After making a slight change to wakeups in ring_buffer_wait()
the system would hang. Spending several hours going on a wild goose
chase I found that the change only triggered the bug because it
changed the timings. The bug was there before the update but never
was triggered.
The poll code has:
rbwork->full_waiters_pending = true;
if (!cpu_buffer->shortest_full ||
cpu_buffer->shortest_full > full)
cpu_buffer->shortest_full = full;
The writer will see full_waiters_pending and check if the ring buffer is
filled over the percentage of the shortest_full value. If it is, it calls
an irq_work to wake up all the waiters.
But the code could get into a circular loop:
CPU 0 CPU 1
----- -----
[ Poll ]
[ shortest_full = 0 ]
rbwork->full_waiters_pending = true;
if (rbwork->full_waiters_pending &&
[ buffer percent ] >
shortest_full) {
rbwork->wakeup_full = true;
[ queue_irqwork ]
cpu_buffer->shortest_full = full;
[ IRQ work ]
if (rbwork->wakeup_full) {
cpu_buffer->shortest_full = 0;
wakeup poll waiters;
[woken]
if ([ buffer percent ] > full)
break;
rbwork->full_waiters_pending = true;
if (rbwork->full_waiters_pending &&
[ buffer percent ] >
shortest_full) {
rbwork->wakeup_full = true;
[ queue_irqwork ]
cpu_buffer->shortest_full = full;
[ IRQ work ]
if (rbwork->wakeup_full) {
cpu_buffer->shortest_full = 0;
wakeup poll waiters;
[woken]
[ Wash, rinse, repeat! ]
The race was triggered when running:
trace-cmd record -p function -m 5000
Which enables function tracing and then creates two files it is writing
into where each is 2500K in size. The -m is a "max file size". When
trace-cmd writes 2500K to one file it then switches to the other, erasing
the old data. To do this, trace-cmd switches between both poll and
the reader using both methods of wake up. The change to the reader wakeup
was able to change the way the poll was woken to trigger this bug.
The second patch is a clean up and also a way to consolidate the logic
of the shortest_full. The read wakeup uses rb_watermark_hit for both
full wakeups and !full wakeups. But since poll uses the same logic for
full wakeups it can just call that function with full set.
Changes since v1:
https://lore.kernel.org/all/20240312115455.666920...@goodmis.org/
- Removed unused 'flags' in ring_buffer_poll_wait() as the spin_lock
is now in rb_watermark_hit().
Steven Rostedt (Google) (2):
ring-buffer: Fix full_waiters_pending in poll
ring-buffer: Reuse rb_watermark_hit() for the poll logic
----
kernel/trace/ring_buffer.c | 32 +++++++++++++++++++-------------
1 file changed, 19 insertions(+), 13 deletions(-)
