Skip to contentSkip to conversation

Try to make SMP booting slightly less fragile

PATCH In review Will Deacon 2 participants 7 messages Aug 27, 2019 arm64

lore
Share & tools
lore

Will Deacon

author maintainer
#0 lore raw

cover Try to make SMP booting slightly less fragile

Hi everyone,

After spending some time investigating SMP boot issues when using (random?) configs from Qian Cai with lots of debug options enabled, I hacked together these two patches which make SMP booting a little more robust.

The whole thing is still a racy mess, but I'm not sure we can do much about that without reworking it to use per-cpu data structures which we can update atomically.

Will

--->8

Will Deacon (3):
arm64: smp: Increase secondary CPU boot timeout value
arm64: smp: Don't enter kernel with NULL stack pointer or task struct
arm64: smp: Treat unknown boot failures as being 'stuck in kernel'

 arch/arm64/kernel/head.S | 8 ++++++++
 arch/arm64/kernel/smp.c  | 4 +++-
 2 files changed, 11 insertions(+), 1 deletion(-)
Signature · 7 lines
2.11.0


_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
author maintainer
#1 lore raw

PATCH 1/3 arm64: smp: Increase secondary CPU boot timeout value

When many debug options are enabled simultaneously (e.g. PROVE_LOCKING, KMEMLEAK, DEBUG_PAGE_ALLOC, KASAN etc), it is possible for us to timeout when attempting to boot a secondary CPU and give up. Unfortunately, the CPU will /eventually/ appear, and sit in the background happily stuck in a recursive exception due to a NULL stack pointer.

Increase the timeout to 5s, which will of course be enough for anybody.

---
 arch/arm64/kernel/smp.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
arch/arm64/kernel/smp.c +1 -1
index 018a33e01b0e..63c7a7682e93 100644
@@ -123,7 +123,7 @@ int __cpu_up(unsigned int cpu, struct task_struct *idle)
 		 * time out.
 		 */
 		wait_for_completion_timeout(&cpu_running,
-					    msecs_to_jiffies(1000));
+					    msecs_to_jiffies(5000));
 
 		if (!cpu_online(cpu)) {
 			pr_crit("CPU%u: failed to come online\n", cpu);
Signature · 7 lines
2.11.0


_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
reviewer Reviewed-by on v1 1/3
#2 lore raw

Re: [PATCH 1/3] arm64: smp: Increase secondary CPU boot timeout value

Quoted 9 lines · Will Deacon

When many debug options are enabled simultaneously (e.g. PROVE_LOCKING, KMEMLEAK, DEBUG_PAGE_ALLOC, KASAN etc), it is possible for us to timeout when attempting to boot a secondary CPU and give up. Unfortunately, the CPU will /eventually/ appear, and sit in the background happily stuck in a recursive exception due to a NULL stack pointer.

Increase the timeout to 5s, which will of course be enough for anybody.

Thanks,
Mark.

Quoted 20 lines
---
 arch/arm64/kernel/smp.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
arch/arm64/kernel/smp.c +1 -1
index 018a33e01b0e..63c7a7682e93 100644
@@ -123,7 +123,7 @@ int __cpu_up(unsigned int cpu, struct task_struct *idle)
 		 * time out.
 		 */
 		wait_for_completion_timeout(&cpu_running,
-					    msecs_to_jiffies(1000));
+					    msecs_to_jiffies(5000));
 
 		if (!cpu_online(cpu)) {
 			pr_crit("CPU%u: failed to come online\n", cpu);
Signature · 1 line
2.11.0

_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

author maintainer
#3 lore raw

PATCH 2/3 arm64: smp: Don't enter kernel with NULL stack pointer or task struct

In reply to Will Deacon

Although SMP bringup is inherently racy, we can significantly reduce the window during which secondary CPUs can unexpectedly enter the kernel by sanity checking the 'stack' and 'task' fields of the 'secondary_data' structure. If the booting CPU gave up waiting for us, then they will have been cleared to NULL and we should spin in a WFE; WFI loop instead.

---
 arch/arm64/kernel/head.S | 8 ++++++++
 arch/arm64/kernel/smp.c  | 1 +
 2 files changed, 9 insertions(+)
arch/arm64/kernel/head.S +8 -0
index 2cdacd1c141b..0baadf335172 100644
@@ -724,14 +724,22 @@ __secondary_switched:
 
 	adr_l	x0, secondary_data
 	ldr	x1, [x0, #CPU_BOOT_STACK]	// get secondary_data.stack
+	cbz	x1, __secondary_too_slow
 	mov	sp, x1
 	ldr	x2, [x0, #CPU_BOOT_TASK]
+	cbz	x2, __secondary_too_slow
 	msr	sp_el0, x2
 	mov	x29, #0
 	mov	x30, #0
 	b	secondary_start_kernel
 ENDPROC(__secondary_switched)
 
