On Tue, Feb 14, 2023 at 04:51:53PM -0500, Gregory Price wrote:
On Tue, Feb 14, 2023 at 09:18:24PM +0000, Jonathan Cameron wrote:
On Tue, 14 Feb 2023 14:01:23 -0500 Gregory Price gregory.price@memverge.com wrote:
Could you test it with TCG (just drop --enable-kvm)? We have a known limitation with x86 instructions running out of CXL emulated memory (side effect of emulating the interleave). You'll need a fix even on TCG for the corner case of an instruction bridging from normal ram to cxl memory. https://lore.kernel.org/qemu-devel/20230206193809.1153124-1-richard.henderso...
Performance will be bad, but so far this is only way we can do it correctly.
Jonathan
Siiiggghh... i had this patch and dropped --enable-kvm, but forgot to drop "accel=kvm" from the -machine line
This was the issue.
And let me tell you, if you numactl --membind=1 python, it is IMPRESSIVELY slow. I wonder if it's even hitting a few 100k instructions a second.
This appears to be the issue. When I get a bit more time, try to dive into the deep dark depths of qemu memory regions to see how difficult a non-mmio fork might be, unless someone else is already looking at it.
~Gregory
Just clarifying one thing: Even with the patch, KVM blows up. Disabling KVM fixes this entirely. I haven't tested without KVM but with the patch, i will do that now.