Re: Possible race during OP-TEE kernel module device probing
Sumit Garg via OP-TEE <[email protected]> Mon, 30 Mar 2026 15:21:12 +0530
| Newsgroups | org.trustedfirmware.lists.op-tee |
|---|---|
| Message-ID | <acpHkOJhmCkMadHj@sumit-xelite> |
On Mon, Mar 30, 2026 at 11:16:59AM +0200, Jens Wiklander wrote: > Hi, > > On Mon, Mar 30, 2026 at 6:07 AM Shf Chen (陳少甫) via OP-TEE > <[email protected]> wrote: > > > > On Fri, 2026-03-27 at 15:31 +0530, Sumit Garg wrote: > > > On Fri, Mar 27, 2026 at 09:01:15AM +0000, Shf Chen (陳少甫) via OP-TEE wrote: > > > > Hi, > > > > > > > > We found a possible race condition issue during OP-TEE kernel driver probing the device. > > > > A NULL pointer dereference exception can happen when another kernel driver open OP-TEE context > > > > with tee_client_open_context() then do a SMC call to OP-TEE. > > > > Below is the exception: > > > > > > How is your kernel driver being probed? Have you registered it as a > > > proper TEE bus driver? Any custom ways to invoke TEE kernel client APIs > > > are surely susceptible to races like the one you mentioned below. > > > > > > -Sumit > > > > Hi, > > > > Our kernel driver does not registered as a TEE bus driver for some reasons, > > and we need to use other API to ensure the OP-TEE driver is probed to > > workaround this issue. But, we think this is error-prone for module developer > > that need to call to OP-TEE. The kernel modules API for OP-TEE is via TEE bus only. > > > > Although this is not the case, I think that after calling tee_device_register(), > > the optee device is shown under `/dev/` filesystem. There is a very small chance > > for userspace program to trigger this race condition, and they don't have clue on > > whether the device is properly being probed yet. > > I think you're on to something. After calling tee_device_register(), > the driver must be able to handle requests, at least without crashing. > optee_enumerate_devices() requires the devices to be registered, so > that must happen after those calls. But other than that, I think the > calls to tee_device_register() should be the last thing in the probe > function. > > Sumit, do you agree? > Sounds reasonable to me if it's something that can be reproduced via the user-space client. -Sumit