Re: [PATCH] net: linkwatch: ignore events for unregistered netdevs

[Jann Horn <jannh@google.com>]

On Mon, Apr 25, 2022 at 5:13 PM Eric Dumazet <edumazet@google.com> wrote:
> On Mon, Apr 25, 2022 at 8:01 AM Jakub Kicinski <kuba@kernel.org> wrote:
> >
> > On Mon, 25 Apr 2022 16:49:34 +0200 Jann Horn wrote:
> > > > Doesn't mean we should make it legal. We can add a warning to catch
> > > > abuses.
> > >
> > > That was the idea with
> > > https://lore.kernel.org/netdev/20220128014303.2334568-1-jannh@google.com/,
> > > but I didn't get any replies when I asked what the precise semantics
> > > of dev_hold() are supposed to be
> > > (https://lore.kernel.org/netdev/CAG48ez1-OyZETvrYAfaHicYW1LbrQUVp=C0EukSWqZrYMej73w@mail.gmail.com/),
> > > so I don't know how to proceed...
> >
> > Yeah, I think after you pointed out that the netdev per cpu refcounting
> > is fundamentally broken everybody decided to hit themselves with the
> > obliviate spell :S
>
> dev_hold() has been an increment of a refcount, and dev_put() a decrement.
>
> Not sure why it is fundamentally broken.

Well, it's not quite a refcount. It's a count that can be incremented
and decremented but can't be read while the device is alive, and then
at some point it turns into a count that can be read and decremented
but can't be incremented (see
https://lore.kernel.org/netdev/CAG48ez1-OyZETvrYAfaHicYW1LbrQUVp=C0EukSWqZrYMej73w@mail.gmail.com/).
Normal refcounts allow anyone who is holding a reference to add
another reference.

> There are specific steps at device dismantles making sure no more
> users can dev_hold()

So you're saying it's intentional that even if you're already holding
a dev_hold() reference, you may not be allowed to call dev_hold()
again?