> On Mar 31, 2015, at 6:51 PM, Leif Hedstrom <zw...@apache.org> wrote: > >> >> On Mar 31, 2015, at 1:39 PM, James Peach <jpe...@apache.org> wrote: >> >> >>> On Mar 31, 2015, at 10:17 AM, zw...@apache.org wrote: >>> >>> Repository: trafficserver >>> Updated Branches: >>> refs/heads/master e13d8a6e8 -> e0de1d6ed >>> >>> >>> TS-3470 traffic_server --help segfaults when SPDY is enabled >>> >>> >> [snip] >>> diff --git a/lib/ts/ink_args.cc b/lib/ts/ink_args.cc >>> index 12ab5c0..91cebd8 100644 >>> --- a/lib/ts/ink_args.cc >>> +++ b/lib/ts/ink_args.cc >>> @@ -323,5 +323,5 @@ usage(const ArgumentDescription *argument_descriptions, >>> unsigned n_argument_desc >>> } >>> fprintf(stderr, " %s\n", argument_descriptions[i].description); >>> } >>> - exit(EX_USAGE); >>> + _exit(EX_USAGE); >>> } >> >> What's the underlying bug here? Note that using _exit() everywhere defeats >> tcmalloc leak detection and seems to be a bit of a cargo cult in our code ... > > > > It’d segfault in SPDY as part of the destructor for some global object (I > think). I didn’t feel it worthwhile to try to fix SPDY, since we’re killing > it, and exiting out of the “usage()” like this hardly can have an impact on > e.g. tcmalloc (would we care to test for memory leaks in “traffic_server > —help”?).
FWIW, we have ~70 instances of _exit(). It's currently not possible to leak check with tcmalloc without hacking the build. > > We can certainly clean up the use of _exit(), I’m all for that. yes that would be good imho J
