On Fri, May 24, 2024 at 11:25 AM Lars Kellogg-Stedman [off-list ref] wrote:

On Thu, May 23, 2024 at 04:39:27PM GMT, Dan Cross wrote:

quoted

On Thu, May 23, 2024 at 2:23 PM Dan Carpenter [off-list ref] wrote:

quoted

The problem is that accept() and ax25_release() are not mirrored pairs.

Right, but my in making this patch I wasn't thinking so much about
accept/ax25_release, which as you say are not necessarily a mirrored
pair...

quoted

It seems clear that this will happen for sockets that have a ref on
the device either via `bind` or via `accept`.

...but rather bind/accept, which *are*. The patch I've submitted gives
us equivalent behavior on the code paths for inbound and outbound
connections.

Without this change, the ax.25 subsystem is completely broken. Maybe we
can come up with a more correct fix down the road, or maybe we'll
refactor all the things, but I would prefer to return the subsystem to a
usable state while we figure that out.

I think the main thrust of my point yesterday was that your
accept/release _are_ paired, via `close`. It seems logical to me that
every active socket associated with a device should hold a ref against
that device, and your change does that.

Since I wrote yesterday, I've enabled a bunch of additional debugging
code in my kernel (ref tracking, memory leak detection, etc). I can
detect no evidence of leaks, despite trying several times to induce
them. Perhaps I'm not looking correctly? Dan Carpenter seems fairly
convinced that there are, but if that were the case, we ought to be
able to detect some evidence of it, no?

Absent evidence to the contrary, I believe that the patch you
submitted for integration is correct. There may be other bugs lurking
in that general area, but if so, they're only tangentially related, if
at all.

        - Dan C.