On Wed, Nov 01, 2023 at 05:32:29PM +0100, Andrew Lunn wrote:
> > > > + for (pos = 0; pos < len; pos += min(sizeof(u32), len - pos)) {
> > > > + u32 word = 0;
> > > > +
> > > > + memcpy(&word, data + pos, min(sizeof(u32), len - pos));
> > >
> > > Rather than do a memcpy, use the get_unaligned_ macros. They might map
> > > to a memcpy(), but some architectures can do unaligned accesses
> > > without problems.
> > >
> >
> > I don't think this is doable for this loop, think we would end up in
> > some funny situation where for the last run we have to copy less than
> > u32. (get_unaligned would always take u32 of data and that would end up
> > reading more than requested) Am I wrong?
>
> Does it happen in practice that the last chunk is not 4 bytes? Since
> this is firmware, its probably produced by some sort of linker, and
> they often round segments to words. Could you take a look at the
> firmware images you have access to and see if this is true?
>
> It could be we do need to keep with the memcpy, but it would be nice
> if we could limit it to words, at least until somebody has a firmware
> which is not word aligned.
>
There are plenty of firmware around so it can be checked by from what I
have, it looks like they are word aligned... Ok I will use the
get_unaligned and add a comment saying that we assume the iram and dram
section are always word aligned.
Is it ok?
--
Ansuel
