Re: [PATCHv3] habanalabs: fix potential spectre v1 gadgets

[Andrew Cooper]

On 02/02/2022 19:11, Jordy Zomer wrote:
> diff --git a/drivers/misc/habanalabs/common/habanalabs_ioctl.c b/drivers/misc/habanalabs/common/habanalabs_ioctl.c
> index 3ba3a8ffda3e..c1cdf712a10d 100644
> --- a/drivers/misc/habanalabs/common/habanalabs_ioctl.c
> +++ b/drivers/misc/habanalabs/common/habanalabs_ioctl.c
> @@ -849,6 +850,7 @@ long hl_ioctl(struct file *filep, unsigned int cmd, unsigned long arg)
>  	}
>  
>  	if ((nr >= HL_COMMAND_START) && (nr < HL_COMMAND_END)) {
> +		nr = array_index_nospec(nr, HL_COMMAND_END);
>  		ioctl = &hl_ioctls[nr];
>  	} else {
>  		dev_err(hdev->dev, "invalid ioctl: pid=%d, nr=0x%02x\n",
> @@ -872,6 +874,7 @@ long hl_ioctl_control(struct file *filep, unsigned int cmd, unsigned long arg)
>  	}
>  
>  	if (nr == _IOC_NR(HL_IOCTL_INFO)) {
> +		nr = array_index_nospec(nr, _IOC_NR(HL_IOCTL_INFO)+1);
>  		ioctl = &hl_ioctls_control[nr];
>  	} else {
>  		dev_err(hdev->dev_ctrl, "invalid ioctl: pid=%d, nr=0x%02x\n",

Sorry for jumping in the middle here, but array_index_nospec() is
finicky to use, and both examples here don't quite behave as intended.

(In reverse order), first, any competent compiler will turn:

    if (nr == _IOC_NR(HL_IOCTL_INFO)) {
        ioctl = &hl_ioctls_control[nr];

into

        ioctl = &hl_ioctls_control[_IOC_NR(HL_IOCTL_INFO)];

which can be evaluated at compile time without reference to nr.  By
using array_index_nospec(), you inhibit this optimisation, forcing the
load to be dependent on nr, and with the extra logic to try and make nr
be safe.


Second, array_index_nospec(nr, HL_COMMAND_END); only protects the upper
bound.  In this case, the valid lower bound is not zero.

i.e. while you've prevented an OoB read of the hl_ioctls[] array as a
whole, there is still a Spectre-v1 gadget where you can speculatively
continue using &hl_ioctls[0] which is out of bounds from the logic's
point of view.

One fix is to use:

    nr = HL_COMMAND_START +
        array_index_nospec(nr, HL_COMMAND_END - HL_COMMAND_START);

while another is to confirm that the subsequent logic is safe to
confused-deputy problems.  This analysis probably needs doing anyway as
there are multiple ioctl implementations that can be selected.

~Andrew