+__secondary_too_slow:
+	wfe
+	wfi
+	b	__secondary_too_slow
+ENDPROC(__secondary_too_slow)
+
 /*
  * The booting CPU updates the failed status @__early_cpu_boot_status,
  * with MMU turned off.
arch/arm64/kernel/smp.c +1 -0
index 63c7a7682e93..1f8aeb77cba5 100644
@@ -136,6 +136,7 @@ int __cpu_up(unsigned int cpu, struct task_struct *idle)
 
 	secondary_data.task = NULL;
 	secondary_data.stack = NULL;
+	__flush_dcache_area(&secondary_data, sizeof(secondary_data));
 	status = READ_ONCE(secondary_data.status);
 	if (ret && status) {
 
Signature · 7 lines
2.11.0


_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
reviewer Reviewed-by on v1 2/3
#4 lore raw

Re: [PATCH 2/3] arm64: smp: Don't enter kernel with NULL stack pointer or task struct

Quoted 8 lines · Will Deacon

Although SMP bringup is inherently racy, we can significantly reduce the window during which secondary CPUs can unexpectedly enter the kernel by sanity checking the 'stack' and 'task' fields of the 'secondary_data' structure. If the booting CPU gave up waiting for us, then they will have been cleared to NULL and we should spin in a WFE; WFI loop instead.

In future I think that we can handle this better using the MPIDR hash and atomics to flip the values and cater for races, but this is better than what we have, so FWIW:

Thanks,
Mark.

Quoted 47 lines
---
 arch/arm64/kernel/head.S | 8 ++++++++
 arch/arm64/kernel/smp.c  | 1 +
 2 files changed, 9 insertions(+)
arch/arm64/kernel/head.S +8 -0
index 2cdacd1c141b..0baadf335172 100644
@@ -724,14 +724,22 @@ __secondary_switched:
 
 	adr_l	x0, secondary_data
 	ldr	x1, [x0, #CPU_BOOT_STACK]	// get secondary_data.stack
+	cbz	x1, __secondary_too_slow
 	mov	sp, x1
 	ldr	x2, [x0, #CPU_BOOT_TASK]
+	cbz	x2, __secondary_too_slow
 	msr	sp_el0, x2
 	mov	x29, #0
 	mov	x30, #0
 	b	secondary_start_kernel
 ENDPROC(__secondary_switched)
 
+__secondary_too_slow:
+	wfe
+	wfi
+	b	__secondary_too_slow
+ENDPROC(__secondary_too_slow)
+
 /*
  * The booting CPU updates the failed status @__early_cpu_boot_status,
  * with MMU turned off.
arch/arm64/kernel/smp.c +1 -0
index 63c7a7682e93..1f8aeb77cba5 100644
@@ -136,6 +136,7 @@ int __cpu_up(unsigned int cpu, struct task_struct *idle)
 
 	secondary_data.task = NULL;
 	secondary_data.stack = NULL;
+	__flush_dcache_area(&secondary_data, sizeof(secondary_data));
 	status = READ_ONCE(secondary_data.status);
 	if (ret && status) {
 
Signature · 1 line
2.11.0

_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

author maintainer
#5 lore raw

PATCH 3/3 arm64: smp: Treat unknown boot failures as being 'stuck in kernel'

In reply to Will Deacon

When we fail to bring a secondary CPU online and it fails in an unknown state, we should assume the worst and increment 'cpus_stuck_in_kernel' so that things like kexec() are disabled.

---
 arch/arm64/kernel/smp.c | 1 +
 1 file changed, 1 insertion(+)
arch/arm64/kernel/smp.c +1 -0
index 1f8aeb77cba5..dc9fe879c279 100644
@@ -147,6 +147,7 @@ int __cpu_up(unsigned int cpu, struct task_struct *idle)
 		default:
 			pr_err("CPU%u: failed in unknown state : 0x%lx\n",
 					cpu, status);
+			cpus_stuck_in_kernel++;
 			break;
 		case CPU_KILL_ME:
 			if (!op_cpu_kill(cpu)) {
Signature · 7 lines
2.11.0


_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
reviewer Reviewed-by on v1 3/3
#6 lore raw

Re: [PATCH 3/3] arm64: smp: Treat unknown boot failures as being 'stuck in kernel'

Quoted 3 lines · Will Deacon

When we fail to bring a secondary CPU online and it fails in an unknown state, we should assume the worst and increment 'cpus_stuck_in_kernel' so that things like kexec() are disabled.

Definitely! I has assumed we already did this, but I see that we don't.

Quoted 1 line

I don't see a nicer way of doing this, so:

Thanks,
Mark.

Quoted 19 lines
---
 arch/arm64/kernel/smp.c | 1 +
 1 file changed, 1 insertion(+)
arch/arm64/kernel/smp.c +1 -0
index 1f8aeb77cba5..dc9fe879c279 100644
@@ -147,6 +147,7 @@ int __cpu_up(unsigned int cpu, struct task_struct *idle)
 		default:
 			pr_err("CPU%u: failed in unknown state : 0x%lx\n",
 					cpu, status);
+			cpus_stuck_in_kernel++;
 			break;
 		case CPU_KILL_ME:
 			if (!op_cpu_kill(cpu)) {
Signature · 1 line
2.11.0

_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel