Re: [PATCH] thunderbolt: Do not RPM complete unrelated subtrees on unplug

[Jonathan Gopel <jgopel@gmail.com>]

Hi Mika,

Thank you for the prompt response, it's much appreciated. I've left a
few questions inline.

On Sun, Oct 04, 2026 at 02:02:55PM +0200, Mika Westerberg wrote:
> Hi,
> 
> On Sat, Oct 03, 2026 at 08:26:04PM -0600, Jonathan Gopel wrote:
> > When a Thunderbolt device is unplugged, that switch and every switch
> > subsequent to it on the bus device list is `complete()`d, not just its
> > child devices. In the case of multiple domains, it may cause, eg, an
> > attempt to `complete()` the uninitialized domain 1 root switch RPM
> > completion on unplug of a switch from domain 0 (though ordering is not
> > guaranteed numerically sorted).
> 
> Good catch!
> 
> > I initially noticed this issue on a Dell XPS9720 when the system would
> > seem to sporadically not wake after unplugging a Thunderbolt dock. I
> > tried multiple kernel versions including v7.2.7 and v7.3-rc4 to see if
> > that resolved the issue, but the issue persisted. Eventually I found
> > that I could consistently generate a kernel Oops reporting a page fault
> > at 0xffff_ffff_ffff_fff8 on unplug from one side of the laptop and went
> > bug hunting. I became suspicious of the `bus_for_each_dev()` walk that I
> > ultimately ended up changing. To confirm this issue, I did an eBPF trace
> > of `bus_for_each_dev()`, `complete_rpm()`, and `complete()` while
> > unplugging in a variety of situations and found that an unplug from
> > domain 0 tries to `complete_rpm()` domain 1's switch, whose RPM
> > completion is not initialized and thus is holding a null wait-list
> > pointer, triggering the page fault.
> > 
> > Here are the key sections of that eBPF trace. Note that HEAD denotes the
> > completion's own wait-list HEAD:
> > 
> > ```
> > RESCAN_ENTER domain=0
> > 
> > COMPLETE rescan_domain=0 target=0-1 target_domain=0
> >          route=0x1 parent=0-0 unplugged=1
> > COMPLETION done=0 head=HEAD next=HEAD prev=HEAD
> > 
> > BUS_WALK domain=0 start=0-1 callback=complete_rpm
> > CALLBACK_ENTER rescan_domain=0 dev=1-0 parent=domain1 is_switch=1
> > 
> > COMPLETE rescan_domain=0 target=1-0 target_domain=1
> >          route=0x0 parent=domain1 unplugged=0
> > COMPLETION done=0 head=HEAD next=0 prev=0
> > 
> >         complete+5
> >         complete_rpm+43
> >         bus_for_each_dev+133
> >         icm_free_unplugged_children+250
> >         icm_rescan_work+42
> > ```
> > 
> > To fix this, I have reused some of the adjacent logic to recursively
> > walk and `complete()` only the unplugged switch and its descendants.
> > 
> > I only have a single system with multiple Thunderbolt domains - the Dell
> > XPS9720. I am consistently able to reproduce the Oops on that system.
> > Since running with this patch, I have not seen the Oops reoccur. I have
> > also tested it on a Thinkpad X1-Gen12, and that system boots and runs
> > well with it, but it only has a single Thunderbolt domain, so it does
> > not exercise the multi-domain failure case.
> 
> This looks like LLM was involved and if that's the case then please add
> Assisted-by: LLM.
> 

I'm happy to add an annotation here, but I was unclear on what the
policy was. I have a few other issues with this system since adding a
TB5 dock to my setup that I may also pursue patching, so if it's ok with
you I'd love to understand the expectation a bit better be expected.
I've read the policy, but I didn't come away with a clear understanding
of what exactly counts as assistance.

I had an LLM do the initial portion of the investigation here, basically
going from symptoms to identifying the Oops, which pointed at the
Thunderbolt unplug code. From there, I wanted to make sure I delivered a
high quality fix, so I did the remainder by hand. Does this warrant the
annotation?

> > 
> > Signed-off-by: Jonathan Gopel <jgopel@gmail.com>
> > ---
> >  drivers/thunderbolt/icm.c | 16 +++++++++-------
> >  1 file changed, 9 insertions(+), 7 deletions(-)
> > 
> > diff --git a/drivers/thunderbolt/icm.c b/drivers/thunderbolt/icm.c
> > index 669807f0eaf8..d6801d8b669b 100644
> > --- a/drivers/thunderbolt/icm.c
> > +++ b/drivers/thunderbolt/icm.c
> > @@ -2068,13 +2068,16 @@ static void icm_unplug_children(struct tb_switch *sw)
> >  	}
> >  }
> >  
> > -static int complete_rpm(struct device *dev, void *data)
> > +static void complete_rpm(struct tb_switch *sw)
> >  {
> > -	struct tb_switch *sw = tb_to_switch(dev);
> > +	struct tb_port *port;}
> 
> So instead of this, just check that sw->tb matches what you pass in..
> 
> >  
> > -	if (sw)
> > -		complete(&sw->rpm_complete);
> > -	return 0;
> > +	complete(&sw->rpm_complete);
> > +
> > +	tb_switch_for_each_port(sw, port) {
> > +		if (tb_port_has_remote(port))
> > +			complete_rpm(port->remote->sw);
> > +	}
> >  }
> >  
> >  static void remove_unplugged_switch(struct tb_switch *sw)
> > @@ -2088,8 +2091,7 @@ static void remove_unplugged_switch(struct tb_switch *sw)
> >  	 * tb_switch_remove() calls pm_runtime_get_sync() that then waits
> >  	 * for it.
> >  	 */
> > -	complete_rpm(&sw->dev, NULL);
> > -	bus_for_each_dev(&tb_bus_type, &sw->dev, NULL, complete_rpm);
> 
> .. here pass sw->tb 

I'm submitting a v2 patch with this revision. If you have time, I would
love to understand the reasoning for preferring one over the other - to
my eye they look equivalent.