Poor performance with vioscsi and scsi-block (AHCI SSD) about kvm-guest-drivers-windows

[tomty89]

Okay figured it out. Turns out to be caused by a bug in libata. Since scsi-block does SCSI passthrough, CmdQue=0 set by libata is exposed to Windows, which apparently prevents command queuing.

CmdQue=0:

CmdQue=1:

It can be fixed with a patch to the kernel like this:

diff --git a/drivers/ata/libata-scsi.c b/drivers/ata/libata-scsi.c
index bfec66f..0a35164 100644
--- a/drivers/ata/libata-scsi.c
+++ b/drivers/ata/libata-scsi.c
@@ -2097,7 +2097,10 @@ static unsigned int ata_scsiop_inq_std(struct ata_scsi_args *args, u8 *rbuf)
                0,
                0x5,    /* claim SPC-3 version compatibility */
                2,
-               95 - 4
+               95 - 4,
+               0,
+               0,
+               2
        };

        VPRINTK("ENTER\n");
-- 

I think it's better to leave this report open until a fix gets into upstream kernel, but sure it's you guys' call.

[tomty89]

As a reference, I also ran the benchmark with the NVMe emulation in qemu. Works pretty well btw.

qemu-system-x86_64 -enable-kvm -cpu host -smp cores=4 -m 4G -net bridge -net nic -full-screen -drive file=disks/win10__.qcow2,format=qcow2,cache=none,aio=native -drive file=/dev/sdb,format=raw,cache=none,aio=native,if=none,id=bench -device nvme,drive=bench,serial=intel530:

[YanVugenfirer]

Thank you for digging into the issue and finding the cause.

I will keep the issue open for now as a reference for others.

[YanVugenfirer]

Fixed in Linux kernel: http://lxr.free-electrons.com/source/drivers/ata/libata-scsi.c#L2099