Re: [PATCH v2 13/27] mfd: twl-core: Clear twl_priv pointer on teardown and probe failure

[Lee Jones <lee@kernel.org>]

On Fri, 30 Oct 2026, Andreas Kemnade wrote:

> On Fri, 30 Oct 2026 10:48:06 +0000
> Lee Jones <lee@kernel.org> wrote:
> 
> > On Wed, 28 Oct 2026, Andreas Kemnade wrote:
> > 
> > > On Tue, 27 Oct 2026 09:30:42 +0000
> > > Lee Jones <lee@kernel.org> wrote:
> > >   
> > > > twl_probe() allocates the file-scope twl_priv state with devm_kzalloc()
> > > > and checks for a non-NULL twl_priv at entry to reject multiple device
> > > > instances.  However, twl_remove() and the early error paths in
> > > > twl_probe() never reset twl_priv to NULL, a failure when adding MFD
> > > > child devices bypasses twl_remove() by jumping to free instead of fail,
> > > > and twl_remove() tears down the IRQ domain and dummy I2C clients before
> > > > child devices are unregistered.  After devres frees the underlying
> > > > memory on probe failure or unbind, twl_priv remains a dangling pointer,
> > > > causing subsequent exported API calls to dereference freed memory and
> > > > blocking future probe attempts with -EBUSY.
> > > > 
> > > > Unregister DT-populated and MFD child devices at the start of
> > > > twl_remove(), ensure mfd_add_devices() failures unwind via fail, and
> > > > reset twl_priv to NULL in twl_remove() and on probe error exit.
> > > > 
> > > > Fixes: 80a97ccd33be ("mfd: twl-core: Collect global variables behind one private structure (global)")
> > > > Closes: https://sashiko.dev/#/bug/linux-73e82f84-94de-4d08-82a4-f22d42a1dedf
> > > > Cc: stable@vger.kernel.org
> > > > Signed-off-by: Lee Jones <lee@kernel.org>  
> > > 
> > > Hmm, couldn't we use devm_add_action_or_reset() for resetting the
> > > twl_priv pointer? Feels more clean for me.  
> > 
> > Nice!  Thanks for the suggestion.
> > 
> > To keep this fix minimal for Stable backporting (and avoid resending a
> > 27-patch series), I'll take this one as-is and send a follow-up patch
> > converting twl_priv to use devm_add_action_or_reset().
> > 
> 
> I think the whole patch can be simplified a lot then, 
> 
> devm_mfd_add_devices can stay.

twl_remove() runs before devres teardown and destroys the IRQ domain,
unregisters the dummy I2C clients and clears twl_priv->ready.

So the MFD child devices still need to be unregistered at the start of
twl_remove() (alongside of_platform_depopulate()) before the parent
transport is torn down beneath them, at least until the rest of
twl_remove() is also converted to devm_*().

-- 
Lee Jones