]> git.proxmox.com Git - mirror_ovs.git/commit - configure.ac
fatal-signal: Catch SIGSEGV and print backtrace.
authorWilliam Tu <u9012063@gmail.com>
Fri, 27 Sep 2019 17:22:55 +0000 (10:22 -0700)
committerBen Pfaff <blp@ovn.org>
Fri, 27 Sep 2019 17:03:24 +0000 (10:03 -0700)
commite2ed6fbeb18cba1235c6a260b754f11ecc9d4cb0
tree27268f8c45340902cf05b023ebb348ab4f9e7275
parent1ca0323e7c29dc7ef5a615c265df0460208f92de
fatal-signal: Catch SIGSEGV and print backtrace.

The patch catches the SIGSEGV signal and prints the backtrace
using libunwind at the monitor daemon. This makes debugging easier
when there is no debug symbol package or gdb installed on production
systems.

The patch works when the ovs-vswitchd compiles even without debug symbol
(no -g option), because the object files still have function symbols.
For example:
 |daemon_unix(monitor)|WARN|SIGSEGV detected, backtrace:
 |daemon_unix(monitor)|WARN|0x0000000000482752 <fatal_signal_handler+0x52>
 |daemon_unix(monitor)|WARN|0x00007fb4900734b0 <killpg+0x40>
 |daemon_unix(monitor)|WARN|0x00007fb49013974d <__poll+0x2d>
 |daemon_unix(monitor)|WARN|0x000000000052b348 <time_poll+0x108>
 |daemon_unix(monitor)|WARN|0x00000000005153ec <poll_block+0x8c>
 |daemon_unix(monitor)|WARN|0x000000000058630a <clean_thread_main+0x1aa>
 |daemon_unix(monitor)|WARN|0x00000000004ffd1d <ovsthread_wrapper+0x7d>
 |daemon_unix(monitor)|WARN|0x00007fb490b3b6ba <start_thread+0xca>
 |daemon_unix(monitor)|WARN|0x00007fb49014541d <clone+0x6d>
 |daemon_unix(monitor)|ERR|1 crashes: pid 122849 died, killed \
    (Segmentation fault), core dumped, restarting

However, if the object files' symbols are stripped, then we can only
get init function plus offset value. This is still useful when trying
to see if two bugs have the same root cause, Example:
 |daemon_unix(monitor)|WARN|SIGSEGV detected, backtrace:
 |daemon_unix(monitor)|WARN|0x0000000000482752 <_init+0x7d68a>
 |daemon_unix(monitor)|WARN|0x00007f5f7c8cf4b0 <killpg+0x40>
 |daemon_unix(monitor)|WARN|0x00007f5f7c99574d <__poll+0x2d>
 |daemon_unix(monitor)|WARN|0x000000000052b348 <_init+0x126280>
 |daemon_unix(monitor)|WARN|0x00000000005153ec <_init+0x110324>
 |daemon_unix(monitor)|WARN|0x0000000000407439 <_init+0x2371>
 |daemon_unix(monitor)|WARN|0x00007f5f7c8ba830 <__libc_start_main+0xf0>
 |daemon_unix(monitor)|WARN|0x0000000000408329 <_init+0x3261>
 |daemon_unix(monitor)|ERR|1 crashes: pid 106155 died, killed \
(Segmentation fault), core dumped, restarting

Most C library functions are not async-signal-safe, meaning that
it is not safe to call them from a signal handler, for example
printf() or fflush(). To be async-signal-safe, the handler only
collects the stack info using libunwind, which is signal-safe, and
issues 'write' to the pipe, where the monitor thread reads and
prints to ovs-vswitchd.log.

Tested-at: https://travis-ci.org/williamtu/ovs-travis/builds/590503433
Signed-off-by: William Tu <u9012063@gmail.com>
Signed-off-by: Ben Pfaff <blp@ovn.org>
.travis.yml
configure.ac
lib/backtrace.c
lib/backtrace.h
lib/daemon-private.h
lib/daemon-unix.c
lib/fatal-signal.c
m4/openvswitch.m4