runtime/trace: TestTraceStress failed #9729
Comments
|
What hardware do you use? |
|
-> uname -a -> cat /proc/cpuinfo processor : 1 |
|
I see, thank you. |
|
On Intel(R) Core(TM) i7-3930K CPU @ 3.20GHz $go version $go test -run TraceStress -v $./all.bash ok runtime/debug 0.012s .... $uname -a $ cat /proc/cpuinfo |
|
It is also failing for me, although with another error message ("failed to parse trace: bogus scanning end"). The same error was reported by one of the build bots: |
|
similar problem |
|
@dvyukov In case it helps, I can reproduce this issue 100% on my temporary Z620 in TOK... but not all on my permanent one back in SFO. SFO (does NOT repro): TOK (DOES repro): Let me know if granting you remote access would be helpful in investigating. I'll have it for at least two more weeks. |
|
I will appreciate access to a machine that can reproduce this. I will contact you offline. |
|
I can reproduce it on your machine. I wrote the following program to investigate it.
But diffs are never larger than ~250.
I hoped this problem is left in stone age. Linux must do its best to synchronize clocks today, so maybe it's bad BIOS or processor microcode. #define _GNU_SOURCE
#include <stdio.h>
#include <stdlib.h>
#include <stdint.h>
#include <pthread.h>
#include <sched.h>
#include <errno.h>
void* thr(void *arg);
uint64_t sig, ticks0, ticks1;
int main() {
cpu_set_t mask0;
if (sched_getaffinity(0, sizeof(mask0), &mask0)) {
perror("sched_getaffinity");
exit(1);
}
int i;
for (i = 0; i < CPU_SETSIZE; i++) {
if (!CPU_ISSET(i, &mask0))
break;
int j;
for (j = 0; j < CPU_SETSIZE; j++) {
if (i == j || !CPU_ISSET(j, &mask0))
break;
printf(".");
fflush(0);
sig = ticks0 = ticks1 = 0;
pthread_t th[2];
pthread_attr_t attr[2];
cpu_set_t mask[2];
CPU_ZERO(&mask[0]);
CPU_SET(i, &mask[0]);
pthread_attr_init(&attr[0]);
pthread_attr_setaffinity_np(&attr[0], sizeof(attr[0]), &mask[0]);
pthread_create(&th[0], &attr[0], thr, (void*)1);
CPU_ZERO(&mask[1]);
CPU_SET(j, &mask[1]);
pthread_attr_init(&attr[1]);
pthread_attr_setaffinity_np(&attr[1], sizeof(mask[1]), &mask[1]);
pthread_create(&th[1], &attr[1], thr, 0);
pthread_join(th[0], 0);
pthread_join(th[1], 0);
pthread_attr_destroy(&attr[0]);
pthread_attr_destroy(&attr[1]);
if (ticks0 != 0)
printf("\n%d-%d: BROKEN %lu -> %lu (diff=%lu)\n", i, j, ticks0, ticks1, ticks0 - ticks1);
}
}
printf("\n");
return 0;
}
uint64_t cputicks() {
uint64_t low, high;
__asm__ __volatile__("lfence; rdtsc" : "=a" (low), "=d" (high));
return (high << 32) | low;
}
void* thr(void *arg) {
int i;
for (i = 0; i < 10000; i++) {
if (arg != 0) {
uint64_t t0 = cputicks();
__atomic_store_n(&sig, 1, __ATOMIC_RELEASE);
while (__atomic_load_n(&sig, __ATOMIC_RELAXED) == 1) {
}
uint64_t t1 = __atomic_load_n(&sig, __ATOMIC_ACQUIRE);
if (t1 <= t0) {
ticks0 = t0;
ticks1 = t1;
}
} else {
while (__atomic_load_n(&sig, __ATOMIC_ACQUIRE) != 1) {
}
__atomic_store_n(&sig, cputicks(), __ATOMIC_RELEASE);
}
}
return 0;
} |
|
We can manully synchronize RDTSC on different cores using RDTSCP instruction. Namely, measure RTT between all cores and offsetting values of RDTSCP by computed deltas. |
See the following issue for context: #9729 (comment) In short, RDTSC can produce skewed results without preceding LFENCE/MFENCE. Information on this matter is very scrappy in the internet. But this is what linux kernel does (see rdtsc_barrier). It also fixes the test program on my machine. Update #9729 Change-Id: I3c1ffbf129fdfdd388bd5b7911b392b319248e68 Reviewed-on: https://go-review.googlesource.com/5033 Reviewed-by: Ian Lance Taylor <[email protected]>
|
FYI the test still 100% fails on this AMD cpu: LKML: "Brunner, Richard": TSC and Power Management Events on AMD Processors "this patch syncs the TSCs every 20 seconds. That is enough to sync |
|
Also saw on OpenBSD-amd64: https://storage.googleapis.com/go-build-log/114bbd11/openbsd-amd64-gce56_7da78022.log |
|
FYI. Tried again today using a revision which includes the "fix cputicks on x86" patch, and it still fails.
This is on the same hardware as last time. |
|
Same error on my very old AMD machine: |
|
Dmitry, reassign as necessary. |
|
I don't believe this is the same problem as #10476. TestTraceStress had real (non-RDTSC-related) synchronization problems that Austin and Rick found and fixed. The failures in TestTraceStressStartStop are different, I believe. They're always about system calls, which are still not traced quite right. To avoid confusion, I'm going to close this and reopen #10476. |
|
Is it fixed? Dockerfile is in https://go-review.googlesource.com/#/c/9077/ In the container, the go version is: devel +3f4de49 Mon Apr 20 15:56:56 2015 +0000 root@0e6edb498432:/go/src/runtime/pprof# go test -run TraceStress -v |
|
No, it is not resolved. |
|
It doesn't fail outside of container (btw that is darwin/amd64). I don't see any failure reports from the build dashboard or my personal ubuntu machine currently. So, it's probably the docker container in my case. |
Yes, it fails on some (arguably misdesigned, see e.g. https://lkml.org/lkml/2005/11/4/173 ) CPUs. Nothing changed since the original report. This is bare Ubuntu LTS 14.04: ../bin/go version |
It is vmware. Inside VM it says Intel Xeon CPU E5-2680 0 @ 2.70GHz, Windows 7.
It looks like my pc has 1 CPU only. Your program seems like it needs multiple CPUs. Isn't it? Can I use Go? Alex |
mikioh
changed the title
-> go version
go version devel +0e80b2e Thu Jan 29 13:07:30 2015 +0000 linux/amd64
-> go test -run TraceStress -v
=== RUN TestTraceStress
--- FAIL: TestTraceStress (1.71s)
trace_test.go:207: failed to parse trace: p 1 is running before start (offset 128960, time 1475855)
FAIL
exit status 1
FAIL runtime/pprof 1.718s
complete trace file: https://drive.google.com/file/d/0B7KKArVsnMd2WldQRDB4VXlOQXc/view?usp=sharing
The text was updated successfully, but these errors were encountered: