Re: [PATCH] tpm: Fix kexec crash due to access to ops NULL pointer (powerpc)

[Tyrel Datwyler <tyreld@linux.ibm.com>]

On 12/11/21 5:28 PM, Stefan Berger wrote:
> Fix the following crash on kexec by checking chip->ops for a NULL pointer
> in tpm_chip_start() and returning an error code if this is the case.
> 
> BUG: Kernel NULL pointer dereference on read at 0x00000060
> Faulting instruction address: 0xc00000000099a06c
> Oops: Kernel access of bad area, sig: 11 [#1]
> ...
> NIP [c00000000099a06c] tpm_chip_start+0x2c/0x140
>  LR [c00000000099a808] tpm_chip_unregister+0x108/0x170
> Call Trace:
> [c0000000188bfa00] [c000000002b03930] fw_devlink_strict+0x0/0x8 (unreliable)
> [c0000000188bfa30] [c00000000099a808] tpm_chip_unregister+0x108/0x170
> [c0000000188bfa70] [c0000000009a3874] tpm_ibmvtpm_remove+0x34/0x130
> [c0000000188bfae0] [c000000000110dbc] vio_bus_remove+0x5c/0xb0
> [c0000000188bfb20] [c0000000009bc154] device_shutdown+0x1d4/0x3a8
> [c0000000188bfbc0] [c000000000196e14] kernel_restart_prepare+0x54/0x70
> 
> The referenced patch below introduced a function to shut down the VIO bus.
> The bus shutdown now calls tpm_del_char_device (via tpm_chip_unregister)
> after a call to tpm_class_shutdown, which already set chip->ops to NULL.
> The crash occurrs when tpm_del_char_device calls tpm_chip_start with the
> chip->ops NULL pointer.

It was unclear to me at first, but IIUC the problem is the ibmvtpm device
receives two shutdown calls, the first is a class shutdown call for TPM devices,
followed by a bus shutdown call for VIO devices.

It appears that the class shutdown routines are meant to be a pre-shutdown
routine as they are defined as class->shutdown_pre(), and it is clearly allowed
to call class->shutdown_pre() followed by one of but not both of the following
bus->shutdown() or driver->shutdown(). Even tpm_class_shutdown() mentions in the
function comment that bus/device shutdown to follow.

> 
> Fixes: 39d0099f9439 ("powerpc/pseries: Add shutdown() to vio_driver and vio_bus")

This patch left implementing each vio driver shutdown routine as an exercise for
the respective maintainers, and instead just big hammers anything that doesn't
have a shutdown routine by calling the driver->remove().

If tpm_class_shutdown() quiecses ibmvtpm enough implementing a no-op
ibmvtpm->shutdown() with a comment saying so suffices.

However, the generic TPM code is still introducing a bug that an attempt to call
tpm_chip_unregister() after tpm_class_shutdown() will crash as mentioned above.

> Signed-off-by: Stefan Berger <stefanb@linux.ibm.com>
> ---
>  drivers/char/tpm/tpm-chip.c | 3 +++
>  1 file changed, 3 insertions(+)
> 
> diff --git a/drivers/char/tpm/tpm-chip.c b/drivers/char/tpm/tpm-chip.c
> index ddaeceb7e109..cca1bde296ee 100644
> --- a/drivers/char/tpm/tpm-chip.c
> +++ b/drivers/char/tpm/tpm-chip.c
> @@ -101,6 +101,9 @@ int tpm_chip_start(struct tpm_chip *chip)
>  {
>  	int ret;
> 
> +	if (!chip->ops)
> +		return -EINVAL;
> +

I wonder if a better fix would to have tpm_del_char_device() check for valid
chip->ops and call tpm_class_shutdown() when the ops are still valid. Calling
tpm_class_shutdown() allows for some code deduplication in tpm_del_char_device().

-Tyrel

>  	tpm_clk_enable(chip);
> 
>  	if (chip->locality == -1) {
>