[PATCH] mach-snapdragon: do carveouts for qcs404 only

[Sam Day]

Salutations Sumit,

On Wednesday, 8 May 2024 at 11:14 AM, Sumit Garg <sumit.garg at linaro.org> wrote:

> 
> 
> Hi Sam,
> 
> On Wed, 8 May 2024 at 00:11, Sam Day me at samcday.com wrote:
> 
> > The newly introduced carve_out_reserved_memory causes issues when
> > U-Boot is chained from the lk2nd bootloader. lk2nd provides a
> > simple-framebuffer device and marks the framebuffer region as no-map in
> > the supplied /reserved-memory. Consequently, the simple_video driver
> > triggers a page fault when it tries to write to this region.
> 
> 
> How does the corresponding Linux kernel driver handle this?

Firstly: I'm something of a middle-man here. I would consider Caleb the authoritative source on the carveouts stuff (since they wrote it) and Nikita Travkin the authority on the simplefb handoff (since he originally wrote it for lk2nd to hand off to Linux simpledrm and then adapted it to work with U-Boot simplefb).

I consulted with Nikita on your first question here. He linked me to this snippet: https://elixir.bootlin.com/linux/v6.9-rc7/source/drivers/gpu/drm/tiny/simpledrm.c#L877

>  Is the
> framebuffer region required to be mapped as normal memory or device
> type or something else?

So I guess based on the link above, it's just mapped as normal uncached memory.

I tried to do something like this a few days ago in U-Boot, but a) it doesn't work and b) I have no idea what I'm doing: https://github.com/samcday/u-boot/commit/c100cb3711ddf5b01601691f3e6a9ec890d9a496

After talking with Caleb about this for a bit, they suggested the patch you see here as what I guess could be considered a "stopgap" solution that hopefully makes it into 2024.07.

> Similarly would normal memory type work for
> all other reserved memory regions marked as no-map?

I'll let Caleb weigh in here. My understanding is that the other regions *should* be marked as PTE_TYPE_FAULT because otherwise drivers might inadvertently speculatively access regions that are very much off-limits, such as TZ app regions.

Cheers,
-Sam

> 
> -Sumit
> 
> > As per Caleb's advice, this simple patch only does the carveouts for the
> > qcs404 SoC for which it was originally designed. The intent is to do the
> > carveouts for more Qualcomm SoCs in future.
> > 
> > ---
> > I'm not sure if it's feasible to get this in for the 2024.07 release,
> > but it'd be great if we could - it's the only thing that breaks U-Boot
> > master on msm8916 devices that chain from lk2nd.
> > 
> > Signed-off-by: Sam Day me at samcday.com
> > ---
> > arch/arm/mach-snapdragon/board.c | 12 +++++++-----
> > 1 file changed, 7 insertions(+), 5 deletions(-)
> > 
> > diff --git a/arch/arm/mach-snapdragon/board.c b/arch/arm/mach-snapdragon/board.c
> > index 3d5994c878..b439a19ec7 100644
> > --- a/arch/arm/mach-snapdragon/board.c
> > +++ b/arch/arm/mach-snapdragon/board.c
> > @@ -467,10 +467,12 @@ void enable_caches(void)
> > gd->arch.tlb_addr = tlb_addr;
> > gd->arch.tlb_size = tlb_size;
> > 
> > - carveout_start = get_timer(0);
> > - /* Takes ~20-50ms on SDM845 /
> > - carve_out_reserved_memory();
> > - debug("carveout time: %lums\n", get_timer(carveout_start));
> > -
> > + / We do the carveouts only for QCS404, for now. /
> > + if (fdt_node_check_compatible(gd->fdt_blob, 0, "qcom,qcs404") == 0) {
> > + carveout_start = get_timer(0);
> > + / Takes ~20-50ms on SDM845 */
> > + carve_out_reserved_memory();
> > + debug("carveout time: %lums\n", get_timer(carveout_start));
> > + }
> > dcache_enable();
> > }
> > 
> > ---
> > base-commit: 1c40dda60f5f7e83a6d6f541cf5a57eb7e8ec43c
> > change-id: 20240507-qcs404-carveout-only-7a15bbf3fd89
> > 
> > Best regards,
> > --
> > Sam Day me at samcday.com