Skip to contentSkip to conversation

arm64: issue ISB after updating system registers

PATCH New Volodymyr Babchuk 7 participants 9 messages Jun 24, 2020 – Jul 7, 2020 Arm (32-bit)

lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 arm64: issue ISB after updating system registers

ARM Architecture reference manual clearly states that PE pipeline should be flushed after any change to system registers. Refer to paragraph "B2.3.5 Memory Barriers" at page B2-92 of "Arm Architecture Reference Manual ARMv8 for ARMv8-A Architecture Profile" (ARM DDI 0487B.a).

Failing to issue instruction memory synchronization barrier can lead to spurious errors, like synchronous exception when accessing FPU registers. This is very prominent on CPUs with long instruction pipeline, like ARM Cortex A72.

This change fixes the following U-Boot panic:

"Synchronous Abort" handler, esr 0x1fe00000
elr: 00000000800948cc lr : 0000000080091e04
x0 : 00000000801ffdc8 x1 : 00000000000000c8
x2 : 00000000800979d4 x3 : 00000000801ffc60
x4 : 00000000801ffd40 x5 : ffffff80ffffffd8
x6 : 00000000801ffd70 x7 : 00000000801ffd70
x8 : 000000000000000a x9 : 0000000000000000
x10: 0000000000000044 x11: 0000000000000000
x12: 0000000000000000 x13: 0000000000000000
x14: 0000000000000000 x15: 0000000000000000
x16: 000000008008b2e0 x17: 0000000000000000
x18: 00000000801ffec0 x19: 00000000800957b0
x20: 00000000000000c8 x21: 00000000801ffdc8
x22: 000000008009909e x23: 0000000000000000
x24: 0000000000000000 x25: 0000000000000000
x26: 0000000000000000 x27: 0000000000000000
x28: 0000000000000000 x29: 00000000801ffc50

 Code: a94417e4 a90217e4 a9051fe6 a90617e4 (3d801fe0)

While executing instruction

str q0, [sp, #112]

in vsnprintf() prologue. This panic was observed only on Cortex A72 so far.

This patch places ISBs on other strategic places as well.

Also, this probably is the right fix for the issue workarounded in the commit 45f41c13 ("ARM: uniphier: add weird workaround code for LD20")

--

Changes from v1:
- Added ISBs under CONFIG_ARMV8_SET_SMPEN and erratas.
- Added Stefano, Julien and Oleksandr

---
 arch/arm/cpu/armv8/start.S | 8 ++++++++
 1 file changed, 8 insertions(+)
arch/arm/cpu/armv8/start.S +8 -0
index 99d126660d..002698b501 100644
@@ -120,6 +120,7 @@ pie_fixup_done:
 	mov	x0, #3 << 20
 	msr	cpacr_el1, x0			/* Enable FP/SIMD */
 0:
+	isb
 
 	/*
 	 * Enable SMPEN bit for coherency.
@@ -132,6 +133,7 @@ pie_fixup_done:
 	mrs     x0, S3_1_c15_c2_1               /* cpuectlr_el1 */
 	orr     x0, x0, #0x40
 	msr     S3_1_c15_c2_1, x0
+	isb
 1:
 #endif
 
@@ -233,6 +235,7 @@ apply_a53_core_errata:
 	/* Enable data cache clean as data cache clean/invalidate */
 	orr	x0, x0, #1 << 44
 	msr	S3_1_c15_c2_0, x0	/* cpuactlr_el1 */
+	isb
 #endif
 	b 0b
 
@@ -247,6 +250,7 @@ apply_a57_core_errata:
 	/* Disable write streaming no-allocate threshold */
 	orr	x0, x0, #3 << 27
 	msr	S3_1_c15_c2_0, x0	/* cpuactlr_el1 */
+	isb
 #endif
 
 #ifdef CONFIG_ARM_ERRATA_826974
@@ -254,6 +258,7 @@ apply_a57_core_errata:
 	/* Disable speculative load execution ahead of a DMB */
 	orr	x0, x0, #1 << 59
 	msr	S3_1_c15_c2_0, x0	/* cpuactlr_el1 */
+	isb
 #endif
 
 #ifdef CONFIG_ARM_ERRATA_833471
@@ -263,6 +268,7 @@ apply_a57_core_errata:
 	    could impact performance. */
 	orr	x0, x0, #1 << 38
 	msr	S3_1_c15_c2_0, x0	/* cpuactlr_el1 */
+	isb
 #endif
 
 #ifdef CONFIG_ARM_ERRATA_829520
@@ -273,6 +279,7 @@ apply_a57_core_errata:
 	    could impact performance. */
 	orr	x0, x0, #1 << 4
 	msr	S3_1_c15_c2_0, x0	/* cpuactlr_el1 */
+	isb
 #endif
 
 #ifdef CONFIG_ARM_ERRATA_833069
@@ -280,6 +287,7 @@ apply_a57_core_errata:
 	/* Disable Enable Invalidates of BTB bit */
 	and	x0, x0, #0xE
 	msr	S3_1_c15_c2_0, x0	/* cpuactlr_el1 */
+	isb
 #endif
 	b 0b
 ENDPROC(apply_core_errata)
Signature · 1 line
2.27.0
#1 lore raw

PATCH v2 1/1 arm64: issue ISB after updating system registers

(+ some Arm folks)

Hi Volodymyr,

Quoted 40 lines · Volodymyr Babchuk

ARM Architecture reference manual clearly states that PE pipeline should be flushed after any change to system registers. Refer to paragraph "B2.3.5 Memory Barriers" at page B2-92 of "Arm Architecture Reference Manual ARMv8 for ARMv8-A Architecture Profile" (ARM DDI 0487B.a).

