Date: Tue, 23 Nov 2010 08:33:29 +0200 From: Andriy Gapon <avg@freebsd.org> To: freebsd-hackers@freebsd.org, "Robert N. M. Watson" <rwatson@freebsd.org> Subject: dtrace/cyclic deadlock Message-ID: <4CEB6039.2040700@freebsd.org>
next in thread | raw e-mail | index | archive | help
I think that I've run into the known issue of dtrace/cyclic deadlock. Just would like to run my understanding and ideas by you. The problem is that the cyclic_fire() callback is executed in the interrupt filter context (and thus with interrupts disabled) and it tries to obtain a spin mutex lock in the cyclic code. At the same time other CPU may execute a thread that holds that spin mutex and uses smp_rendezvous_cpus() to perform a synchronous function invocation on the first CPU. So, CPU #1 can not make forward progress because it is spinning on the spin-lock and CPU #2 can not make forward progress because it can not interrupt CPU #1. I think that the problem was introduced during the porting of the code. On (Open)Solaris there are no spin-locks in this code, all data structures are per-CPU and data coherency is ensured by (1) accessing the data only from the CPU to which it belongs; and (2) using some modern-day spl*() equivalent[?] to block interrupts. I think that this is quite similar to what we do for per-CPU caches in UMA and so the same approach should work here. That is, as in (Open)Solaris, the data should be accessed only from the owning CPU and spinlock_enter()/spinlock_exit() should be used to prevent races between non-interrupt code and nested interrupt code. What do you think? Thanks! -- Andriy Gapon
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?4CEB6039.2040700>