Message ID | 1381226311-23489-1-git-send-email-stefanha@redhat.com |
---|---|
State | New |
Headers | show |
Am 08.10.2013 um 11:58 hat Stefan Hajnoczi geschrieben: > glib versions prior to 2.31.0 require an explicit g_thread_init() call > to enable multi-threading. > > Failure to initialize threading causes glib to take single-threaded code > paths without synchronization. For example, the g_slice allocator will > crash due to race conditions. > > Fix this for all QEMU tool programs (qemu-nbd, qemu-io, qemu-img) by > moving the g_thread_init() call from vl.c:main() into a new > osdep.c:thread_init() constructor function. > > thread_init() has __attribute__((constructor)) and is automatically > invoked by the runtime during startup. > > We can now drop the "simple" trace backend's g_thread_init() call since > thread_init() already called it. > > Note that we must keep coroutine-gthread.c's g_thread_init() call which > is located in a constructor function. There is no guarantee for > constructor function ordering so thread_init() may only be called later. The glib documentation says: Since version 2.24, calling g_thread_init() multiple times is allowed, but nothing happens except for the first call. I take that this means previously it wasn't allowed. qemu's configure checks for a minimum version of 2.12, so we seems to support glib versions that don't allow g_thread_init() to be called multiple times. Do we need to protect against this? Kevin
Il 08/10/2013 14:25, Kevin Wolf ha scritto: > The glib documentation says: > > Since version 2.24, calling g_thread_init() multiple times is > allowed, but nothing happens except for the first call. > > I take that this means previously it wasn't allowed. qemu's configure > checks for a minimum version of 2.12, so we seems to support glib > versions that don't allow g_thread_init() to be called multiple times. > > Do we need to protect against this? I think that's the point of the "if (!g_thread_supported ())" tests. Paolo
Am 08.10.2013 um 15:08 hat Paolo Bonzini geschrieben: > Il 08/10/2013 14:25, Kevin Wolf ha scritto: > > The glib documentation says: > > > > Since version 2.24, calling g_thread_init() multiple times is > > allowed, but nothing happens except for the first call. > > > > I take that this means previously it wasn't allowed. qemu's configure > > checks for a minimum version of 2.12, so we seems to support glib > > versions that don't allow g_thread_init() to be called multiple times. > > > > Do we need to protect against this? > > I think that's the point of the "if (!g_thread_supported ())" tests. Ah yes, I think you're right. Not the best function name I've ever seen that glib uses there, but okay. Kevin
On Tue, Oct 08, 2013 at 03:58:24PM +0200, Kevin Wolf wrote: > Am 08.10.2013 um 15:08 hat Paolo Bonzini geschrieben: > > Il 08/10/2013 14:25, Kevin Wolf ha scritto: > > > The glib documentation says: > > > > > > Since version 2.24, calling g_thread_init() multiple times is > > > allowed, but nothing happens except for the first call. > > > > > > I take that this means previously it wasn't allowed. qemu's configure > > > checks for a minimum version of 2.12, so we seems to support glib > > > versions that don't allow g_thread_init() to be called multiple times. > > > > > > Do we need to protect against this? > > > > I think that's the point of the "if (!g_thread_supported ())" tests. > > Ah yes, I think you're right. Not the best function name I've ever seen > that glib uses there, but okay. That's correct, g_thread_support() is there to initialize once only. Stefan
