Re: [Lse-tech] lockmeter fixes -- any interest??

[John H. <ha...@sg...>]

From: "Dipankar Sarma" <dip...@in...>

> > From: "Dipankar Sarma" <dip...@in...>
> > > Does lockmeter handle atomic_dec_and_lock() ? If I remember
correctly,
> > > it didn't. Fixing this would be handy.

> Oh, I can see the aggregate of all calls to atomic_dec_and_lock(),
> but that is not very useful. Is it not possible to treat
> atomic_dec_and_lock() as another locking primitive for instrumentation
> and show the stats broken up by calling routines ?

The problem is a generic Lockmeter shortcoming:  a specific spin_lock()
call that locks more than one spinlock.

I tried making the i386 atomic_dec_and_lock() an inline (in order to
have each caller of atomic_dec_and_lock have the spin_lock() inside that
caller's procedure), but that didn't work.  When I hacked individual
"static inline" versions of atomic_dec_and_lock() for each particular
caller, then those were nicely inline'd.  That's not a general long-term
solution, however, since atomic_dec_and_lock() calls can come and go,
and it would complicate the Lockmeter patch.

John Hawkes