[RFC,1/1] rasdaemon: Fix poll() on per_cpu trace_pipe_raw do not retun

Commit Message

Shiju Jose Feb. 2, 2023, 6:23 p.m. UTC

From: Shiju Jose <[e-mail]>

The error events are not received in the rasdaemon since kernel 6.1-rc6.
This issue is firstly detected and reported, when testing the CXL error
events in the rasdaemon.

Debugging showed, poll() on trace_pipe_raw in the ras-events.c do not return.
and this issue is seen after the commit
42fb0a1e84ff525ebe560e2baf9451ab69127e2b ("tracing/ring-buffer: Have
polling block on watermark").

This also verified using the test application for poll()
and select() on trace_pipe_raw.

There is also a bug reported on this issue, 
https://lore.kernel.org/all/[e-mail]/

This issue occurs for the per_cpu case, when calling the ring_buffer_poll_wait(),
in kernel/trace/ring_buffer.c, with the buffer_percent > 0 and then wait until the
percentage of pages are available. The default value set for the buffer_percent is 50
in the kernel/trace/trace.c.

As a fix, rasdaemon set buffer_percent = 0 through the
/sys/kernel/debug/tracing/instances/rasdaemon/buffer_percent, then the task will
wake up as soon as data is added to any of the specific cpu buffer and
poll() on per_cpu/cpuX/trace_pipe_raw does not block indefinitely.

Depends on kernel patch
tracing: Fix poll() and select() do not work on per_cpu trace_pipe and trace_pipe_raw

Signed-off-by: Shiju Jose <[e-mail]>
---
 ras-events.c | 29 +++++++++++++++++++++++++++++
 1 file changed, 29 insertions(+)

Patch

diff --git a/ras-events.c b/ras-events.c
index 3691311..72225a9 100644
--- a/ras-events.c
+++ b/ras-events.c
@@ -383,6 +383,8 @@  static int read_ras_event_all_cpus(struct pthread_data *pdata,
 	int warnonce[n_cpus];
 	char pipe_raw[PATH_MAX];
 	int legacy_kernel = 0;
+	int fd;
+	char buf[2];
 #if 0
 	int need_sleep = 0;
 #endif
@@ -402,6 +404,33 @@  static int read_ras_event_all_cpus(struct pthread_data *pdata,
 		return -ENOMEM;
 	}
 
+	/* Fix for poll() do not return on per_cpu trace_pipe and
+	 * trace_pipe_raw with the default buffer_percent in the kernel trace,
+	 * introduced by the following change in the kernel
+	 * https://lore.kernel.org/all/[e-mail]/T/#u.
+	 * Set buffer_percent to 0 so that poll() returned immediately
+	 * when the trace data is available in the ras per_cpu trace pipe_raw */
+	fd = open_trace(pdata[0].ras, "buffer_percent", O_RDWR);
+	if (fd < 0) {
+		log(TERM, LOG_ERR, "Can't open trace buffer_percent\n");
+		return -1;
+	}
+	size = write(fd, "0", 1);
+	if (size <= 0) {
+		log(TERM, LOG_ERR, "Can't set rasdaemon trace buffer_percent\n");
+		return -1;
+	}
+	size = read(fd, buf, 1);
+	if (size <= 0) {
+		log(TERM, LOG_ERR, "Can't read rasdaemon trace buffer_percent\n");
+		return -1;
+	}
+	if (atoi(buf) != 0) {
+		log(TERM, LOG_ERR, "Can't verify rasdaemon trace buffer_percent\n");
+		return -1;
+	}
+	close(fd);
+
 	for (i = 0; i < (n_cpus + 1); i++)
 		fds[i].fd = -1;