Net modules are not loaded in VMs when using the 2.6.38 kernel

[joanna]

When I log into a VM and manually do modprobe iwlagn it works. But not automatically as it was with previous kernel...

This applies not only to the default netvm, but also to all other VMs -- so they do not get eth0 interface by default -- one needs to manually load xennet to get the eth0 interface, even though the corresponding vifX.Y interfaces are created on the corresponding backend domain right from the beginning.

[joanna]

This might also be a race-condition -- e.g. the kernel's udev scripts might be executing before the actual virtual device gets connected? (In case of netvm: before the pcifront starts seeing the pci device?)

[marmarek]

Probably this is race condition between udev and mounting /lib/modules...

[rafal]

In fact, udev is started before mounting local filesystems (and now /lib/modules is on local filesystem) - so it is a race likely to end fatally.
Perhaps enabling udev-post init script will retry search for failed modules ? I am not sure it is what it is meant for.

[marmarek]

Actually I've fixed it by mounting /lib/modules in initramfs (pre-pivot). But not tested yet...

[marmarek]

fixed in rel 4