On Mon, Oct 11, 2010 at 04:52:31PM +0100, Stefan Hajnoczi wrote: > 2010/10/11 Gleb Natapov <g...@redhat.com>: > > On Mon, Oct 11, 2010 at 01:48:09PM +0100, Stefan Hajnoczi wrote: > >> On Mon, Oct 11, 2010 at 12:16 PM, Bernhard Kohl <bernhard.k...@nsn.com> > >> wrote: > >> > Am 11.10.2010 12:18, schrieb ext Gleb Natapov: > >> >> > >> >> Currently if VM is started with multiple disks it is almost impossible > >> >> to > >> >> guess which one of them will be used as boot device especially if there > >> >> is a mix of ATA/virtio/SCSI devices. Essentially BIOS decides the order > >> >> and without looking into the code you can't tell what the order will > >> >> be (and in qemu-kvm if boot=on is used it brings even more havoc). We > >> >> should allow fine-grained control of boot order from qemu command line, > >> >> or as a minimum control what device will be used for booting. > >> >> > >> >> To do that along with inventing syntax to specify boot order on qemu > >> >> command line we need to communicate boot order to seabios via fw_cfg > >> >> interface. For that we need to have a way to unambiguously specify a > >> >> disk from qemu to seabios. PCI bus address is not enough since not all > >> >> devices are PCI (do we care about them?) and since one PCI device may > >> >> control more then one disk (ATA slave/master, SCSI LUNs). We can do what > >> >> EDD specification does. Describe disk as: > >> >> bus type (isa/pci), > >> >> address on a bus (16 bit base address for isa, b/s/f for pci) > >> >> device type (ATA/SCSI/VIRTIO) > >> >> device path (slave/master for ATA, LUN for SCSI, nothing for virtio) > >> >> > >> >> Will it cover all use cased? Any other ideas? > >> > > >> > I think this also applies to network booting via gPXE. Usually our VMs > >> > have 4 NICs, mixed virtio-net and PCI pass-through. 2 of the NICs shall > >> > be used for booting, even if there are hard disks or floppy disks > >> > connected. This scenario is currently almost impossible to configure. > >> > >> Here is a gPXE to support fw_cfg. You can pass gPXE script files from > >> the host to gPXE inside the guest. This means you can boot specific > >> NICs: > >> http://patchwork.ozlabs.org/patch/43777/ > >> > >> Just wanted to post the link because it is related to the gPXE side of > >> this discussion. > >> > > Don't we load gPXE for each NIC and seabios passes PCI device to boot from > > when it invokes one of them? > > SeaBIOS may do that but gPXE internally just probes all PCI devices. > It does not take advantage of the PCI bus/addr/fn that was passed to > the option ROM. A gPXE instance will try booting from each available > NIC in sequence.
Ah, thanks for clarification. Looks like gPXE does the wrong thing here. Can this behaviour be changed by compile time option? -- Gleb.