[PATCH 1/2] bloblist: Introduce BLOBLIST_PRIOR_STAGE options

[Tom Rini]

On Sun, Dec 15, 2024 at 05:25:24PM -0700, Simon Glass wrote:
> Hi Tom,
> 
> On Fri, 13 Dec 2024 at 10:07, Tom Rini <trini at konsulko.com> wrote:
> >
> > On Fri, Dec 13, 2024 at 07:32:24AM -0700, Simon Glass wrote:
> > > Hi Tom,
> > >
> > > On Thu, 12 Dec 2024 at 08:11, Tom Rini <trini at konsulko.com> wrote:
> > > >
> > > > Introduce an option to control if we expect that a prior stage has
> > > > created a bloblist already and thus it is safe to scan the address. We
> > > > need to have this be configurable because we do not (and cannot) know
> if
> > > > the address at CONFIG_BLOBLIST_ADDR is in regular DRAM or some special
> > > > on-chip memory and so it may or may not be safe to read from this
> > > > address this early.
> > > >
> > > > Signed-off-by: Tom Rini <trini at konsulko.com>
> > > > ---
> > > > Cc: Simon Glass <sjg at chromium.org>
> > > > ---
> > > >  common/Kconfig | 32 ++++++++++++++++++++++++++++++++
> > > >  lib/fdtdec.c   | 11 +++--------
> > > >  2 files changed, 35 insertions(+), 8 deletions(-)
> > > >
> > >
> > > Since this is the essentially the same as my OF_BLOBLIST patch, which
> > > was rejected:
> > >
> > > NAK
> > >
> > > The issue is not whether some 'previous stage' set up a bloblist. For
> > > example, if VPL_BLOBLIST_PRIOR_STAGE is enabled, that presumably means
> > > that TPL set it up. But it doesn't mean that TPL put the FDT in there.
> >
> > This is wrong. The root problem is saying that the bloblist is in
> > possibly uninitialized memory. The code is quite happy to NOT find the
> > device tree in the bloblist and continue searching other paths.
> 
> But until the bloblist is set up, it doesn't exist.

An almost philosophical statement, yes. How should the generic code know
if it exists, or does not exist? That is the question.

> We should not be
> looking for a bloblist that we know is not there yet. The bloblist is set
> up by bloblist_init(). You seem to be imputing your own semantics for
> bloblist.

We don't know if the bloblist exists or not. That's the problem. It
could already exist. That's the whole reason we're in this particular
argument.

> In SPL, init happens in board_init_r(), i.e. after RAM is set up. So don't
> look in the bloblist until it is set up! It doesn't exist.

Unless of course it was setup before. We don't know if it was setup
before or not until we look. A big part of the problem is that for prior
to U-Boot bloblist we don't need to use main DRAM, the bloblist is tiny
and we have some other persistent memory available. Unfortunately, x86
does not.

> In U-Boot proper, we look for it before relocation, so we can relocate and
> expand it for use with ACPI tables, etc.

I'm not sure that's relevant. This means we've taken what we were passed
at a fixed address and allocated more space for it and are using it
somewhere else.

> Take a look at this patch, too:
> 
> 3d6531803e1 bloblist: Support initing from multiple places

I'm not sure it's relevant. It does show we have a problem of not
knowing if the bloblist exists, or not. And I don't think we're solving
that right today (nor does v1 of my patch).

> > An alternative here would be to better document the requirements of
> > BLOBLIST_ADDR, and then just disable it on chromebook_coral until
> > someone is inclined to move DDR init in to TPL, or someone finds a
> > non-main-memory address to use for BLOBLIST_ADDR or switches to using
> > BLOBLIST_ALLOC.
> 
> Coral's TPL runs in cache-as-ram, with a 30KB limit on code size. It cannot
> do DDR init in TPL because the DDR blob is large (160KB?) and there is a
> ton of setup to do first, in any case. When CAR goes away, its memory
> vanishes, so you need to have copied the bloblist somewhere else (the
> bloblist holds the MRC data which needs to be written to SPI flash later
> on, when possible). This all works perfectly and there is really no need to
> change anything.

Ah, so there's some of the details finally. And oh goodness, we're
writing to the SPI flash on each boot? That's not good for longevity...

> Again, you seem to be imputing semantics to bloblist which don't exist.

Unfortunately some things were missed along the way, and are still
missing.

-- 
Tom
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 659 bytes
Desc: not available
URL: <https://lists.denx.de/pipermail/u-boot/attachments/20241216/c9375872/attachment.sig>