Message ID | 1352210901-1923-9-git-send-email-hdegoede@redhat.com |
---|---|
State | New |
Headers | show |
Il 06/11/2012 15:08, Hans de Goede ha scritto: > Our ehci code has is capable of significantly lowering the wakeup rate > for the hcd emulation while the device is idle. It is possible to add > similar code ot the uhci emulation, but that simply is not there atm, > and there is no reason why a (virtual) usb device can not be a USB-2 device. > > Making usb-hid devices connect to the emulated ehci controller instead > of the emulated uhci controller on vms which have both lowers the cpuload > for a fully idle vm from 20% to 2-3% (on my laptop). You need this to be dependent on the machine version. Otherwise the USB paths may change and migration will break. Paolo > Signed-off-by: Hans de Goede <hdegoede@redhat.com> > --- > hw/usb/dev-hid.c | 9 ++++++--- > 1 file changed, 6 insertions(+), 3 deletions(-) > > diff --git a/hw/usb/dev-hid.c b/hw/usb/dev-hid.c > index 69f89ff..96ba0c0 100644 > --- a/hw/usb/dev-hid.c > +++ b/hw/usb/dev-hid.c > @@ -97,7 +97,7 @@ static const USBDescIface desc_iface_mouse = { > .bEndpointAddress = USB_DIR_IN | 0x01, > .bmAttributes = USB_ENDPOINT_XFER_INT, > .wMaxPacketSize = 4, > - .bInterval = 0x0a, > + .bInterval = 8, > }, > }, > }; > @@ -127,7 +127,7 @@ static const USBDescIface desc_iface_tablet = { > .bEndpointAddress = USB_DIR_IN | 0x01, > .bmAttributes = USB_ENDPOINT_XFER_INT, > .wMaxPacketSize = 8, > - .bInterval = 0x0a, > + .bInterval = 8, > }, > }, > }; > @@ -158,7 +158,7 @@ static const USBDescIface desc_iface_keyboard = { > .bEndpointAddress = USB_DIR_IN | 0x01, > .bmAttributes = USB_ENDPOINT_XFER_INT, > .wMaxPacketSize = 8, > - .bInterval = 0x0a, > + .bInterval = 8, > }, > }, > }; > @@ -224,6 +224,7 @@ static const USBDesc desc_mouse = { > .iSerialNumber = STR_SERIALNUMBER, > }, > .full = &desc_device_mouse, > + .high = &desc_device_mouse, > .str = desc_strings, > }; > > @@ -237,6 +238,7 @@ static const USBDesc desc_tablet = { > .iSerialNumber = STR_SERIALNUMBER, > }, > .full = &desc_device_tablet, > + .high = &desc_device_tablet, > .str = desc_strings, > }; > > @@ -250,6 +252,7 @@ static const USBDesc desc_keyboard = { > .iSerialNumber = STR_SERIALNUMBER, > }, > .full = &desc_device_keyboard, > + .high = &desc_device_keyboard, > .str = desc_strings, > }; > >
On 11/06/12 15:08, Hans de Goede wrote: > Our ehci code has is capable of significantly lowering the wakeup rate > for the hcd emulation while the device is idle. It is possible to add > similar code ot the uhci emulation, but that simply is not there atm, > and there is no reason why a (virtual) usb device can not be a USB-2 device. We should probably add usb-tablet2 (+friends) devices as this is a guest-visible change. We could also try tricks with properties but I guess separate devices are easier. cheers, Gerd
Hi, On 11/08/2012 04:36 PM, Gerd Hoffmann wrote: > On 11/06/12 15:08, Hans de Goede wrote: >> Our ehci code has is capable of significantly lowering the wakeup rate >> for the hcd emulation while the device is idle. It is possible to add >> similar code ot the uhci emulation, but that simply is not there atm, >> and there is no reason why a (virtual) usb device can not be a USB-2 device. > > We should probably add usb-tablet2 (+friends) devices as this is a > guest-visible change. We could also try tricks with properties but I > guess separate devices are easier. A quick status