On Fri, Jan 29, 2016 at 04:13:30PM +0900, Sergey Senozhatsky wrote: > On (01/29/16 15:54), Byungchul Park wrote: > > On Fri, Jan 29, 2016 at 09:27:03AM +0900, Sergey Senozhatsky wrote: > > > > > > well, the stack is surely limited, but on every > > > spin_dump()->spin_lock() recursive call it does another > > > round of > > > > > > u64 loops = loops_per_jiffy * HZ; > > > > > > for (i = 0; i < loops; i++) { > > > if (arch_spin_trylock(&lock->raw_lock)) ^^^ this is a trylock.
> > > return; > > > __delay(1); > > > } > > > > > > so if you have 1000 spin_dump()->spin_lock() then, well, > > > something has been holding the lock for '1000 * loops_per_jiffy * HZ'. > > > > Or the printk() is heavily called and the lock is congested. > > well, isn't it the case that ticket-based locking assumes at least > some sort of fairness? how many cpus do you have there? you can It's true for a congestion between arch_spin_lock()s, not trylock(). And I cannot remember how many online cpus there are, since I frequently change the number. Sorry. But the range is from 2 to 4. > have `num_online_cpus() - 1' tasks spinning on the spin lock and > 1 owning the spin lock... if your lock is in correct state (no > before/after spinlock debug errors) even most unlucky task should > get the lock eventually... > > -ss