cvs commit: src/sys/kern lwkt_ipiq.c

[Matthew Dillon]

dillon      2005/07/23 00:17:42 PDT

DragonFly src repository

  Modified files:
    sys/kern             lwkt_ipiq.c 
  Log:
  Fix a critical bug in the IPI messaging code, effecting SMP systems.  In
  order to avoid placing a load fence in the FIFO scanning loop the FIFO write
  index is cached and the code then loops the read index until it reaches
  the cached write index, rather then real-time write index.  However, if a
  FIFO full condition occurs during the callback AND additional IPIQ messages
  are queued to the current cpu by a remote cpu at the same time, a recursive
  call to lwkt_process_ipiq*() can occur and advance the read index past
  the cached write index of the parent processing loop.
  
  An exact comparison against the cached write index was being used which
  resulted in the parent processing loop blowing past the actual write
  index and re-executing stale IPI messages.  Fix the comparison.
  
  The nature of this bug combined with other bugs in the token code and the
  sockbuf code (which were causing crashes far more often) made this a
  particularly nasty problem to find, with it taking upwards of a week to
  generate a crash and the crash occuring at the worst place imagineable
  (a hard IPI interrupt) and doing terrible things (re-executing a stale
  IPI message).  It took KTR logging on both the sending and receiving
  side of the IPI code to nail the problem.
  
  Very special thanks to Peter Avalos and David rhodus for their debugging
  help.  And, most especially, David Rhodus for helping track this down over
  the last *THREE* months.
  
  Reported-by: David Rhodus, Peter Avalos, YONETANI Tomokazu, Tomaz Borstnar
  Special-thanks-to: David Rhodus and Peter Avalos
  
  Revision  Changes    Path
  1.15      +5 -1      src/sys/kern/lwkt_ipiq.c


http://www.dragonflybsd.org/cvsweb/src/sys/kern/lwkt_ipiq.c.diff?r1=1.14&r2=1.15&f=u