Fix SIGSEGV due to CPU/Real profiler

azat · azat · commit 94041d193cc1 · 2024-05-15T18:22:59.000+02:00
The problem was due to incorrect unwinding due from signal handlers,
which leads to incorrect DWARF (FDE/CIE) interpretation.

After this patch I was not able to reproduce the crash for couple of
hours, while before it was very stable (I've reduced the minimal
threshold for query_profiler_real_time_period_ns), using simply:

    $ clickhouse-benchmark --port 19000 -q "SELECT * FROM remote('127.{1..10}', system, one)" --query_profiler_real_time_period_ns=1

Note, I'm using here remote() for fibers, that has stack with guard
pages that helps with reproducing the crash more faster.

P.S. I also have another implementation of this fix, without patching
unwind and using info from signal context directly, and even though it
is better, because you don't need to trip extra frames and you can use
all the 45 frames for something useful, it is too complex, so let's go
with a simpler patch first, and I think it could be even backported.

Signed-off-by: Azat Khuzhin &lt;a.khuzhin@semrush.com&gt;
diff --git a/contrib/libunwind b/contrib/libunwind
@@ -1 +1 @@
-Subproject commit 40d8eadf96b127d9b22d53ce7a4fc52aaedea965
+Subproject commit ba533a7246a2686b0552061809612f503804d26b
