ARM: socfpga: Clear PL310 early in SPL
Subsystems (3)
PATCH 1/1 ARM: socfpga: Clear PL310 early in SPL
On SoCFPGA Gen5 systems, it can rarely happen that a reboot from Linux will result in stale data in PL310 L2 cache controller. Even if the L2 cache controller is disabled via the CTRL register CTRL_EN bit, those data can interfere with operation of devices using DMA, like e.g. the DWMMC controller. This can in turn cause e.g. SPL to fail reading data from SD/MMC.
The obvious solution here would be to fully reset the L2 cache controller via the reset manager MPUMODRST L2 bit, however this causes bus hang even if executed entirely from L1 I-cache to avoid generating any bus traffic through the L2 cache controller.
This patch thus configures and enables the L2 cache controller very early in the SPL boot process, clears the L2 cache and disables the L2 cache controller again.
The reason for doing it in SPL is because we need to avoid accessing any of the potentially stale data in the L2 cache, and we are certain any of the stale data will be below the OCRAM address range. To further reduce bus traffic during the L2 cache invalidation, we enable L1 I-cache and run the invalidation code entirely out of the L1 I-cache.
--- arch/arm/mach-socfpga/spl_gen5.c | 55 ++++++++++++++++++++++++++++++++ 1 file changed, 55 insertions(+)
arch/arm/mach-socfpga/spl_gen5.c +55 -0
@@ -63,6 +63,60 @@ u32 spl_boot_mode(const u32 boot_device) }#endif +static void socfpga_pl310_clear (void ) +{ +u32 mask =0xff , ena =0 ; + +icache_enable (); + +/* Disable the L2 cache */ +clrbits_le32 (&pl310->pl310_ctrl,L2X0_CTRL_EN ); + +writel (0x111 , &pl310->pl310_tag_latency_ctrl); +writel (0x121 , &pl310->pl310_data_latency_ctrl); + +/* enable BRESP, instruction and data prefetch, full line of zeroes */ +setbits_le32 (&pl310->pl310_aux_ctrl, +L310_AUX_CTRL_DATA_PREFETCH_MASK | +L310_AUX_CTRL_INST_PREFETCH_MASK | +L310_SHARED_ATT_OVERRIDE_ENABLE ); + +/* Enable the L2 cache */ + ena =readl (&pl310->pl310_ctrl); + ena |=L2X0_CTRL_EN ; + +/* +* Invalidate the PL310 L2 cache. Keep the invalidation code +* entirely in L1 I-cache to avoid any bus traffic through +* the L2. +*/ + asmvolatile ( +".align 5 \n" +" b 3f \n" +"1: str %1, [%4] \n" +" dsb \n" +" isb \n" +" str %0, [%2] \n" +" dsb \n" +" isb \n" +"2: ldr %0, [%2] \n" +" cmp %0, #0 \n" +" bne 2b \n" +" str %0, [%3] \n" +" dsb \n" +" isb \n" +" b 4f \n" +"3: b 1b \n" +"4: nop \n" + :"+r" (mask),"+r" (ena) + :"r" (&pl310->pl310_inv_way), +"r" (&pl310->pl310_cache_sync),"r" (&pl310->pl310_ctrl) + :"memory" ,"cc" ); + +/* Disable the L2 cache */ +clrbits_le32 (&pl310->pl310_ctrl,L2X0_CTRL_EN ); +} +void board_init_f (ulong dummy) {const struct cm_config *cm_default_cfg =cm_get_default_config ();
@@ -85,6 +139,7 @@ void board_init_f(ulong dummy)memset (__bss_start,0 , __bss_end - __bss_start);socfpga_sdram_remap_zero (); +socfpga_pl310_clear ();debug ("Freezing all I/O banks\n" );/* freeze all IO banks */
Signature · 1 line
2.19.2
PATCH 1/1 ARM: socfpga: Clear PL310 early in SPL
Quoted 23 lines · Marek Vasut
On SoCFPGA Gen5 systems, it can rarely happen that a reboot from Linux will result in stale data in PL310 L2 cache controller. Even if the L2 cache controller is disabled via the CTRL register
CTRL_ENbit, those data can interfere with operation of devices using DMA, like e.g. the DWMMC controller. This can in turn cause e.g. SPL to fail reading data from SD/MMC.The obvious solution here would be to fully reset the L2 cache controller via the reset manager MPUMODRST L2 bit, however this causes bus hang even if executed entirely from L1 I-cache to avoid generating any bus traffic through the L2 cache controller.
This patch thus configures and enables the L2 cache controller very early in the SPL boot process, clears the L2 cache and disables the L2 cache controller again.
The reason for doing it in SPL is because we need to avoid accessing any of the potentially stale data in the L2 cache, and we are certain any of the stale data will be below the OCRAM address range. To further reduce bus traffic during the L2 cache invalidation, we enable L1 I-cache and run the invalidation code entirely out of the L1 I-cache.
Quoted 3 lines
Signed-off-by Marek Vasut <[email protected]>Cc Dalon Westergreen <[email protected]>Cc Dinh Nguyen <[email protected]>
We've tested this and it seems to fix the issue, so:
However, I don't really get why clearing the L2 cache later in U-Boot isn't good enough.
Also, wouldn't this affect other platforms as well?
Regards,
Simon
Quoted 79 lines
--- arch/arm/mach-socfpga/spl_gen5.c | 55 ++++++++++++++++++++++++++++++++ 1 file changed, 55 insertions(+)arch/arm/mach-socfpga/spl_gen5.c +55
-1@@ -63,6 +63,60 @@ u32 spl_boot_mode(const u32 boot_device) }#endif +static void socfpga_pl310_clear (void ) +{ +u32 mask =0xff , ena =0 ; + +icache_enable (); + +/* Disable the L2 cache */ +clrbits_le32 (&pl310->pl310_ctrl,L2X0_CTRL_EN ); + +writel (0x111 , &pl310->pl310_tag_latency_ctrl); +writel (0x121 , &pl310->pl310_data_latency_ctrl); + +/* enable BRESP, instruction and data prefetch, full line of zeroes */ +setbits_le32 (&pl310->pl310_aux_ctrl, +L310_AUX_CTRL_DATA_PREFETCH_MASK | +L310_AUX_CTRL_INST_PREFETCH_MASK | +L310_SHARED_ATT_OVERRIDE_ENABLE ); + +/* Enable the L2 cache */ + ena =readl (&pl310->pl310_ctrl); + ena |=L2X0_CTRL_EN ; + +/* +* Invalidate the PL310 L2 cache. Keep the invalidation code +* entirely in L1 I-cache to avoid any bus traffic through +* the L2. +*/ + asmvolatile ( +".align 5 \n" +" b 3f \n" +"1: str %1, [%4] \n" +" dsb \n" +" isb \n" +" str %0, [%2] \n" +" dsb \n" +" isb \n" +"2: ldr %0, [%2] \n" +" cmp %0, #0 \n" +" bne 2b \n" +" str %0, [%3] \n" +" dsb \n" +" isb \n" +" b 4f \n" +"3: b 1b \n" +"4: nop \n" + :"+r" (mask),"+r" (ena) + :"r" (&pl310->pl310_inv_way), +"r" (&pl310->pl310_cache_sync),"r" (&pl310->pl310_ctrl) + :"memory" ,"cc" ); + +/* Disable the L2 cache */ +clrbits_le32 (&pl310->pl310_ctrl,L2X0_CTRL_EN ); +} +void board_init_f (ulong dummy) {const struct cm_config *cm_default_cfg =cm_get_default_config ();@@ -85,6 +139,7 @@ void board_init_f(ulong dummy)memset (__bss_start,0 , __bss_end - __bss_start);socfpga_sdram_remap_zero (); +socfpga_pl310_clear ();debug ("Freezing all I/O banks\n" );/* freeze all IO banks */ --2.19.2
PATCH 1/1 ARM: socfpga: Clear PL310 early in SPL
Quoted 28 lines · Simon Goldschmidt
Quoted 23 lines · Marek Vasut
On SoCFPGA Gen5 systems, it can rarely happen that a reboot from Linux will result in stale data in PL310 L2 cache controller. Even if the L2 cache controller is disabled via the CTRL register
CTRL_ENbit, those data can interfere with operation of devices using DMA, like e.g. the DWMMC controller. This can in turn cause e.g. SPL to fail reading data from SD/MMC.The obvious solution here would be to fully reset the L2 cache controller via the reset manager MPUMODRST L2 bit, however this causes bus hang even if executed entirely from L1 I-cache to avoid generating any bus traffic through the L2 cache controller.
