Re: [Qemu-devel] qemu usb-linux.c vl.h hw/esp.c hw/lsi53c895a.c ...

[Fabrice Bellard <fabrice@bellard.org>]

Hi,

OK for the SCSI async support. I am making changes to have a better low 
level handling of CDROMs (both real host CDROMs and emulated ones), but 
I don't know when it will be ready. The changes mostly consists in 
adding a method in the block devices to support packet CD commands.

About async USB, maybe some simplifications are possible. When you have 
a real USB controller, there is always a single current request and you 
don't neep to have provisions to handle several ones (of course more 
complicated host controllers could do that, but it is not useful to 
model it now). Moreover, the devices must reply to the request 
immediately so the request are asynchronous just because it takes a few 
uS to send the data, not because the device blocks (I am not sure it is 
clear for you).

So in QEMU, async USB is not needed in practise for emulated devices. 
For MSD you don't need it for example even if you are using SCSI AIO 
because the host keep polling the MSD storage device until there is an 
SCSI reply.

Async USB is only really needed in the case where QEMU speaks to a host 
device. In this case the operation to send a TD may block (ideally at 
most 1 ms) and it is interesting to hide that to the guest. In this 
case, we must make compromises to hide this fact to the guest. Your 
proposal seems good for that but USB frames will be longer than 1 ms in 
case of host devices (and I am ready to accept that as it is better than 
the previous solution !).

The more complicated implementation would be to suppress USB async as 
you did it but to export to the generic USB layers that list of TD that 
the controller maintains. Then it is possible to submit several requests 
at once to the host devices while having a consistent low level model.

Regards,

Fabrice.