David Miller wrote:

From: "Paul E. McKenney" <redacted>
Date: Sun, 25 Feb 2007 17:52:30 -0800

quoted

Why doesn't the traditional hash table of locks work here?  Use the
cache-line address as input to the hash function, take the corresponding
lock, do the compare-and-exchange by hand, and then release the lock.
What am I missing here?  Address aliasing do to memory being mapped into
multiple locations or something?  (In that case, use only the portion
of the address within the page, right?)

It doesn't protect against pure stores.

Use atomic_cmpxchg (or atomic_long_cmpxchg). The atomic_t type ensures
that unprotected stores don't happen (unless you try to do something
stupid, but you can break all other atomic_ ops like that anyway).

Even ignoring that you'll also need to disable interrupts, always,
which therefore will always perform very poorly compared to an
open-coded spinlock variant.

But on those architectures, the same argument could also be said about
a lot of atomic_ operations in the generic kernel code (ie. just use
spinlocks).

There are zero gains to cmpxchg(), in my opinion, you dirty and
grab exclusive access to a cacheline with cmpxchg() just the
same as a spinlock, so the real cost is the same.

There is zero justification for using this primitive in generic
code.

IMO it is useful for when you are memory constrained (and bit locks
are no good because they require an atomic unlock). For example we
could actually cut down on a reasonable number of atomic ops by adding
a spinlock to struct page. However I suspect this wouldn't go down too
well.

With that said, our use of atomic operations in the kernel is probably
not an optimal balance for SMP sparc32 and parisc (ie. those platforms
might be a little faster with more spinlocks and less atomics). Does
it really matter, though?

-- 
SUSE Labs, Novell Inc.
Send instant messages to your online friends http://au.messenger.yahoo.com