This patch thus configures and enables the L2 cache controller very early in the SPL boot process, clears the L2 cache and disables the L2 cache controller again.
The reason for doing it in SPL is because we need to avoid accessing any of the potentially stale data in the L2 cache, and we are certain any of the stale data will be below the OCRAM address range. To further reduce bus traffic during the L2 cache invalidation, we enable L1 I-cache and run the invalidation code entirely out of the L1 I-cache.
Quoted 3 lines
Signed-off-by Marek Vasut <[email protected]>Cc Dalon Westergreen <[email protected]>Cc Dinh Nguyen <[email protected]>
Quoted 6 lines
We've tested this and it seems to fix the issue, so:
Tested-by Simon Goldschmidt <[email protected]>However, I don't really get why clearing the L2 cache later in U-Boot isn't good enough.
Because when U-Boot is running, it is already running from RAM, which is on different PL310 master than OCRAM, and you're likely to hit the corrupted cache lines on the DRAM master which is primed by prior Linux operation. Such cacheline can be hit between the code enabling the PL310 and the code invalidating it, which is a C code, using stack and calling functions, thus accessing memory which would likely reside in different PL310 cachelines. If one of those cachelines contains invalid/corrupted data, they can be provided to the CPU before the cacheline is invalidated.
To further reduce the likelihood of hitting any such cache line, the code which enables the PL310 and invalidates it runs from primed L1 icache lines, thus generating no bus traffic at all.
Quoted 1 line
Also, wouldn't this affect other platforms as well?
I had the same concern, however I suspect it might have to do with the reset implementation on SoCFPGA, which doesn't clear the L2CC, while reset implementations on other SoCs likely do clear the L2CC. I am however inquiring with Altera/Intel about this.
[...]
Signature · 2 lines
Best regards, Marek Vasut
PATCH 1/1 ARM: socfpga: Clear PL310 early in SPL
Am Di., 19. Feb. 2019, 14:45 hat Marek Vasut <[email protected]> geschrieben:
Quoted 48 lines · Simon Goldschmidt
Quoted 10 lines · Marek Vasut
Quoted 9 lines
On SoCFPGA Gen5 systems, it can rarely happen that a reboot from Linux will result in stale data in PL310 L2 cache controller. Even if the L2 cache controller is disabled via the CTRL register
CTRL_ENbit, those data can interfere with operation of devices using DMA, like e.g. the DWMMC controller. This can in turn cause e.g. SPL to fail reading data from SD/MMC.The obvious solution here would be to fully reset the L2 cache
controller
Quoted 1 line
Quoted 1 line
via the reset manager MPUMODRST L2 bit, however this causes bus hang
even
Quoted 4 lines
Quoted 4 lines
if executed entirely from L1 I-cache to avoid generating any bus traffic through the L2 cache controller.
This patch thus configures and enables the L2 cache controller very
early
Quoted 13 lines
Quoted 9 lines
in the SPL boot process, clears the L2 cache and disables the L2 cache controller again.
The reason for doing it in SPL is because we need to avoid accessing any of the potentially stale data in the L2 cache, and we are certain any of the stale data will be below the OCRAM address range. To further reduce bus traffic during the L2 cache invalidation, we enable L1 I-cache and run the invalidation code entirely out of the L1 I-cache.
Quoted 3 lines
Signed-off-by Marek Vasut <[email protected]>Cc Dalon Westergreen <[email protected]>Cc Dinh Nguyen <[email protected]>Quoted 6 lines
We've tested this and it seems to fix the issue, so:
Tested-by Simon Goldschmidt <[email protected]>However, I don't really get why clearing the L2 cache later in U-Boot isn't good enough.
Because when U-Boot is running, it is already running from RAM, which is on different PL310 master than OCRAM, and you're likely to hit the corrupted cache lines on the DRAM master which is primed by prior Linux operation. Such cacheline can be hit between the code enabling the PL310 and the code invalidating it, which is a C code, using stack and calling functions, thus accessing memory which would likely reside in different PL310 cachelines. If one of those cachelines contains invalid/corrupted data, they can be provided to the CPU before the cacheline is invalidated.
Quoted 4 lines
To further reduce the likelihood of hitting any such cache line, the code which enables the PL310 and invalidates it runs from primed L1 icache lines, thus generating no bus traffic at all.
Ah, right. I somehow didn't realize that invalidating is done after enabling:-)
Quoted 7 lines
Also, wouldn't this affect other platforms as well?
I had the same concern, however I suspect it might have to do with the reset implementation on SoCFPGA, which doesn't clear the L2CC, while reset implementations on other SoCs likely do clear the L2CC. I am however inquiring with Altera/Intel about this.
Ok, thanks for the explanation!
Regards,
Simon
PATCH 1/1 ARM: socfpga: Clear PL310 early in SPL
Quoted 66 lines · Simon Goldschmidt
Am Di., 19. Feb. 2019, 14:45 hat Marek Vasut <[email protected] <mailto:[email protected]>> geschrieben:
On 2/19/19 2:28 PM, Simon Goldschmidt wrote:
> On Tue, Feb 19, 2019 at 1:53 PM Marek Vasut <[email protected] <mailto:[email protected]>> wrote:
>>
>> On SoCFPGA Gen5 systems, it can rarely happen that a reboot from Linux
>> will result in stale data in PL310 L2 cache controller. Even if the L2
>> cache controller is disabled via the CTRL registerCTRL_ENbit, those >> data can interfere with operation of devices using DMA, like e.g. the >> DWMMC controller. This can in turn cause e.g. SPL to fail reading data
>> from SD/MMC.
>>
>> The obvious solution here would be to fully reset the L2 cache controller
>> via the reset manager MPUMODRST L2 bit, however this causes bus hang even
>> if executed entirely from L1 I-cache to avoid generating any bus traffic
>> through the L2 cache controller.
>>
>> This patch thus configures and enables the L2 cache controller very early
>> in the SPL boot process, clears the L2 cache and disables the L2 cache
>> controller again.
>>
>> The reason for doing it in SPL is because we need to avoid accessing any
>> of the potentially stale data in the L2 cache, and we are certain any of
>> the stale data will be below the OCRAM address range. To further reduce
>> bus traffic during the L2 cache invalidation, we enable L1 I-cache and
>> run the invalidation code entirely out of the L1 I-cache. >>
>> Signed-off-by: Marek Vasut <marex at denx.de <mailto:[email protected]>> >> Cc: Dalon Westergreen <[email protected] <mailto:[email protected]>>
>> Cc: Dinh Nguyen <dinguyen at kernel.org <mailto:[email protected]>> >
> We've tested this and it seems to fix the issue, so:
>
> Tested-by: Simon Goldschmidt <[email protected] <mailto:[email protected]>>
>
> However, I don't really get why clearing the L2 cache later in U-Boot > isn't good enough.Because when U-Boot is running, it is already running from RAM, which is on different PL310 master than OCRAM, and you're likely to hit the corrupted cache lines on the DRAM master which is primed by prior Linux operation. Such cacheline can be hit between the code enabling the PL310 and the code invalidating it, which is a C code, using stack and calling functions, thus accessing memory which would likely reside in different PL310 cachelines. If one of those cachelines contains invalid/corrupted data, they can be provided to the CPU before the cacheline is invalidated.
