On Fri, Oct 11, 2002, Johannes Erdfelt <[EMAIL PROTECTED]> wrote: > On Fri, Oct 11, 2002, Dan Streetman <[EMAIL PROTECTED]> wrote: > > On Fri, 11 Oct 2002, Johannes Erdfelt wrote: > > >On Fri, Oct 11, 2002, Dan Streetman <[EMAIL PROTECTED]> wrote: > > >> -for places where a td->link was written, added check to make sure HC > > >> hadn't already wrote that td->link back to the td's qh->element. > > > > > >This sounds racy. > > > > I think it actually fixes a race. There are only 2 places, in in > > uhci_append_queued_urb and uhci_delete_queued_urb. My mental picture of > > the race is this: > > > > QH1 -(element)-> TD -> <whatever...QH or TERM> > > > > So, let's say the HC is currently processing the TD. It finishes, reads > > TD->link, and writes it back into QH->element. Then, TD->link is updated > > by the HCD. But, TD is already out of the frame list! So whatever was > > put there really should be in QH->element, and whatever's in QH->element > > shouldn't be there. > > > > You can check the 2 places (which do essentially the same thing) in the > > patch to see if you agree. They both have a comment "If the HC..." > > Oh, that's not a problem. We'll be removing that finished URB (QH + > TD's) very shortly and we'll put the new URB that just lost the race into > the schedule. > > It'll get serviced. So there is a race, but it's a self fixing one. > > The HC could also retire the last TD in between setting td->link and > updating qh->element as well. The HC caches the entire TD for the duration > of the packet being sent to the device. That's a lot of CPU cycles. > > I'm just hesitant of reading/writing shared fields when we know that we > don't own it exclusively.
I forgot that we can fix this too. Georg had a clever method in usb-uhci that used a bottom QH, which was always at the end of the last URB. It would be passed on to the next queued URB for it's head. This would fix this race. I don't think the race is common enough to really need to do this. JE ------------------------------------------------------- This sf.net email is sponsored by:ThinkGeek Welcome to geek heaven. http://thinkgeek.com/sf _______________________________________________ [EMAIL PROTECTED] To unsubscribe, use the last form field at: https://lists.sourceforge.net/lists/listinfo/linux-usb-devel
