Messages in this thread | | | Date | Mon, 19 Dec 2016 11:20:24 +0100 | From | Peter Zijlstra <> | Subject | Re: [RFC][PATCH] spinlock_debug: report spinlock lockup from unlock |
| |
On Sun, Dec 18, 2016 at 01:19:11AM +0900, Sergey Senozhatsky wrote: > There is a race window between the point when __spin_lock_debug() > detects spinlock lockup and the time when CPU that caused the > lockup receives its backtrace interrupt. > > Before __spin_lock_debug() triggers all_cpu_backtrace() it calls > spin_dump() to printk() the current state of the lock and CPU > backtrace. These printk() calls can take some time to print the > messages to serial console, for instance (we are not talking > about console_unlock() loop and a flood of messages from other > CPUs, but just spin_dump() printk() and serial console). > > All those preparation steps can give CPU that caused the lockup > enough time to run away, so when it receives a backtrace interrupt > it can look completely innocent. > > The patch extends `struct raw_spinlock' with additional variable > that stores jiffies of successful do_raw_spin_lock() and checks > in debug_spin_unlock() whether the spin_lock has been locked for > too long. So we will have a reliable backtrace from CPU that > locked up and a reliable backtrace from CPU that caused the > lockup.
But why? Also, why jiffies, that's a horrible source of time.
| |