Re: [rfc] lock bitops

[David Howells <dhowells@redhat.com>]

Nick Piggin <npiggin@suse.de> wrote:

> This patch (along with the subsequent one to optimise unlock_page) reduces
> the overhead of lock_page/unlock_page (measured with page faults and a patch
> to lock the page in the fault handler) by about 425 cycles on my 2-way G5.

Seems reasonable, though test_and_set_lock_bit() might be a better name.

> +There are two special bitops with lock barrier semantics (acquire/release,
> +same as spinlocks).

You should update Documentation/memory-barriers.txt also.

>  #define TestSetPageLocked(page)		\
>  		test_and_set_bit(PG_locked, &(page)->flags)
> +#define TestSetPageLocked_Lock(page)		\
> +		test_and_set_bit_lock(PG_locked, &(page)->flags)

Can we get away with just moving TestSetPageLocked() to the new function
rather than adding another accessor?  Or how about LockPageLocked() and
UnlockPageLocked() rather than SetPageLocked_Lock() that last looks wrong
somehow.

The FRV changes look reasonable, btw.

David