Description of problem: When tracing multiple threads in a process using ptrace the Fedora Core kernel only delivers an exit notification through wait for the initial thread. Vanilla Linux kernels deliver an exit notification for each traced thread. Version-Release number of selected component (if applicable): 2.6.19-1.2911.6.5.fc6 How reproducible: Every time. Steps to Reproduce: 1. Untar the attached test. 2. cd exit-test 3. make 4. ./monitor Actual results: Prints only one exit notification. Expected results: Should print two exit notifications. Additional info: For a less synthetic testcase, strace is also affected by this. Try compiling this program with "gcc -fopenmp -o foo foo.c": int main(int argc, char **argv) { int i; #pragma omp parallel for for (i = 0; i < 1; i++) ; return 0; } Then trace it with "strace -f ./foo". Strace reports multiple detaches for the same thread or bogus SIGSEGVs: ... [pid 4586] futex(0x886b048, FUTEX_WAKE, 1) = 1 [pid 4585] <... futex resumed> ) = 0 [pid 4585] futex(0x886b048, FUTEX_WAKE, 1) = 0 [pid 4585] exit_group(0) = ? Process 4585 detached Process 4586 detached Process 4585 detached ... [pid 4589] futex(0x19d830, FUTEX_WAIT, 0, NULL <unfinished ...> [pid 4588] <... futex resumed> ) = 0 [pid 4588] exit_group(0) = ? [pid 4589] <... futex resumed> ) = 9 [pid 4589] --- SIGSEGV (Segmentation fault) @ 0 (0) ---
Created attachment 150042 [details] Test case
Please test 2.6.20-1.2925.fc6 It's in Fedora testing now and should be released today.
Looks like this is still a problem.
Created attachment 150103 [details] test case contained in one source file This is the version of the test case I'm using.
Created attachment 150106 [details] fix patch I've fixed the bug and the attached patch is now part of the "upstream" utrace patches.
I've committed the new utrace patch to the rawhide kernel cvs. The attached patch can be folded into the utrace patch for fc6.
Is this bug fixed? Kernel 2933 is out with this fix and kernel 2937 or better (from http://people.redhat.com/cebbert) has more fixes.
Yes, the bug is fixed in both 2933 and 2937.