Fix: work-around glibc __nptl_setxid vs clone hang
authorMathieu Desnoyers <mathieu.desnoyers@efficios.com>
Thu, 12 Apr 2012 01:07:13 +0000 (21:07 -0400)
committerDavid Goulet <dgoulet@efficios.com>
Mon, 16 Apr 2012 15:27:24 +0000 (11:27 -0400)
commit8cf66a2b6551d770a8aedf1301bcec2270041785
treec1dda6a8f9129a887f95193a9711fa4c37185b9d
parent422361c276c49e1239194398f31f8842432da536
Fix: work-around glibc __nptl_setxid vs clone hang

hash table resize threads exit end up setting a "locked" state within
libc pthread, which deadlocks with seteuid/setegid called from the
cloned process in runas.c when runas() is called exactly when a resize
thread exits.

Temporarily fix this issue by adding a mutex cross this resize
operation, which holds mutual exclusion with runas() usage.

We should investigate whether we want to properly call exec() from the
runas.c clone child before touching any non-async-signal-safe libc call.
However, given that this change is more intrusive, let's first use this
mutex-based work-around.

Before this fix, running 1000 instances of "demo-trace 300" with
sessiond running as root, and:

lttng create
lttng enable-event -u -a
lttng start

would sometimes lead to consumerd hang with the following clone child
backtrace:

setxid_mark_thread (cmdp=<optimized out>, t=0x7f52dd47c700)
    at allocatestack.c:995
995 allocatestack.c: No such file or directory.
(gdb) bt full
    at allocatestack.c:995
        ch = <optimized out>
    at allocatestack.c:1088
        t = 0x80
        signalled = <optimized out>
        result = <optimized out>
        runp = 0x7f52dd47c9c0
    at ../sysdeps/unix/sysv/linux/setegid.c:44
        __p = 0xfffffffffffffe00
        __cmd = {syscall_no = 119, id = {-1, 1000, -1}, cntr = 0}
        result = <optimized out>
        data = 0x7f52e66e1930
        writelen = <optimized out>
        writeleft = <optimized out>
        index = <optimized out>
        sendret = {i = 0, c = "\000\000\000"}
        ret = <optimized out>
        __func__ = "child_run_as"
    at ../sysdeps/unix/sysv/linux/x86_64/clone.S:112
No locals.
No symbol table info available.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
src/common/hashtable/rculfhash.c
src/common/runas.c
src/common/runas.h
This page took 0.026605 seconds and 4 git commands to generate.