Hi Mathias,
On Thu, Feb 28, 2019 at 11:18:58AM +0200, Mathias Nyman wrote:
> Hi
>
> On 19.2.2019 17.41, Suwan Kim wrote:
> > Patch "USB: HCD: support giveback of URB in tasklet context"
> > introduced giveback of urb in tasklet context. [1] This patch was
> > applied to ehci but not xhci. [2] It significantly reduces the hard
> > irq time of xhci. Especially for the uvc driver, the hard irq including
> > the uvc complete function runs quite long (about 300-350us in my
> > Thinkpad s440 laptop with webcam) but applying this patch reduces
> > the hard irq time of xhci to about 18-30us.
>
> Sorry about the late reply.
> Can you recall any reason why this wasn't applied to xhci back then?
Ming Lei who is the author of the giveback-tasklet patch in usb core
worked with Alan stern when he was working on the patch, and he only
replaced giveback of ehci. I don't know why he didn't replace xhci.
After some time, somebody asked him why this patch did not apply to
xhci. And there was no apparent reason.[1] IMO, no one seems
interested in xhci...
> xhci is doing a lot in hard interrupt context, and reducing that would
> be a good idea. Another option to look at is using threaded interrupts
> for xhci.
When the giveback-tasklet patch was working, the author experimented
with comparing threaded interrupt and tasklet. At that time, tasklet
showed better performance than threaded interrupt.[2]
> We might however be opening a can of worms with this, the impact is unknown.
>
> How much testing was done with URB return in tasklet for SS devices?
>
> Would be nice to test this out on a bit wider audience before applying it.
Unfortunately, SS devices I have are only usb mass storage devices.
When I tested with a USB mass storage device, both xhci with tasklet
and without tasklet(urb complete in hard IRQ) showed similar performance.
[USB-mass storage TEST]
- Testbed is i5-7600 and two mass storage devices (usb flash memory,
external hard drive) are used.
Test is executed 10 times and figure out the average speed
- dd if=/dev/sdN of=/dev/null iflag=direct bs=1G count=1
- device1 : Sandisk Ultra Flair USB 3.0 32GB
- device2 : WD My Passport 2TB (external hard drive)
- xhci without tasklet
- device1 - 129.727MB/s
- device2 - 103.667MB/s
- xhci with tasklet
- device1 - 103.2MB/s
- device2 - 103.692MB/s
I have only one high-speed isochronous type device that is built-in
webcam in my laptop (Thinkpad s440, i5-4210U) In this case, it is
difficult to measure the performance. So i checked the overrun/underrun
event in xhci. Until now, no overrun/underrun event has occurred
in actual use.
Regards
Suwan Kim