Quoted 7 lines
To further reduce the likelihood of hitting any such cache line, the code which enables the PL310 and invalidates it runs from primed L1 icache lines, thus generating no bus traffic at all.
Ah, right. I somehow didn't realize that invalidating is done after enabling:-)
Right. If it could be done before (like e.g. by whacking the L2CC reset), that'd be fantastic.
Quoted 9 lines
Quoted 1 line
Also, wouldn't this affect other platforms as well?
I had the same concern, however I suspect it might have to do with the reset implementation on SoCFPGA, which doesn't clear the L2CC, while reset implementations on other SoCs likely do clear the L2CC. I am however inquiring with Altera/Intel about this.
Ok, thanks for the explanation!
Sure
Signature · 2 lines
Best regards, Marek Vasut
PATCH 1/1 ARM: socfpga: Clear PL310 early in SPL
Am 19.02.2019 um 15:00 schrieb Marek Vasut:
Quoted 45 lines · Simon Goldschmidt
Quoted 44 lines
Am Di., 19. Feb. 2019, 14:45 hat Marek Vasut <[email protected] <mailto:[email protected]>> geschrieben:
On 2/19/19 2:28 PM, Simon Goldschmidt wrote:
> On Tue, Feb 19, 2019 at 1:53 PM Marek Vasut <[email protected] <mailto:[email protected]>> wrote:
>>
>> On SoCFPGA Gen5 systems, it can rarely happen that a reboot from Linux
>> will result in stale data in PL310 L2 cache controller. Even if the L2
>> cache controller is disabled via the CTRL registerCTRL_ENbit, those >> data can interfere with operation of devices using DMA, like e.g. the >> DWMMC controller. This can in turn cause e.g. SPL to fail reading data
>> from SD/MMC.
>>
>> The obvious solution here would be to fully reset the L2 cache controller
>> via the reset manager MPUMODRST L2 bit, however this causes bus hang even
>> if executed entirely from L1 I-cache to avoid generating any bus traffic
>> through the L2 cache controller.
>>
>> This patch thus configures and enables the L2 cache controller very early
>> in the SPL boot process, clears the L2 cache and disables the L2 cache
>> controller again.
>>
>> The reason for doing it in SPL is because we need to avoid accessing any
>> of the potentially stale data in the L2 cache, and we are certain any of
>> the stale data will be below the OCRAM address range. To further reduce
>> bus traffic during the L2 cache invalidation, we enable L1 I-cache and
>> run the invalidation code entirely out of the L1 I-cache. >>
>> Signed-off-by: Marek Vasut <marex at denx.de <mailto:[email protected]>>
Quoted 8 lines
Quoted 8 lines
>> Cc: Dalon Westergreen <[email protected] <mailto:[email protected]>>
>> Cc: Dinh Nguyen <dinguyen at kernel.org <mailto:[email protected]>> >
> We've tested this and it seems to fix the issue, so:
>
> Tested-by: Simon Goldschmidt <[email protected] <mailto:[email protected]>>
Seems like I was a little fast there: the patch does not apply cleanly to master. It did apply to our tree, obviously.
It's missing an include to <asm/pl310.h> and 'pl310' is undefined. Probably a result of rebasing it...
Quoted 19 lines
Quoted 19 lines
>
> However, I don't really get why clearing the L2 cache later in U-Boot > isn't good enough.Because when U-Boot is running, it is already running from RAM, which is on different PL310 master than OCRAM, and you're likely to hit the corrupted cache lines on the DRAM master which is primed by prior Linux operation. Such cacheline can be hit between the code enabling the PL310 and the code invalidating it, which is a C code, using stack and calling functions, thus accessing memory which would likely reside in different PL310 cachelines. If one of those cachelines contains invalid/corrupted data, they can be provided to the CPU before the cacheline is invalidated.
To further reduce the likelihood of hitting any such cache line, the code which enables the PL310 and invalidates it runs from primed L1 icache lines, thus generating no bus traffic at all.
Quoted 5 lines
Quoted 2 lines
Ah, right. I somehow didn't realize that invalidating is done after enabling:-)
Right. If it could be done before (like e.g. by whacking the L2CC reset), that'd be fantastic.
I haven't yet found a way to cleanly reproduce this on my socrates, so unfortunately, I currently cannot test any suggestions, right now :-(
Wouldn't it be better to place this function in some central position near the cache initialization code instead of running it from SPL? Or is it required to run it in SPL?
Also, it seems this patch enables the ICACHE in SPL, which wasn't enabled before. Not a bad thing, but might this have side effects?
Regards,
Simon
Quoted 13 lines
Quoted 10 lines
> Also, wouldn't this affect other platforms as well?
I had the same concern, however I suspect it might have to do with the reset implementation on SoCFPGA, which doesn't clear the L2CC, while reset implementations on other SoCs likely do clear the L2CC. I am however inquiring with Altera/Intel about this.
Ok, thanks for the explanation!
Sure
PATCH 1/1 ARM: socfpga: Clear PL310 early in SPL
Quoted 94 lines · Simon Goldschmidt
Am 19.02.2019 um 15:00 schrieb Marek Vasut:
Quoted 51 lines · Simon Goldschmidt
Quoted 50 lines
Am Di., 19. Feb. 2019, 14:45 hat Marek Vasut <[email protected] <mailto:[email protected]>> geschrieben:
On 2/19/19 2:28 PM, Simon Goldschmidt wrote:
> On Tue, Feb 19, 2019 at 1:53 PM Marek Vasut <[email protected] <mailto:[email protected]>> wrote:
>>
>> On SoCFPGA Gen5 systems, it can rarely happen that a reboot from Linux
>> will result in stale data in PL310 L2 cache controller. Even if the L2
>> cache controller is disabled via the CTRL registerCTRL_ENbit, those
>> data can interfere with operation of devices using DMA, like e.g. the
>> DWMMC controller. This can in turn cause e.g. SPL to fail reading
data
>> from SD/MMC.
>>
>> The obvious solution here would be to fully reset the L2 cache controller
>> via the reset manager MPUMODRST L2 bit, however this causes bus hang even
>> if executed entirely from L1 I-cache to avoid generating any bus traffic
>> through the L2 cache controller.
>>
>> This patch thus configures and enables the L2 cache controller very early
>> in the SPL boot process, clears the L2 cache and disables the L2 cache
>> controller again.
>>
>> The reason for doing it in SPL is because we need to avoid accessing any
>> of the potentially stale data in the L2 cache, and we are certain
any of
>> the stale data will be below the OCRAM address range. To further reduce
>> bus traffic during the L2 cache invalidation, we enable L1 I-cache and
>> run the invalidation code entirely out of the L1 I-cache. >>
>> Signed-off-by: Marek Vasut <[email protected] <mailto:[email protected]>>
>> Cc: Dalon Westergreen <[email protected]Quoted 8 lines
Quoted 8 lines
<mailto:[email protected]>>
>> Cc: Dinh Nguyen <[email protected] <mailto:[email protected]>>
>
> We've tested this and it seems to fix the issue, so:
>
> Tested-by: Simon Goldschmidt <[email protected] <mailto:[email protected]>>Seems like I was a little fast there: the patch does not apply cleanly to master. It did apply to our tree, obviously.
