code_api|pthreads.ptsig failing on x86 Travis #2921
Comments
The test occasionally hit the case where a async signal was received while the main thread is exiting. The patch adds a global var that waits for until the signal has been received. Fixes #2921
|
There are 2 separate problems with this test. First problem. The 2 threads it's spawning each are sending a signal. Occasionally the signal is delayed enough such that the receiving main thread is already on its way to exit and we are receiving the signal w/o a dcontext. Second problem. The 2 threads are getting spawned right after another, and it appears that occasionally the first child thread is sending it so early that the signal is received in the main thread while it is cloning the second child thread. Specifically, while it is cloning/memcpy its tls memory with its tls magic TLS_MAGIC_INVALID. DR's signal handler code only handles the case where a SUSPEND signal arrives while a thread is in this temporary state. Other signals cause error above. |
|
Aso xref #2400 |
Fixes an asynch signal arriving late when thread is on its way to exit by blocking all signals during exit. This is ok if because we're past the app's thread_exit system call. When doing a detach, the app's signal mask is restored before going native. For terminate events using kill, we are not blocking the kill signal. Suspend signal is also excluded because a detaching thread may try to synchronize with an exiting threads and as long as the signal is getting delivered to this thread, we need to reply to it from the signal handler. Adds support to the signal handler for an asynch signal arriving in the middle of a clone system call within DR, while the spawning thread's tls magic is invalid. We're now detecting this condition in the signal handler, but we need to be careful in the future that any thread has not been added to the global all_threads list as long as its tls magic is invalid. Fixes #2921
…one syscall (#3364) Fixes an asynch signal arriving late when thread is on its way to exit by blocking all signals during exit. This is ok, because we're at the app's thread_exit system call. When doing a detach, the app's signal mask is restored before going native. For terminate events using kill, we are not blocking the kill signal. The suspend signal is also excluded, because a detaching thread may try to synchronize with an exiting thread and as long as the signal is getting delivered to this thread, we need to reply to it from the signal handler. Adds support to the signal handler for an asynch signal arriving in the middle of a clone system call or temporarily-native thread, while the spawning thread's tls magic sentinel is invalid. We're now detecting this condition in the signal handler by making assumptions as stated in this patch. Fixes #2921
Seen once in #2906
The text was updated successfully, but these errors were encountered: