[U-Boot] [PATCH 4/5] tools: mkimage: fix imximage header size

[Albert ARIBAUD]

Hello Stefan (and sorry for the duplicate),

On Wed, 15 Jul 2015 12:41:59 +0200, Stefan Agner <stefan at agner.ch>
wrote:
> Hi Albert,
> 
> On 2015-07-15 09:54, Albert ARIBAUD wrote:
> > Hello Stefano,
> > 
> > On Wed, 15 Jul 2015 09:19:55 +0200, Stefano Babic <sbabic at denx.de>
> > wrote:
> > 
> >> > I think my patch back then solves the issue nicer. The struct dcd_v2_t
> >> > is not the reason the whole header is not aligned by 8-byte, it is a
> >> > problem of the boot_data_t header which is 12 bytes long. So inserting
> >> > the padding just after struct boot_data_t (or inside of struct
> >> > boot_data_t) seems to be more appropriate.
> >>
> >> I see. Albert, can you test on your board if Stefan's patch solves your
> >> issue, too ? I could then revert this one and apply Stefan's.
> > 
> > I can test it, but I'm sure that boot data does not need size alignment
> > since in my case there is no such alignment and Vybrid boots. To me
> > this means aligning boot_data_t size would introduce a constraint that
> > is not really there.
> 
> I quote here from both emails, since its about the same topic.
> 
> > After reading the U-Boot and Freescale discussions, IIUC you have come
> > to the conclusion that the missing 4 bytes were in boot_data_t because
> > it was the only structure in the header which was not 8-bytes-aligned.
> > 
> > However, the available documentation does not specify this constraint,
> > and from a more experimental vewpoint, my patch adds 4 bytes at the end
> > of the overall header, thus leaving boot data at 12 bytes (therefore
> > leaving the dcd_table not 8-byte-aligned) and yet Vybrid can boot.
> 
> Yes, I agree on that. Also the boot data seems not to have the 8-byte
> alignment requirement (see the graphics I posted on the Freescale
> community article).
> 
> > To me, this proves that the size alignment problem is not with
> > boot_data_t but with the overall header.
> 
> I agree. Still, boot_data_t is the "offender", hence I would prefer that
> solution over your solution.
> 
> > If Stefan would test my patch as well, I reckon he would find it to
> > work as well as his.
> > 
> > So we have two patches which fix Vybrid booting:
> > 
> > - one which aligns the boot_data_t /size/ but keeps its /offset/
> >   unaligned;
> > 
> > - one which aligns the the boot_data_t /offset/ but keeps its /size/
> >   unaligned.
> 
> I don't get this classification. The complete header struct shows that
> dcd_v2_t is _after_ boot_data_t.
> 
> typedef struct {
> 	flash_header_v2_t fhdr;
> 	boot_data_t boot_data;
> 	dcd_v2_t dcd_table;
> } imx_header_v2_t;
> 
> Your patch changes the size of dcd_v2_t, hence compensating the
> "unaligned" size of boot_data_t.

Compensating the unaligned size of the overall header for sure, since
I'm adding a last field in the last structure. But precisely since I'm
adding 4 bytes at the very end of the construct, we cannot tell whether
I fixed the 'size' of the flash header, boot data or DCD.

What we can tell, though, is that Vybrid will boot as long as we
increase the size of the overall header to an 8-byte boundary, and also
that it will boot regardless to whether these 4 bytes are added between
boot data and DCD or after DCD.

(note, btw, that in the Vybrid RM, the IVT itself, and the DCD, contain
their own size explicitly in their tag-length-version headers, but the
boot data does not contain its own size, nor is it contained elsewhere.
What we call 'the size of the boot data' is 'the difference between the
boot data and dcd addresses', but that's a definition for convenience.)

> As far as I see we have these two patches which fix Vybrid booting:
> 
> - one which changes the boot_data_t /size/ which keeps the offset
>   of dcd_v2_t and the image aligned.
> 
> - one which changes the dcd_v2_t /size/ which compensates the unaligned
>   size of boot_data_t and keeps the image aligned.
>  
> > Seems to me that the conclusion is that the actual alignment of
> > boot_data_t does not matter and that only the alignment of the
> > whole imx_header_v2_t size (and, consequently, offset) matters.
> > 
> > How about just adding an attribute((align(8))) to imx_header_v2_t?
> 
> Hm this sounds tempting, but it does not seem to work here. I think
> because it only aligns the beginning of imx_header_v2_t in to 8-byte,
> however it does not align the size of the whole struct to 8 bytes, I
> guess? Hence the header size is still "unaligned".

Correct -- my fault, this aligns the address, not size, and there is
no attribute that will control size alignment.

So we really need to add manual padding.

And if we want to not introduce any artificial constaints, I suggest we
define the padding as a separate field in the overall structure, i.e.:

typedef struct {
        flash_header_v2_t fhdr;
        boot_data_t boot_data;
        dcd_v2_t dcd_table;
	uint32_t padding; /* make size an 8-byte multiple */
} imx_header_v2_t;

I think this is the 'least inexact' solution: it does not enforce any
address or size alignment constraint that is not defined in the RM, and
it does show that the constraint is not on the boot data or DCD but on
the header as a whole.

Functionally, it is identical to what I did, so I'm pretty sure it
works. :)

> >> It remains doubious how the ROMs on Freescale are interpretating the
> >> header, but we can only test it.
> > 
> > See my other answer. We could prove it by disassembling the ROM code.
> > Any volunteer? :)
> 
> Hehe, not me :-)

:)

> --
> Stefan

Amicalement,
-- 
Albert.