Re: [PATCH v3 1/1] Shared memory uio_pci driver

["Michael S. Tsirkin" <mst@redhat.com>]

On Mon, Mar 29, 2010 at 11:59:24PM +0300, Avi Kivity wrote:
> On 03/28/2010 10:48 PM, Cam Macdonell wrote:
>> On Sat, Mar 27, 2010 at 11:48 AM, Avi Kivity<avi@redhat.com>  wrote:
>>    
>>> On 03/26/2010 07:14 PM, Cam Macdonell wrote:
>>>      
>>>>        
>>>>> I'm not familiar with the uio internals, but for the interface, an
>>>>> ioctl()
>>>>> on the fd to assign an eventfd to an MSI vector.  Similar to ioeventfd,
>>>>> but
>>>>> instead of mapping a doorbell to an eventfd, it maps a real MSI to an
>>>>> eventfd.
>>>>>
>>>>>          
>>>> uio will never support ioctls.
>>>>        
>>> Why not?
>>>      
>> Perhaps I spoke too strongly, but it was rejected before
>>
>> http://thread.gmane.org/gmane.linux.kernel/756481
>>
>> With a compelling case perhaps it could be added.
>>    
>
> Ah, the usual "ioctls are ugly, go away".
>
> It could be done via sysfs:
>
>   $ cat /sys/.../msix/max-interrupts
>   256

This one can be discovered with existing sysfs.

>   $ echo 4 > /sys/.../msix/allocate
>   $ # subdirectories 0 1 2 3 magically appear
>   $ # bind fd 13 to msix

There's no way to know, when qemu starts, how many vectors will be used
by driver.  So I think we can just go ahead and allocate as many vectors
as supported by device at the moment when the first eventfd is bound.

>   $ echo 13 > /sys/.../msix/2/bind-fd

I think that this one can't work, both fd 13 and sysfs file will get
closed when /bin/echo process exits.

>   $ # from now on, msix interrupt 2 will call eventfd_signal() on fd 13
>
> Call me old fashioned, but I prefer ioctls.

I think we will have to use ioctl or irqcontrol because of lifetime
issues outlines above.

> -- 
> Do not meddle in the internals of kernels, for they are subtle and quick to panic.