Failing to issue instruction memory synchronization barrier can lead to spurious errors, like synchronous exception when accessing FPU registers. This is very prominent on CPUs with long instruction pipeline, like ARM Cortex A72.

This change fixes the following U-Boot panic:

"Synchronous Abort" handler, esr 0x1fe00000
elr: 00000000800948cc lr : 0000000080091e04
x0 : 00000000801ffdc8 x1 : 00000000000000c8
x2 : 00000000800979d4 x3 : 00000000801ffc60
x4 : 00000000801ffd40 x5 : ffffff80ffffffd8
x6 : 00000000801ffd70 x7 : 00000000801ffd70
x8 : 000000000000000a x9 : 0000000000000000
x10: 0000000000000044 x11: 0000000000000000
x12: 0000000000000000 x13: 0000000000000000
x14: 0000000000000000 x15: 0000000000000000
x16: 000000008008b2e0 x17: 0000000000000000
x18: 00000000801ffec0 x19: 00000000800957b0
x20: 00000000000000c8 x21: 00000000801ffdc8
x22: 000000008009909e x23: 0000000000000000
x24: 0000000000000000 x25: 0000000000000000
x26: 0000000000000000 x27: 0000000000000000
x28: 0000000000000000 x29: 00000000801ffc50

  Code: a94417e4 a90217e4 a9051fe6 a90617e4 (3d801fe0)

While executing instruction

str q0, [sp, #112]

in vsnprintf() prologue. This panic was observed only on Cortex A72 so far.

Quoted 7 lines

This patch places ISBs on other strategic places as well.

Also, this probably is the right fix for the issue workarounded in the commit 45f41c13 ("ARM: uniphier: add weird workaround code for LD20")

Please use julien at xen.org :).

Quoted 4 lines
CC Tom Rini <[email protected]>
CC Masahiro Yamada <[email protected]>
CC Stefano Stabellini <[email protected]>
Quoted 79 lines

--

Changes from v1:
- Added ISBs under CONFIG_ARMV8_SET_SMPEN and erratas.
- Added Stefano, Julien and Oleksandr

---
  arch/arm/cpu/armv8/start.S | 8 ++++++++
  1 file changed, 8 insertions(+)
arch/arm/cpu/armv8/start.S +8 -0
index 99d126660d..002698b501 100644
@@ -120,6 +120,7 @@ pie_fixup_done:
  	mov	x0, #3 << 20
  	msr	cpacr_el1, x0			/* Enable FP/SIMD */
  0:
+	isb
 
  	/*
  	 * Enable SMPEN bit for coherency.
@@ -132,6 +133,7 @@ pie_fixup_done:
  	mrs     x0, S3_1_c15_c2_1               /* cpuectlr_el1 */
  	orr     x0, x0, #0x40
  	msr     S3_1_c15_c2_1, x0
+	isb
  1:
  #endif
 
@@ -233,6 +235,7 @@ apply_a53_core_errata:
  	/* Enable data cache clean as data cache clean/invalidate */
  	orr	x0, x0, #1 << 44
  	msr	S3_1_c15_c2_0, x0	/* cpuactlr_el1 */
+	isb
  #endif
  	b 0b
 
@@ -247,6 +250,7 @@ apply_a57_core_errata:
  	/* Disable write streaming no-allocate threshold */
  	orr	x0, x0, #3 << 27
  	msr	S3_1_c15_c2_0, x0	/* cpuactlr_el1 */
+	isb
  #endif
 
  #ifdef CONFIG_ARM_ERRATA_826974
@@ -254,6 +258,7 @@ apply_a57_core_errata:
  	/* Disable speculative load execution ahead of a DMB */
  	orr	x0, x0, #1 << 59
  	msr	S3_1_c15_c2_0, x0	/* cpuactlr_el1 */
+	isb
  #endif
 
  #ifdef CONFIG_ARM_ERRATA_833471
@@ -263,6 +268,7 @@ apply_a57_core_errata:
  	    could impact performance. */
  	orr	x0, x0, #1 << 38
  	msr	S3_1_c15_c2_0, x0	/* cpuactlr_el1 */
+	isb
  #endif
 
