2015-06-12 8:41 GMT+02:00 Willy Tarreau <[email protected]>: > On Fri, Jun 12, 2015 at 08:35:15AM +0200, Willy Tarreau wrote: >> On Fri, Jun 12, 2015 at 08:27:34AM +0200, joris dedieu wrote: >> (...) >> > All those one are innocents. Including sys/limits.h on common/compat.h >> > correct those differences and not the segfault. >> >> OK. Please use limits.h instead as sys/limits.h is not present on all >> systems. >> >> I'm thinking about something else you can test, which is to verify all >> objects sizes. You need to build with and without the patch, then you >> can diff the outputs and check whether some structures would have a >> different size. This could help spot the culprit : >> >> $ find . -name '*.o' | sort | xargs nm --size-sort > > When doing this here, I found something troubling : > > (joris.out is your version with the revert) > > --- normal.out 2015-06-12 08:37:00.828256778 +0200 > +++ joris.out 2015-06-12 08:36:41.509255648 +0200 > @@ -91,6 +91,7 @@ > > ./src/auth.o: > 00000004 B userlist > +00000006 C localtimezone > 00000041 T auth_find_userlist > 00000050 T check_group > 00000088 T check_user > @@ -1002,7 +1003,7 @@ > 00000020 T tcp_req_cont_keywords_register > 00000020 T tcp_res_cont_keywords_register > 00000027 t port_range_release_port > -00000028 r CSWTCH.124 > +00000028 r CSWTCH.122 > 00000039 t bind_parse_defer_accept > 00000039 t bind_parse_transparent > 00000039 T tcp_get_src > @@ -1545,7 +1546,7 @@ > 00000008 C date > 00000008 C now > 00000008 C start_date > -00000008 b tv_offset.4129 > +00000008 b tv_offset.3989 > 00000024 T _tv_isgt > 00000024 T _tv_isle > 0000002f T _tv_ms_elapsed > @@ -1562,6 +1563,7 @@ > 0000023f T tv_update_date > > ./src/uri_auth.o: > +00000006 C localtimezone > 0000001e T stats_set_flag > 0000001e T stats_set_refresh > 0000005a T stats_set_realm > > As you can see, "localtimezone" moves to another file. Guess what ? > it's an array of 6 chars which is defined in standard.h ! That's > completely bogus, I wonder where that comes from! > > I would not be surprized if you have a similarly named variable on > FreeBSD which is of a smaller size and which gets overwritten when > this variable is modified. > > Please could you try to rename it as a quick test ? It only appears > at 3 locations : > > $ git grep -n localtimezone > include/common/standard.h:827:char localtimezone[6]; > src/haproxy.c:588: strftime(localtimezone, 6, "%z", &curtime); > src/standard.c:2340: memcpy(dst, localtimezone, 5); // timezone
Done and unfortunately it stills segfaulting. But despite Level3 high quality route filtering near lunchtime, I think, I'm not too far. Has you suggested I looked at nm output and found things that make me think on struct proxy size. I found that there was a link with _COMMON_COMPAT_H defined or not and finally with CONFIG_HAP_TRANSPARENT. Commenting lines 101 to 108 on include/common/compat.h without other modifications stop segfaulting. I will try to continue this way. Best Regards Joris > > Willy >
