Message ID | 1443690640-7568-1-git-send-email-sjoerd.simons@collabora.co.uk |
---|---|
State | New |
Headers | show |
Hi Sjoerd, On 01-10-15 11:10, 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> > > --- > 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. I actually have a patch series pending for this: http://patchwork.ozlabs.org/patch/517191/ http://patchwork.ozlabs.org/patch/517194/ http://patchwork.ozlabs.org/patch/517193/ http://patchwork.ozlabs.org/patch/517195/ http://patchwork.ozlabs.org/patch/517196/ (I've omitted 2 uninteresting patches) Your review of / input on this series would be appreciated. > 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. I agree that a minimal rockchip only fix likely is best at this time, however your fix seems wrong: > arch/arm/mach-rockchip/board-spl.c | 4 ++++ > 1 file changed, 4 insertions(+) > > diff --git a/arch/arm/mach-rockchip/board-spl.c b/arch/arm/mach-rockchip/board-spl.c > index a241d96..5daced7 100644 > --- a/arch/arm/mach-rockchip/board-spl.c > +++ b/arch/arm/mach-rockchip/board-spl.c > @@ -217,6 +217,10 @@ 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; > } > > static int setup_led(void) SPL_STACK_R_ADDR is where the stack will be put by spl_relocate_stack_gd so now you've the stack and the heap overlapping. You likely will want to make CONFIG_SPL_STACK_R_ADDR be something like 0x80000 and then do: gd->malloc_base = 0x00000000; gd->malloc_limit = CONFIG_SPL_STACK_R_ADDR; gd->malloc_ptr = 0; Here to stop the 2 from overlapping, things are likely working with your patch since you're not resetting gd->malloc_ptr, so you keep space for the stack for whatever that is set to at that point. Regards, Hans
Hey Hans, On Thu, 2015-10-01 at 12:08 +0200, Hans de Goede wrote: > Hi Sjoerd, > > On 01-10-15 11:10, 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> > > > > --- > > 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. > > I actually have a patch series pending for this: > > http://patchwork.ozlabs.org/patch/517191/ > http://patchwork.ozlabs.org/patch/517194/ > http://patchwork.ozlabs.org/patch/517193/ > http://patchwork.ozlabs.org/patch/517195/ > http://patchwork.ozlabs.org/patch/517196/ > > (I've omitted 2 uninteresting patches) > > Your review of / input on this series would be appreciated. Cool, I'll try to make some time to give that a closer look. > > 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. > > I agree that a minimal rockchip only fix likely is best at this time, > however your fix seems wrong: > > > arch/arm/mach-rockchip/board-spl.c | 4 ++++ > > 1 file changed, 4 insertions(+) > > > > diff --git a/arch/arm/mach-rockchip/board-spl.c b/arch/arm/mach > > -rockchip/board-spl.c > > index a241d96..5daced7 100644 > > --- a/arch/arm/mach-rockchip/board-spl.c > > +++ b/arch/arm/mach-rockchip/board-spl.c > > @@ -217,6 +217,10 @@ 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; > > } > > > > static int setup_led(void) > > SPL_STACK_R_ADDR is where the stack will be put by > spl_relocate_stack_gd > so now you've the stack and the heap overlapping. If i'm not mistaken the stack grows downward, while the heap grows upwards so there shouldn't be a conflict. In my understanding the memory layout after spl_relocate_stack_gd should look something like this 0x0 . <misc, other> . CONFIG_SPL_STACK_ADDR_R - sizeof(gd_t): relocated Stack pointer (growing downwards) CONFIG_SPL_STACK_ADDR_R - sizeof(gd_t): global data CONFIG_SPL_STACK_ADDR_R : Start of heap (growing upward> CONFIG_SPL_STACK_ADDR_R + CONFIG_SYS_MALLOC_F_LEN: End of heap I'm pretty sure that's correct, well either that, or i'm missing something obvious and spl_relocate_stack_gd doesn't make any sense (as it als puts the new stack pointer to start at the gd location) :) > You likely will want to make CONFIG_SPL_STACK_R_ADDR be something > like 0x80000 > and then do: > > gd->malloc_base = 0x00000000; > gd->malloc_limit = CONFIG_SPL_STACK_R_ADDR; > gd->malloc_ptr = 0; > > Here to stop the 2 from overlapping, things are likely working with > your patch since you're not resetting gd->malloc_ptr, so you keep > space > for the stack for whatever that is set to at that point. Actually i've been tempted to put the malloc_base at 0x0, but that would mean the first malloc_simple call will return 0 so is indistinguishable from an error.
