Re: Possible race in spi-tegra114.ko

[Thierry Reding]

On Mon, Jul 24, 2017 at 03:30:36PM +0300, Anton Volkov wrote:
> Hello.
> While searching for races in Linux kernel I've come across
> drivers/spi/spi-tegra114.ko module. Here is the question that I came up with
> while analysing results. Lines are given using the info from Linux v4.12.
> 
> Consider the following case:
> Thread 1:                                  Thread 2:
> tegra_spi_probe
>   -> request_threaded_irq
>       (spi-tegra114.c: line 1070)
>   ------------interrupt comes------------- tegra_spi_isr_thread
>                                              -> handle_dma_based_xfer
>                                                   ->
> tegra_spi_start_dma_based_transfer
>                                                        ->
> tegra_spi_copy_client_txbuf_to_spi_txbuf
>                                                             ->
> dma_sync_single_for_cpu(tspi->dev, tspi->tx_dma_phys, ...) (spi-tegra114.c:
> line 368)
>   -> tspi->tx_dma_phys = ...
>      (spi-tegra114.c: line 621 or 625)
> 
> If the situation above is feasible then the value of tspi->tx_dma_phys
> (which is 0) is passed to dma_sync_single_for_cpu and further down the
> callstack. This is probably not good.
> Note that other fields of tspi structure can also be affected using the
> similar templates.
> Similar cases can occur also for tspi structure in
> drivers/spi/spi-tegra20-slink.ko module and for tsd structure in
> drivers/spi/spi-tegra20-sflash.ko module.
> So the question is: is there a possibility of interrupt triggering before
> the registration of master (spi-tegra114.c: line 1125) or a write to
> tspi->def_command1_reg (spi-tegra114.c: line 1121)?
> 
> Thank you for your time.

I suspect that it would be possible for an interrupt to be raised if the
bootloader had left it enabled and it was pending when the bootloader
passed control to the kernel. I'm not aware of any reports of this ever
happening, but that might just mean we've only dealt with reasonably
sane bootloaders so far.

I think the easiest would probably be to move the request_threaded_irq()
call to after everything required by the interrupt handler has been
initialized.

So the correct place would probably be right before the call to
pm_runtime_enable().

Laxman, any comments?

Thierry

Attachment: signature.asc
Description: PGP signature