On Thu, 2013-05-02 at 10:17 +0100, Chris Wilson wrote: > On Thu, May 02, 2013 at 12:00:03PM +0300, Imre Deak wrote: > > Due to possible scheduling latencies wait_event_timeout doesn't > > guarantee a non-zero return value, even if the condition becomes true > > before the specified timeout expires. Thus we can incorrectly signal a > > timeout and abort a DP AUX transaction. > > > > If wait_event_timeout returns 0, it's guaranteed that at least the > > specified timeout (minus one jiffies, see below) had passed, so we can > > fix this by checking the condition explicitly in this case. > > > > Also the timeout that wait_event_timeout() is guaranteed to wait if the > > condition doesn't become true is one less jiffies than what is passed to > > it as a parameter. This is because the absolute expiration time in > > schedule_timeout() may be calculated at a moment close to the next > > scheduling tick, when jiffies is incremented. So make sure we pass always > > a jiffies value of 2 or greater. Here this makes a difference only for > > HZ=100. > > > > This fixes DP AUX errors I saw during booting on an ILK. > > > > This should ideally be fixed in wait_event_timeout(), but that can take > > a while. Until that's done use this fix as a band-aid. > > As we have 3 such vulnerable callsite in our driver alone, perhaps we > should push for your general fix. Yes, I've sent a patch for wait_event_timeout() to start the discussion, I'm not sure how long it will take.. --Imre
