git: kernel - Remove bottlenecks related to mass pipe closures

[Matthew Dillon]

commit 5eab490ea55e4eabb73d3ecb5cd179f745d78196
Author: Matthew Dillon <dillon at apollo.backplane.com>
Date:   Sun Aug 13 19:48:28 2017 -0700

    kernel - Remove bottlenecks related to mass pipe closures
    
    * These changes fix teardown latencies involved with destroying a large
      number of pipe()s at the same time, for example when a large number
      of processes are killed simultaneously.  Testing with 900,000 processes
      chaining 900,000 pipe()s, the system could become unusable for upwards
      of an hour when the test is killed with ^C prior to these changes.
    
      After these changes the system balks for less than 30 seconds in the
      900,000 process case.
    
    * Wrap pipe teardowns in a pcpu mutex.  Such teardowns only occur when the
      normal pcpu pipe cache / hysteresis gets blown out due to a large number
      of closures occurring simultaneously and are not typically in the
      critical path.
    
      The pcpu mutex allows good concurrency while also limiting the number of
      processes which can be tearing down pipe KVM at the same time.  If
      we have say 900,000 processes all exiting, this mechanism limits the
      teardown to (ncpus) concurrent processes.
    
      The mutex uses a queued lock (whereas the lockmgr spams wakeup()s),
      so we use a mutex to avoid a degenerate O(N^2) wakeup situation.
      This removes an enormous amount of overhead during the teardown.
    
    * When removing areas from the kernel_map, issue a pmap_remove() call
      on the kernel_pmap before cleaning up the underlying VM object just as
      we do for other types of maps.  This was originally not done because the
      kernel_pmap can be extremely sparse and iteration could be inefficient.
    
      However, the current pmap implementation handles sparse pmaps just
      fine.  Removing the pages in bulk reduces the number of global IPIs
      that wind up being issued during the removal, greatly improving
      performance when a large volume of kernel resource is being destroyed
      all at once.
    
    * Increase the zalloc burst from 4 pages to 32 pages.  When adding pages,
      use more relaxed vm_page_alloc() flags after the 4th page to try to
      avoid blowing up the emergency free page cache.

Summary of changes:
 sys/kern/sys_pipe.c | 32 ++++++++++++++++++++++++++++++--
 sys/vm/vm_map.c     |  7 +++++++
 sys/vm/vm_zone.c    | 21 +++++++++++++++------
 sys/vm/vm_zone.h    |  2 +-
 4 files changed, 53 insertions(+), 9 deletions(-)

http://gitweb.dragonflybsd.org/dragonfly.git/commitdiff/5eab490ea55e4eabb73d3ecb5cd179f745d78196


-- 
DragonFly BSD source repository