Re: [v10] Userspace patches for PCI device assignment

[Anthony Liguori <anthony@codemonkey.ws>]

muli@il.ibm.com wrote:
> This patchset enables device assignment for KVM hosts for PCI
> devices. It uses the Intel IOMMU by default if available.
>   

This is still broken wrt CPU hotplug.  If a new CPU comes online, it 
won't have it's ioperm mask set correct and it will crash QEMU.  You 
need to keep track of the current ioperm settings and whenever a new 
VCPU comes online, issue an ioperm.

Regards,

Anthony Liguori

> Changes from v9->v10:
>
> - really emove nr_assigned_devices and device_assignment_enabled and
>   other cleanups
>
> Changes from v8->v9 in no particular order:
>
> - fixed DEBUG, removed uneeded headers
> - introduce USE_KVM_DEVICE_ASSIGNMENT modeled after KVM_USE_PIT, per
>   aliguori's suggestion
> - only call term_printf from the monitor
> - implement markmc's suggestions: don't overflow dev->regions, fix
>   disable_iommu bogosity, fix usage and comment with regards to
>   -pcidevice format, and some other bits
>
> Changes from v7->v8 in in particular order:
>
> - various formatting fixes, DEBUG cleanups, cast removals, etc.
> - s/strncpy/snprintf/
> - split initialization in two phases per aliguori's suggestion
> - bail out on errors when we can't limp on
> - do ioperm on every cpu and vcpu (Weidong Han)
> - use pwrite/pread where applicable
> - split r_virtbase into different fields for memory and IO
> - fix destruction of MMIO regions (Disheng Su and Weidong Han)
>
> Changes from v6->v7 in no particular order:
>
> - formatting changes: adhere to qemu style
> - use strncmp, strncpy etc. instead of the insecure ones
> - move from array to linked list
> - change iopl() to ioperm() (Weidong Han)
> - other small changes as suggested during the review of v6.
>