It's missing an include to <
asm/pl310.h> and 'pl310' is undefined. Probably a result of rebasing it...Quoted 27 lines
Quoted 27 lines
>
> However, I don't really get why clearing the L2 cache later in U-Boot
> isn't good enough.Because when U-Boot is running, it is already running from RAM, which is
on different PL310 master than OCRAM, and you're likely to hit the corrupted cache lines on the DRAM master which is primed by prior Linux
operation. Such cacheline can be hit between the code enabling the PL310
and the code invalidating it, which is a C code, using stack and calling
functions, thus accessing memory which would likely reside in different
PL310 cachelines. If one of those cachelines contains invalid/corrupted
data, they can be provided to the CPU before the cacheline is invalidated.To further reduce the likelihood of hitting any such cache line, the
code which enables the PL310 and invalidates it runs from primed L1 icache lines, thus generating no bus traffic at all.
Quoted 8 lines
Quoted 5 lines
Quoted 2 lines
Ah, right. I somehow didn't realize that invalidating is done after enabling:-)
Right. If it could be done before (like e.g. by whacking the L2CC reset), that'd be fantastic.
I haven't yet found a way to cleanly reproduce this on my socrates, so unfortunately, I currently cannot test any suggestions, right now :-(
Did you ever have this hang happen on SoCrates ?
Quoted 3 lines
Wouldn't it be better to place this function in some central position near the cache initialization code instead of running it from SPL? Or is it required to run it in SPL?
It's required in SPL as soon as possible, to avoid accessing the DRAM, see above. The current location is close to the rest of PL310 init in SPL.
Quoted 2 lines
Also, it seems this patch enables the ICACHE in SPL, which wasn't enabled before. Not a bad thing, but might this have side effects?
Actually, I-Cache is enabled in SPL:
[...]
Signature · 2 lines
Best regards, Marek Vasut
PATCH 1/1 ARM: socfpga: Clear PL310 early in SPL
Am 19.02.2019 um 17:00 schrieb Marek Vasut:
Quoted 98 lines · Simon Goldschmidt
Quoted 97 lines
Am 19.02.2019 um 15:00 schrieb Marek Vasut:
Quoted 51 lines · Simon Goldschmidt
Quoted 50 lines
Am Di., 19. Feb. 2019, 14:45 hat Marek Vasut <[email protected] <mailto:[email protected]>> geschrieben:
On 2/19/19 2:28 PM, Simon Goldschmidt wrote:
> On Tue, Feb 19, 2019 at 1:53 PM Marek Vasut <[email protected] <mailto:[email protected]>> wrote:
>>
>> On SoCFPGA Gen5 systems, it can rarely happen that a reboot from Linux
>> will result in stale data in PL310 L2 cache controller. Even if the L2
>> cache controller is disabled via the CTRL registerCTRL_EN
bit, those
>> data can interfere with operation of devices using DMA, like
e.g. the
>> DWMMC controller. This can in turn cause e.g. SPL to fail
reading
data
>> from SD/MMC.
>>
>> The obvious solution here would be to fully reset the L2 cache controller
>> via the reset manager MPUMODRST L2 bit, however this causes bus hang even
>> if executed entirely from L1 I-cache to avoid generating any bus traffic
>> through the L2 cache controller.
>>
>> This patch thus configures and enables the L2 cache controller very early
>> in the SPL boot process, clears the L2 cache and disables the L2 cache
>> controller again.
>>
>> The reason for doing it in SPL is because we need to avoid accessing any
>> of the potentially stale data in the L2 cache, and we are
certain
any of
>> the stale data will be below the OCRAM address range. To further reduce
>> bus traffic during the L2 cache invalidation, we enable L1 I-cache and
>> run the invalidation code entirely out of the L1 I-cache. >>
>> Signed-off-by: Marek Vasut <[email protected]
<mailto:[email protected]>>
>> Cc: Dalon Westergreen <[email protected]Quoted 8 lines
Quoted 8 lines
<mailto:[email protected]>>
>> Cc: Dinh Nguyen <[email protected]
<mailto:[email protected]>>
>
> We've tested this and it seems to fix the issue, so:
>
> Tested-by: Simon Goldschmidt <[email protected] <mailto:[email protected]>>Seems like I was a little fast there: the patch does not apply cleanly to master. It did apply to our tree, obviously.
It's missing an include to <
asm/pl310.h> and 'pl310' is undefined. Probably a result of rebasing it...Quoted 30 lines
Quoted 29 lines
>
> However, I don't really get why clearing the L2 cache later in
U-Boot
> isn't good enough.Because when U-Boot is running, it is already running from RAM,
which is
on different PL310 master than OCRAM, and you're likely to hit the corrupted cache lines on the DRAM master which is primed by
prior Linux
operation. Such cacheline can be hit between the code enabling
the PL310
and the code invalidating it, which is a C code, using stack and
calling
functions, thus accessing memory which would likely reside in
different
PL310 cachelines. If one of those cachelines contains
invalid/corrupted
data, they can be provided to the CPU before the cacheline is invalidated.To further reduce the likelihood of hitting any such cache line,
the
code which enables the PL310 and invalidates it runs from primed L1 icache lines, thus generating no bus traffic at all.Ah, right. I somehow didn't realize that invalidating is done after enabling:-)
Quoted 7 lines
Quoted 5 lines
Quoted 2 lines
Right. If it could be done before (like e.g. by whacking the L2CC reset), that'd be fantastic.
I haven't yet found a way to cleanly reproduce this on my socrates, so unfortunately, I currently cannot test any suggestions, right now :-(
Did you ever have this hang happen on SoCrates ?
I think so, yes.
What I did was running some lmb tests loading from mmc, removing the sdcard to write an updated version of U-Boot, replacing it and then typing 'reset'. It would then sometimes hang in cache initialization - I added debug prints throughout cache initialization and was debug prints from one function and then nothing more after a simple function call...
That would very much fit to your explanation. However, I cannot reproduce this currently :-(
Quoted 7 lines
Quoted 4 lines
Wouldn't it be better to place this function in some central position near the cache initialization code instead of running it from SPL? Or is it required to run it in SPL?
It's required in SPL as soon as possible, to avoid accessing the DRAM, see above. The current location is close to the rest of PL310 init in SPL.
OK, I just did not understand why this is necessary before accessing DRAM.
Quoted 6 lines
Quoted 3 lines
Also, it seems this patch enables the ICACHE in SPL, which wasn't enabled before. Not a bad thing, but might this have side effects?
Actually, I-Cache is enabled in SPL:
Oh, right. Unless disabled via config option.
Regards,
Simon
PATCH 1/1 ARM: socfpga: Clear PL310 early in SPL
Quoted 119 lines · Simon Goldschmidt
Am 19.02.2019 um 17:00 schrieb Marek Vasut:
Quoted 108 lines · Simon Goldschmidt
Quoted 107 lines
Am 19.02.2019 um 15:00 schrieb Marek Vasut:
Quoted 59 lines · Simon Goldschmidt
Quoted 58 lines
Am Di., 19. Feb. 2019, 14:45 hat Marek Vasut <[email protected] <mailto:[email protected]>> geschrieben:
On 2/19/19 2:28 PM, Simon Goldschmidt wrote:
> On Tue, Feb 19, 2019 at 1:53 PM Marek Vasut <[email protected] <mailto:[email protected]>> wrote:
>>
>> On SoCFPGA Gen5 systems, it can rarely happen that a reboot from
Linux
>> will result in stale data in PL310 L2 cache controller. Even if
the L2
>> cache controller is disabled via the CTRL registerCTRL_ENbit, those
>> data can interfere with operation of devices using DMA, like e.g. the
>> DWMMC controller. This can in turn cause e.g. SPL to fail reading
data
>> from SD/MMC.
