On 05/25/2015 02:58 PM, Lukas Slebodnik wrote:
On (22/05/15 11:02), Pavel Březina wrote:
> On 05/20/2015 02:20 PM, Lukas Slebodnik wrote:
>> ehlo,
>>
>> attached patch fixes a crash if return code of proxy_child is not 0.
>> You can reproduce it with sssd in non-root mode or
>> you can replace proxy_child with bash script.
>>
>> I'm not sure about the name of tevent_immediate callback
>> so feel free to propose better name.
>>
>> LS
>
> Hi,
> the whole proxy auth code does not really follow tevent style but I think
> your name is good enough.
> However, I don't see why it should crash, can you
> explain it in more detail please?
a) proxy_child_init_send is called
-- request with state "struct pc_init_ctx" is created
-- process is forked
-- parent register signal SIGCHLD with clallback pc_init_sig_handler
b) child return non-zero
-- parent recieve pc_init_sig_handler
c) pc_init_sig_handler is executed
-- unused argument "__siginfo" contains NON null data
-- tevent_req_error(req, EIO) is called
-- proxy_child_init_done callback is called
-- request is released (including "struct tevent_signal" and all related
data)
-- unused argument "__siginfo" is NULL now
d) function pc_init_sig_handler finished
-- we are back in tevent code from singnal handler
-- tevent wants to do some clean-up and dereference NULL in memset
It is not clear from back trace but the 1st argument of memset is NULL.
(gdb) bt
#0 0x00007fd7ba400505 in memset (__len=<optimized out>, __ch=<optimized
out>, __dest=<optimized out>) at /usr/include/bits/string3.h:84
#1 tevent_common_check_signal (ev=0x7fd7bfddf670) at ../tevent_signal.c:459
#2 0x00007fd7ba40228c in epoll_event_loop (tvalp=0x7fff85536430,
epoll_ev=0x7fd7bfddf8b0) at ../tevent_epoll.c:647
#3 epoll_event_loop_once (ev=<optimized out>, location=<optimized out>) at
../tevent_epoll.c:926
#4 0x00007fd7ba4007d7 in std_event_loop_once (ev=0x7fd7bfddf670, location=0x7fd7bdb417c3
"src/util/server.c:668") at ../tevent_standard.c:114
#5 0x00007fd7ba3fcfbd in _tevent_loop_once (ev=ev@entry=0x7fd7bfddf670,
location=location@entry=0x7fd7bdb417c3 "src/util/server.c:668") at
../tevent.c:530
#6 0x00007fd7ba3fd15b in tevent_common_loop_wait (ev=0x7fd7bfddf670,
location=0x7fd7bdb417c3 "src/util/server.c:668") at ../tevent.c:634
#7 0x00007fd7ba400777 in std_event_loop_wait (ev=0x7fd7bfddf670, location=0x7fd7bdb417c3
"src/util/server.c:668") at ../tevent_standard.c:140
#8 0x00007fd7bdb29343 in server_loop (main_ctx=0x7fd7bfde0ac0) at src/util/server.c:668
#9 0x00007fd7be39ca42 in main (argc=8, argv=<optimized out>) at
src/providers/data_provider_be.c:2909
BTW it's easily reproducible with proxy provider. I can give you an access
to such machine.
I hope it is clear now.
LS
Hi,
I'm sorry, but I don't think this is the problem. I'm not saying the
solution may not be right, but the cause is different.
1) siginfo is not attached to tevent_signal context, so freeing
tevent_signal does not do anything with siginfo
static struct tevent_sig_state {
struct tevent_common_signal_list *sig_handlers[TEVENT_NUM_SIGNALS+1];
struct sigaction *oldact[TEVENT_NUM_SIGNALS+1];
struct tevent_sigcounter signal_count[TEVENT_NUM_SIGNALS+1];
struct tevent_sigcounter got_signal;
#ifdef SA_SIGINFO
/* with SA_SIGINFO we get quite a lot of info per signal */
siginfo_t *sig_info[TEVENT_NUM_SIGNALS+1];
struct tevent_sigcounter sig_blocked[TEVENT_NUM_SIGNALS+1];
#endif
} *sig_state;
2) it is a ring buffer, that is the reason why it is cleared in tevent
after sighandler is finished
3) freeing tevent_signal inside the handler is a valid operation, you
can see a pretty clever check whether the structure was freed or not in
the tevent code
/*
* We have to be careful to not touch "se"
* after it was deleted in its handler. Thus
* we allocate a child whose destructor will
* tell by nulling out itself that its parent
* is gone.
*/
exists = talloc(se, struct tevent_se_exists);
if (exists == NULL) {
continue;
}
exists->myself = &exists;
talloc_set_destructor(
exists, tevent_se_exists_destructor);
4) so it is definitely not supposed to crash because you free
tevent_signal inside the handler
5) I'm not sure why it crashes, however we can stop using flag
SA_SIGINFO since it is unused anyway, it should do the trick
6) freeing tevent_signal causes to run destructor of pc_init_ctx which
sends kill(SIGKILL). Maybe it triggers some tevent bug?