Michael Butler wrote:
1) the ata-raid driver currently leaks ata_composite and ata_request structures into "neverland" in a mirrored configuration. This can be observed using "sysctl -a | grep ^ata_" and noting the increasing "in-use" count as time goes on. Eventually, this causes the kernel to run out of memory. This is fixed by tracking the request counts on each composite request.
Looks pretty much on the spot, I'll look this one over and get it committed once I'm satisfied with it fixing the bug, thanks a bunch for hunting this one down !
2) another part of this patch is to ata-queue where a channel lock is asserted in a (hopefully rare) rebuild process even if the dependencies flag is set (we're waiting for a read). Moving the test for a dependency outside of the lock saves waiting on it when nothing can be done. A small nit with near negligible impact but, when you're waiting for a rebuild ...
I dont think you can measure this actually, but it doesn't hurt at least to move the lock outside.
3) another part of this patch is to ata-raid where the choice of drive from which to read favours one side of a mirror even when both drives are near the block(s) we want. Because the mirror is on another channel on the Highpoint controller, it performs (marginally :-() better when we toggle between them.
Hmm, well, depends on what sort of behavior one wants to optimise for. I have a few other patches for optimisations but havn't decided what to eventually use yet. Guess its time for me to run some tests on this..
I'll look into get this integrated, again thanks for digging in and doing the hard work of finding the problem(s) !!
-Søren _______________________________________________ freebsd-stable@freebsd.org mailing list http://lists.freebsd.org/mailman/listinfo/freebsd-stable To unsubscribe, send any mail to "[EMAIL PROTECTED]"
