Few fixes to the libvirt.xml:

[mykaul]

Move to virtio (hard disk, NICs), SCSI (cd-rom)
Removed USB controller (not needed), Balloon driver.
Added virtio-rng controller.

I did not remove the graphics and replaced with a console, though
it can be done as well.

Changed default iothreads from threads to native.

Tested on Fedora 26, but should work on EL7 hosts as well.

[zakame]

This looks better than my PR #52 🎉 hope it gets in! 😻

[zakame]

Seems to break on my system:

θ61° [zakame:~] % docker-machine create --driver=kvm --engine-registry-mirror=http://192.168.122.1:50000 --swarm-master swarm-manager-1
Running pre-create checks...
Creating machine...
(swarm-manager-1) Copying /home/zakame/.docker/machine/cache/boot2docker.iso to /home/zakame/.docker/machine/machines/swarm-manager-1/boot2docker.iso...
(swarm-manager-1) Creating SSH key...
(swarm-manager-1) Failed to start: virError(Code=1, Domain=10, Message='internal error: process exited while connecting to monitor: 2017-08-10T03:23:14.589561Z qemu-kvm: -drive file=/home/zakame/.docker/machine/machines/swarm-manager-1/swarm-manager-1.img,format=raw,if=none,id=drive-virtio-disk0,aio=native: aio=native was specified, but it requires cache.direct=on, which was not specified.')
Error creating machine: Error in driver during machine creation: virError(Code=1, Domain=10, Message='internal error: process exited while connecting to monitor: 2017-08-10T03:23:14.589561Z qemu-kvm: -drive file=/home/zakame/.docker/machine/machines/swarm-manager-1/swarm-manager-1.img,format=raw,if=none,id=drive-virtio-disk0,aio=native: aio=native was specified, but it requires cache.direct=on, which was not specified.')

Setting kvm-cache-mode set to none or directsync, or any other mode with explicit cache.direct=on got it to work again.

See also https://bugzilla.redhat.com/show_bug.cgi?id=1086704

[mykaul]

OK, reverted the change to switch the default IO mode from threads to native and kept it on threads.

[gbraad]

This however would also need to be tested with other guest OS, besides Fedora and CentOS/RHEL. HOW about minikube, or boot2docker ISOs?

[zakame]

@gbraad it should work on minikube/boot2docker based machines as I'm using this PR myself, on a Slackware host

[mykaul]

@gbraad - this PR came from me trying to use minikube with bookt2docker ISO and being a bit concerned about the old (vintage?) devices that were used. It works OK with minikube 0.21.0, and on top of it I've ran K8S 1.7.0, 1.7.2, 1.7.3.

[mykaul]

ping?

[mykaul]

Anything I can help with to push this?

[afbjorklund]

I thought virtio devices were normally called /dev/vda and not /dev/sda ?
At least that is what I got, when I did the same thing (virtio) with qemu-kvm...

There was also an interesting side effect, in that their order got reversed.
That is: without virtio I get /dev/sda1 (A) and with I get dev/vdb1 (B) :

-cdrom boot2docker.iso disk.qcow2

Disk /dev/sda: 20 GB, 20971524608 bytes, 40960009 sectors
2549 cylinders, 255 heads, 63 sectors/track
Units: cylinders of 16065 * 512 = 8225280 bytes

Device  Boot StartCHS    EndCHS        StartLBA     EndLBA    Sectors  Size Id Type
/dev/sda1    123,0,1     1023,254,63    1975995   40949684   38973690 18.5G 83 Linux
/dev/sda2    0,1,1       122,254,63          63    1975994    1975932  964M 82 Linux swap

Partition table entries are not in disk order
Note: sector size is 2048 (not 512)
Disk /dev/sr0: 45 MB, 47185920 bytes, 92160 sectors
11 cylinders, 64 heads, 32 sectors/track
Units: cylinders of 2048 * 2048 = 4194304 bytes

Device   Boot StartCHS    EndCHS        StartLBA     EndLBA    Sectors  Size Id Type
/dev/sr0p1 *  0,0,1       44,63,32             0      92159      92160  180M 17 Hidden HPFS/NTFS

-drive file=boot2docker.iso,index=2,media=cdrom,if=virtio -drive file=disk.qcow2,index=0,media=disk,if=virtio

Disk /dev/vda: 45 MB, 47185920 bytes, 92160 sectors
45 cylinders, 64 heads, 32 sectors/track
Units: cylinders of 2048 * 512 = 1048576 bytes

Device  Boot StartCHS    EndCHS        StartLBA     EndLBA    Sectors  Size Id Type
/dev/vda1 *  0,0,1       44,63,32             0      92159      92160 45.0M 17 Hidden HPFS/NTFS
Disk /dev/vdb: 20 GB, 20971524608 bytes, 40960009 sectors
40634 cylinders, 16 heads, 63 sectors/track
Units: cylinders of 1008 * 512 = 516096 bytes

Device  Boot StartCHS    EndCHS        StartLBA     EndLBA    Sectors  Size Id Type
/dev/vdb1    1023,15,63  1023,15,63     1954512   40959071   39004560 18.5G 83 Linux
/dev/vdb2    0,1,1       1023,15,63          63    1954511    1954449  954M 82 Linux swap

Partition table entries are not in disk order

[mykaul]

I've changed (in my patch) the CDROM device to use SCSI, which is therefore using /dev/sdX. If that's an issue, we can revert it back to IDE (or SATA). I wanted it to be quick as we read from it (and IDE is somewhat slow). But I really have not touched this patch for months as it did not get any attention whatsoever.

[afbjorklund]

I haven't benchmarked sda vs. vdb, both seemed to be "fast enough"...
Made my own driver instead, since this doesn't work without libvirt access.