Re: [PATCH v2] EDAC/thunderx: Fix some potential buffer overflow in thunderx_ocx_com_threaded_isr()

[Kees Cook]

On Sat, Oct 21, 2023 at 07:13:51PM +0200, Christophe JAILLET wrote:
> strncat() usage in thunderx_ocx_com_threaded_isr() is wrong.
> The size given to strncat() is the maximum number of bytes that can be
> written, excluding the trailing NULL.
> 
> Here, the size of the 'msg' buffer is used (i.e. OCX_MESSAGE_SIZE), not
> the space that is remaining.
> The space for the ending NULL is also not taken into account.
> 
> in order to fix it:
>    - call decode_register() before the snprintf() calls
>    - use scnprintf() instead of snprintf() and compute, in the 'remaining'
>      variable, the space that is still available in the 'msg' buffer
>    - add a %s at the end of the format strings and append directly the
>      result of decode_register() stored in 'other'
>    - write directly at the right position in the 'msg' buffer when
>      appending some data in the for loop.
> 
> Doing so, all usages of strncat() are removed.
> 
> Fixes: 41003396f932 ("EDAC, thunderx: Add Cavium ThunderX EDAC driver")
> Signed-off-by: Christophe JAILLET <christophe.jaillet@xxxxxxxxxx>
> ---
> This is cross-compile tested only.
> Review with care.
> 
> v2: remove some other erroneous usage of strncat()
> ---
>  drivers/edac/thunderx_edac.c | 44 ++++++++++++++++--------------------
>  1 file changed, 20 insertions(+), 24 deletions(-)
> 
> diff --git a/drivers/edac/thunderx_edac.c b/drivers/edac/thunderx_edac.c
> index b9c5772da959..62e1e120178b 100644
> --- a/drivers/edac/thunderx_edac.c
> +++ b/drivers/edac/thunderx_edac.c
> @@ -1111,6 +1111,7 @@ static irqreturn_t thunderx_ocx_com_threaded_isr(int irq, void *irq_id)
>  
>  	unsigned long tail;
>  	struct ocx_com_err_ctx *ctx;
> +	size_t remaining;
>  	int lane;
>  	char *msg;
>  	char *other;
> @@ -1127,27 +1128,26 @@ static irqreturn_t thunderx_ocx_com_threaded_isr(int irq, void *irq_id)
>  				ARRAY_SIZE(ocx->com_err_ctx));
>  		ctx = &ocx->com_err_ctx[tail];
>  
> -		snprintf(msg, OCX_MESSAGE_SIZE, "%s: OCX_COM_INT: %016llx",
> -			ocx->edac_dev->ctl_name, ctx->reg_com_int);
> -
>  		decode_register(other, OCX_OTHER_SIZE,
>  				ocx_com_errors, ctx->reg_com_int);
>  
> -		strncat(msg, other, OCX_MESSAGE_SIZE);
> +		remaining = OCX_MESSAGE_SIZE;
> +		remaining -= scnprintf(msg, remaining, "%s: OCX_COM_INT: %016llx%s",
> +				       ocx->edac_dev->ctl_name, ctx->reg_com_int,
> +				       other);

As the replacements get longer, I would encourage you to use seq_buf
instead -- it does all the length math internally. For example:

	seq_buf s;

        msg = kmalloc(OCX_MESSAGE_SIZE, GFP_KERNEL);
	seq_buf_init(&s, msg, OCX_MESSAGE_SIZE);


	seq_buf_printf(&s, "%s: OCX_COM_INT: %016llx%s",
		       ocx->edac_dev->ctl_name,
		       ctx->reg_com_int, other);

	...
  		for (lane = 0; lane < OCX_RX_LANES; lane++)
  			if (ctx->reg_com_int & BIT(lane)) {
				...

				seq_buf_printf(&s, "\n\tOCX_LNE_INT[%02d]: %016llx OCX_LNE_STAT11[%02d]: %016llx%s",
	                                       lane, ctx->reg_lane_int[lane],
	                                       lane, ctx->reg_lane_stat11[lane],
	                                       other);
		}

	...
	seq_buf_terminate(&s);


etc...

Though I think the seq_buf API could use some improvement: it should
always maintainer a %NUL-terminated string so the final
seq_buf_terminate() isn't needed...

-- 
Kees Cook