>>
>> The obvious solution here would be to fully reset the L2 cache
controller
>> via the reset manager MPUMODRST L2 bit, however this causes bus
hang even
>> if executed entirely from L1 I-cache to avoid generating any bus
traffic
>> through the L2 cache controller.
>>
>> This patch thus configures and enables the L2 cache controller
very early
>> in the SPL boot process, clears the L2 cache and disables the L2
cache
>> controller again.
>>
>> The reason for doing it in SPL is because we need to avoid accessing any
>> of the potentially stale data in the L2 cache, and we are certain
any of
>> the stale data will be below the OCRAM address range. To further
reduce
>> bus traffic during the L2 cache invalidation, we enable L1 I-cache and
>> run the invalidation code entirely out of the L1 I-cache. >>
>> Signed-off-by: Marek Vasut <[email protected] <mailto:[email protected]>>
>> Cc: Dalon Westergreen <[email protected]Quoted 8 lines
Quoted 8 lines
<mailto:[email protected]>>
>> Cc: Dinh Nguyen <[email protected] <mailto:[email protected]>>
>
> We've tested this and it seems to fix the issue, so:
>
> Tested-by: Simon Goldschmidt <[email protected] <mailto:[email protected]>>Seems like I was a little fast there: the patch does not apply cleanly to master. It did apply to our tree, obviously.
It's missing an include to <
asm/pl310.h> and 'pl310' is undefined. Probably a result of rebasing it...Quoted 32 lines
Quoted 31 lines
>
> However, I don't really get why clearing the L2 cache later in U-Boot
> isn't good enough.Because when U-Boot is running, it is already running from RAM, which is
on different PL310 master than OCRAM, and you're likely to hit the
corrupted cache lines on the DRAM master which is primed by prior Linux
operation. Such cacheline can be hit between the code enabling the PL310
and the code invalidating it, which is a C code, using stack and calling
functions, thus accessing memory which would likely reside in different
PL310 cachelines. If one of those cachelines contains invalid/corrupted
data, they can be provided to the CPU before the cacheline is invalidated.To further reduce the likelihood of hitting any such cache line, the
code which enables the PL310 and invalidates it runs from primed L1
icache lines, thus generating no bus traffic at all.Ah, right. I somehow didn't realize that invalidating is done after enabling:-)
Quoted 7 lines
Quoted 5 lines
Quoted 2 lines
Right. If it could be done before (like e.g. by whacking the L2CC reset), that'd be fantastic.
I haven't yet found a way to cleanly reproduce this on my socrates, so unfortunately, I currently cannot test any suggestions, right now :-(
Did you ever have this hang happen on SoCrates ?
I think so, yes.
Quoted 8 lines
What I did was running some lmb tests loading from mmc, removing the sdcard to write an updated version of U-Boot, replacing it and then typing 'reset'. It would then sometimes hang in cache initialization - I added debug prints throughout cache initialization and was debug prints from one function and then nothing more after a simple function call...
That would very much fit to your explanation. However, I cannot reproduce this currently :-(
So basically a board which uses DMA (like the DWMMC DMA) is likely to be affected. Surely, the DMA does trigger PL310 operations.
Quoted 9 lines
Quoted 7 lines
Quoted 3 lines
Wouldn't it be better to place this function in some central position near the cache initialization code instead of running it from SPL? Or is it required to run it in SPL?
It's required in SPL as soon as possible, to avoid accessing the DRAM, see above. The current location is close to the rest of PL310 init in SPL.
OK, I just did not understand why this is necessary before accessing DRAM.
:)
Quoted 8 lines
Quoted 6 lines
Quoted 2 lines
Also, it seems this patch enables the ICACHE in SPL, which wasn't enabled before. Not a bad thing, but might this have side effects?
Actually, I-Cache is enabled in SPL:
Oh, right. Unless disabled via config option.
Yes
Signature · 2 lines
Best regards, Marek Vasut
PATCH 1/1 ARM: socfpga: Clear PL310 early in SPL
Am 19.02.2019 um 17:39 schrieb Marek Vasut:
Quoted 136 lines · Simon Goldschmidt
Quoted 119 lines
Am 19.02.2019 um 17:00 schrieb Marek Vasut:
Quoted 108 lines · Simon Goldschmidt
Quoted 107 lines
Am 19.02.2019 um 15:00 schrieb Marek Vasut:
Quoted 59 lines · Simon Goldschmidt
Quoted 58 lines
Am Di., 19. Feb. 2019, 14:45 hat Marek Vasut <[email protected] <mailto:[email protected]>> geschrieben:
On 2/19/19 2:28 PM, Simon Goldschmidt wrote:
> On Tue, Feb 19, 2019 at 1:53 PM Marek Vasut <[email protected] <mailto:[email protected]>> wrote:
>>
>> On SoCFPGA Gen5 systems, it can rarely happen that a
reboot from
Linux
>> will result in stale data in PL310 L2 cache controller.
Even if
the L2
>> cache controller is disabled via the CTRL registerCTRL_EN
bit, those
>> data can interfere with operation of devices using DMA, like
e.g. the
>> DWMMC controller. This can in turn cause e.g. SPL to fail
reading
data
>> from SD/MMC.
>>
>> The obvious solution here would be to fully reset the L2
cache
controller
>> via the reset manager MPUMODRST L2 bit, however this
causes bus
hang even
>> if executed entirely from L1 I-cache to avoid generating
any bus
traffic
>> through the L2 cache controller.
>>
>> This patch thus configures and enables the L2 cache
controller
very early
>> in the SPL boot process, clears the L2 cache and disables
the L2
cache
>> controller again.
>>
>> The reason for doing it in SPL is because we need to avoid accessing any
>> of the potentially stale data in the L2 cache, and we are
certain
any of
>> the stale data will be below the OCRAM address range. To
further
reduce
>> bus traffic during the L2 cache invalidation, we enable L1 I-cache and
>> run the invalidation code entirely out of the L1 I-cache. >>
>> Signed-off-by: Marek Vasut <[email protected]
<mailto:[email protected]>>
>> Cc: Dalon Westergreen <[email protected]Quoted 8 lines
Quoted 8 lines
<mailto:[email protected]>>
>> Cc: Dinh Nguyen <[email protected]
<mailto:[email protected]>>
>
> We've tested this and it seems to fix the issue, so:
>
> Tested-by: Simon Goldschmidt <[email protected] <mailto:[email protected]>>Seems like I was a little fast there: the patch does not apply cleanly to master. It did apply to our tree, obviously.
It's missing an include to <
asm/pl310.h> and 'pl310' is undefined. Probably a result of rebasing it...Quoted 32 lines
Quoted 31 lines
>
> However, I don't really get why clearing the L2 cache later in
U-Boot
> isn't good enough.Because when U-Boot is running, it is already running from RAM,
which is
on different PL310 master than OCRAM, and you're likely to
hit the
corrupted cache lines on the DRAM master which is primed by
prior Linux
operation. Such cacheline can be hit between the code enabling
the PL310
and the code invalidating it, which is a C code, using stack and
calling
functions, thus accessing memory which would likely reside in
different
PL310 cachelines. If one of those cachelines contains
invalid/corrupted
data, they can be provided to the CPU before the cacheline is invalidated.To further reduce the likelihood of hitting any such cache line,
the
code which enables the PL310 and invalidates it runs from
primed L1
icache lines, thus generating no bus traffic at all.Ah, right. I somehow didn't realize that invalidating is done after enabling:-)
Quoted 7 lines
Quoted 5 lines
Quoted 2 lines
Right. If it could be done before (like e.g. by whacking the L2CC reset), that'd be fantastic.
I haven't yet found a way to cleanly reproduce this on my socrates, so unfortunately, I currently cannot test any suggestions, right now :-(
Did you ever have this hang happen on SoCrates ?
