• Steven Rostedt (Red Hat)'s avatar
    sched/rt: Simplify the IPI based RT balancing logic · 4bdced5c
    Steven Rostedt (Red Hat) authored
    When a CPU lowers its priority (schedules out a high priority task for a
    lower priority one), a check is made to see if any other CPU has overloaded
    RT tasks (more than one). It checks the rto_mask to determine this and if so
    it will request to pull one of those tasks to itself if the non running RT
    task is of higher priority than the new priority of the next task to run on
    the current CPU.
    
    When we deal with large number of CPUs, the original pull logic suffered
    from large lock contention on a single CPU run queue, which caused a huge
    latency across all CPUs. This was caused by only having one CPU having
    overloaded RT tasks and a bunch of other CPUs lowering their priority. To
    solve this issue, commit:
    
      b6366f04 ("sched/rt: Use IPI to trigger RT task push migration instead of pulling")
    
    changed the way to request a pull. Instead of grabbing the lock of the
    overloaded CPU's runqueue, it simply sent an IPI to that CPU to do the work.
    
    Although the IPI logic worked very well in removing the large latency build
    up, it still could suffer from a large number of IPIs being sent to a single
    CPU. On a 80 CPU box, I measured over 200us of processing IPIs. Worse yet,
    when I tested this on a 120 CPU box, with a stress test that had lots of
    RT tasks scheduling on all CPUs, it actually triggered the hard lockup
    detector! One CPU had so many IPIs sent to it, and due to the restart
    mechanism that is triggered when the source run queue has a priority status
    change, the CPU spent minutes! processing the IPIs.
    
    Thinking about this further, I realized there's no reason for each run queue
    to send its own IPI. As all CPUs with overloaded tasks must be scanned
    regardless if there's one or many CPUs lowering their priority, because
    there's no current way to find the CPU with the highest priority task that
    can schedule to one of these CPUs, there really only needs to be one IPI
    being sent around at a time.
    
    This greatly simplifies the code!
    
    The new approach is to have each root domain have its own irq work, as the
    rto_mask is per root domain. The root domain has the following fields
    attached to it:
    
      rto_push_work	 - the irq work to process each CPU set in rto_mask
      rto_lock	 - the lock to protect some of the other rto fields
      rto_loop_start - an atomic that keeps contention down on rto_lock
    		    the first CPU scheduling in a lower priority task
    		    is the one to kick off the process.
      rto_loop_next	 - an atomic that gets incremented for each CPU that
    		    schedules in a lower priority task.
      rto_loop	 - a variable protected by rto_lock that is used to
    		    compare against rto_loop_next
      rto_cpu	 - The cpu to send the next IPI to, also protected by
    		    the rto_lock.
    
    When a CPU schedules in a lower priority task and wants to make sure
    overloaded CPUs know about it. It increments the rto_loop_next. Then it
    atomically sets rto_loop_start with a cmpxchg. If the old value is not "0",
    then it is done, as another CPU is kicking off the IPI loop. If the old
    value is "0", then it will take the rto_lock to synchronize with a possible
    IPI being sent around to the overloaded CPUs.
    
    If rto_cpu is greater than or equal to nr_cpu_ids, then there's either no
    IPI being sent around, or one is about to finish. Then rto_cpu is set to the
    first CPU in rto_mask and an IPI is sent to that CPU. If there's no CPUs set
    in rto_mask, then there's nothing to be done.
    
    When the CPU receives the IPI, it will first try to push any RT tasks that is
    queued on the CPU but can't run because a higher priority RT task is
    currently running on that CPU.
    
    Then it takes the rto_lock and looks for the next CPU in the rto_mask. If it
    finds one, it simply sends an IPI to that CPU and the process continues.
    
    If there's no more CPUs in the rto_mask, then rto_loop is compared with
    rto_loop_next. If they match, everything is done and the process is over. If
    they do not match, then a CPU scheduled in a lower priority task as the IPI
    was being passed around, and the process needs to start again. The first CPU
    in rto_mask is sent the IPI.
    
    This change removes this duplication of work in the IPI logic, and greatly
    lowers the latency caused by the IPIs. This removed the lockup happening on
    the 120 CPU machine. It also simplifies the code tremendously. What else
    could anyone ask for?
    
    Thanks to Peter Zijlstra for simplifying the rto_loop_start atomic logic and
    supplying me with the rto_start_trylock() and rto_start_unlock() helper
    functions.
    Signed-off-by: default avatarSteven Rostedt (VMware) <rostedt@goodmis.org>
    Signed-off-by: default avatarPeter Zijlstra (Intel) <peterz@infradead.org>
    Cc: Clark Williams <williams@redhat.com>
    Cc: Daniel Bristot de Oliveira <bristot@redhat.com>
    Cc: John Kacur <jkacur@redhat.com>
    Cc: Linus Torvalds <torvalds@linux-foundation.org>
    Cc: Mike Galbraith <efault@gmx.de>
    Cc: Peter Zijlstra <peterz@infradead.org>
    Cc: Scott Wood <swood@redhat.com>
    Cc: Thomas Gleixner <tglx@linutronix.de>
    Link: http://lkml.kernel.org/r/20170424114732.1aac6dc4@gandalf.local.homeSigned-off-by: default avatarIngo Molnar <mingo@kernel.org>
    4bdced5c
rt.c 62.2 KB