Re: [PATCH] libfdt: fdt_check_full: Add can_assume(PERFECT) check

On Tue, May 26, 2026 at 10:38:59PM -0600, Tom Rini wrote:
> On Wed, May 27, 2026 at 02:23:20PM +1000, David Gibson wrote:
> > On Tue, May 26, 2026 at 02:30:22PM -0600, Tom Rini wrote:
> > > In this function from fdt_check.c we have (reasonably and as the name
> > > implies) a number of checks on the DTB. However, there are cases where
> > > we may wish to assume that we have been given a perfect DTB already and
> > > do nothing here. Add a test for can_assume(PERFECT) as the first check
> > > in this function and if true, perform no checks.
> > > 
> > > Signed-off-by: Tom Rini <[email protected]>
> > > ---
> > > Along the lines of the patches I posted back in December, in U-Boot SPL
> > > we just don't have the space for this check much of the time and so have
> > > always omitted it (going back to at least when Simon posted the initial
> > > patch to make libfdt/fdt_check.c here). This is another case where it's
> > > a noticeable size win for us. I had missed this change in particular
> > > because we had in turn missed catching up on fdt_check_full being moved
> > > out of fdt_ro.c and in to fdt_check.c.
> > 
> > I'm not necessarily against this, but I have some misgivings.
> > 
> > fdt_check_full() is (deliberately) not called from anywhere else in
> > libfdt - it's intended to allow the user to explicitly do a full
> > validity check on the tree.  Given that meaning, I'm not sure it's
> > wise to turn it into a no-op based on the assume flags.
> > 
> > Your comment seems to imply that the issue here is size - simply
> > having this function compiled - rather than being too expensive when
> > (explicitly) called.  That's a little surprising to me - it's in its
> > own compilation unit, specifically so that the linker can omit it if
> > it's not used.  Is there something unusual about your build
> > environment that's not letting that happen?
> 
> So, we have code like this:
>         /* Get the total space reserved for FDT in blob */
>         live_fdt = bloblist_get_blob(BLOBLISTT_CONTROL_FDT, &blob_size);
>         if (live_fdt != gd->fdt_blob)
>                 return -ENOENT;
> 
>         ret = fdt_check_full(live_fdt, blob_size);
>         if (ret)
>                 return fdtdec_ret_to_errno(ret);
> 
> And this is compiled on TPL, SPL and full U-Boot builds. On the first
> two, we're just too space constrained to do this check. So it's not the
> linker doing the right thing or not, it's avoiding having to #if the
> code directly (or rather, CONFIG_VAL(...)). My line of thinking was that
> since ASSUME_PERFECT is that everything is really perfect, this is the
> way to go. Yes, it's a little odd to have both "call the validation
> function" and "the validation function does not validate" but that's
> just because in the second case, we explicitly configured ourself to not
> validate anything.
> 
> And FWIW, that's really how we use the assume mask in U-Boot, either
> 0xff or 0x0. It's a case where we're either passing along the tree we
> bundled with ourself (and so we can assume it's fine) or it's passed
> along (and we verify).

Hey, just checking if there's any further discussion here? I want to try
and get the delta between U-Boot and upstream down to zero if we can.
Thanks!

-- 
Tom