Bug 216369
| Summary: | trace_read_pipe peek a PADDING event and trying to print it | ||
|---|---|---|---|
| Product: | Tracing/Profiling | Reporter: | Jiazi Li (jiazi.li) |
| Component: | Ftrace | Assignee: | Steven Rostedt (rostedt) |
| Status: | NEW --- | ||
| Severity: | normal | ||
| Priority: | P1 | ||
| Hardware: | ARM | ||
| OS: | Linux | ||
| Kernel Version: | 4.19、5.10 | Subsystem: | |
| Regression: | No | Bisected commit-id: | |
|
Description
Jiazi Li
2022-08-16 06:39:26 UTC
hi,
This issue has happened many times. Every time peek_next_entry return a padding event at end of page.
Our reader task poll trace_pipe and other files.
Maybe this issue will triggered in the following scenarios:
1. reader task wakeup by other file.
2. after read data from other file, call ring_buffer_poll_wait and find ring_buffer no empty because read != commit,
but because of cpu out-of-order execution, next event has not been set to padding type.
3. in ring_buffer_peek, expired data in this event show it is not a padding event, and return event to peek_next_entry.
4. event is set to padding type and detected in rb_event_data, then print a warning.
5. print_trace_line use error type to print this event, eventually crash.
Maybe we can add smp_wb to make ture padding type will be set first:
--- a/kernel/trace/ring_buffer.c
+++ b/kernel/trace/ring_buffer.c
@@ -2681,6 +2681,7 @@ rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
rb_reset_tail(cpu_buffer, tail, info);
+ smp_wb();
/* Commit what we have for now. */
rb_end_commit(cpu_buffer);
/* rb_end_commit() decs committing */
|