update on this patchset. a) I promised to resend it without this patch Friday afternoon at kvm-forum, which I went to do directly after saying goodbye, but when I unsuspended my laptop it broke... (looks like the cable between the motherboard and screen has a loose contact), so it got delayed a bit. b) While working on fixing all the comments on the set yesterday to send a v2, I realized at the very last moment that handling interrupts packet async, rather then by nakking them, causes problems with migration, as the following can happen: 1) hcd submits a packet to device, device returns USB_RET_ASYNC 2) device gets some data, fills the packet, calls complete() on it 3) For ehci complete() calls a bh, which immediately walks the async schedule, but if not enough time has passed since the last bh, to process periodic frames, or if the interrupt endpoint has an interval of more then 1 ms, and this periodic frame it is not scheduled, the completion will not get written back to guest memory! 4) If we now migrate, the completion never gets written back to guest memory and the interrupt packet is effectively lost Notice that the same can happen with non async interrupt packets on the uhci code... I was planning on modifying that to be more like the ehci code anyways to lower the CPU-load (amount of wakeups), but that will not necessarily fix this, so this needs some more thinking. I hope to have a v2 of the wakeup reduction for interrupt endpoints patchset by the end of the day ... Regards, Hans
diff --git a/hw/usb/dev-hid.c b/hw/usb/dev-hid.c index 69f89ff..96ba0c0 100644 --- a/hw/usb/dev-hid.c +++ b/hw/usb/dev-hid.c @@ -97,7 +97,7 @@ static const USBDescIface desc_iface_mouse = { .bEndpointAddress = USB_DIR_IN | 0x01, .bmAttributes = USB_ENDPOINT_XFER_INT, .wMaxPacketSize = 4, - .bInterval = 0x0a, + .bInterval = 8, }, }, }; @@ -127,7 +127,7 @@ static const USBDescIface desc_iface_tablet = { .bEndpointAddress = USB_DIR_IN | 0x01, .bmAttributes = USB_ENDPOINT_XFER_INT, .wMaxPacketSize = 8, - .bInterval = 0x0a, + .bInterval = 8, }, }, }; @@ -158,7 +158,7 @@ static const USBDescIface desc_iface_keyboard = { .bEndpointAddress = USB_DIR_IN | 0x01, .bmAttributes = USB_ENDPOINT_XFER_INT, .wMaxPacketSize = 8, - .bInterval = 0x0a, + .bInterval = 8, }, }, }; @@ -224,6 +224,7 @@ static const USBDesc desc_mouse = { .iSerialNumber = STR_SERIALNUMBER, }, .full = &desc_device_mouse, + .high = &desc_device_mouse, .str = desc_strings, }; @@ -237,6 +238,7 @@ static const USBDesc desc_tablet = { .iSerialNumber = STR_SERIALNUMBER, }, .full = &desc_device_tablet, + .high = &desc_device_tablet, .str = desc_strings, }; @@ -250,6 +252,7 @@ static const USBDesc desc_keyboard = { .iSerialNumber = STR_SERIALNUMBER, }, .full = &desc_device_keyboard, + .high = &desc_device_keyboard, .str = desc_strings, };
Our ehci code has is capable of significantly lowering the wakeup rate for the hcd emulation while the device is idle. It is possible to add similar code ot the uhci emulation, but that simply is not there atm, and there is no reason why a (virtual) usb device can not be a USB-2 device. Making usb-hid devices connect to the emulated ehci controller instead of the emulated uhci controller on vms which have both lowers the cpuload for a fully idle vm from 20% to 2-3% (on my laptop). Signed-off-by: Hans de Goede <hdegoede@redhat.com> --- hw/usb/dev-hid.c | 9 ++++++--- 1 file changed, 6 insertions(+), 3 deletions(-)