  #ifdef CONFIG_ARM_ERRATA_829520
@@ -273,6 +279,7 @@ apply_a57_core_errata:
  	    could impact performance. */
  	orr	x0, x0, #1 << 4
  	msr	S3_1_c15_c2_0, x0	/* cpuactlr_el1 */
+	isb
  #endif
 
  #ifdef CONFIG_ARM_ERRATA_833069
@@ -280,6 +287,7 @@ apply_a57_core_errata:
  	/* Disable Enable Invalidates of BTB bit */
  	and	x0, x0, #0xE
  	msr	S3_1_c15_c2_0, x0	/* cpuactlr_el1 */
+	isb
  #endif
  	b 0b
  ENDPROC(apply_core_errata)
Signature · 1 line
Julien Grall
#2 lore raw

PATCH v2 1/1 arm64: issue ISB after updating system registers

Quoted 47 lines · Julien Grall

(+ some Arm folks)

Hi Volodymyr,

Quoted 35 lines · Volodymyr Babchuk

ARM Architecture reference manual clearly states that PE pipeline should be flushed after any change to system registers. Refer to paragraph "B2.3.5 Memory Barriers" at page B2-92 of "Arm Architecture Reference Manual ARMv8 for ARMv8-A Architecture Profile" (ARM DDI 0487B.a).
Failing to issue instruction memory synchronization barrier can lead to spurious errors, like synchronous exception when accessing FPU registers. This is very prominent on CPUs with long instruction pipeline, like ARM Cortex A72.
This change fixes the following U-Boot panic:
"Synchronous Abort" handler, esr 0x1fe00000
elr: 00000000800948cc lr : 0000000080091e04
x0 : 00000000801ffdc8 x1 : 00000000000000c8
x2 : 00000000800979d4 x3 : 00000000801ffc60
x4 : 00000000801ffd40 x5 : ffffff80ffffffd8
x6 : 00000000801ffd70 x7 : 00000000801ffd70
x8 : 000000000000000a x9 : 0000000000000000
x10: 0000000000000044 x11: 0000000000000000
x12: 0000000000000000 x13: 0000000000000000
x14: 0000000000000000 x15: 0000000000000000
x16: 000000008008b2e0 x17: 0000000000000000
x18: 00000000801ffec0 x19: 00000000800957b0
x20: 00000000000000c8 x21: 00000000801ffdc8
x22: 000000008009909e x23: 0000000000000000
x24: 0000000000000000 x25: 0000000000000000
x26: 0000000000000000 x27: 0000000000000000
x28: 0000000000000000 x29: 00000000801ffc50

 Code: a94417e4 a90217e4 a9051fe6 a90617e4 (3d801fe0)

While executing instruction
str q0, [sp, #112]
in vsnprintf() prologue. This panic was observed only on Cortex A72 so far.
This patch places ISBs on other strategic places as well. Also, this probably is the right fix for the issue workarounded in the commit 45f41c13 ("ARM: uniphier: add weird workaround code for LD20")

Quoted 2 lines
Quoted 8 lines

Please use julien at xen.org :).

Quoted 4 lines
CC Tom Rini <[email protected]>
CC Masahiro Yamada <[email protected]>
CC Stefano Stabellini <[email protected]>

Reviewed-by Bertrand Marquis <[email protected]>

Quoted 64 lines
Quoted 64 lines

--
Changes from v1:
- Added ISBs under CONFIG_ARMV8_SET_SMPEN and erratas.
- Added Stefano, Julien and Oleksandr

---
 arch/arm/cpu/armv8/start.S | 8 ++++++++
 1 file changed, 8 insertions(+)
arch/arm/cpu/armv8/start.S +2 -0
index 99d126660d..002698b501 100644
@@ -120,6 +120,7 @@ pie_fixup_done:
 movx0, #3 << 20
 msrcpacr_el1, x0/* Enable FP/SIMD */
 0:
+isb
   /*
  * Enable SMPEN bit for coherency.
@@ -132,6 +133,7 @@ pie_fixup_done:
 mrs     x0, S3_1_c15_c2_1               /* cpuectlr_el1 */
 orr     x0, x0, #0x40
 msr     S3_1_c15_c2_1, x0
+isb
 1:
 #endif
 @@ -233,6 +235,7 @@ apply_a53_core_errata:

/* Enable data cache clean as data cache clean/invalidate */ orrx0, x0, #1 << 44
msrS3_1_c15_c2_0, x0/* cpuactlr_el1 */

