Re: [PATCH] generalized spin_lock_bit

[William Lee Irwin III <wli@holomorphy.com>]

On 20 Jul 2002, Robert Love wrote:
>> The attached patch implements bit-sized spinlocks via the following
>> interfaces:

On Sat, Jul 20, 2002 at 01:40:22PM -0700, Linus Torvalds wrote:
> In particular, with the current pte_chain_lock() interface, it will be
> _trivial_ to turn that bit in page->flags to be instead a hash based on
> the page address into an array of spinlocks. Which is a lot more portable
> than the current code.
> (The current code works, but look at what it generates on old sparcs, for
> example).

I was hoping to devolve the issue of the implementation of it to arch
maintainers by asking for this. I was vaguely aware that the atomic bit
operations are implemented via hashed spinlocks on PA-RISC and some
others, so by asking for the right primitives to come back up from arch
code I hoped those who spin elsewhere might take advantage of their
window of exclusive ownership.

Would saying "Here is an address, please lock it, and if you must flip
a bit, use this bit" suffice? I thought it might give arch code enough
room to wiggle, but is it enough?


Thanks,
Bill