On Tue, Oct 08, 2013 at 11:58:31AM +0200, Stefan Hajnoczi wrote: > glib versions prior to 2.31.0 require an explicit g_thread_init() call > to enable multi-threading. > > Failure to initialize threading causes glib to take single-threaded code > paths without synchronization. For example, the g_slice allocator will > crash due to race conditions. > > Fix this for all QEMU tool programs (qemu-nbd, qemu-io, qemu-img) by > moving the g_thread_init() call from vl.c:main() into a new > osdep.c:thread_init() constructor function. > > thread_init() has __attribute__((constructor)) and is automatically > invoked by the runtime during startup. > > We can now drop the "simple" trace backend's g_thread_init() call since > thread_init() already called it. > > Note that we must keep coroutine-gthread.c's g_thread_init() call which > is located in a constructor function. There is no guarantee for > constructor function ordering so thread_init() may only be called later. > > Reported-by: Mario de Chenno <mario.dechenno@unina2.it> > Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> > --- > trace/simple.c | 9 --------- > util/osdep.c | 18 ++++++++++++++++++ > vl.c | 8 -------- > 3 files changed, 18 insertions(+), 17 deletions(-) Applied to my block tree, we need this for QEMU 2.0: https://github.com/stefanha/qemu/commits/block Stefan
diff --git a/trace/simple.c b/trace/simple.c index 1e3f691..7833309 100644 --- a/trace/simple.c +++ b/trace/simple.c @@ -404,15 +404,6 @@ bool trace_backend_init(const char *events, const char *file) { GThread *thread; - if (!g_thread_supported()) { -#if !GLIB_CHECK_VERSION(2, 31, 0) - g_thread_init(NULL); -#else - fprintf(stderr, "glib threading failed to initialize.\n"); - exit(1); -#endif - } - #if !GLIB_CHECK_VERSION(2, 31, 0) trace_available_cond = g_cond_new(); trace_empty_cond = g_cond_new(); diff --git a/util/osdep.c b/util/osdep.c index 62072b4..e29ba6f 100644 --- a/util/osdep.c +++ b/util/osdep.c @@ -437,6 +437,24 @@ int socket_init(void) return 0; } +/* Ensure that glib is running in multi-threaded mode */ +static void __attribute__((constructor)) thread_init(void) +{ + if (!g_thread_supported()) { +#if !GLIB_CHECK_VERSION(2, 31, 0) + /* Old versions of glib require explicit initialization. Failure to do + * this results in the single-threaded code paths being taken inside + * glib. For example, the g_slice allocator will not be thread-safe + * and cause crashes. + */ + g_thread_init(NULL); +#else + fprintf(stderr, "glib threading failed to initialize.\n"); + exit(1); +#endif + } +} + #ifndef CONFIG_IOVEC /* helper function for iov_send_recv() */ static ssize_t diff --git a/vl.c b/vl.c index 983cdc6..e2dee8e 100644 --- a/vl.c +++ b/vl.c @@ -2857,14 +2857,6 @@ int main(int argc, char **argv, char **envp) error_set_progname(argv[0]); g_mem_set_vtable(&mem_trace); - if (!g_thread_supported()) { -#if !GLIB_CHECK_VERSION(2, 31, 0) - g_thread_init(NULL); -#else - fprintf(stderr, "glib threading failed to initialize.\n"); - exit(1); -#endif - } module_call_init(MODULE_INIT_QOM);
glib versions prior to 2.31.0 require an explicit g_thread_init() call to enable multi-threading. Failure to initialize threading causes glib to take single-threaded code paths without synchronization. For example, the g_slice allocator will crash due to race conditions. Fix this for all QEMU tool programs (qemu-nbd, qemu-io, qemu-img) by moving the g_thread_init() call from vl.c:main() into a new osdep.c:thread_init() constructor function. thread_init() has __attribute__((constructor)) and is automatically invoked by the runtime during startup. We can now drop the "simple" trace backend's g_thread_init() call since thread_init() already called it. Note that we must keep coroutine-gthread.c's g_thread_init() call which is located in a constructor function. There is no guarantee for constructor function ordering so thread_init() may only be called later. Reported-by: Mario de Chenno <mario.dechenno@unina2.it> Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> --- trace/simple.c | 9 --------- util/osdep.c | 18 ++++++++++++++++++ vl.c | 8 -------- 3 files changed, 18 insertions(+), 17 deletions(-)