arch/arm/cpu/armv8/start.S +5 -0
+isb
 #endif
 b 0b
 @@ -247,6 +250,7 @@ apply_a57_core_errata:
 /* Disable write streaming no-allocate threshold */
 orrx0, x0, #3 << 27
 msrS3_1_c15_c2_0, x0/* cpuactlr_el1 */
+isb
 #endif
   #ifdef CONFIG_ARM_ERRATA_826974
@@ -254,6 +258,7 @@ apply_a57_core_errata:
 /* Disable speculative load execution ahead of a DMB */
 orrx0, x0, #1 << 59
 msrS3_1_c15_c2_0, x0/* cpuactlr_el1 */
+isb
 #endif
   #ifdef CONFIG_ARM_ERRATA_833471
@@ -263,6 +268,7 @@ apply_a57_core_errata:
     could impact performance. */
 orrx0, x0, #1 << 38
 msrS3_1_c15_c2_0, x0/* cpuactlr_el1 */
+isb
 #endif
   #ifdef CONFIG_ARM_ERRATA_829520
@@ -273,6 +279,7 @@ apply_a57_core_errata:
     could impact performance. */
 orrx0, x0, #1 << 4
 msrS3_1_c15_c2_0, x0/* cpuactlr_el1 */
+isb
 #endif
   #ifdef CONFIG_ARM_ERRATA_833069
@@ -280,6 +287,7 @@ apply_a57_core_errata:
 /* Disable Enable Invalidates of BTB bit */
 andx0, x0, #0xE
Quoted 8 lines

arch/arm/cpu/armv8/start.S apply_a57_core_errata: · line 293

Quoted 5 lines
arch/arm/cpu/armv8/start.S +1 -0
 msrS3_1_c15_c2_0, x0/* cpuactlr_el1 */
+isb
 #endif
 b 0b
 ENDPROC(apply_core_errata)

--
Julien Grall

IMPORTANT NOTICE: The contents of this email and any attachments are confidential and may also be privileged. If you are not the intended recipient, please notify the sender immediately and do not disclose the contents to any other person, use it for any purpose, or store or copy the information in any medium. Thank you.

#3 lore raw

PATCH v2 1/1 arm64: issue ISB after updating system registers

In reply to Volodymyr Babchuk

On 24/06/2020 02:05, Volodymyr Babchuk wrote:

Hi Volodymyr,

thanks for the find and for taking care! And thanks Julien for forwarding!

Quoted 2 lines

ARM Architecture reference manual clearly states that PE pipeline should be flushed after any change to system registers.

Don't want to be pedantic here, but this does not affect all system registers. Some clearly don't need an ISB, some are self-synchronising. The manual speaks of "*Examples* of context-changing operations ..." Could you change this to "... after changes to some system registers."?

Quoted 6 lines

Refer to
paragraph "B2.3.5 Memory Barriers" at page B2-92 of "Arm Architecture Reference Manual ARMv8 for ARMv8-A Architecture Profile" (ARM DDI 0487B.a).

Failing to issue instruction memory synchronization barrier can lead

It's just "instruction synchronization barrier". This is not really about memory, it's about making sure that the processor re-fetches the next instructions and "re-evaluates" them again, to take the changed system state into account. (Flushing the pipeline is just one possible way to implement this, btw.)

Quoted 38 lines

to spurious errors, like synchronous exception when accessing FPU registers. This is very prominent on CPUs with long instruction pipeline, like ARM Cortex A72.

This change fixes the following U-Boot panic:

"Synchronous Abort" handler, esr 0x1fe00000
elr: 00000000800948cc lr : 0000000080091e04
x0 : 00000000801ffdc8 x1 : 00000000000000c8
x2 : 00000000800979d4 x3 : 00000000801ffc60
x4 : 00000000801ffd40 x5 : ffffff80ffffffd8
x6 : 00000000801ffd70 x7 : 00000000801ffd70
x8 : 000000000000000a x9 : 0000000000000000
x10: 0000000000000044 x11: 0000000000000000
x12: 0000000000000000 x13: 0000000000000000
x14: 0000000000000000 x15: 0000000000000000
x16: 000000008008b2e0 x17: 0000000000000000
x18: 00000000801ffec0 x19: 00000000800957b0
x20: 00000000000000c8 x21: 00000000801ffdc8
x22: 000000008009909e x23: 0000000000000000
x24: 0000000000000000 x25: 0000000000000000
x26: 0000000000000000 x27: 0000000000000000
x28: 0000000000000000 x29: 00000000801ffc50

 Code: a94417e4 a90217e4 a9051fe6 a90617e4 (3d801fe0)

While executing instruction

str q0, [sp, #112]

in vsnprintf() prologue. This panic was observed only on Cortex A72 so far.

This patch places ISBs on other strategic places as well.

Also, this probably is the right fix for the issue workarounded in the commit 45f41c13 ("ARM: uniphier: add weird workaround code for LD20")

Quoted 6 lines

I checked the A57 ACTLR_EL1 and ECTLR_EL1 bits affected below, and indeed an ISB seems to be in order here. Since this is the initialization code path, it wouldn't hurt anyway ;-)

Cheers,
Andre

Quoted 78 lines

--

Changes from v1:
- Added ISBs under CONFIG_ARMV8_SET_SMPEN and erratas.
- Added Stefano, Julien and Oleksandr

---
 arch/arm/cpu/armv8/start.S | 8 ++++++++
 1 file changed, 8 insertions(+)
arch/arm/cpu/armv8/start.S +8 -0
index 99d126660d..002698b501 100644
@@ -120,6 +120,7 @@ pie_fixup_done:
 	mov	x0, #3 << 20
 	msr	cpacr_el1, x0			/* Enable FP/SIMD */
 0:
+	isb
 
