calloc(3) initialises all memory with zero, so you don't need explicitly assign was_restored to 0.

Also this is related to all below "was_restored = 0" changes.

Instead allocating dev_info, I would like to suggest to use one of the following:

1. Use "sc" that already allocated in the line 2345, i.e. put some fields to sc's struct for every PCI type.
2. Introduce a functions like:
   
   snapshot_pci_device_register(pi)"; for pci devices snapshot_device_register(cb, data, name); for non-pci devices

And replace all init code for all devices with those functions.

This make code simpler and incapsulate all allocations, etc. Thus each decl "struct vm_snapshot_dev_info *dev_info" and code below will be removed in xxx_init() functions.

If meta_data always should be not NULL, just place assert() here.

I would suggest to call pci_pause()/pci_resume() unconditionally for each PCI device and if pe_pause is NULL do not do anything. The same for pci_resume().
That will simplify initialisation and if PCI device has implemented pe_pause/pe_resume, a handler will be called during suspend/resume. If PCI device does not implement pe_pause/pe_resume, just ignore it.

Probably you should correct a code to conform TODO's requirements and then remove TODO comment.

Is this comment still required?

What is this function currently used for? It only calls lookup_devs().

Probably lookup_devs() should be renamed with walk_and_restore_devices()

commented code - bad style ?

Put fprintf under ifdef DEBUG or even replace with dprintf?

Some of the information will need to be used outside the context of the device itself - the static array of user devices has been removed, so the snapshot component must keep track of the functions that need to be called when performing a snapshot or restore. Also, for devices like atkbdc, the snapshot_pci_device_register will likely not be compatible.

Before having support for atkbdc, the static userspace device array was also not required, since you could discover most devices through the PCI bus; however, if other device types are to be supported, a more generic approach is required.

However, I agree that some of the logic, like allocating the structure and performing a part of the initialization could be done separately.

Changing an order of vm_snapshot_kern_struct() and vm_snapsshot_user_devs() makes kernel structure corruption on resume and host panic. For example, I got panic at restore_guest_fpustate() when vcpu->guest_xcr0 became corrupted:

http://bxr.su/FreeBSD/sys/amd64/vmm/vmm.c#1134

This is due to vm_snapshot_kern_struct() resets "offset" to 0 while vm_snapsshot_user_devs() uses lseek(SEEK_CUR).