Re: [PATCH v3 1/6] tpm: Move buffer handling from static inlines to real functions

[Jerry Snitselaar]

On Thu, Oct 26, 2023 at 01:55:55PM -0400, James Bottomley wrote:
> On Thu, 2023-10-26 at 10:10 -0700, Jerry Snitselaar wrote:
> > On Wed, Oct 25, 2023 at 08:35:55PM +0300, Jarkko Sakkinen wrote:
> > > On Wed Oct 25, 2023 at 12:03 PM EEST, Jerry Snitselaar wrote:
> > > > Reviewed-by: Jerry Snitselaar <jsnitsel@xxxxxxxxxx>
> > > 
> > > On Wed, 2023-10-25 at 02:03 -0700, Jerry Snitselaar wrote:
> > > > Reviewed-by: Jerry Snitselaar <jsnitsel@xxxxxxxxxx>
> > > > 
> > > 
> > > Thanks I'll add it to the next round.
> > > 
> > > For the tpm_buf_read(), I was thinking along the lines of:
> > > 
> > > /**
> > >  * tpm_buf_read() - Read from a TPM buffer
> > >  * @buf:        &tpm_buf instance
> > >  * @pos:        position within the buffer
> > >  * @count:      the number of bytes to read
> > >  * @output:     the output buffer
> > >  *
> > >  * Read bytes from a TPM buffer, and update the position. Returns
> > > false when the
> > >  * amount of bytes requested would overflow the buffer, which is
> > > expected to
> > >  * only happen in the case of hardware failure.
> > >  */
> > > static bool tpm_buf_read(const struct tpm_buf *buf, off_t *pos,
> > > size_t count, void *output)
> > > {
> > >         off_t next = *pos + count;
> > > 
> > >         if (next >= buf->length) {
> > >                 pr_warn("%s: %lu >= %lu\n", __func__, next,
> > > *offset);
> > >                 return false;
> > >         }
> > > 
> > >         memcpy(output, &buf->data[*pos], count);
> > >         *offset = next;
> > >         return true;
> > > }
> > > 
> > > BR, Jarkko
> > > 
> > 
> > Then the callers will check, and return -EIO?
> 
> Really, no, why would we do that?
> 
> The initial buffer is a page and no TPM currently can have a command
> that big, so if the buffer overflows, it's likely a programming error
> (failure to terminate loop or something) rather than a runtime one (a
> user actually induced a command that big and wanted it to be sent to
> the TPM).  The only reason you might need to check is the no-alloc case
> and you passed in a much smaller buffer, but even there, I would guess
> it will come down to a coding fault not a possible runtime error.
> 
> James
> 

I was clarifying based on this exchange below between Jarkko and Mario discussing patch 5/6.
>From https://lore.kernel.org/all/CWGM2YH00DJ3.JKSYNNEWVRW4@suppilovahvero/ :


> > In the overflow case wouldn't you want to pass an error code up instead 
> > of just showing a WARN trace?
> >
> > The helper functions can't tell the difference, and the net outcome is 
> > going to be that if there is overflow you get a warning trace in the 
> > kernel log and whatever garbage "value" happened to have going to the 
> > caller.  It's a programmer error but it's also unpredictable what 
> > happens here.
> >
> > I think it's cleaner to have callers of 
> > tpm_buf_read_u8/tpm_buf_read_u16/tpm_buf_read_u32 to to be able to know 
> > something wrong happened.
>
> I think you have a fair point here and I also think it is also a bigger
> issue for the response parsing than programmer error. I.e. faulty or
> malicious TPM could return corrupted data, which makes WARN() wrong
> choice.
>
> So, as a corrective measure I think it should be pr_warn() instead, and
> instead of returning u8/u16/u32, all functions should return 'ssize_t'
> and -EIO in the case of overflow.
>
> Thank you, it was a really good catch.
>
> BR, Jarkko