> I'd love to hear this sequence. Certainly regression testing never generated
> this sequence yet but yes that doesn't mean anything. Note that your slow
> path is very different than mine.
One of my testcases fell over on it...
> I don't feel the need of any xchg to enforce additional serialization.
I don't use XCHG anywhere... do you mean CMPXCHG?
> yours plays with cmpxchg in a way that I still cannot understand
It tries not to let the "active count" transition 1->0 happen if it can avoid
it (ie: it would rather wake someone up and not decrement the count). It also
only calls __rwsem_do_wake() if the caller manages to transition the active
This avoids another subtle bug that I think I have a sequence written down for
> Infact eax will be changed because it will be clobbered by the slow path, so
> I have to. Infact you are not using the +a like I do there and you don't
> save EAX explicitly on the stack I think that's "your" bug.
Not so... my down-failed slowpath functions return sem in EAX.
> Again it's not a performance issue, the "+a" (sem) is a correctness issue
> because the slow path will clobber it.
There must be a performance issue too, otherwise our read up/down fastpaths
are the same. Which clearly they're not.
> About the reason I'm faster than you in the down_write fast path is that I
> can do the subl instead of the cmpxchg, you say this is my big fault, I
> think my algorithm allows me to do that, but maybe I'm wrong.
I used to do that.
> Unfortunatly I "have" to put the pushl there because I don't want to save
> %ecx in the fast path (if I declare ecx clobbered it's even worse no?).
It benchmarked faster without it for me. I suspect this will be different on
different CPUs anyway.
I'm going to have to have a play with Intel's VTUNE program and see what it
> I said on 64bit archs. Of course on x86-64 there is xaddq and the rex
But the instructions you've specified aren't 64-bit.
> It isn't mandatory, if you don't want the xchgadd infrastructure then you
> don't set even CONFIG_RWSEM_XCHGADD, no?
My point is that mine isn't mandatory either... You just don't set the XADD
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to firstname.lastname@example.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
This archive was generated by hypermail 2b29 : Mon Apr 30 2001 - 21:00:10 EST