• Andrew Morton's avatar
    [PATCH] separate locking for vfsmounts · 91b79ba7
    Andrew Morton authored
    From: Maneesh Soni <maneesh@in.ibm.com>
    
    While path walking we do follow_mount or follow_down which uses
    dcache_lock for serialisation.  vfsmount related operations also use
    dcache_lock for all updates. I think we can use a separate lock for
    vfsmount related work and can improve path walking.
    
    The following two patches does the same. The first one replaces
    dcache_lock with new vfsmount_lock in namespace.c. The lock is
    local to namespace.c and is not required outside. The second patch
    uses RCU to have lock free lookup_mnt(). The patches are quite simple
    and straight forward.
    
    The lockmeter reults show reduced contention, and lock acquisitions
    for dcache_lock while running dcachebench* on a 4-way SMP box
    
        SPINLOCKS         HOLD            WAIT
        UTIL  CON    MEAN(  MAX )   MEAN(  MAX )(% CPU)     TOTAL NOWAIT SPIN RJECT  NAME
    
      baselkm-2569:
        20.7% 20.9%  0.5us( 146us)  2.9us( 144us)(0.81%)  31590840 79.1% 20.9%    0%  dcache_lock
      mntlkm-2569:
        14.3% 13.6%  0.4us( 170us)  2.9us( 187us)(0.42%)  23071746 86.4% 13.6%    0%  dcache_lock
    
    We get more than 8% improvement on 4-way SMP and 44% improvement on 16-way
    NUMAQ while runing dcachebench*.
    
    		Average (usecs/iteration)	Std. Deviation
    		(lower is better)
    4-way SMP
      2.5.69	15739.3				470.90
      2.5.69-mnt	14459.6				298.51
    
    16-way NUMAQ
      2.5.69	120426.5			363.78
      2.5.69-mnt	 63225.8			427.60
    
    *dcachebench is a microbenchmark written by Bill Hartner and is available at
    http://www-124.ibm.com/developerworks/opensource/linuxperf/dcachebench/dcachebench.html
    
     vfsmount_lock.patch
     -------------------
     - Patch for replacing dcache_lock with new vfsmount_lock for all mount
       related operation. This removes the need to take dcache_lock while
       doing follow_mount or follow_down operations in path walking.
    
    I re-ran dcachebench with 2.5.70 as base on 16-way NUMAQ box.
    
                    	Average (usecs/iteration)       Std. Deviation
                    	(lower is better)
    16-way NUMAQ
    2.5.70 				120710.9		 	230.67
     + vfsmount_lock.patch  	65209.6				242.97
        + lookup_mnt-rcu.patch 	64042.3				416.61
    
    So just the lock splitting (vfsmount_lock.patch) gives almost similar benifits
    91b79ba7
namei.c 54.2 KB