Thank you for your contribution! Sashiko AI review found no new issues
in this patch, but noticed 1 related pre-existing issue it does not
introduce:
- [High] ring-buffer: atomic counter underflow in ring_buffer_reset()
[1]
--
commit fbf1784e6b7f3202f4d6898704fb8180140b6da5
Author: Vincent Donnefort [off-list ref]
ring-buffer: Always read buffer->buffers after testing the cpumask
This commit moves the load of buffer->buffers[cpu] after the
cpumask_test_cpu() check in ring_buffer_peek(), ring_buffer_reset_cpu(),
and ring_buffer_read_page() to avoid reading a stale per-CPU buffer
pointer when racing with trace_rb_cpu_prepare().
quoted hunk ↗ jump to hunk
diff --git a/kernel/trace/ring_buffer.c b/kernel/trace/ring_buffer.c
index 1fd28a72e6ee1..bd8126f6353b7 100644
--- a/kernel/trace/ring_buffer.c
+++ b/kernel/trace/ring_buffer.c
[ ... ]
quoted hunk ↗ jump to hunk
@@ -6740,11 +6742,13 @@ static void reset_disabled_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
*/
void ring_buffer_reset_cpu(struct trace_buffer *buffer, int cpu)
{
- struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
+ struct ring_buffer_per_cpu *cpu_buffer;
if (!cpumask_test_cpu(cpu, buffer->cpumask))
return;
+ cpu_buffer = buffer->buffers[cpu];
+
/* prevent another thread from changing buffer sizes */
mutex_lock(&buffer->mutex);
[Severity: High]
This problem wasn't introduced by this patch, but can a concurrent
trace_rb_cpu_prepare() race with ring_buffer_reset() [1] and
underflow cpu_buffer->record_disabled and cpu_buffer->resize_disabled?
Unlike ring_buffer_reset_online_cpus(), which marks each disabled
cpu_buffer with RESET_BIT before sleeping in synchronize_rcu(),
ring_buffer_reset() iterates over buffer->cpumask twice without tracking
which CPUs were disabled in the first loop or holding cpus_read_lock():
CPU0 CPU1
ring_buffer_reset()
for_each_buffer_cpu(buffer, cpu)
atomic_inc(&cpu_buffer->resize_disabled);
atomic_inc(&cpu_buffer->record_disabled);
synchronize_rcu();
trace_rb_cpu_prepare()
rb_allocate_cpu_buffer()
cpumask_set_cpu(cpu, buffer->cpumask);
for_each_buffer_cpu(buffer, cpu)
reset_disabled_cpu_buffer(cpu_buffer);
atomic_dec(&cpu_buffer->record_disabled); <-- underflows to -1
atomic_dec(&cpu_buffer->resize_disabled); <-- underflows to -1
For the newly prepared CPU, reset_disabled_cpu_buffer() runs while
recording is still enabled and both atomic counters underflow from 0 to
-1, corrupting the per-CPU ring buffer state.
[1] https://sashiko.dev/#/bug/linux-2557ac4d-25af-4832-a2c9-c12af32c7a04
--
Sashiko AI review · https://sashiko.dev/#/patchset/20261112141440.3050124-1-vdonnefort@google.com?part=1