Hi Steve !
On 18/11/2024 21:20, Steven Rostedt wrote:
[ Added Tomas as he knows this code better than I do ]
Thanks !
On Mon, 18 Nov 2024 11:11:48 +0100
Jean-Michel Hautbois <jeanmichel.hautbois@xxxxxxxxxx> wrote:
Hi Steve,
On 15/11/2024 20:55, Steven Rostedt wrote:
On Fri, 15 Nov 2024 16:33:06 +0100
Jean-Michel Hautbois <jeanmichel.hautbois@xxxxxxxxxx> wrote:
Hi Steve,
On 15/11/2024 16:25, Steven Rostedt wrote:
On Fri, 15 Nov 2024 09:26:07 +0100
Jean-Michel Hautbois <jeanmichel.hautbois@xxxxxxxxxx> wrote:
Nevertheless it sounds like a really high latency for wake_up().
I have a custom driver which basically gets an IRQ, and calls wake_up on
a read() call. This wake_up() on a high cpu usage can be more than 1ms !
Even with a fifo/99 priority for my kernel thread !
I don't know if it rings any bell ?
I can obviously do more tests if it can help getting down to the issue :-).
Try running timerlat.
Thanks !
Here is what I get:
# echo timerlat > current_tracer
# echo 1 > events/osnoise/enable
# echo 25 > osnoise/stop_tracing_total_us
# tail -10 trace
bash-224 [000] d.h.. 153.268917: #77645 context irq timer_latency 45056 ns
bash-224 [000] dnh.. 153.268987: irq_noise: timer:206 start 153.268879083 duration 93957 ns
bash-224 [000] d.... 153.269056: thread_noise: bash:224 start 153.268905324 duration 71045 ns
timerlat/0-271 [000] ..... 153.269103: #77645 context thread timer_latency 230656 ns
bash-224 [000] d.h.. 153.269735: irq_noise: timer:206 start 153.269613847 duration 103558 ns
bash-224 [000] d.h.. 153.269911: #77646 context irq timer_latency 40640 ns
bash-224 [000] dnh.. 153.269982: irq_noise: timer:206 start 153.269875367 duration 93190 ns
bash-224 [000] d.... 153.270053: thread_noise: bash:224 start 153.269900969 duration 72709 ns
timerlat/0-271 [000] ..... 153.270100: #77646 context thread timer_latency 227008 ns
timerlat/0-271 [000] ..... 153.270155: timerlat_main: stop tracing hit on cpu 0
It looks awful, right ?
awful is relative ;-) If that was on x86, I would say it was bad.
Also check out rtla (in tools/trace/rtla).
Thanks ! I knew it only by name, so I watched a presentation recorded
during OSS summit given by Daniel Bristot de Oliveira who wrote it and
it is really impressive !
I had to modify the source code a bit, as it does not compile with my
uclibc toolchain:
diff --git a/tools/tracing/rtla/Makefile.rtla
b/tools/tracing/rtla/Makefile.rtla
index cc1d6b615475..b22016a88d09 100644
--- a/tools/tracing/rtla/Makefile.rtla
+++ b/tools/tracing/rtla/Makefile.rtla
@@ -15,7 +15,7 @@ $(call allow-override,LD_SO_CONF_PATH,/etc/ld.so.conf.d/)
$(call allow-override,LDCONFIG,ldconfig)
export CC AR STRIP PKG_CONFIG LD_SO_CONF_PATH LDCONFIG
-FOPTS := -flto=auto -ffat-lto-objects -fexceptions
-fstack-protector-strong \
+FOPTS := -flto=auto -ffat-lto-objects -fexceptions \
-fasynchronous-unwind-tables -fstack-clash-protection
WOPTS := -O -Wall -Werror=format-security
-Wp,-D_FORTIFY_SOURCE=2 \
-Wp,-D_GLIBCXX_ASSERTIONS -Wno-maybe-uninitialized
I'm not sure what the consequence of the above would be. Perhaps Daniel
just copied this from other code?
diff --git a/tools/tracing/rtla/src/timerlat_u.c
b/tools/tracing/rtla/src/timerlat_u.c
index 01dbf9a6b5a5..92ad2388b123 100644
--- a/tools/tracing/rtla/src/timerlat_u.c
+++ b/tools/tracing/rtla/src/timerlat_u.c
@@ -15,10 +15,16 @@
#include <pthread.h>
#include <sys/wait.h>
#include <sys/prctl.h>
+#include <sys/syscall.h>
#include "utils.h"
#include "timerlat_u.h"
+static inline pid_t gettid(void)
+{
+ return syscall(SYS_gettid);
+}
+
/*
* This is the user-space main for the tool timerlatu/ threads.
*
diff --git a/tools/tracing/rtla/src/utils.c b/tools/tracing/rtla/src/utils.c
index 9ac71a66840c..b754dc1016a4 100644
--- a/tools/tracing/rtla/src/utils.c
+++ b/tools/tracing/rtla/src/utils.c
@@ -229,6 +229,9 @@ long parse_ns_duration(char *val)
#elif __s390x__
# define __NR_sched_setattr 345
# define __NR_sched_getattr 346
+#elif __m68k__
+# define __NR_sched_setattr 349
+# define __NR_sched_getattr 350
#endif
#define SCHED_DEADLINE 6
But it is not enough, as executing rtla fails with a segfault.
I can dump a core, but I could not manage to build gdb for my board so I
can't debug it (I don't know how to debug a coredump without gdb !).
printf()! That's how I debug things without gdb ;-)
Indeed printf gave me clues !
It appears to be a bug in libtracefs (v1.8.1). rtla segfaults when
calling tracefs_local_events() in trace_instance_init().
Debugging libtracefs pointed me to the load_events() function, and the
segfault happens after tep_parse_event() is called for
"/sys/kernel/debug/tracing/events/vmscan/mm_vmscan_write_folio/format".
Going through the calls I get to event_read_print_args().
I changed libtraceevent log level to get the warnings, and it says:
libtraceevent: Resource temporarily unavailable
unknown op '.'
Segmentation fault
JM