Hi, On 01-10-15 13:25, Sjoerd Simons wrote: > Hey Hans, > > On Thu, 2015-10-01 at 12:08 +0200, Hans de Goede wrote: >> Hi Sjoerd, >> >> On 01-10-15 11:10, 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> >>> >>> --- >>> 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. >> >> I actually have a patch series pending for this: >> >> http://patchwork.ozlabs.org/patch/517191/ >> http://patchwork.ozlabs.org/patch/517194/ >> http://patchwork.ozlabs.org/patch/517193/ >> http://patchwork.ozlabs.org/patch/517195/ >> http://patchwork.ozlabs.org/patch/517196/ >> >> (I've omitted 2 uninteresting patches) >> >> Your review of / input on this series would be appreciated. > > Cool, I'll try to make some time to give that a closer look. > >> > 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. >> >> I agree that a minimal rockchip only fix likely is best at this time, >> however your fix seems wrong: >> >>> arch/arm/mach-rockchip/board-spl.c | 4 ++++ >>> 1 file changed, 4 insertions(+) >>> >>> diff --git a/arch/arm/mach-rockchip/board-spl.c b/arch/arm/mach >>> -rockchip/board-spl.c >>> index a241d96..5daced7 100644 >>> --- a/arch/arm/mach-rockchip/board-spl.c >>> +++ b/arch/arm/mach-rockchip/board-spl.c >>> @@ -217,6 +217,10 @@ 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; >>> } >>> >>> static int setup_led(void) >> >> SPL_STACK_R_ADDR is where the stack will be put by >> spl_relocate_stack_gd >> so now you've the stack and the heap overlapping. > > If i'm not mistaken the stack grows downward, while the heap grows > upwards so there shouldn't be a conflict. In my understanding the > memory layout after spl_relocate_stack_gd should look something like > this > > 0x0 > . > <misc, other> > . > CONFIG_SPL_STACK_ADDR_R - sizeof(gd_t): relocated Stack pointer (growing downwards) > CONFIG_SPL_STACK_ADDR_R - sizeof(gd_t): global data > CONFIG_SPL_STACK_ADDR_R : Start of heap (growing upward> > CONFIG_SPL_STACK_ADDR_R + CONFIG_SYS_MALLOC_F_LEN: End of heap > > I'm pretty sure that's correct, well either that, or i'm missing > something obvious and spl_relocate_stack_gd doesn't make any sense (as > it als puts the new stack pointer to start at the gd location) :) Ah yes you're right, and since the stack grows downwards I guess that CONFIG_SPL_STACK_ADDR_R already is not 0 for rockchip :) You should probably still reset gd->malloc_ptr to 0, otherwise the new DRAM heap will begin at CONFIG_SPL_STACK_ADDR_R + gd->malloc_ptr and it will be only CONFIG_SYS_MALLOC_F_LEN - gd->malloc_ptr bytes large. Regards, Hans
diff --git a/arch/arm/mach-rockchip/board-spl.c b/arch/arm/mach-rockchip/board-spl.c index a241d96..5daced7 100644 --- a/arch/arm/mach-rockchip/board-spl.c +++ b/arch/arm/mach-rockchip/board-spl.c @@ -217,6 +217,10 @@ 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; } 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. arch/arm/mach-rockchip/board-spl.c | 4 ++++ 1 file changed, 4 insertions(+)