Hi, Why would SIGVTALRM ever cause IO#sysread to throw EINTR? I'm having a problem where a Net::HTTP connection is repeatedly dying because its IO#sysread call is being interrupted and raising EINTR. I have attached strace to the process, and verified that the signal being raised is SIGVTALRM. What I don't understand is why SIGVTALRM would ever interrupt a system call. the setitimer(2) documentation says that the ITIMER_VIRTUAL timer *only* decrements when the process is executing. So why would the signal fire when I'm blocked on a system call (eg. when my process is *not* executing)? I notice that up until recently, signal handlers for all signals except SIGVTALRM were installed with the SA_RESTART flag, which would make read(2) retry instead of returning EINTR when a signal was received. (I'm still using a version of Ruby that has this behavior). Why did Ruby specifically *not* set SA_RESTART for SIGVTALRM? I'm guessing the rationale has something to do with allowing Ruby to run its thread scheduler. But this doesn't make sense to me for two reasons: 1. the Ruby interpreter already takes measures to ensure that it never blocks on a system call if there are other runnable threads. So when would it ever be blocked on a system call, but have other threads it wants to schedule by receiving SIGVTALRM? 2. again, according to the setitimer(2) documentation, it doesn't make sense that SIGVTALRM would ever be raised when blocked on a system call. So why should not setting SA_RESTART have any effect? Any help is much appreciated! Josh P.S. Does the change in 1.8.6 of not setting SA_RESTART with sigaction(2) mean that Ruby scripts now need to handle EINTR in many places they did not previously have to?
on 2007-05-02 19:50
on 2007-05-03 19:41
On Thu, May 03, 2007 at 02:50:04AM +0900, Joshua Haberman wrote: > Hi, > > Why would SIGVTALRM ever cause IO#sysread to throw EINTR? I'm having > a problem where a Net::HTTP connection is repeatedly dying because its > IO#sysread call is being interrupted and raising EINTR. I have > attached strace to the process, and verified that the signal being > raised is SIGVTALRM. Which platform are you running on and which version of ruby are you running? Do you have a small test case that can reproduce the problem? > P.S. Does the change in 1.8.6 of not setting SA_RESTART with > sigaction(2) mean that Ruby scripts now need to handle EINTR in many > places they did not previously have to? This appears to be the case, but I haven't tested it. Looks like maybe this was changed in response to [ruby-talk:133002]. Incidentally, one of my coworkers has recently found some cases in the python standard library that didn't handle EINTR correctly. It's easy to neglect if you aren't thinking about it. Paul