| Message ID | 1443725506-20587-1-git-send-email-sjoerd.simons@collabora.co.uk |
|---|---|
| State | Superseded |
| Headers | show |
Hi, On 01-10-15 20:51, Sjoerd Simons wrote: > When malloc_base initially gets setup in the SPL it is based on the > current (early) stack pointer, which for rockchip is pointing into SRAM. > This means simple memory allocations happen in SRAM space, which is > somewhat unfortunate. Specifically a bounce buffer for the mmc allocated > in SRAM space seems to cause the mmc engine to stall/fail causing > timeouts and a failure to load the main u-boot image. > > To resolve this, reconfigure the malloc_base to start at the relocated > stack pointer after DRAM has been setup. > > For reference, things did work fine on rockchip before 596380db was > merged to fix memalign_simple due to a combination of rockchip SDRAM > starting at address 0 and the dw_mmc driver not checking errors from > bounce_buffer_start. As a result, when a bounce buffer needed to be > allocated mem_align simple would fail and return NULL. The mmc driver > ignored the error and happily continued with the bounce buffer address > being set to 0, which just happened to work fine.. > > Signed-off-by: Sjoerd Simons <sjoerd.simons@collabora.co.uk> Looks good to me, with the caveat that we really should do something better (and revert this one) for v2016.01: Reviewed-by: Hans de Goede <hdegoede@redhat.com> Regards, Hans > > --- > A potentially better fix for this issue would be to reconfigure the > malloc_base in spl_relocate_stack_gd following the same steps as is done > for the initial setup. However at this point in the release cycle i > preferred to do a minimal rockchip only fix (so those boards become > bootable again) for this issue to minimize the potential impact on other > boards. > > Changes in v2: > - Also gd->malloc_ptr to 0 to not waste any potential space > > arch/arm/mach-rockchip/board-spl.c | 5 +++++ > 1 file changed, 5 insertions(+) > > diff --git a/arch/arm/mach-rockchip/board-spl.c b/arch/arm/mach-rockchip/board-spl.c > index a241d96..982a1d9 100644 > --- a/arch/arm/mach-rockchip/board-spl.c > +++ b/arch/arm/mach-rockchip/board-spl.c > @@ -217,6 +217,11 @@ void board_init_f(ulong dummy) > debug("DRAM init failed: %d\n", ret); > return; > } > + > + /* Now that DRAM is initialized setup base pointer for simple malloc > + * into RAM */ > + gd->malloc_base = CONFIG_SPL_STACK_R_ADDR; > + gd->malloc_ptr = 0; > } > > static int setup_led(void) >
Hi, On 1 October 2015 at 12:51, Sjoerd Simons <sjoerd.simons@collabora.co.uk> wrote: > When malloc_base initially gets setup in the SPL it is based on the > current (early) stack pointer, which for rockchip is pointing into SRAM. > This means simple memory allocations happen in SRAM space, which is > somewhat unfortunate. Specifically a bounce buffer for the mmc allocated > in SRAM space seems to cause the mmc engine to stall/fail causing > timeouts and a failure to load the main u-boot image. > > To resolve this, reconfigure the malloc_base to start at the relocated > stack pointer after DRAM has been setup. > > For reference, things did work fine on rockchip before 596380db was > merged to fix memalign_simple due to a combination of rockchip SDRAM > starting at address 0 and the dw_mmc driver not checking errors from > bounce_buffer_start. As a result, when a bounce buffer needed to be > allocated mem_align simple would fail and return NULL. The mmc driver > ignored the error and happily continued with the bounce buffer address > being set to 0, which just happened to work fine.. > > Signed-off-by: Sjoerd Simons <sjoerd.simons@collabora.co.uk> > > --- > A potentially better fix for this issue would be to reconfigure the > malloc_base in spl_relocate_stack_gd following the same steps as is done > for the initial setup. However at this point in the release cycle i > preferred to do a minimal rockchip only fix (so those boards become > bootable again) for this issue to minimize the potential impact on other > boards. > > Changes in v2: > - Also gd->malloc_ptr to 0 to not waste any potential space > > arch/arm/mach-rockchip/board-spl.c | 5 +++++ > 1 file changed, 5 insertions(+) Acked-by: Simon Glass <sjg@chromium.org> > > diff --git a/arch/arm/mach-rockchip/board-spl.c b/arch/arm/mach-rockchip/board-spl.c > index a241d96..982a1d9 100644 > --- a/arch/arm/mach-rockchip/board-spl.c > +++ b/arch/arm/mach-rockchip/board-spl.c > @@ -217,6 +217,11 @@ void board_init_f(ulong dummy) > debug("DRAM init failed: %d\n", ret); > return; > } > + > + /* Now that DRAM is initialized setup base pointer for simple malloc > + * into RAM */ Little nit - can you fix the comment style? Or I can do that when applying if you prefer. > + gd->malloc_base = CONFIG_SPL_STACK_R_ADDR; > + gd->malloc_ptr = 0; > } > > static int setup_led(void) > -- > 2.5.3 > Regards, Simon
diff --git a/arch/arm/mach-rockchip/board-spl.c b/arch/arm/mach-rockchip/board-spl.c index a241d96..982a1d9 100644 --- a/arch/arm/mach-rockchip/board-spl.c +++ b/arch/arm/mach-rockchip/board-spl.c @@ -217,6 +217,11 @@ void board_init_f(ulong dummy) debug("DRAM init failed: %d\n", ret); return; } + + /* Now that DRAM is initialized setup base pointer for simple malloc + * into RAM */ + gd->malloc_base = CONFIG_SPL_STACK_R_ADDR; + gd->malloc_ptr = 0; } static int setup_led(void)
When malloc_base initially gets setup in the SPL it is based on the current (early) stack pointer, which for rockchip is pointing into SRAM. This means simple memory allocations happen in SRAM space, which is somewhat unfortunate. Specifically a bounce buffer for the mmc allocated in SRAM space seems to cause the mmc engine to stall/fail causing timeouts and a failure to load the main u-boot image. To resolve this, reconfigure the malloc_base to start at the relocated stack pointer after DRAM has been setup. For reference, things did work fine on rockchip before 596380db was merged to fix memalign_simple due to a combination of rockchip SDRAM starting at address 0 and the dw_mmc driver not checking errors from bounce_buffer_start. As a result, when a bounce buffer needed to be allocated mem_align simple would fail and return NULL. The mmc driver ignored the error and happily continued with the bounce buffer address being set to 0, which just happened to work fine.. Signed-off-by: Sjoerd Simons <sjoerd.simons@collabora.co.uk> --- A potentially better fix for this issue would be to reconfigure the malloc_base in spl_relocate_stack_gd following the same steps as is done for the initial setup. However at this point in the release cycle i preferred to do a minimal rockchip only fix (so those boards become bootable again) for this issue to minimize the potential impact on other boards. Changes in v2: - Also gd->malloc_ptr to 0 to not waste any potential space arch/arm/mach-rockchip/board-spl.c | 5 +++++ 1 file changed, 5 insertions(+)