lkml.org 
[lkml]   [2010]   [Nov]   [27]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
    /
    Date
    From
    SubjectRe: [thisops uV2 02/10] vmstat: Optimize zone counter modifications through the use of this cpu operations
    * Christoph Lameter (cl@linux.com) wrote:
    > this cpu operations can be used to slightly optimize the function. The
    > changes will avoid some address calculations and replace them with the
    > use of the percpu segment register.
    >
    > If one would have this_cpu_inc_return and this_cpu_dec_return then it
    > would be possible to optimize inc_zone_page_state and dec_zone_page_state even
    > more.

    Then we might want to directly target the implementation with
    this_cpu_add_return/this_cpu_sub_return (you implement these in patch 03), which
    would not need to disable preemption on the fast path. I think we already
    discussed this in the past. The reason eludes me at the moment, but I remember
    discussing that changing the increment/decrement delta to the nearest powers of
    two would let us deal with overflow cleanly. But it's probably too early in the
    morning for me to wrap my head around the issue at the moment.

    Thanks,

    Mathieu

    >
    > V1->V2:
    > - Fix __dec_zone_state overflow handling
    > - Use s8 variables for temporary storage.
    >
    > Signed-off-by: Christoph Lameter <cl@linux.com>
    >
    > ---
    > mm/vmstat.c | 56 ++++++++++++++++++++++++++++++++------------------------
    > 1 file changed, 32 insertions(+), 24 deletions(-)
    >
    > Index: linux-2.6/mm/vmstat.c
    > ===================================================================
    > --- linux-2.6.orig/mm/vmstat.c 2010-11-24 13:38:34.000000000 -0600
    > +++ linux-2.6/mm/vmstat.c 2010-11-24 15:03:08.000000000 -0600
    > @@ -167,18 +167,20 @@ static void refresh_zone_stat_thresholds
    > void __mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
    > int delta)
    > {
    > - struct per_cpu_pageset *pcp = this_cpu_ptr(zone->pageset);
    > -
    > - s8 *p = pcp->vm_stat_diff + item;
    > + struct per_cpu_pageset * __percpu pcp = zone->pageset;
    > + s8 * __percpu p = pcp->vm_stat_diff + item;
    > long x;
    > + long t;
    > +
    > + x = delta + __this_cpu_read(*p);
    >
    > - x = delta + *p;
    > + t = __this_cpu_read(pcp->stat_threshold);
    >
    > - if (unlikely(x > pcp->stat_threshold || x < -pcp->stat_threshold)) {
    > + if (unlikely(x > t || x < -t)) {
    > zone_page_state_add(x, zone, item);
    > x = 0;
    > }
    > - *p = x;
    > + __this_cpu_write(*p, x);
    > }
    > EXPORT_SYMBOL(__mod_zone_page_state);
    >
    > @@ -221,16 +223,19 @@ EXPORT_SYMBOL(mod_zone_page_state);
    > */
    > void __inc_zone_state(struct zone *zone, enum zone_stat_item item)
    > {
    > - struct per_cpu_pageset *pcp = this_cpu_ptr(zone->pageset);
    > - s8 *p = pcp->vm_stat_diff + item;
    > -
    > - (*p)++;
    > + struct per_cpu_pageset * __percpu pcp = zone->pageset;
    > + s8 * __percpu p = pcp->vm_stat_diff + item;
    > + s8 v, t;
    > +
    > + __this_cpu_inc(*p);
    > +
    > + v = __this_cpu_read(*p);
    > + t = __this_cpu_read(pcp->stat_threshold);
    > + if (unlikely(v > t)) {
    > + s8 overstep = t >> 1;
    >
    > - if (unlikely(*p > pcp->stat_threshold)) {
    > - int overstep = pcp->stat_threshold / 2;
    > -
    > - zone_page_state_add(*p + overstep, zone, item);
    > - *p = -overstep;
    > + zone_page_state_add(v + overstep, zone, item);
    > + __this_cpu_write(*p, - overstep);
    > }
    > }
    >
    > @@ -242,16 +247,19 @@ EXPORT_SYMBOL(__inc_zone_page_state);
    >
    > void __dec_zone_state(struct zone *zone, enum zone_stat_item item)
    > {
    > - struct per_cpu_pageset *pcp = this_cpu_ptr(zone->pageset);
    > - s8 *p = pcp->vm_stat_diff + item;
    > -
    > - (*p)--;
    > -
    > - if (unlikely(*p < - pcp->stat_threshold)) {
    > - int overstep = pcp->stat_threshold / 2;
    > + struct per_cpu_pageset * __percpu pcp = zone->pageset;
    > + s8 * __percpu p = pcp->vm_stat_diff + item;
    > + s8 v, t;
    > +
    > + __this_cpu_dec(*p);
    > +
    > + v = __this_cpu_read(*p);
    > + t = __this_cpu_read(pcp->stat_threshold);
    > + if (unlikely(v < - t)) {
    > + s8 overstep = t >> 1;
    >
    > - zone_page_state_add(*p - overstep, zone, item);
    > - *p = overstep;
    > + zone_page_state_add(v - overstep, zone, item);
    > + __this_cpu_write(*p, overstep);
    > }
    > }
    >
    >

    --
    Mathieu Desnoyers
    Operating System Efficiency R&D Consultant
    EfficiOS Inc.
    http://www.efficios.com


    \
     
     \ /
      Last update: 2010-11-27 15:51    [W:4.099 / U:0.060 seconds]
    ©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site