[PATCH v2] thunderbolt: Do not RPM complete unrelated subtrees on unplug

[Jonathan Gopel <jgopel@gmail.com>]

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).

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
```

This patch gates the `complete()` on an additional condition - that the
Thunderbolt domain matches the domain of the device that was unplugged.

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.

Assisted by: LLM
Signed-off-by: Jonathan Gopel <jgopel@gmail.com>
---
 drivers/thunderbolt/icm.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/drivers/thunderbolt/icm.c b/drivers/thunderbolt/icm.c
index 669807f0eaf8..c1e6b2ea63cb 100644
--- a/drivers/thunderbolt/icm.c
+++ b/drivers/thunderbolt/icm.c
@@ -2071,9 +2071,9 @@ static void icm_unplug_children(struct tb_switch *sw)
 static int complete_rpm(struct device *dev, void *data)
 {
 	struct tb_switch *sw = tb_to_switch(dev);
-
-	if (sw)
+	if (sw && sw->tb == (struct tb *)data) {
 		complete(&sw->rpm_complete);
+	}
 	return 0;
 }
 
@@ -2088,8 +2088,8 @@ 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);
+	complete_rpm(&sw->dev, sw->tb);
+	bus_for_each_dev(&tb_bus_type, &sw->dev, sw->tb, complete_rpm);
 	tb_switch_remove(sw);
 
 	pm_runtime_mark_last_busy(parent);
-- 
2.55.0