3月21日周一 00:24写道：
> > >
> > > When a process exits, /proc/${pid}, and /proc/${pid}/net dentries are flushed.
> > > However some leaf dentries like /proc/${pid}/net/arp_cache aren't.
> > > That's because respective PDEs have proc_misc_d_revalidate() hook which
> > > returns 1 and leaves dentries/inodes in the LRU.
> > >
> > > Force revalidation/lookup on everything under /proc/${pid}/net by inheriting
> > > proc_net_dentry_ops.
> > >
> > > Fixes: c6c75deda813 ("proc: fix lookup in /proc/net subdirectories after setns(2)")
> > > Reported-by: hui li <juanfengpy@xxxxxxxxx>
> > > Signed-off-by: Alexey Dobriyan <adobriyan@xxxxxxxxx>
> > > ---
> > >
> > > fs/proc/generic.c | 4 ++++
> > > fs/proc/proc_net.c | 3 +++
> > > 2 files changed, 7 insertions(+)
> > >
> > > --- a/fs/proc/generic.c
> > > +++ b/fs/proc/generic.c
> > > @@ -448,6 +448,10 @@ static struct proc_dir_entry *__proc_create(struct proc_dir_entry **parent,
> > > proc_set_user(ent, (*parent)->uid, (*parent)->gid);
> > >
> > > ent->proc_dops = &proc_misc_dentry_ops;
> > > + /* Revalidate everything under /proc/${pid}/net */
> > > + if ((*parent)->proc_dops == &proc_net_dentry_ops) {
> > > + pde_force_lookup(ent);
> > > + }
> > >
> > > out:
> > > return ent;
> > > --- a/fs/proc/proc_net.c
> > > +++ b/fs/proc/proc_net.c
> > > @@ -376,6 +376,9 @@ static __net_init int proc_net_ns_init(struct net *net)
> > >
> > > proc_set_user(netd, uid, gid);
> > >
> > > + /* Seed dentry revalidation for /proc/${pid}/net */
> > > + pde_force_lookup(netd);
> > > +
> > > err = -EEXIST;
> > > net_statd = proc_net_mkdir(net, "stat", netd);
> > > if (!net_statd)
>
> > proc_misc_dentry_ops is a general ops for dentry under /proc, except
> > for "/proc/${pid}/net"，other dentries may also use there own ops too,
> > so I think change proc_misc_d_delete may be better?
> > see patch under: https://lkml.org/lkml/2022/3/17/319
>
> I don't think so.
>
> proc_misc_d_delete covers "everything else" part under /proc/ and
> /proc/net which are 2 separate trees. Now /proc/net/ requires
> revalidation because of
>
> commit c6c75deda81344c3a95d1d1f606d5cee109e5d54
> proc: fix lookup in /proc/net subdirectories after setns(2)
>
> so the bug is that the above commit was applied only partially.
> In particular, /proc/*/net/stat/arp_cache was created with
> proc_create_seq_data(), avoiding proc_net_* APIs.
>
> And there is probably the same "lookup after setns find wrong file"
> if you search hard enough in /proc/*/net/
>
> This is the logic. Please test on your systems.
