Revert "linux: workaround to avoid deadlock inside dl_iterate_phdr in glibc" #57059
Conversation
… glibc (…" This reverts commit 1f05953.
|
It looks like FreeBSD is already aware of this bug in their libc: https://reviews.freebsd.org/D47350, so I don't think we should revert the fix just because it triggers a known bug in FreeBSD. |
|
Is the |
The changes should be applied only to Linux then, otherwise we're knowingly introducing a regression.
It is! At least I think it is... Does this suffice as a check?: julia> libc = dlopen("/lib/libc.so.7");
julia> dlsym(libc, :_rtld_atfork_pre)
Ptr{Nothing}(0x0000000082f76b28) |
| { | ||
| callback_t callback = {f, ctx}; | ||
| dl_iterate_phdr(with_dl_iterate_phdr_lock, &callback); |
There was a problem hiding this comment.
On FreeBSD, can you just replace this call with int rtld_locks[MAX_RTLD_LOCKS]; jl_lock_profile(); _rtld_atfork_pre(rtld_locks); f(ctx); _rtld_atfork_post(rtld_locks); jl_unlock_profile();
There was a problem hiding this comment.
Looks like even though _rtld_atfork_pre is accessible from libc, there's no corresponding header that's included in a standard system installation that has a prototype for it nor a definition for MAX_RTLD_LOCKS. 😔
There was a problem hiding this comment.
It's here, if you have this rtld_lock.h header, or you can just copy-paste the relevant lines from it:
https://github.com/lattera/freebsd/blob/401a161083850a9a4ce916f37520c084cff1543b/libexec/rtld-elf/rtld_lock.h#L48
There was a problem hiding this comment.
Right, what I mean is that it's in the system source files (as in your link) but those aren't included in a standard installation, so I don't have the definitions available from an existing file on my machine. I'm compiling now with this change:
diff --git a/src/signals-unix.c b/src/signals-unix.c
index 91d3378068..e1f1d70a80 100644
--- a/src/signals-unix.c
+++ b/src/signals-unix.c
@@ -41,6 +41,12 @@
#include <sys/event.h>
#endif
+#if defined(_OS_FREEBSD_)
+#define MAX_RTLS_LOCKS 8
+extern void _rtld_atfork_pre(int*);
+extern void _rtld_atfork_post(int*);
+#endif
+
// 8M signal stack, same as default stack size (though we barely use this)
static const size_t sig_stack_size = 8 * 1024 * 1024;
@@ -325,8 +331,17 @@ static int with_dl_iterate_phdr_lock(struct dl_phdr_info *info, size_t size, voi
void jl_with_stackwalk_lock(void (*f)(void*), void *ctx)
{
+#if defined(_OS_FREEBSD_)
+ int rtld_locks[MAX_RTLD_LOCKS];
+ jl_lock_profile();
+ _rtld_atfork_pre(rtld_locks);
+ f(ctx);
+ _rtld_atfork_post(rtld_locks);
+ jl_unlock_profile();
+#else
callback_t callback = {f, ctx};
dl_iterate_phdr(with_dl_iterate_phdr_lock, &callback);
+#endif
}
#if defined(_OS_LINUX_) && (defined(_CPU_X86_64_) || defined(_CPU_X86_))Any idea what tests need to be run to trigger this specific deadlock or should I just run all of them?
There was a problem hiding this comment.
Looks like that didn't fix it, I still hit a deadlock.
Reverts #57035, fixes #57058, reopens #57017.