On Thu, Sep 26, 2002 at 02:48:53AM +0200, Dirk-Willem van Gulik wrote: > I could not agree more. > > The original patch used 'time' - but that is expensive to calculate. If > you make a histogram of conseq EINTR over a few weeks you find that > sequences longer than 5-10 are very rare on a healthy machine. But once > something runs out of control (say a PHP/sql connector) it shoots in the > hundreds and thousands for just a single process quickly. > > Suggestions for a better metric (Short of making it self learning ;-).
But, that's heavily dependent on the OS. Solaris has a tendancy to have much higher EINTR occurances than other OSes. So, I'd prefer to let the OS tell us when something has gone afoul rather than trying to second-guess it when the error only means "You were interrupted - try again." So, I don't think there is a metric that can work (without fail) for this case. -- justin