| Message ID | 20210204203951.52105-13-marcan@marcan.st (mailing list archive) |
|---|---|
| State | Changes Requested |
| Headers | show |
| Series | Apple M1 SoC platform bring-up | expand |
On Thu, Feb 4, 2021 at 9:39 PM Hector Martin <marcan@marcan.st> wrote: > > This is a hack. I do not expect this to be merged as-is. > > The problem: on Apple ARM platforms, SoC MMIO needs to use nGnRnE > mappings: writes using nGnRE are blackholed. This seems to be by design, > and there doesn't seem to be any fabric configuration or other bit we > can flip to make the problem go away. > > Particularly tricky is that this affects earlycon, which uses fixmap, > which all gets initialized before any of the usual cpufeatures / > alternatives stuff. So we need to take care of fixmap very early. > > Options I can think of: > > (1) Unconditionally use nGnRnE on all platforms for fixmap IO. Maybe > this is actually fine? I suspect it might break some PCI-based > earlycons? > > (2) Deal with this special case in the earlycon code, since that seems > to be the only user that matters on these platforms. Since the > IO mapping is done in earlycon.c, this will require some cooperation > with samsung_tty.c so earlycon knows when it needs to do this. Note > that doing it with DT properties will break cmdline-only earlycon > config (which otherwise works fine on this driver). > > (3) This patch, but do something saner, like use a specific DT flag to > trigger this mode instead of a platform match. > > Any other ideas? I think we should decide the approach for the normal ioremap() path first. Once we have solved that, there is likely an obvious answer for this one as well. My feeling is that we'll end up with (2). Arnd
diff --git a/arch/arm64/include/asm/fixmap.h b/arch/arm64/include/asm/fixmap.h index 4335800201c9..999351ce84df 100644 --- a/arch/arm64/include/asm/fixmap.h +++ b/arch/arm64/include/asm/fixmap.h @@ -94,7 +94,15 @@ enum fixed_addresses { #define FIXADDR_SIZE (__end_of_permanent_fixed_addresses << PAGE_SHIFT) #define FIXADDR_START (FIXADDR_TOP - FIXADDR_SIZE) -#define FIXMAP_PAGE_IO __pgprot(PROT_DEVICE_nGnRE) +/* + * We use nGnRE by default, but some platforms require nGnRnE for MMIO. + */ +extern bool arm64_use_ne_io; + +#define FIXMAP_PAGE_IO_DEFAULT __pgprot(PROT_DEVICE_nGnRE) +#define FIXMAP_PAGE_IO_STRICT __pgprot(PROT_DEVICE_nGnRnE) +#define FIXMAP_PAGE_IO (arm64_use_ne_io ? FIXMAP_PAGE_IO_STRICT \ + : FIXMAP_PAGE_IO_DEFAULT) void __init early_fixmap_init(void); diff --git a/arch/arm64/kernel/setup.c b/arch/arm64/kernel/setup.c index c18aacde8bb0..cd2dc3bdbae4 100644 --- a/arch/arm64/kernel/setup.c +++ b/arch/arm64/kernel/setup.c @@ -56,6 +56,8 @@ static struct resource *standard_resources; phys_addr_t __fdt_pointer __initdata; +bool arm64_use_ne_io; + /* * Standard memory resources */ @@ -197,6 +199,16 @@ static void __init setup_machine_fdt(phys_addr_t dt_phys) pr_info("Machine model: %s\n", name); dump_stack_set_arch_desc("%s (DT)", name); + +#ifdef CONFIG_ARCH_APPLE + /* + * Apple SoCs need to use nGnRnE mappings for MMIO, and this needs + * to be detected before earlycon is initialized. + */ + if (of_flat_dt_is_compatible(of_get_flat_dt_root(), + "AAPL,arm-platform")) + arm64_use_ne_io = true; +#endif } static void __init request_standard_resources(void)
This is a hack. I do not expect this to be merged as-is. The problem: on Apple ARM platforms, SoC MMIO needs to use nGnRnE mappings: writes using nGnRE are blackholed. This seems to be by design, and there doesn't seem to be any fabric configuration or other bit we can flip to make the problem go away. Particularly tricky is that this affects earlycon, which uses fixmap, which all gets initialized before any of the usual cpufeatures / alternatives stuff. So we need to take care of fixmap very early. Options I can think of: (1) Unconditionally use nGnRnE on all platforms for fixmap IO. Maybe this is actually fine? I suspect it might break some PCI-based earlycons? (2) Deal with this special case in the earlycon code, since that seems to be the only user that matters on these platforms. Since the IO mapping is done in earlycon.c, this will require some cooperation with samsung_tty.c so earlycon knows when it needs to do this. Note that doing it with DT properties will break cmdline-only earlycon config (which otherwise works fine on this driver). (3) This patch, but do something saner, like use a specific DT flag to trigger this mode instead of a platform match. Any other ideas? Signed-off-by: Hector Martin <marcan@marcan.st> --- arch/arm64/include/asm/fixmap.h | 10 +++++++++- arch/arm64/kernel/setup.c | 12 ++++++++++++ 2 files changed, 21 insertions(+), 1 deletion(-)