I think so, yes.
Quoted 8 lines
What I did was running some lmb tests loading from mmc, removing the sdcard to write an updated version of U-Boot, replacing it and then typing 'reset'. It would then sometimes hang in cache initialization - I added debug prints throughout cache initialization and was debug prints from one function and then nothing more after a simple function call...
That would very much fit to your explanation. However, I cannot reproduce this currently :-(
So basically a board which uses DMA (like the DWMMC DMA) is likely to be affected. Surely, the DMA does trigger PL310 operations.
Quoted 4 lines
Quoted 4 lines
Quoted 3 lines
Wouldn't it be better to place this function in some central position near the cache initialization code instead of running it from SPL? Or is it required to run it in SPL?
Quoted 7 lines
Quoted 5 lines
Quoted 3 lines
It's required in SPL as soon as possible, to avoid accessing the DRAM, see above. The current location is close to the rest of PL310 init in SPL.
OK, I just did not understand why this is necessary before accessing DRAM.
:)
Sorry, I phrased that wrong. I *do* (still) not understand why this is necessary before accessing DRAM. Sorry if I'm annoying you with that, but I'd like to understand.
Does the DWMMC DMA really trigger PL310 operations? I thought DMA would only affect cache if it goes through ACP/SCU?
Regards,
Simon
Quoted 12 lines
Quoted 9 lines
Quoted 7 lines
Quoted 3 lines
Also, it seems this patch enables the ICACHE in SPL, which wasn't enabled before. Not a bad thing, but might this have side effects?
Actually, I-Cache is enabled in SPL:
Oh, right. Unless disabled via config option.
Yes
PATCH 1/1 ARM: socfpga: Clear PL310 early in SPL
Quoted 149 lines · Simon Goldschmidt
Am 19.02.2019 um 17:39 schrieb Marek Vasut:
Quoted 148 lines · Simon Goldschmidt
Quoted 126 lines
Am 19.02.2019 um 17:00 schrieb Marek Vasut:
Quoted 115 lines · Simon Goldschmidt
Quoted 114 lines
Am 19.02.2019 um 15:00 schrieb Marek Vasut:
Quoted 61 lines · Simon Goldschmidt
Quoted 60 lines
Am Di., 19. Feb. 2019, 14:45 hat Marek Vasut <[email protected] <mailto:[email protected]>> geschrieben:
On 2/19/19 2:28 PM, Simon Goldschmidt wrote:
> On Tue, Feb 19, 2019 at 1:53 PM Marek Vasut <[email protected] <mailto:[email protected]>> wrote:
>>
>> On SoCFPGA Gen5 systems, it can rarely happen that a reboot from
Linux
>> will result in stale data in PL310 L2 cache controller. Even if
the L2
>> cache controller is disabled via the CTRL registerCTRL_ENbit, those
>> data can interfere with operation of devices using DMA, like
e.g. the
>> DWMMC controller. This can in turn cause e.g. SPL to fail reading
data
>> from SD/MMC.
>>
>> The obvious solution here would be to fully reset the L2 cache
controller
>> via the reset manager MPUMODRST L2 bit, however this causes bus
hang even
>> if executed entirely from L1 I-cache to avoid generating any bus
traffic
>> through the L2 cache controller.
>>
>> This patch thus configures and enables the L2 cache controller
very early
>> in the SPL boot process, clears the L2 cache and disables the L2
cache
>> controller again.
>>
>> The reason for doing it in SPL is because we need to avoid accessing any
>> of the potentially stale data in the L2 cache, and we are certain
any of
>> the stale data will be below the OCRAM address range. To further
reduce
>> bus traffic during the L2 cache invalidation, we enable L1 I-cache and
>> run the invalidation code entirely out of the L1 I-cache. >>
>> Signed-off-by: Marek Vasut <[email protected] <mailto:[email protected]>>
>> Cc: Dalon Westergreen <[email protected] <mailto:[email protected]>>Quoted 8 lines
Quoted 8 lines
>> Cc: Dinh Nguyen <[email protected] <mailto:[email protected]>>
>
> We've tested this and it seems to fix the issue, so:
>
> Tested-by: Simon Goldschmidt <[email protected] <mailto:[email protected]>>Seems like I was a little fast there: the patch does not apply cleanly to master. It did apply to our tree, obviously.
It's missing an include to <
asm/pl310.h> and 'pl310' is undefined. Probably a result of rebasing it...Quoted 37 lines
Quoted 36 lines
>
> However, I don't really get why clearing the L2 cache later in
U-Boot
> isn't good enough.Because when U-Boot is running, it is already running from RAM,
which is
on different PL310 master than OCRAM, and you're likely to hit the
corrupted cache lines on the DRAM master which is primed by prior Linux
operation. Such cacheline can be hit between the code enabling
the PL310
and the code invalidating it, which is a C code, using stack and
calling
functions, thus accessing memory which would likely reside in different
PL310 cachelines. If one of those cachelines contains invalid/corrupted
data, they can be provided to the CPU before the cacheline is invalidated.To further reduce the likelihood of hitting any such cache line,
the
code which enables the PL310 and invalidates it runs from primed L1
icache lines, thus generating no bus traffic at all.Ah, right. I somehow didn't realize that invalidating is done after enabling:-)
Quoted 7 lines
Quoted 5 lines
Quoted 2 lines
Right. If it could be done before (like e.g. by whacking the L2CC reset), that'd be fantastic.
I haven't yet found a way to cleanly reproduce this on my socrates, so unfortunately, I currently cannot test any suggestions, right now :-(
Did you ever have this hang happen on SoCrates ?
I think so, yes.
Quoted 8 lines
What I did was running some lmb tests loading from mmc, removing the sdcard to write an updated version of U-Boot, replacing it and then typing 'reset'. It would then sometimes hang in cache initialization - I added debug prints throughout cache initialization and was debug prints from one function and then nothing more after a simple function call...
That would very much fit to your explanation. However, I cannot reproduce this currently :-(
So basically a board which uses DMA (like the DWMMC DMA) is likely to be affected. Surely, the DMA does trigger PL310 operations.
Quoted 9 lines
Quoted 8 lines
Quoted 4 lines
Wouldn't it be better to place this function in some central position near the cache initialization code instead of running it from SPL? Or is
it required to run it in SPL?It's required in SPL as soon as possible, to avoid accessing the DRAM, see above. The current location is close to the rest of PL310 init in SPL.
Quoted 8 lines
Quoted 4 lines
Quoted 2 lines
OK, I just did not understand why this is necessary before accessing DRAM.
:)
Sorry, I phrased that wrong. I *do* (still) not understand why this is necessary before accessing DRAM. Sorry if I'm annoying you with that, but I'd like to understand.
Because we know that once the L2 cache is enabled and until it's content is invalidated, it may return stale entries (*). The stale entries in the L2 cache can only come from a cacheable area. Linux treats the DRAM as cacheable area, the registers and OCRAM are not cacheable. Therefore, if we do not fetch anything from DRAM until the L2 cache is invalidated, we will not get any stale entries in return, and thus we are safe.
The best way to assure that we don't access the DRAM is to execute from OCRAM, so to do the invalidation in SPL. The best way to assure we won't generate any traffic on the bus is to execute the code from primed L1 I-cache lines.
* U-Boot is complex code and spans multiple L2 cache lines. The code which enables L2 cache can very well reside in a different L2 cache line than the code which invalidates the L2 cache. Furthermore, the code also spans multiple L1 I-cache lines.
The code which enables the L2 cache is likely in active L1 I-cache line, however the code which invalidates the L2 cache is not in L1 I-cache yet, so it needs to be fetched. But (!) now the L2 cache is enabled, so the system first checks whether the code might be in the L2 cache ; if there is a stale entry in just the right L2 cache line, the L2 cache will refill the L1 I-cache with undefined stale content, which the CPU will happily execute and crash.
