On Wed, 2 Nov 2022 10:49:26 +0000 Leonid Ravich <leonid.ravich@xxxxxxxxxxxxxxxx> wrote: > after investigation looks like there is a broken assumption of rb_end_commit and rb_start_commit are on same CPU (no migration) during trace > looking in ib_mad trace in include/trace/events/ib_mad.h there is a call to create_mad_addr_info during TP assign which sleeps (at least under mlx5 ib driver) Wait, WHAT? > > so the scenario looks : > > rb_start_commit - buffer 0xffffa0984777e400 CPU 1 pid 1368 > switch from 1368 to 1605 > thread 1368 () is migrating from 1 to 0 > rb_end_commit - buffer 0xffffa09847c14a00 CPU 0 pid 1368 That's a MAJOR bug with the use case of tracepoints! > > before starting throwing some patch into the the air I would like to align with you the approach we should take here. > > my suggestion here : > - ftrace infra should verify no migration happen (end and start happens on same CPU) in case not we will throw warning for the issue . The scheduler should have. On entering the ring buffer code ring_buffer_lock_reserver() it disables preemption and does not re-enable it until ring_buffer_unlock_commit(). The only way to migrate is if you re-enable preemption. WHICH IS A BUG! > - ftrace users will be responsible to avoid migration during TP assign (RDMA in my case) Yes, it is responsible to do that. Just like it's up to users of spin_locks to not enable preemption and migrate while holding one. Entering the ring buffer code is the same as holding a spin_lock except that the ring buffer code allows for reentrancy. > > please let me know what do you think I think that whatever enabled preemption in the ib_mad trace event needs to be fixed. -- Steve
