Re: [PATCH] spl: Align device tree blob address at 8-byte boundary

[Tom Rini <trini@konsulko.com>]

[-- Attachment #1: Type: text/plain, Size: 5130 bytes --]

On Mon, Jul 12, 2021 at 05:51:29PM +0200, Marek Vasut wrote:
> On 7/12/21 5:43 PM, Tom Rini wrote:
> > On Mon, Jul 12, 2021 at 05:38:33PM +0200, Marek Vasut wrote:
> > > On 7/12/21 5:15 PM, Tom Rini wrote:
> > > > On Mon, Jul 12, 2021 at 01:36:14PM +0800, Bin Meng wrote:
> > > > > On Mon, Jul 12, 2021 at 1:21 PM Reuben Dowle <reuben.dowle@4rf.com> wrote:
> > > > > > 
> > > > > > I submitted an almost identical patch. See https://github.com/u-boot/u-boot/commit/eb39d8ba5f0d1468b01b89a2a464d18612d3ea76
> > > > > > 
> > > > > > This patch eventually had to be reverted (https://github.com/u-boot/u-boot/commit/5675ed7cb645f5ec13958726992daeeed16fd114), because it was causing issues on some platforms that had FIT on 32 bit boundary. However I continue to use it in production code, as without it the boot on my platform aborts.
> > > > > > 
> > > > > > I don't have time to investigate why this was happening, but you need to check this code won't just cause exactly the same faults.
> > > > > 
> > > > > Thanks for your information.
> > > > > 
> > > > > +Marek who did the revert
> > > > > 
> > > > > The revert commit message says:
> > > > > 
> > > > >       "The commit breaks booting of fitImage by SPL, the system simply
> > > > > hangs. This is because on arm32, the fitImage and all of its content
> > > > > can be aligned to 4 bytes and U-Boot expects just that."
> > > > > 
> > > > > I don't understand this. If an address is aligned to 8, it is already
> > > > > aligned to 4, so how did this commit make the system hang on arm32?
> > > > 
> > > > I think this had something to do with embedding contents somewhere in
> > > > the image?  There is a thread on the ML from then but I don't know how
> > > > informative it will end up being.
> > > 
> > > If I recall this correctly, DT node alignment is 4 byte and that is what DTC
> > > emits. If you have fitImage with embedded data, you basically end up with
> > > 
> > > / {
> > >   prop1 = "string1";
> > >   prop2 = "string2";
> > > };
> > > 
> > > where the "string2" is aligned to 4 bytes. And that is what U-Boot expects
> > > when it tries to access those data in-place in SPL.
> > > 
> > > The problem with the reverted patch was that it made U-Boot assume the
> > > alignment is 8 bytes, and that actually works only if you use fitImage with
> > > external data (mkimage -E), but with embedded data (mkimage default) not so
> > > much. That caused off-by-4 error in some cases and that made the SPL hang.
> > > 
> > > > > Note, as I indicated in this patch, now with libfdt 1.6.1, the
> > > > > alignment to 8 byte is a must-have. So we have to do such alignment
> > > > > anyway.
> > > > > 
> > > > > @Tom may fill in why libfdt commit commit 5e735860c478 ("libfdt: Check
> > > > > for 8-byte address alignment in fdt_ro_probe_()") was made to have the
> > > > > 8-byte alignment requirement.
> > > > 
> > > > Note that it's not so much since libfdt 1.6.1 but that since always the
> > > > device tree has required 8 byte alignment.
> > > 
> > > DT alignment was always 4 byte , no ?
> > 
> > I'm pretty sure, no, 8 byte base alignment is a pretty much always
> > thing.  I don't have a reference handy but I also know I couldn't have
> > convinced dgibson to add the check otherwise.
> 
> DTSpec rev 0.3 says the following and I think you got confused by section
> 5.6 which talks about alignment of the entire tree, not its nodes:
> 
> 5.4 Structure Block
> "
> Each token in the structure block, and thus the structure block itself,
> shall be located at a 4-byte aligned offset from the
> beginning of the devicetree blob (see 5.6).
> "
> 
> 5.4.2 Tree structure
> "
> For each property of the node:
> ...
> – FDT_PROP token
> ...
> * [zeroed padding bytes to align to a 4-byte boundary]
> "
> 
> 5.5 Strings Block
> "
> The strings block contains strings representing all the property names used
> in the tree. These null terminated strings are
> simply concatenated together in this section, and referred to from the
> structure block by an offset into the strings block.
> The strings block has no alignment constraints and may appear at any offset
> from the beginning of the devicetree blob.
> "
> 
> 5.6 Alignment
> "
> As described in the previous sections, the structure and strings blocks
> shall have aligned offsets from the beginning of
> the devicetree blob. To ensure the in-memory alignment of the blocks, it is
> sufficient to ensure that the devicetree as a
> whole is loaded at an address aligned to the largest alignment of any of the
> subblocks, that is, to an 8-byte boundary.

Right.  A device tree must start at an 8-byte boundary and U-Boot was
violating this both with:
- All of the boards that use fdt_high=0xffffffff to disable relocation,
  and then then place things at arbitrary spots in memory that may or
  may not violate these requirements.
- Perhaps some of the FIT internals where we have a device tree inside a
  device tree?  And we need to fixup whatever we're doing there that's
  wrong.

-- 
Tom

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 659 bytes --]