Re: [PATCH V2 16/20] accel/amdxdna: Finalize runtime PM before acquiring dev_lock on removal

[Eva Crystal]

└─$ sed -n '/^You are right/,$p'
/home/iviel/xdna-audit/aie4-kmq-v2-reply-16-retraction.md
You are right, and I withdraw the claim.

The step I missed is pci_device_remove(), which calls
pm_runtime_get_sync() at drivers/pci/pci-driver.c:527 before
drv->remove() at :535 and only drops it at :536 once remove returns.
Once that get succeeds, the device is RPM_ACTIVE for all of
amdxdna_remove(), so pm_runtime_forbid() takes the early return in
rpm_resume() at drivers/base/power/runtime.c:820 and never fetches the
->runtime_resume callback at :936 at all. No second dev_lock
acquisition, no deadlock.

The AIE2 deadlock claim is therefore wrong, and the Fixes:
1aa82181a3c2, Cc: stable and split-to-drm-misc-fixes suggestions go
with it. I withdraw the drm_dev_unplug() question too, since nothing
resumes on that path.

I had not reproduced this before sending. Sorry for the noise.

Eva Crystal (0xiviel)
XSource Security
https://xsourcesec.com