git: DragonFly_RELEASE_2_8 kernel - Fix serious issue w/ smp_invltlb(), plus other issues (2)

Matthew Dillon dillon at crater.dragonflybsd.org
Thu Oct 28 14:23:45 PDT 2010


commit dd4ef322d7c93551990f53da73fb477e8db39ee8
Author: Matthew Dillon <dillon at apollo.backplane.com>
Date:   Thu Oct 28 12:36:32 2010 -0700

    kernel - Fix serious issue w/ smp_invltlb(), plus other issues (2)
    
    It looks like there are a number of additional cases where kernel threads
    can wind up running with interrupts physically disabled, which can create
    serious problems for smp_invltlb().
    
    It is not entirely clear how these deadlocks occur since the IPI code does
    a forced "STI" if it would otherwise loop, but there are several other
    hard loops that do not:  lwkt_switch looping on a vm_token, spin locks,
    and probably a few other places.
    
    We want interrupts to be enabled in all cases where these sorts of loops
    occur in order to be able to service Xinvltlb via smp_invltlb() as well
    as to prevent the LAPIC interrupt queue from filling up and livelocking
    something that it shouldn't.
    
    * lwkt_preempt() must save, zero, and restore gd->gd_intr_nesting_level
      when issuing a preemption.  Otherwise it will improperly panic on
      an assertion if the preempting interrupt thread code tries to switch out.
      It is perfectly acceptable for the preempting thread to block (it just
      switches back to the thread that got preempted).
    
      Why the assertion was not occuring before I do not know but it is
      probably related to threads getting stuck in 'cli' mode.  The additional
      changes below appear to significantly increase the number of interrupt
      thread preemptions which succeed (lwkt.preempt_{hit,miss} counters).
    
    * STI prior to running ithread_fast_handler() from all IDTVECs related
      to device interrupts.
    
    * STI in Xcpustop, Xipiq, and Xtimer.  These functions can call more
      complex C code and doing so with interrupts disabled may prevent
      Xinvltlb (via smp_invltlb()) from being executed, deadlocking the
      system.
    
    * Reorder a mfence().  Probably not needed but do it anyway.

Summary of changes:
 sys/kern/lwkt_thread.c                |    8 +++++++
 sys/platform/pc32/apic/apic_vector.s  |    9 +++++++
 sys/platform/pc32/i386/mp_machdep.c   |   38 +++++++++++++++++++++++++++++++++
 sys/platform/pc32/icu/icu_vector.s    |    1 +
 sys/platform/pc64/apic/apic_vector.s  |    9 +++++++
 sys/platform/pc64/icu/icu_vector.s    |    1 +
 sys/platform/pc64/x86_64/mp_machdep.c |   38 ++++++++++++++++++++++++++++----
 7 files changed, 99 insertions(+), 5 deletions(-)

http://gitweb.dragonflybsd.org/dragonfly.git/commitdiff/dd4ef322d7c93551990f53da73fb477e8db39ee8


-- 
DragonFly BSD source repository





More information about the Commits mailing list