We recently made a patch in the Intel DRM driver:
http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commit;h=b60678a75d44fa9d5969f79781bd856ad5858609
This switches from non-MSI to MSI-mode.
It "should" have had no effect, but our concern is that what we're
really seeing is interrupt sharing troubles. We've had several people
say that they got an interrupt flood when plugging in a USB stick or
external disk. Switching the graphics driver to MSI mode "cures" the
bug, but that sure seems like a work-around rather than a bug fix to me.
When our driver is active, it can generate a lot of interrupts,
potentially thousands per second. Is it possible that the UHCI or EHCI
drivers could get confused if checking for interrupt status this often?
Eric noted that the USB driver appears to not check and ACK interrupt
status unconditionally, preferring to check the software state
beforehand. I'm wondering if this may open up a potential race between
hardware state change and ISR bit setting.
--
[email protected]
<added linux-usb list, that's the proper place for it...)
On Thu, Dec 11, 2008 at 08:35:37PM -0800, Keith Packard wrote:
> We recently made a patch in the Intel DRM driver:
>
> http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commit;h=b60678a75d44fa9d5969f79781bd856ad5858609
>
> This switches from non-MSI to MSI-mode.
>
> It "should" have had no effect, but our concern is that what we're
> really seeing is interrupt sharing troubles. We've had several people
> say that they got an interrupt flood when plugging in a USB stick or
> external disk. Switching the graphics driver to MSI mode "cures" the
> bug, but that sure seems like a work-around rather than a bug fix to me.
>
> When our driver is active, it can generate a lot of interrupts,
> potentially thousands per second. Is it possible that the UHCI or EHCI
> drivers could get confused if checking for interrupt status this often?
>
> Eric noted that the USB driver appears to not check and ACK interrupt
> status unconditionally, preferring to check the software state
> beforehand. I'm wondering if this may open up a potential race between
> hardware state change and ISR bit setting.
Check what software state? It shouldn't cause any race issues, but I'll
defer to David and Alan to concur this or not.
thanks,
greg k-h
On Sat, 13 Dec 2008, Greg KH wrote:
> <added linux-usb list, that's the proper place for it...)
>
> On Thu, Dec 11, 2008 at 08:35:37PM -0800, Keith Packard wrote:
> > We recently made a patch in the Intel DRM driver:
> >
> > http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commit;h=b60678a75d44fa9d5969f79781bd856ad5858609
> >
> > This switches from non-MSI to MSI-mode.
> >
> > It "should" have had no effect, but our concern is that what we're
> > really seeing is interrupt sharing troubles. We've had several people
> > say that they got an interrupt flood when plugging in a USB stick or
> > external disk.
Which drivers are using the IRQ that gets flooded? Can we see a dmesg
log from a kernel built with CONFIG_USB_DEBUG enabled?
> > Switching the graphics driver to MSI mode "cures" the
> > bug, but that sure seems like a work-around rather than a bug fix to me.
> >
> > When our driver is active, it can generate a lot of interrupts,
> > potentially thousands per second. Is it possible that the UHCI or EHCI
> > drivers could get confused if checking for interrupt status this often?
I don't think so.
> > Eric noted that the USB driver
Which USB driver? Does this refer to usb_hcd_irq()?
> > appears to not check and ACK interrupt
> > status unconditionally, preferring to check the software state
> > beforehand. I'm wondering if this may open up a potential race between
> > hardware state change and ISR bit setting.
Not if the drivers are written correctly. As far as I know, they are.
Alan Stern