 	/*
 	 * Enable SMPEN bit for coherency.
@@ -132,6 +133,7 @@ pie_fixup_done:
 	mrs     x0, S3_1_c15_c2_1               /* cpuectlr_el1 */
 	orr     x0, x0, #0x40
 	msr     S3_1_c15_c2_1, x0
+	isb
 1:
 #endif
 
@@ -233,6 +235,7 @@ apply_a53_core_errata:
 	/* Enable data cache clean as data cache clean/invalidate */
 	orr	x0, x0, #1 << 44
 	msr	S3_1_c15_c2_0, x0	/* cpuactlr_el1 */
+	isb
 #endif
 	b 0b
 
@@ -247,6 +250,7 @@ apply_a57_core_errata:
 	/* Disable write streaming no-allocate threshold */
 	orr	x0, x0, #3 << 27
 	msr	S3_1_c15_c2_0, x0	/* cpuactlr_el1 */
+	isb
 #endif
 
 #ifdef CONFIG_ARM_ERRATA_826974
@@ -254,6 +258,7 @@ apply_a57_core_errata:
 	/* Disable speculative load execution ahead of a DMB */
 	orr	x0, x0, #1 << 59
 	msr	S3_1_c15_c2_0, x0	/* cpuactlr_el1 */
+	isb
 #endif
 
 #ifdef CONFIG_ARM_ERRATA_833471
@@ -263,6 +268,7 @@ apply_a57_core_errata:
 	    could impact performance. */
 	orr	x0, x0, #1 << 38
 	msr	S3_1_c15_c2_0, x0	/* cpuactlr_el1 */
+	isb
 #endif
 
 #ifdef CONFIG_ARM_ERRATA_829520
@@ -273,6 +279,7 @@ apply_a57_core_errata:
 	    could impact performance. */
 	orr	x0, x0, #1 << 4
 	msr	S3_1_c15_c2_0, x0	/* cpuactlr_el1 */
+	isb
 #endif
 
 #ifdef CONFIG_ARM_ERRATA_833069
@@ -280,6 +287,7 @@ apply_a57_core_errata:
 	/* Disable Enable Invalidates of BTB bit */
 	and	x0, x0, #0xE
 	msr	S3_1_c15_c2_0, x0	/* cpuactlr_el1 */
+	isb
 #endif
 	b 0b
 ENDPROC(apply_core_errata)
#4 lore raw

PATCH v2 1/1 arm64: issue ISB after updating system registers

In reply to Volodymyr Babchuk
Quoted 38 lines · Volodymyr Babchuk

ARM Architecture reference manual clearly states that PE pipeline should be flushed after any change to system registers. Refer to paragraph "B2.3.5 Memory Barriers" at page B2-92 of "Arm Architecture Reference Manual ARMv8 for ARMv8-A Architecture Profile" (ARM DDI 0487B.a).

Failing to issue instruction memory synchronization barrier can lead to spurious errors, like synchronous exception when accessing FPU registers. This is very prominent on CPUs with long instruction pipeline, like ARM Cortex A72.

This change fixes the following U-Boot panic:

"Synchronous Abort" handler, esr 0x1fe00000
elr: 00000000800948cc lr : 0000000080091e04
x0 : 00000000801ffdc8 x1 : 00000000000000c8
x2 : 00000000800979d4 x3 : 00000000801ffc60
x4 : 00000000801ffd40 x5 : ffffff80ffffffd8
x6 : 00000000801ffd70 x7 : 00000000801ffd70
x8 : 000000000000000a x9 : 0000000000000000
x10: 0000000000000044 x11: 0000000000000000
x12: 0000000000000000 x13: 0000000000000000
x14: 0000000000000000 x15: 0000000000000000
x16: 000000008008b2e0 x17: 0000000000000000
x18: 00000000801ffec0 x19: 00000000800957b0
x20: 00000000000000c8 x21: 00000000801ffdc8
x22: 000000008009909e x23: 0000000000000000
x24: 0000000000000000 x25: 0000000000000000
x26: 0000000000000000 x27: 0000000000000000
x28: 0000000000000000 x29: 00000000801ffc50

 Code: a94417e4 a90217e4 a9051fe6 a90617e4 (3d801fe0)

While executing instruction

str q0, [sp, #112]

Quoted 7 lines

in vsnprintf() prologue. This panic was observed only on Cortex A72 so far.

This patch places ISBs on other strategic places as well.

Also, this probably is the right fix for the issue workarounded in the commit 45f41c13 ("ARM: uniphier: add weird workaround code for LD20")

Thanks for addressing this issue.
Currently, I do not have a board in hand to test this. (I do not commute to the office due to COVID-19 these days...)

I have another SoC board, but it does not integrate CA72. I have ever seen this problem only on CA72.

Eventually, I will go to the office, and I can test this. But, you do not need to wait for my test if other people review it.

Thank you.

