[U-Boot] [PATCH V2] spl: Weed out CONFIG_SYS_TEXT_BASE usage

[Simon Goldschmidt]

On Tue, Aug 14, 2018 at 8:17 PM Marek Vasut <marex at denx.de> wrote:
>
> On 08/14/2018 02:56 PM, Simon Goldschmidt wrote:
> > On Tue, Aug 14, 2018 at 2:05 PM Marek Vasut <marex at denx.de> wrote:
> >>
> >> On 08/14/2018 12:22 PM, Simon Goldschmidt wrote:
> >>> On Tue, Aug 14, 2018 at 12:10 PM Marek Vasut <marex at denx.de> wrote:
> >>>>
> >>>> On 08/14/2018 11:51 AM, Simon Goldschmidt wrote:
> >>>>> On Tue, Aug 14, 2018 at 11:27 AM Marek Vasut <marex at denx.de> wrote:
> >>>>>>
> >>>>>> The SPL loaders assume that the CONFIG_SYS_TEXT_BASE memory location
> >>>>>> is available and can be corrupted by loading ie. uImage or fitImage
> >>>>>> headers there. Sometimes it could be beneficial to load the headers
> >>>>>> elsewhere, ie. if CONFIG_SYS_TEXT_BASE is not yet writable while we
> >>>>>> still want to parse the image headers in some local onchip memory to
> >>>>>> ie. extract firmware from that image.
> >>>>>>
> >>>>>> Add the possibility to override the location where the headers get
> >>>>>> loaded by introducing new function, spl_get_load_buffer() which takes
> >>>>>> two arguments -- offset from the CONFIG_SYS_TEXT_BASE and size of the
> >>>>>> data that are to be loaded there -- and returns a valid buffer address
> >>>>>> or hangs the system. The default behavior is the same as before, add
> >>>>>> the offset to CONFIG_SYS_TEXT_BASE and return that address. User can
> >>>>>> override the weak spl_get_load_buffer() function though.
> >>>>>>
> >>>>>> Signed-off-by: Marek Vasut <marex at denx.de>
> >>>>>> Cc: Tom Rini <trini at konsulko.com>
> >>>>>> ---
> >>>>>> V2: Fix build issues on multiple boards due to incorrect pointer casting
> >>>>
> >>>> [...]
> >>>>
> >>>>>> diff --git a/common/spl/spl_ram.c b/common/spl/spl_ram.c
> >>>>>> index e594beaeaa..619b39a537 100644
> >>>>>> --- a/common/spl/spl_ram.c
> >>>>>> +++ b/common/spl/spl_ram.c
> >>>>>> @@ -63,8 +63,9 @@ static int spl_ram_load_image(struct spl_image_info *spl_image,
> >>>>>>                          * No binman support or no information. For now, fix it
> >>>>>>                          * to the address pointed to by U-Boot.
> >>>>>>                          */
> >>>>>> -                       u_boot_pos = CONFIG_SYS_TEXT_BASE -
> >>>>>> -                                       sizeof(struct image_header);
> >>>>>> +                       header = spl_get_load_buffer(-sizeof(*header),
> >>>>>> +                                                    sizeof(*header));
> >>>>>> +
> >>>>>
> >>>>> Using "spl_get_load_buffer()" here seems to be a bit misleading, as
> >>>>> the address is used for "execute-in-place", not for loading.
> >>>>
> >>>> Do you have a better solution ? Instead of hard-coding the load buffer
> >>>> address with some macro, this uses function which could be overridden.
> >>>> Whether it's XIP or not has nothing to do with it.
> >>>
> >>> I meant the name is a bit misleading as it implies loading. But since
> >>> the preferred way to do this is using binman, it's probably not worth
> >>> adding yet another weak function for RAM boot?
> >>
> >> Do you have a better name that fits all the other usecases ? This
> >> function just gets you buffer into which you can load the image.
> >
> > Not really. I just wonder if you have to override the location for
> > some board, RAM booting might not work any more as it relies on a
> > fixed address, not some generic buffer.
>
> I do, yeah, the board is not upstream completely yet though, so I am
> just sending this as a cleanup.
>
> > Maybe we could add the boot device to your new weak function? If we
> > add some comment to the new weak function, that would have made it
> > much more clear for me how to boot U-Boot from FPGA OnChip RAM when I
> > tried some months ago :-)
>
> This really just gives you a buffer. I don't need to know which boot
> media is used. If there is a usecase, sure, it can be added later.

I may have a use case for one of our custom boards, but it's probably
not worth doing that now.

>
> >>>>>>                 }
> >>>>>>                 header = (struct image_header *)map_sysmem(u_boot_pos, 0);
> >>>>>>
> >>>>>> diff --git a/common/spl/spl_spi.c b/common/spl/spl_spi.c
> >>>>>> index ba60a3a3c5..e10cf0124f 100644
> >>>>>> --- a/common/spl/spl_spi.c
> >>>>>> +++ b/common/spl/spl_spi.c
> >>>>>> @@ -88,8 +88,7 @@ static int spl_spi_load_image(struct spl_image_info *spl_image,
> >>>>>>                 return -ENODEV;
> >>>>>>         }
> >>>>>>
> >>>>>> -       /* use CONFIG_SYS_TEXT_BASE as temporary storage area */
> >>>>>> -       header = (struct image_header *)(CONFIG_SYS_TEXT_BASE);
> >>>>>> +       header = spl_get_load_buffer(-sizeof(*header), 0x40);
> >>>>>
> >>>>> Shouldn't the first argument be 0 here instead of -sizeof(*header)?
> >>>>
> >>>> No, because then the payload doesn't end up at CONFIG_SYS_TEXT_BASE .
> >>>
> >>> Sorry, I haven't studied the code around the patch, only the patch.
> >>> And I still think "header" has changed from "CONFIG_SYS_TEXT_BASE" to
> >>> "CONFIG_SYS_TEXT_BASE - sizeof(*header)" with your patch. That might
> >>> be a change required to get it work, I don't know that. But as this
> >>> isn' mentioned in the commit message, to me it seemed like a copy and
> >>> paste error or something.
> >>
> >> I suspect it's the SPI that's weird. Look at the surrounding code, IMO
> >> this is how it should be.
> >
> > Reading the code, I guess the exact location of 'header' is not
> > important. So the code should still work after applying your patch,
> > even if it changes the location of 'header'.
>
> The thing is, the payload (ie. uboot) is linked against the TEXT_BASE,
> so putting it at TEXT_BASE + offset can cause trouble.

Right. All I wanted to say is I saw something that changed but from
the commit message, it didn't seem like it was intended.
The rest seems fine to me right now, so if it's of any use:

Reviewed-by: Simon Goldschmidt <simon.k.r.goldschmidt at gmail.com>