Commits on Jun 11, 2020

1. Fix incorrect cpu.sleeper accounting …

   Under contention for the CPU lock, you could end up in a scenario where
   waiters on the cpu.sem would be awoken earlier than they should have been.
   
   The problem crosses two functions. lkl_cpu_get and lkl_cpu_put:
   
   https://github.com/lsds/lkl/blob/master/arch/lkl/kernel/cpu.c#L95
   https://github.com/lsds/lkl/blob/master/arch/lkl/kernel/cpu.c#L113
   
   The handling of "cpu.sleepers" is incorrect. "cpu.sleepers" should be
   incremented when a thread is waiting to get the cpu.sem so that in lkl_cpu_put,
   it can call sem_up to wake a sleeper if any exists.
   
   However what it is currently doing is in lkl_cpu_get is incrementing the
   sleepers thread count every time it can't get the cpu lock.
   
   https://github.com/lsds/lkl/blob/master/arch/lkl/kernel/cpu.c#L102
   
   So, if a thread fails to get the lock more than once it will increment the
   count of sleeping threads more than once. This means as it stands, in lkl_cpu_put,
   that one thread trying to get the lock several times can account for multiple calls
   to sem_up. Given that it is awoken each time in lkl_cpu_get, that is wrong.
   
   The correct logic is to decrement the sleepers count after the sem_down returns in
   lkl_cpu_get and not in lkl_cpu_put. That will correctly match up cpu.sleepers with
   the number of threads waiting on cpu.sem.

   

   SeanTAllen committed Jun 11, 2020
   Configuration menu

   • View commit details
   • Copy full SHA for 52964a5
   • Browse repository at this point

   Copy the full SHA

   52964a5 View commit details

   Browse the repository at this point in the history