[U-Boot] [PATCH 17/17] spi: Avoid using malloc() in a critical function

[Simon Glass]

Hi Urja,

On Sun, 19 May 2019 at 11:04, Urja Rannikko <urjaman at gmail.com> wrote:
>
> Hi,
>
> On Sat, May 18, 2019 at 6:05 PM Simon Glass <sjg at chromium.org> wrote:
> >
> > In general we should avoid calling malloc() and free() repeatedly in
> > U-Boot lest we turn it into tianocore. In SPL this can make SPI flash
> > unusable since free() is often a nop and allocation space is limited.
> >
> > In any case, these seems no need for malloc() since the number of bytes
> > is very small, perhaps less than 8.
> >
> > Signed-off-by: Simon Glass <sjg at chromium.org>
> > Fixes: d13f5b254a (spi: Extend the core to ease integration of SPI
> >         memory controllers)
> >
> > ---
> >
> >  drivers/spi/spi-mem.c | 15 +++++++++++----
> >  1 file changed, 11 insertions(+), 4 deletions(-)
> >
> > diff --git a/drivers/spi/spi-mem.c b/drivers/spi/spi-mem.c
> > index b86eee75bcb..7aabebeff5f 100644
> > --- a/drivers/spi/spi-mem.c
> > +++ b/drivers/spi/spi-mem.c
> > @@ -201,7 +201,6 @@ int spi_mem_exec_op(struct spi_slave *slave, const struct spi_mem_op *op)
> >         unsigned int pos = 0;
> >         const u8 *tx_buf = NULL;
> >         u8 *rx_buf = NULL;
> > -       u8 *op_buf;
> >         int op_len;
> >         u32 flag;
> >         int ret;
> > @@ -338,7 +337,17 @@ int spi_mem_exec_op(struct spi_slave *slave, const struct spi_mem_op *op)
> >         }
> >
> >         op_len = sizeof(op->cmd.opcode) + op->addr.nbytes + op->dummy.nbytes;
> > -       op_buf = calloc(1, op_len);
> > +
> > +       /*
> > +        * Avoid using malloc() here so that we can use this code in SPL where
> > +        * simple malloc may be used. That implementation does not allow free()
> > +        * so repeated calls to this code can exhaust the space.
> > +        *
> > +        * The value of op_len is small, since it does not include the actual
> > +        * data being sent, only the op-code and address. In fact, it should be
> > +        * possible to just use a small fixed value here instead of op_len.
> > +        */
> > +       u8 op_buf[op_len];
>
> I'd say just make this a fixed buffer instead of a VLA - less code
> space bloat and potential stack problems in case of nonsensical
> inputs.
>
> As for the size, 8 bytes would be fine and actually leave some margin:
> the most i would expect for op_len is 1 + 4 + 1 = 6 bytes (the lowest
> would be 1+3+0 and the usual 1+3+1).

Is it possible to add a build-time assert for this somewhere?

Is VLA very large array? How large is very large?

I think the stack is much better place than a fixed buffer for
something only 8 bytes in size.

Regards,
Simon