Messages in this thread | | | Date | Tue, 29 Jul 2008 18:41:16 -0400 | From | Mathieu Desnoyers <> | Subject | Re: module-placed markers/tracepoints |
| |
* Frank Ch. Eigler (fche@redhat.com) wrote: > Hi - > > Some locals are wondering -- is there code for (or need for new code > for) incrementing module reference counts while markers and/or > tracepoints resident in modules have active clients? > > - FChE
Hi Frank,
Probe module unloading is supposed to be done automatically assuming the following module unload behavior :
1 - Module unload code takes the stop machine lock. 2 - It is assumed that the module exit function is called at that time, while no preempt-disabled section is executed. This function unregisters the callbacks. 3 - Given that preemption is disabled around tracepoint/marker calls, no calls are left to a freed module.
Actually, I notice that it might not be the case. I took for granted what (I can't remember who ? Rusty ?) told me about module unload code. If we look at sys_delete_module() :
/* Stop the machine so refcounts can't move and disable module. */ ret = try_stop_module(mod, flags, &forced); if (ret != 0) goto out;
/* Never wait if forced. */ if (!forced && module_refcount(mod) != 0) wait_for_zero_refcount(mod);
mutex_unlock(&module_mutex); /* Final destruction now noone is using it. */ if (mod->exit != NULL) mod->exit();
So we see that mod->exit(), which I believe is responsible for calling the module exit() functions, is called after the stop_machine, not during the stop machine.
Therefore, what would be needed here is to add a synchronize_sched() after mod->exit() in module.c or after each tracepoint/marker unregister in marker.c/tracepoint.c. However, I'd really prefer to add this to module.c since adding this for _every_ unregister will slow down processing or multiple probe unregistration too much.
Rusty, am I understanding that correctly ?
Mathieu
-- Mathieu Desnoyers OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F BA06 3F25 A8FE 3BAE 9A68
| |