Skip to contentSkip to conversation

KGDB: add smp_mb() in synchronisation during exception handler exit

PATCH In review Will Deacon 2 participants 3 messages Mar 9, 2010 – Mar 10, 2010

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 KGDB: add smp_mb() in synchronisation during exception handler exit

KGDB uses atomic variables and busy-wait loops to co-ordinate between multiple CPUs on an SMP system. When an exception is handled, the primary CPU executes kgdb_handle_exception() whilst the others execute kgdb_wait.

There comes a point when the waiters are waiting for the primary CPU to finish:

	/* Wait till primary CPU is done with debugging */
(1)	while (atomic_read(&passive_cpu_wait[cpu]))
		cpu_relax();

	/* Do important KGDB stuff */

	/* Signal the primary CPU that we are done: */
	atomic_set(&cpu_in_kgdb[cpu], 0);

In parallel to this, the primary CPU is doing:

	for (i = NR_CPUS-1; i >= 0; i--)
		atomic_set(&passive_cpu_wait[i], 0);
	/*
	 * Wait till all the CPUs have quit
	 * from the debugger.
	 */
	for_each_online_cpu(i) {
(1)		while (atomic_read(&cpu_in_kgdb[i]))
			cpu_relax();
	}

There is a potential deadlock situation at point (1) because the previous writes to the passive_cpu_wait variables by the primary CPU may not yet be visible to the other CPUs [for instance, they may be sitting in the local store buffer]. This means that the waiter CPUs will never exit the while loop and therefore never write to the cpu_in_kgdb variables, which the primary CPU is blocked on. Furthermore, because the primary CPU is aggressively performing reads, the store buffer may not necessarily drain so the system will deadlock.

This deadlock has been experienced on a quad-core ARM11MPCore platform.

The following patch addresses the issue by adding a memory barrier to the primary CPU before the polling loop, therefore forcing the previous atomic_sets to be visible before waiting for the waiters to finish.

Cc KGDB Mailing List <[email protected]>
Cc Catalin Marinas <[email protected]>
Cc Russell King - ARM Linux <[email protected]>
Cc linux-arm-kernel at lists.infradead.org
---
 kernel/kgdb.c |    1 +
 1 files changed, 1 insertions(+), 0 deletions(-)
kernel/kgdb.c +1 -0
index 761fdd2..ee7694b 100644
@@ -1537,6 +1537,7 @@ acquirelock:
 		 * Wait till all the CPUs have quit
 		 * from the debugger.
 		 */
+		smp_mb();
 		for_each_online_cpu(i) {
 			while (atomic_read(&cpu_in_kgdb[i]))
 				cpu_relax();
Signature · 1 line
1.6.3.3
#1 lore raw

On 03/09/2010 11:20 AM, Will Deacon was caught saying:

Quoted 57 lines

KGDB uses atomic variables and busy-wait loops to co-ordinate between multiple CPUs on an SMP system. When an exception is handled, the primary CPU executes kgdb_handle_exception() whilst the others execute kgdb_wait.

There comes a point when the waiters are waiting for the primary CPU to finish:

	/* Wait till primary CPU is done with debugging */
(1)	while (atomic_read(&passive_cpu_wait[cpu]))
		cpu_relax();

	/* Do important KGDB stuff */

	/* Signal the primary CPU that we are done: */
	atomic_set(&cpu_in_kgdb[cpu], 0);

In parallel to this, the primary CPU is doing:

	for (i = NR_CPUS-1; i>= 0; i--)
		atomic_set(&passive_cpu_wait[i], 0);
	/*
	 * Wait till all the CPUs have quit
	 * from the debugger.
	 */
	for_each_online_cpu(i) {
(1)		while (atomic_read(&cpu_in_kgdb[i]))
			cpu_relax();
	}

There is a potential deadlock situation at point (1) because the previous writes to the passive_cpu_wait variables by the primary CPU may not yet be visible to the other CPUs [for instance, they may be sitting in the local store buffer]. This means that the waiter CPUs will never exit the while loop and therefore never write to the cpu_in_kgdb variables, which the primary CPU is blocked on. Furthermore, because the primary CPU is aggressively performing reads, the store buffer may not necessarily drain so the system will deadlock.

This deadlock has been experienced on a quad-core ARM11MPCore platform.

The following patch addresses the issue by adding a memory barrier to the primary CPU before the polling loop, therefore forcing the previous atomic_sets to be visible before waiting for the waiters to finish.

Cc KGDB Mailing List<[email protected]>
Cc Catalin Marinas<[email protected]>
Cc Russell King - ARM Linux<[email protected]>
Cc linux-arm-kernel at lists.infradead.org
---
  kernel/kgdb.c |    1 +
  1 files changed, 1 insertions(+), 0 deletions(-)
kernel/kgdb.c +0 -0
index 761fdd2..ee7694b 100644

kernel/kgdb.c acquirelock: · line 1543

Quoted 7 lines
kernel/kgdb.c +1 -0
@@ -1537,6 +1537,7 @@ acquirelock:
  		 * Wait till all the CPUs have quit
  		 * from the debugger.
  		 */
+		smp_mb();
  		for_each_online_cpu(i) {
  			while (atomic_read(&cpu_in_kgdb[i]))
  				cpu_relax();

Doesn't this have the same issue if this cpu gets to the while prior to the other cpu doing its write. I would think the "smp_mb()" should be in the while loop not prior to it.

Signature · 1 line
George Anzinger   george at wildturkeyranch.net
#2 lore raw

PATCH 1/1 KGDB: add smp_mb() in synchronisation during exception handler exit

Hi George,

Quoted 8 lines
Quoted 8 lines
kernel/kgdb.c +0 -0
index 761fdd2..ee7694b 100644
@@ -1537,6 +1537,7 @@ acquirelock:
  		 * Wait till all the CPUs have quit
  		 * from the debugger.
  		 */
Quoted 8 lines

kernel/kgdb.c acquirelock: · line 1543

Quoted 5 lines
kernel/kgdb.c +1 -0
+		smp_mb();
  		for_each_online_cpu(i) {
  			while (atomic_read(&cpu_in_kgdb[i]))
  				cpu_relax();

Doesn't this have the same issue if this cpu gets to the while prior to the other cpu doing its write. I would think the "smp_mb()" should be in the while loop not prior to it.

I don't think so. The deadlock in question is caused because the aggressive reading in the polling loop prevents the write-buffer from being able to drain. For a dual-core system the setup can be simplified like this:

Initially: x = 1, y = 1;

CPU1                        CPU0
====                        ====
while (y == 1);             y = 0;
x = 0;                      while (x == 1); /* This loop prevents y from

leaving the store buffer */

So the patch inserts a memory barrier on CPU0 between the two lines, forcing the update to y to become visible before the reads to x occur. CPU1 will then trundle along happily and update x. The absence of a further polling loop means that the update to x will become visible to CPU0 at some future time.

Cheers,

Will