Quoted 85 lines

--

Changes from v1:
- Added ISBs under CONFIG_ARMV8_SET_SMPEN and erratas.
- Added Stefano, Julien and Oleksandr

---
 arch/arm/cpu/armv8/start.S | 8 ++++++++
 1 file changed, 8 insertions(+)
arch/arm/cpu/armv8/start.S +8 -1
index 99d126660d..002698b501 100644
@@ -120,6 +120,7 @@ pie_fixup_done:
        mov     x0, #3 << 20
        msr     cpacr_el1, x0                   /* Enable FP/SIMD */
 0:
+       isb
 
        /*
         * Enable SMPEN bit for coherency.
@@ -132,6 +133,7 @@ pie_fixup_done:
        mrs     x0, S3_1_c15_c2_1               /* cpuectlr_el1 */
        orr     x0, x0, #0x40
        msr     S3_1_c15_c2_1, x0
+       isb
 1:
 #endif
 
@@ -233,6 +235,7 @@ apply_a53_core_errata:
        /* Enable data cache clean as data cache clean/invalidate */
        orr     x0, x0, #1 << 44
        msr     S3_1_c15_c2_0, x0       /* cpuactlr_el1 */
+       isb
 #endif
        b 0b
 
@@ -247,6 +250,7 @@ apply_a57_core_errata:
        /* Disable write streaming no-allocate threshold */
        orr     x0, x0, #3 << 27
        msr     S3_1_c15_c2_0, x0       /* cpuactlr_el1 */
+       isb
 #endif
 
 #ifdef CONFIG_ARM_ERRATA_826974
@@ -254,6 +258,7 @@ apply_a57_core_errata:
        /* Disable speculative load execution ahead of a DMB */
        orr     x0, x0, #1 << 59
        msr     S3_1_c15_c2_0, x0       /* cpuactlr_el1 */
+       isb
 #endif
 
 #ifdef CONFIG_ARM_ERRATA_833471
@@ -263,6 +268,7 @@ apply_a57_core_errata:
            could impact performance. */
        orr     x0, x0, #1 << 38
        msr     S3_1_c15_c2_0, x0       /* cpuactlr_el1 */
+       isb
 #endif
 
 #ifdef CONFIG_ARM_ERRATA_829520
@@ -273,6 +279,7 @@ apply_a57_core_errata:
            could impact performance. */
        orr     x0, x0, #1 << 4
        msr     S3_1_c15_c2_0, x0       /* cpuactlr_el1 */
+       isb
 #endif
 
 #ifdef CONFIG_ARM_ERRATA_833069
@@ -280,6 +287,7 @@ apply_a57_core_errata:
        /* Disable Enable Invalidates of BTB bit */
        and     x0, x0, #0xE
        msr     S3_1_c15_c2_0, x0       /* cpuactlr_el1 */
+       isb
 #endif
        b 0b
 ENDPROC(apply_core_errata)
--
Quoted 1 line

2.27.0

--
Best Regards
Masahiro Yamada

#5 lore raw

PATCH v2 1/1 arm64: issue ISB after updating system registers

Hi.

Quoted 54 lines · Masahiro Yamada
Quoted 38 lines · Volodymyr Babchuk

ARM Architecture reference manual clearly states that PE pipeline should be flushed after any change to system registers. Refer to paragraph "B2.3.5 Memory Barriers" at page B2-92 of "Arm Architecture Reference Manual ARMv8 for ARMv8-A Architecture Profile" (ARM DDI 0487B.a).

Failing to issue instruction memory synchronization barrier can lead to spurious errors, like synchronous exception when accessing FPU registers. This is very prominent on CPUs with long instruction pipeline, like ARM Cortex A72.

This change fixes the following U-Boot panic:

"Synchronous Abort" handler, esr 0x1fe00000
elr: 00000000800948cc lr : 0000000080091e04
x0 : 00000000801ffdc8 x1 : 00000000000000c8
x2 : 00000000800979d4 x3 : 00000000801ffc60
x4 : 00000000801ffd40 x5 : ffffff80ffffffd8
x6 : 00000000801ffd70 x7 : 00000000801ffd70
x8 : 000000000000000a x9 : 0000000000000000
x10: 0000000000000044 x11: 0000000000000000
x12: 0000000000000000 x13: 0000000000000000
x14: 0000000000000000 x15: 0000000000000000
x16: 000000008008b2e0 x17: 0000000000000000
x18: 00000000801ffec0 x19: 00000000800957b0
x20: 00000000000000c8 x21: 00000000801ffdc8
x22: 000000008009909e x23: 0000000000000000
x24: 0000000000000000 x25: 0000000000000000
x26: 0000000000000000 x27: 0000000000000000
x28: 0000000000000000 x29: 00000000801ffc50

 Code: a94417e4 a90217e4 a9051fe6 a90617e4 (3d801fe0)

While executing instruction

str q0, [sp, #112]

Quoted 7 lines

in vsnprintf() prologue. This panic was observed only on Cortex A72 so far.

This patch places ISBs on other strategic places as well.

Also, this probably is the right fix for the issue workarounded in the commit 45f41c13 ("ARM: uniphier: add weird workaround code for LD20")

Thanks for addressing this issue.
Currently, I do not have a board in hand to test this. (I do not commute to the office due to COVID-19 these days...)

Quoted 9 lines

I have another SoC board, but it does not integrate CA72. I have ever seen this problem only on CA72.

Eventually, I will go to the office, and I can test this. But, you do not need to wait for my test if other people review it.

Thank you.

Today I tested this patch.

Yes, it fixes the CA72 problem on my board.

After this patch is picked up,
I will revert the ugly workaround:

http://patchwork.ozlabs.org/project/uboot/patch/20200706042304.15853-1-yamada.masahiro at socionext.com/

Interestingly, I observe this problem
only on U-Boot running at EL1.

Anyway, this fix makes it work at
any exception level.

--
Best Regards
Masahiro Yamada

#6 lore raw

PATCH v2 1/1 arm64: issue ISB after updating system registers

Quoted 69 lines · Masahiro Yamada

Hi.

Quoted 54 lines · Masahiro Yamada
Quoted 38 lines · Volodymyr Babchuk

ARM Architecture reference manual clearly states that PE pipeline should be flushed after any change to system registers. Refer to paragraph "B2.3.5 Memory Barriers" at page B2-92 of "Arm Architecture Reference Manual ARMv8 for ARMv8-A Architecture Profile" (ARM DDI 0487B.a).

Failing to issue instruction memory synchronization barrier can lead to spurious errors, like synchronous exception when accessing FPU registers. This is very prominent on CPUs with long instruction pipeline, like ARM Cortex A72.

This change fixes the following U-Boot panic:

"Synchronous Abort" handler, esr 0x1fe00000
elr: 00000000800948cc lr : 0000000080091e04
x0 : 00000000801ffdc8 x1 : 00000000000000c8
x2 : 00000000800979d4 x3 : 00000000801ffc60
x4 : 00000000801ffd40 x5 : ffffff80ffffffd8
x6 : 00000000801ffd70 x7 : 00000000801ffd70
x8 : 000000000000000a x9 : 0000000000000000
x10: 0000000000000044 x11: 0000000000000000
x12: 0000000000000000 x13: 0000000000000000
x14: 0000000000000000 x15: 0000000000000000
x16: 000000008008b2e0 x17: 0000000000000000
x18: 00000000801ffec0 x19: 00000000800957b0
x20: 00000000000000c8 x21: 00000000801ffdc8
x22: 000000008009909e x23: 0000000000000000
x24: 0000000000000000 x25: 0000000000000000
x26: 0000000000000000 x27: 0000000000000000
x28: 0000000000000000 x29: 00000000801ffc50

 Code: a94417e4 a90217e4 a9051fe6 a90617e4 (3d801fe0)

While executing instruction

str q0, [sp, #112]

Quoted 7 lines

in vsnprintf() prologue. This panic was observed only on Cortex A72 so far.

This patch places ISBs on other strategic places as well.

Also, this probably is the right fix for the issue workarounded in the commit 45f41c13 ("ARM: uniphier: add weird workaround code for LD20")

Thanks for addressing this issue.
Currently, I do not have a board in hand to test this. (I do not commute to the office due to COVID-19 these days...)

Quoted 9 lines

I have another SoC board, but it does not integrate CA72. I have ever seen this problem only on CA72.

Eventually, I will go to the office, and I can test this. But, you do not need to wait for my test if other people review it.

Thank you.

Quoted 6 lines

Today I tested this patch.

Yes, it fixes the CA72 problem on my board.

Now that MW is open, I want to see this patch in mainline because it fixes the CA-72 problem in the right way.

There was no object except two minor comments from Andre.

Can Tom fix up the commit description when he picks it up?

"... after any change to system registers."

     --->

"... after changes to some system registers."

"instruction memory synchronization barrier"

    --->

"instruction synchronization barrier"

Thanks.

Signature · 2 lines
Best Regards
Masahiro Yamada
#7 lore raw

PATCH v2 1/1 arm64: issue ISB after updating system registers

Quoted 89 lines · Masahiro Yamada
Quoted 69 lines · Masahiro Yamada

Hi.

Quoted 54 lines · Masahiro Yamada
Quoted 38 lines · Volodymyr Babchuk

ARM Architecture reference manual clearly states that PE pipeline should be flushed after any change to system registers. Refer to paragraph "B2.3.5 Memory Barriers" at page B2-92 of "Arm Architecture Reference Manual ARMv8 for ARMv8-A Architecture Profile" (ARM DDI 0487B.a).

Failing to issue instruction memory synchronization barrier can lead to spurious errors, like synchronous exception when accessing FPU registers. This is very prominent on CPUs with long instruction pipeline, like ARM Cortex A72.

