lkml.org 
[lkml]   [2012]   [Feb]   [25]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
    /
    Date
    From
    SubjectRe: [PATCH v3 00/21] mm: lru_lock splitting
    KAMEZAWA Hiroyuki wrote:
    > On Thu, 23 Feb 2012 17:51:36 +0400
    > Konstantin Khlebnikov<khlebnikov@openvz.org> wrote:
    >
    >> v3 changes:
    >> * inactive-ratio reworked again, now it always calculated from from scratch
    >> * hierarchical pte reference bits filter in memory-cgroup reclaimer
    >> * fixed two bugs in locking, found by Hugh Dickins
    >> * locking functions slightly simplified
    >> * new patch for isolated pages accounting
    >> * new patch with lru interleaving
    >>
    >> This patchset is based on next-20120210
    >>
    >> git: https://github.com/koct9i/linux/commits/lruvec-v3
    >>
    >
    > I wonder.... I just wonder...if we can split a lruvec in a zone into small
    > pieces of lruvec and have splitted LRU-lock per them, do we need per-memcg-lrulock ?

    What per-memcg-lrulock? I don't have it.
    last patch splits lruvecs in memcg with the same factor.

    >
    > It seems per-memcg-lrulock can be much bigger lock than small-lruvec-lock.
    > (depends on configuraton) and much more complicated..and have to take care
    > of many things.. If unit of splitting can be specified by boot option,
    > it seems admins can split a big memcg's per-memcg-lru lock into more small pieces.

    lruvec count per memcg can be arbitrary and changeable if cgroup is empty.
    This is not in this patch, but it's really easy.

    >
    > BTW, how to think of default size of splitting ? I wonder splitting lru into
    > the number of cpus per a node can be a choice. Each cpu may have a chance to
    > set prefered-pfn-range at page allocation with additional patches.

    If we rework page to memcg linking and add direct lruvec-id into page->flags,
    we will able to change lruvec before inserting page to lru.
    Thus each cpu will always insert pages into its own lruvec in zone.
    I have not thought about races yet, but this would be perfect solution.

    >
    > Thanks,
    > -Kame
    >
    >
    >> ---
    >>
    >> Konstantin Khlebnikov (21):
    >> memcg: unify inactive_ratio calculation
    >> memcg: make mm_match_cgroup() hirarchical
    >> memcg: fix page_referencies cgroup filter on global reclaim
    >> memcg: use vm_swappiness from target memory cgroup
    >> mm: rename lruvec->lists into lruvec->pages_lru
    >> mm: lruvec linking functions
    >> mm: add lruvec->pages_count
    >> mm: unify inactive_list_is_low()
    >> mm: add lruvec->reclaim_stat
    >> mm: kill struct mem_cgroup_zone
    >> mm: move page-to-lruvec translation upper
    >> mm: push lruvec into update_page_reclaim_stat()
    >> mm: push lruvecs from pagevec_lru_move_fn() to iterator
    >> mm: introduce lruvec locking primitives
    >> mm: handle lruvec relocks on lumpy reclaim
    >> mm: handle lruvec relocks in compaction
    >> mm: handle lruvec relock in memory controller
    >> mm: add to lruvec isolated pages counters
    >> memcg: check lru vectors emptiness in pre-destroy
    >> mm: split zone->lru_lock
    >> mm: zone lru vectors interleaving
    >>
    >>
    >> include/linux/huge_mm.h | 3
    >> include/linux/memcontrol.h | 75 ------
    >> include/linux/mm.h | 66 +++++
    >> include/linux/mm_inline.h | 19 +-
    >> include/linux/mmzone.h | 39 ++-
    >> include/linux/swap.h | 6
    >> mm/Kconfig | 16 +
    >> mm/compaction.c | 31 +--
    >> mm/huge_memory.c | 14 +
    >> mm/internal.h | 204 +++++++++++++++++
    >> mm/ksm.c | 2
    >> mm/memcontrol.c | 343 +++++++++++-----------------
    >> mm/migrate.c | 2
    >> mm/page_alloc.c | 70 +-----
    >> mm/rmap.c | 2
    >> mm/swap.c | 217 ++++++++++--------
    >> mm/vmscan.c | 534 ++++++++++++++++++++++++--------------------
    >> mm/vmstat.c | 6
    >> 18 files changed, 932 insertions(+), 717 deletions(-)
    >>
    >> --
    >> Signature
    >>
    >



    \
     
     \ /
      Last update: 2012-02-25 06:33    [W:0.038 / U:3.096 seconds]
    ©2003-2016 Jasper Spaans. hosted at Digital OceanAdvertise on this site