Re: [PATCH] iommu/vt-d: replace snprintf with scnprintf in dmar_latency_snapshot()

[Will Deacon <will@kernel.org>]

On Wed, Jul 23, 2025 at 04:28:54AM -0400, Seyediman Seyedarab wrote:
> On 25/07/22 06:58PM, Will Deacon wrote:
> > On Tue, Jul 22, 2025 at 09:11:17AM -0400, Seyediman Seyedarab wrote:
> > > snprintf returns the number of bytes that would have been written,
> > > not the number actually written to the buffer. When accumulating
> > > the byte count with the return value of snprintf, this can cause
> > > the offset to exceed the actual buffer size if truncation occurs.
> > > 
> > > The byte count is passed to seq_puts() in latency_show_one() with-
> > > out checking for truncation.
> > > 
> > > Replace snprintf with scnprintf, ensuring the buffer offset stays
> > > within bound.
> > > 
> > > Signed-off-by: Seyediman Seyedarab <ImanDevel@gmail.com>
> > > ---
> > >  drivers/iommu/intel/perf.c | 6 +++---
> > >  1 file changed, 3 insertions(+), 3 deletions(-)

[...]

> > Should the check of the return value in latency_show_one() also be
> > adjusted so that 'ret <= 0' is an error? I couldn't convince myself
> > that the string in 'debug_buf' is always null-terminated if ret == 0.
> > 
> IMO, that's not necessary. 'bytes' can't be less than zero that's
> for sure (AFAIK, scnprintf() doesn't have any case where it returns
> a negative number).
> As for being zero, in every scnprintf() call, 'size - bytes' would
> have to be == 0. (or size > INT_MAX, but still you get zero, not a
> negative number as an error)
> 
> In latency_show_one(), the 'size' is DEBUG_BUFFER_SIZE, and
> 'bytes' in the first run is 0. So, 'size - bytes' == DEBUG_BUFFER_SIZE.
> Since 'latency_counter_names' and 'latency_type_names' are arrays of
> string literals, 'bytes' is guaranteed to be increased in the first
> iteration, even if the rest become zero (which won't happen, since
> they are smaller than DEBUG_BUFFER_SIZE).
> 
> So, the case of zero is impossible, unless you want a bulletproof
> check for future implementations where the function might be rewritten.

Thanks, so that sounds like the error check in latency_show_one() is dead
code in which case dmar_latency_snapshot() should just have a return type
of 'void'?

Will