* Oleg Nesterov (o...@redhat.com) wrote: > On 03/16, Mathieu Desnoyers wrote: > > > > utrace_add_engine() > > set_notify_resume(target); > > > > ok, so this is where the TIF_NOTIFY_RESUME thread flag is set. I notice > > that it is set asynchronously with the execution of the target thread > > (as I do with my TIF_KERNEL_TRACE thread flag). > > > > However, on x86_64, _TIF_DO_NOTIFY_MASK is only tested in > > entry_64.S > > > > int_signal: > > and > > retint_signal: > > > > code paths. However, if there is no syscall tracing to do upon syscall > > entry, the thread flags are not re-read at syscall exit and you will > > miss the syscall exit returning from your target thread if this thread > > was blocked while you set its TIF_NOTIFY_RESUME. > > Afaics, TIF_NOTIFY_RESUME is not needed to trace syscall entry/exit. > If engine wants the syscall tracing, utrace_set_events(UTRACE_SYSCALL_xxx) > sets TIF_SYSCALL_TRACE. And syscall_trace_enter/syscall_trace_leave call > tracehook_report_syscall_xxx(). > > Oleg.
I recall that TIF_SYSCALL_TRACE also suffers from the same problem as TIF_NOTIFY_RESUME if set asynchronously with the target thread's execution at least on x86_64 and arm. Do you take care to stop the target thread in utrace_set_events ? Mathieu > -- Mathieu Desnoyers OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F BA06 3F25 A8FE 3BAE 9A68