[Qemu-devel] Re: qemu svn r5281 on FreeBSD - slow usb, vmwarevga, screen updates... (now updated to r5313)

[Juergen Lock]

On Tue, Sep 30, 2008 at 09:51:33AM -0500, Anthony Liguori wrote:
> Juergen Lock wrote:
>> In article <48DE5256.5000101 at codemonkey.ws> you write:
>>   
>>> [...]
>>>     
>> 
>>   
>>>>> The one thing that really tripped me up with the whole aio kld-module 
>>>>> thing.  Perhaps we should detect the presence of the module at run time 
>>>>> and disable aio?  I assume kldload can only be run as root?
>>>>>             
>>>> Yes.  Atm the ports print a warning when aio is not loaded:
>>>>         
>>> Yeah, I don't think this is enough.  I'd rather see AIO be disabled when 
>>> modfind("aio") is not available (printing a warning along with that would 
>>> be fine).  A non-privileged user cannot load the aio module so it's not 
>>> very useful to tell them to load it.
>>>     
>> 
>> OK so how about the following?  (only tested with a raw image, but if
>> the way its disabled for OpenBSD works for all of them this should as well.)
>> 
>>  Oh and am I right qemu-img doesn't use aio?  If it actually does we may
>> want to add the same check there instead of just disabling it.  (I kept it
>> enabled for qemu-nbd since thats not built on FreeBSD anyway.)
>>   
> 
> Disabling aio for everyone is not the right thing if posix-aio is broken.
> 
Well, I went after what is done for the OpenBSD case (CONFIG_AIO not set),
i.e. tell bdrv_register() to set bdrv_aio_read & frieds to bdrv_aio_read_em
etc for bdrv_raw if aio is not loaded.  I found one bug tho, the same
should be done for bdrv_host_device, i.e. in block.c
	bdrv_register(&bdrv_host_device, 0);
should be
	bdrv_register(&bdrv_host_device, emulate_aio);
too.  The 0 for the others there mean don't emulate i.e. keep aio enabled...

 Or are you talking about qemu-img?  If that would in fact benefit from
using aio like this too we could just add the same test as in vl.c.
(Or we could move the test to bdrv_init(), I just didn't want to print
the warning from in there.)

> What would be better is in block-raw-posix.c, to have a one type check of 
> modfind() (if we're FreeBSD), and if it fails, set a flag that forces the 
> aio routines to call bdrv_aio_{read,write}_em.
> 
 You mean runtime checks every time a raw aio fn is called (even if just
a flag?)  That's what I was trying to avoid... :)

 Regards,
	Juergen