From: Theodore Ts'o <ty...@mi...> - 2003-07-28 15:10:27
|
On Sun, Jul 27, 2003 at 12:21:35PM -0500, Ray Bryant wrote: > Hanna, > > David Hansen mentioned something at the OLS performance tools bof about > there being a problem with lockmeter (sometimes lock hold times are > being reported that are wildly out of whack). Are you aware of such a > problem (this was in 2.6-test). > > If so, I'd appreciate being pointed at a test case and I will see what I > can do about this. Well, there was a problem I discovered where some of the newer spinlock primitives weren't getting instrumented, and so code sequences of the form: while (queue_not_empty) ... spin_lock_irq() ... spin_unlock() ... spin_lock() ... spin_unlock_irqresture() ... do something that takes a long, long time.... } where the *_irq, *_irqreqstore functions were not being instrumented. I believe Hanna fixed this, but I don't know how well the fix was propagated; maybe some folks were using an unfixed version, perhaps? Another possibility is that some other spinlock variant is still not being instrumented and Hanna and I both missed it. I'd certainly start by auditing spinlock.h after applying the lockmeter patch as an initial starting point. - Ted |