Does it make sense now ? Just keep asking until it does make sense ...
Quoted 1 line
Does the DWMMC DMA really trigger PL310 operations?
No
Quoted 1 line
I thought DMA would only affect cache if it goes through ACP/SCU?
Correct
I suspect shuffling data between the SD card (DWMMC puts data directly in DRAM using DMA) and CPU (which looks at the DRAM through multiple levels of caches, so it has to invalidate/refill/flush the caches to access the data delivered by DWMMC DMA in DRAM) just primes the L2 cache with the right kind of entries to trigger the issue.
[...]
Signature · 2 lines
Best regards, Marek Vasut
PATCH 1/1 ARM: socfpga: Clear PL310 early in SPL
Am 19.02.2019 um 21:07 schrieb Marek Vasut:
Quoted 184 lines · Simon Goldschmidt
Quoted 149 lines
Am 19.02.2019 um 17:39 schrieb Marek Vasut:
Quoted 148 lines · Simon Goldschmidt
Quoted 126 lines
Am 19.02.2019 um 17:00 schrieb Marek Vasut:
Quoted 115 lines · Simon Goldschmidt
Quoted 114 lines
Am 19.02.2019 um 15:00 schrieb Marek Vasut:
Quoted 61 lines · Simon Goldschmidt
Quoted 60 lines
Am Di., 19. Feb. 2019, 14:45 hat Marek Vasut <[email protected] <mailto:[email protected]>> geschrieben:
On 2/19/19 2:28 PM, Simon Goldschmidt wrote:
> On Tue, Feb 19, 2019 at 1:53 PM Marek Vasut <[email protected] <mailto:[email protected]>> wrote:
>>
>> On SoCFPGA Gen5 systems, it can rarely happen that a
reboot from
Linux
>> will result in stale data in PL310 L2 cache controller.
Even if
the L2
>> cache controller is disabled via the CTRL registerCTRL_EN
bit, those
>> data can interfere with operation of devices using DMA,
like
e.g. the
>> DWMMC controller. This can in turn cause e.g. SPL to fail
reading
data
>> from SD/MMC.
>>
>> The obvious solution here would be to fully reset the L2
cache
controller
>> via the reset manager MPUMODRST L2 bit, however this
causes bus
hang even
>> if executed entirely from L1 I-cache to avoid generating
any bus
traffic
>> through the L2 cache controller.
>>
>> This patch thus configures and enables the L2 cache
controller
very early
>> in the SPL boot process, clears the L2 cache and disables
the L2
cache
>> controller again.
>>
>> The reason for doing it in SPL is because we need to avoid accessing any
>> of the potentially stale data in the L2 cache, and we are
certain
any of
>> the stale data will be below the OCRAM address range. To
further
reduce
>> bus traffic during the L2 cache invalidation, we enable L1 I-cache and
>> run the invalidation code entirely out of the L1 I-cache. >>
>> Signed-off-by: Marek Vasut <[email protected]
<mailto:[email protected]>>
>> Cc: Dalon Westergreen <[email protected] <mailto:[email protected]>>Quoted 8 lines
Quoted 8 lines
>> Cc: Dinh Nguyen <[email protected]
<mailto:[email protected]>>
>
> We've tested this and it seems to fix the issue, so:
>
> Tested-by: Simon Goldschmidt
<[email protected]
<mailto:[email protected]>>Seems like I was a little fast there: the patch does not apply cleanly to master. It did apply to our tree, obviously.
It's missing an include to <
asm/pl310.h> and 'pl310' is undefined. Probably a result of rebasing it...Quoted 37 lines
Quoted 36 lines
>
> However, I don't really get why clearing the L2 cache
later in
U-Boot
> isn't good enough.Because when U-Boot is running, it is already running from
RAM,
which is
on different PL310 master than OCRAM, and you're likely to
hit the
corrupted cache lines on the DRAM master which is primed by
prior Linux
operation. Such cacheline can be hit between the code
enabling
the PL310
and the code invalidating it, which is a C code, using
stack and
calling
functions, thus accessing memory which would likely reside in
different
PL310 cachelines. If one of those cachelines contains
invalid/corrupted
data, they can be provided to the CPU before the cacheline is invalidated.To further reduce the likelihood of hitting any such cache
line,
the
code which enables the PL310 and invalidates it runs from
primed L1
icache lines, thus generating no bus traffic at all.Ah, right. I somehow didn't realize that invalidating is done after enabling:-)
Quoted 7 lines
Quoted 5 lines
Quoted 2 lines
Right. If it could be done before (like e.g. by whacking the L2CC reset), that'd be fantastic.
I haven't yet found a way to cleanly reproduce this on my socrates, so unfortunately, I currently cannot test any suggestions, right now :-(
Did you ever have this hang happen on SoCrates ?
I think so, yes.
Quoted 8 lines
What I did was running some lmb tests loading from mmc, removing the sdcard to write an updated version of U-Boot, replacing it and then typing 'reset'. It would then sometimes hang in cache initialization - I added debug prints throughout cache initialization and was debug prints from one function and then nothing more after a simple function call...
That would very much fit to your explanation. However, I cannot reproduce this currently :-(
So basically a board which uses DMA (like the DWMMC DMA) is likely to be affected. Surely, the DMA does trigger PL310 operations.
Quoted 9 lines
Quoted 8 lines
Quoted 4 lines
Wouldn't it be better to place this function in some central position near the cache initialization code instead of running it from SPL? Or is
it required to run it in SPL?It's required in SPL as soon as possible, to avoid accessing the DRAM, see above. The current location is close to the rest of PL310 init in SPL.
Quoted 8 lines
Quoted 4 lines
Quoted 2 lines
OK, I just did not understand why this is necessary before accessing DRAM.
:)
Sorry, I phrased that wrong. I *do* (still) not understand why this is necessary before accessing DRAM. Sorry if I'm annoying you with that, but I'd like to understand.
Because we know that once the L2 cache is enabled and until it's content is invalidated, it may return stale entries (*). The stale entries in the L2 cache can only come from a cacheable area. Linux treats the DRAM as cacheable area, the registers and OCRAM are not cacheable. Therefore, if we do not fetch anything from DRAM until the L2 cache is invalidated, we will not get any stale entries in return, and thus we are safe.
The best way to assure that we don't access the DRAM is to execute from OCRAM, so to do the invalidation in SPL. The best way to assure we won't generate any traffic on the bus is to execute the code from primed L1 I-cache lines.
* U-Boot is complex code and spans multiple L2 cache lines. The code which enables L2 cache can very well reside in a different L2 cache line than the code which invalidates the L2 cache. Furthermore, the code also spans multiple L1 I-cache lines.
The code which enables the L2 cache is likely in active L1 I-cache line, however the code which invalidates the L2 cache is not in L1 I-cache yet, so it needs to be fetched. But (!) now the L2 cache is enabled, so the system first checks whether the code might be in the L2 cache ; if there is a stale entry in just the right L2 cache line, the L2 cache will refill the L1 I-cache with undefined stale content, which the CPU will happily execute and crash.
Quoted 1 line
Does it make sense now ? Just keep asking until it does make sense ...
Yes, absolutely! Thanks for the excellent explanation!
I hadn't payed attention to the fact that SPL runs from OCRAM and concentrated on DWMMC. Because what kept us racking our brains was the fact that this hang only occurs on one platform booting from eMMC, but not on a similar platform booting from SPI-NOR.
And yes, like you I'm wondering who the hell should be able to write mpumodrst.l2 if not the CPU... I also tried this but failed ;-)
Regards,
Simon
Quoted 17 lines
Quoted 1 line
Does the DWMMC DMA really trigger PL310 operations?