This change fixes the following U-Boot panic:

"Synchronous Abort" handler, esr 0x1fe00000
elr: 00000000800948cc lr : 0000000080091e04
x0 : 00000000801ffdc8 x1 : 00000000000000c8
x2 : 00000000800979d4 x3 : 00000000801ffc60
x4 : 00000000801ffd40 x5 : ffffff80ffffffd8
x6 : 00000000801ffd70 x7 : 00000000801ffd70
x8 : 000000000000000a x9 : 0000000000000000
x10: 0000000000000044 x11: 0000000000000000
x12: 0000000000000000 x13: 0000000000000000
x14: 0000000000000000 x15: 0000000000000000
x16: 000000008008b2e0 x17: 0000000000000000
x18: 00000000801ffec0 x19: 00000000800957b0
x20: 00000000000000c8 x21: 00000000801ffdc8
x22: 000000008009909e x23: 0000000000000000
x24: 0000000000000000 x25: 0000000000000000
x26: 0000000000000000 x27: 0000000000000000
x28: 0000000000000000 x29: 00000000801ffc50

 Code: a94417e4 a90217e4 a9051fe6 a90617e4 (3d801fe0)

While executing instruction

str q0, [sp, #112]

Quoted 7 lines

in vsnprintf() prologue. This panic was observed only on Cortex A72 so far.

This patch places ISBs on other strategic places as well.

Also, this probably is the right fix for the issue workarounded in the commit 45f41c13 ("ARM: uniphier: add weird workaround code for LD20")

Thanks for addressing this issue.
Currently, I do not have a board in hand to test this. (I do not commute to the office due to COVID-19 these days...)

Quoted 9 lines

I have another SoC board, but it does not integrate CA72. I have ever seen this problem only on CA72.

Eventually, I will go to the office, and I can test this. But, you do not need to wait for my test if other people review it.

Thank you.

Quoted 6 lines

Today I tested this patch.

Yes, it fixes the CA72 problem on my board.

Now that MW is open, I want to see this patch in mainline because it fixes the CA-72 problem in the right way.

There was no object except two minor comments from Andre.

Can Tom fix up the commit description when he picks it up?

Quoted 8 lines

"... after any change to system registers."

     --->

"... after changes to some system registers."

"instruction memory synchronization barrier"

    --->

"instruction synchronization barrier"

Yes, thanks. I've been wanting to put this directly to master (and not part of the next merge) for easier bisecting should there be a problem. I'm running this (and a few other risky patches) through CI now.

Signature · 8 lines
Tom
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 659 bytes
Desc: not available
URL: <https://lists.denx.de/pipermail/u-boot/attachments/20200707/78e7d62b/attachment.sig>
#8 lore raw

PATCH v2 1/1 arm64: issue ISB after updating system registers

In reply to Volodymyr Babchuk
Quoted 45 lines · Volodymyr Babchuk

ARM Architecture reference manual clearly states that PE pipeline should be flushed after any change to system registers. Refer to paragraph "B2.3.5 Memory Barriers" at page B2-92 of "Arm Architecture Reference Manual ARMv8 for ARMv8-A Architecture Profile" (ARM DDI 0487B.a).

Failing to issue instruction memory synchronization barrier can lead to spurious errors, like synchronous exception when accessing FPU registers. This is very prominent on CPUs with long instruction pipeline, like ARM Cortex A72.

This change fixes the following U-Boot panic:

"Synchronous Abort" handler, esr 0x1fe00000
elr: 00000000800948cc lr : 0000000080091e04
x0 : 00000000801ffdc8 x1 : 00000000000000c8
x2 : 00000000800979d4 x3 : 00000000801ffc60
x4 : 00000000801ffd40 x5 : ffffff80ffffffd8
x6 : 00000000801ffd70 x7 : 00000000801ffd70
x8 : 000000000000000a x9 : 0000000000000000
x10: 0000000000000044 x11: 0000000000000000
x12: 0000000000000000 x13: 0000000000000000
x14: 0000000000000000 x15: 0000000000000000
x16: 000000008008b2e0 x17: 0000000000000000
x18: 00000000801ffec0 x19: 00000000800957b0
x20: 00000000000000c8 x21: 00000000801ffdc8
x22: 000000008009909e x23: 0000000000000000
x24: 0000000000000000 x25: 0000000000000000
x26: 0000000000000000 x27: 0000000000000000
x28: 0000000000000000 x29: 00000000801ffc50

 Code: a94417e4 a90217e4 a9051fe6 a90617e4 (3d801fe0)

While executing instruction

str q0, [sp, #112]

in vsnprintf() prologue. This panic was observed only on Cortex A72 so far.

This patch places ISBs on other strategic places as well.

Also, this probably is the right fix for the issue workarounded in the commit 45f41c13 ("ARM: uniphier: add weird workaround code for LD20")

Quoted 9 lines

Applied to u-boot/master, thanks!

Signature · 8 lines
Tom
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 659 bytes
Desc: not available
URL: <https://lists.denx.de/pipermail/u-boot/attachments/20200707/e839c49b/attachment.sig>