On Wed, Oct 31, 2018 at 02:49:26PM +0100, Peter Zijlstra wrote: > On Tue, Oct 30, 2018 at 03:18:43PM -0700, Douglas Anderson wrote: > > Looking closely at it, it seems like a really bad idea to be calling > > local_irq_enable() in kgdb_roundup_cpus(). If nothing else that seems > > like it could violate spinlock semantics and cause a deadlock. > > > > Instead, let's use a private csd alongside > > smp_call_function_single_async() to round up the other CPUs. Using > > smp_call_function_single_async() doesn't require interrupts to be > > enabled so we can remove the offending bit of code. > > You might want to mention that the only reason this isn't a deadlock > itself is because there is a timeout on waiting for the slaves to > check-in. dbg_master_lock must be owned to call kgdb_roundup_cpus() so the calls to smp_call_function_single_async() should never deadlock the calling CPU unless there has been a previous failure to round up (e.g. cores that cannot react to the round up signal). When there is a failure to round up when we resume, there is a window (before whatever locks that prevented the IPI being serviced are released) during which the system will deadlock if the debugger is re entered. I don't think there is any point trying to round up a CPU that did not previously respond... it should still have an IPI pending. The deadlock can be eliminated by getting the round up code to avoid CPUs whose csd->flags are non-zero either by checking the flag in the kgdb code or adding something like smp_trycall_function_single_async(). Daniel.