No
Quoted 1 line
I thought DMA would only affect cache if it goes through ACP/SCU?
Correct
I suspect shuffling data between the SD card (DWMMC puts data directly in DRAM using DMA) and CPU (which looks at the DRAM through multiple levels of caches, so it has to invalidate/refill/flush the caches to access the data delivered by DWMMC DMA in DRAM) just primes the L2 cache with the right kind of entries to trigger the issue.
[...]
PATCH 1/1 ARM: socfpga: Clear PL310 early in SPL
Quoted 209 lines · Simon Goldschmidt
Am 19.02.2019 um 21:07 schrieb Marek Vasut:
Quoted 208 lines · Simon Goldschmidt
Quoted 173 lines
Am 19.02.2019 um 17:39 schrieb Marek Vasut:
Quoted 172 lines · Simon Goldschmidt
Quoted 153 lines
Am 19.02.2019 um 17:00 schrieb Marek Vasut:
Quoted 134 lines · Simon Goldschmidt
Quoted 133 lines
Am 19.02.2019 um 15:00 schrieb Marek Vasut:
Quoted 72 lines · Simon Goldschmidt
Quoted 71 lines
Am Di., 19. Feb. 2019, 14:45 hat Marek Vasut <[email protected] <mailto:[email protected]>> geschrieben:
On 2/19/19 2:28 PM, Simon Goldschmidt wrote:
> On Tue, Feb 19, 2019 at 1:53 PM Marek Vasut <[email protected]
<mailto:[email protected]>> wrote:
>>
>> On SoCFPGA Gen5 systems, it can rarely happen that a reboot from
Linux
>> will result in stale data in PL310 L2 cache controller. Even if
the L2
>> cache controller is disabled via the CTRL registerCTRL_EN
bit, those
>> data can interfere with operation of devices using DMA, like
e.g. the
>> DWMMC controller. This can in turn cause e.g. SPL to fail
reading
data
>> from SD/MMC.
>>
>> The obvious solution here would be to fully reset the L2
cache
controller
>> via the reset manager MPUMODRST L2 bit, however this causes bus
hang even
>> if executed entirely from L1 I-cache to avoid generating
any bus
traffic
>> through the L2 cache controller.
>>
>> This patch thus configures and enables the L2 cache controller
very early
>> in the SPL boot process, clears the L2 cache and disables
the L2
cache
>> controller again.
>>
>> The reason for doing it in SPL is because we need to avoid
accessing any
>> of the potentially stale data in the L2 cache, and we are
certain
any of
>> the stale data will be below the OCRAM address range. To
further
reduce
>> bus traffic during the L2 cache invalidation, we enable L1
I-cache and
>> run the invalidation code entirely out of the L1 I-cache.
>>
>> Signed-off-by: Marek Vasut <[email protected] <mailto:[email protected]>>
>> Cc: Dalon Westergreen <[email protected] <mailto:[email protected]>>Quoted 8 lines
Quoted 8 lines
>> Cc: Dinh Nguyen <[email protected] <mailto:[email protected]>>
>
> We've tested this and it seems to fix the issue, so:
>
> Tested-by: Simon Goldschmidt <[email protected] <mailto:[email protected]>>Seems like I was a little fast there: the patch does not apply cleanly
to master. It did apply to our tree, obviously.It's missing an include to <
asm/pl310.h> and 'pl310' is undefined. Probably a result of rebasing it...Quoted 43 lines
Quoted 37 lines
>
> However, I don't really get why clearing the L2 cache later in
U-Boot
> isn't good enough.Because when U-Boot is running, it is already running from RAM,
which is
on different PL310 master than OCRAM, and you're likely to hit the
corrupted cache lines on the DRAM master which is primed by
prior Linux
operation. Such cacheline can be hit between the code enabling
the PL310
and the code invalidating it, which is a C code, using stack and
calling
functions, thus accessing memory which would likely reside in
different
PL310 cachelines. If one of those cachelines contains invalid/corrupted
data, they can be provided to the CPU before the cacheline is
invalidated.To further reduce the likelihood of hitting any such cache line,
the
code which enables the PL310 and invalidates it runs from primed L1
icache lines, thus generating no bus traffic at all.Quoted 3 lines
Ah, right. I somehow didn't realize that invalidating is done after
enabling:-)Right. If it could be done before (like e.g. by whacking the L2CC reset), that'd be fantastic.
Quoted 6 lines
Quoted 4 lines
I haven't yet found a way to cleanly reproduce this on my socrates, so
unfortunately, I currently cannot test any suggestions, right now :-(Did you ever have this hang happen on SoCrates ?
I think so, yes.
What I did was running some lmb tests loading from mmc, removing the sdcard to write an updated version of U-Boot, replacing it and then typing 'reset'. It would then sometimes hang in cache initialization - I
added debug prints throughout cache initialization and was debug prints
from one function and then nothing more after a simple function call...Quoted 2 lines
That would very much fit to your explanation. However, I cannot reproduce this currently :-(
So basically a board which uses DMA (like the DWMMC DMA) is likely to be
affected. Surely, the DMA does trigger PL310 operations.Quoted 11 lines
Quoted 10 lines
Quoted 5 lines
Wouldn't it be better to place this function in some central position
near the cache initialization code instead of running it from SPL? Or is
it required to run it in SPL?It's required in SPL as soon as possible, to avoid accessing the DRAM,
see above. The current location is close to the rest of PL310 init in SPL.Quoted 8 lines
Quoted 4 lines
Quoted 2 lines
OK, I just did not understand why this is necessary before accessing DRAM.
:)
Sorry, I phrased that wrong. I *do* (still) not understand why this is necessary before accessing DRAM. Sorry if I'm annoying you with that, but I'd like to understand.
Because we know that once the L2 cache is enabled and until it's content is invalidated, it may return stale entries (*). The stale entries in the L2 cache can only come from a cacheable area. Linux treats the DRAM as cacheable area, the registers and OCRAM are not cacheable. Therefore, if we do not fetch anything from DRAM until the L2 cache is invalidated, we will not get any stale entries in return, and thus we are safe.
The best way to assure that we don't access the DRAM is to execute from OCRAM, so to do the invalidation in SPL. The best way to assure we won't generate any traffic on the bus is to execute the code from primed L1 I-cache lines.
* U-Boot is complex code and spans multiple L2 cache lines. The code which enables L2 cache can very well reside in a different L2 cache line than the code which invalidates the L2 cache. Furthermore, the code also spans multiple L1 I-cache lines.
The code which enables the L2 cache is likely in active L1 I-cache line, however the code which invalidates the L2 cache is not in L1 I-cache yet, so it needs to be fetched. But (!) now the L2 cache is enabled, so the system first checks whether the code might be in the L2 cache ; if there is a stale entry in just the right L2 cache line, the L2 cache will refill the L1 I-cache with undefined stale content, which the CPU will happily execute and crash.
Quoted 8 lines
Quoted 1 line
Does it make sense now ? Just keep asking until it does make sense ...
Yes, absolutely! Thanks for the excellent explanation!
I hadn't payed attention to the fact that SPL runs from OCRAM and concentrated on DWMMC. Because what kept us racking our brains was the fact that this hang only occurs on one platform booting from eMMC, but not on a similar platform booting from SPI-NOR.
I think that's because the CQSPI doesn't do DMA into the memory, but rather exposes the SPI NOR as a memory mapped chunk of register area.
Quoted 2 lines
And yes, like you I'm wondering who the hell should be able to write mpumodrst.l2 if not the CPU... I also tried this but failed ;-)
Something in the FPGA, a well-programmed PL330 DMA or any other L3 master I think.
Signature · 2 lines
Best regards, Marek Vasut