[Qemu-devel] Re: [PATCH 0/8] Various USB fixes and improvements (update 2)

[Max Krasnyansky <maxk@kernel.org>]

Anthony Liguori wrote:
> Any objections to applying this series?  It seems like the consensus is 
> that OHCI support is better long term but this series seems pretty sane 
> and self-contained.

I'm actually having seconds thought on the OHCI vs UHCI. You probably 
saw my reply to Paul. New UHCI code is fairly clean and simple, and is 
working well. But I definitely think that we should fix OHCI too, and at 
some point we need to add EHCI (ie USB 2.0 stuff).

I have more patches coming. I made isoc transactions work on Friday 
which made MS VX-3000 USB camera totally usable (will send a patch in a 
couple of hours). And I'm working on making control transactions async 
too which should shave off last bits of bad latency (2-3milliseconds) 
from the USB code. Initially I thought control stuff is not very 
frequent but this stupid USB webcam drivers does one very often.
So yeah, it be nice if this stuff is merged soonish :).

btw If you're using git as a front end to the SVN I can push my git tree 
somewhere to kernel.org so that you can just pull the whole thing.

Max

> 
> Regards,
> 
> Anthony Liguori
> 
> Max Krasnyansky wrote:
>> This is an updated version of the USB patches I sent out yesterday.
>> It includes changes and fixes suggested by Anthony.
>>
>> This time I also did more testing with XP running on top of QEMU/KQEMU
>> (used to be QEMU/KVM).
>>
>> Max Krasnyansky (8):
>>   husb: support for USB host device auto disconnect.
>>   husb: support for USB host device auto connect.
>>   usb:  generic packet handler cleanup and documentation
>>   uhci: rewrite UHCI emulator, fully async operation with multiple
>>         outstanding transactions
>>   husb: rewrite Linux host USB layer, fully async operation
>>   husb: remove disconnect detection timer
>>   husb: fixup printfs and stuff based on the review comments
>>   uhci: fixes for save/load-vm
>>
>>  hw/usb-uhci.c |  905 
>> ++++++++++++++++++++++++++++++++++-----------------------
>>  hw/usb.c      |  265 +++++++++--------
>>  hw/usb.h      |   37 +++-
>>  usb-linux.c   |  645 +++++++++++++++++++++++-----------------
>>  vl.c          |   83 +++---
>>  5 files changed, 1138 insertions(+), 797 deletions(-)
>>
>> ----
>> Here is the original description.
>>
>> This patch series started when I tried to share USB ports between four 
>> instances of Windows XP running on the same Linux box (under KVM).
>> I quickly realized that current USB support is not very flexible. We do
>> not handle devices disconnects, there is not way to assign certain USB
>> ports to VM instance, etc.
>>
>> Once I fixed that I discovered that USB devices that I absolutely need
>> in the VMs (Xilinx and Altera USB dongles) do not really work with
>> QEMU. VMs were getting stuck, applications unhappy, etc.
>> So I endded up rewriting UHCI and Linux host USB layers to make them 
>> fully async and to support multiple outstanding transactions.
>>
>> The result is quite nice. We can now assign USB buses to VM instances 
>> and devices are automatically connected to the VMs. Just do     
>> usb_add host:N.*
>> in the console or -usbdevice command line option (N is the bus number).
>> Also when device is disconnected from the host it's automatically removed
>> from the guest.
>>
>> Host USB devices operate in fully async mode (except the control 
>> transfers).
>> All the stalls and jerkiness due to long synchronous transactions is 
>> gone.
>> I can easily hook up four different USB devices (mouse, CF card 
>> reader, phone, Xilinx dongle) and everything is working perfectly. 
>> Mouse movements
>> are silky smooth :).
>>
>> I did some profiling with OProfile and we seems to be doing ok while 
>> XP is pumping ~10 MBytes over USB (reported by one of the apps I'm 
>> using). UHCI stuff is well below VNC for example.
>>
>> There is more work to be done (async control transfers for example). 
>> But I think this is way better than what we have now and is ready for 
>> more testing
>> by wider audience.
>> Most of the testing so far was done with KVM flavor of QEMU. I did 
>> test generic i386-softmmu target a bit, it's too slow for any serious 
>> testing with XP. I did full compile (all targets) too and it went fine.
>>   
>