lkml.org 
[lkml]   [2012]   [Mar]   [28]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
    Patch in this message
    /
    Date
    From
    Subject[RFC][PATCH 3/6] memcg: add PageCgroupReset()

    A commit "memcg: simplify LRU handling by new rule" removes PCG_ACCT_LRU.
    and the bug introduced by it was fixed by "memcg: fix GPF when cgroup removal
    races with last exit"

    This was for reducing flags on pc->flags....Now, we have 3bits of flags.
    but this patch adds a new flag, I'm sorry. (Considering alignment of
    kmalloc(), we'll able to have 5 bits..)

    This patch adds PCG_RESET which is similar to PCG_ACCT_LRU. This is set
    when mem_cgroup_add_lru_list() finds we cannot trust the pc's mem_cgroup.

    The reason why this patch adds a (renamed) flag again is for merging
    pc->flags and pc->mem_cgroup. Assume pc's mem_cgroup is encoded as

    mem_cgroup = pc->flags & ~0x7

    Updating multiple bits of pc->flags without talking lock_page_cgroup()
    is very dangerous. And mem_cgroup_add_lru_list() updates pc->mem_cgroup
    without taking lock. Then I add RESET bit. After this, pc_to_mem_cgroup()
    is written as

    if (PageCgroupReset(pc))
    return root_mem_cgroup;
    return pc->mem_cgroup;

    This update of Reset bit can be done in atomic by set_bit(). And
    cleared when USED bit is set.

    Considering kmalloc()'s alignment, having 4bits of flags will be ok....

    Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
    ---
    include/linux/page_cgroup.h | 15 ++++++++-------
    mm/memcontrol.c | 5 +++--
    2 files changed, 11 insertions(+), 9 deletions(-)

    diff --git a/include/linux/page_cgroup.h b/include/linux/page_cgroup.h
    index 2707809..3f3b4ff 100644
    --- a/include/linux/page_cgroup.h
    +++ b/include/linux/page_cgroup.h
    @@ -8,6 +8,7 @@ enum {
    PCG_LOCK, /* Lock for pc->mem_cgroup and following bits. */
    PCG_USED, /* this object is in use. */
    PCG_MIGRATION, /* under page migration */
    + PCG_RESET, /* have been reset to root_mem_cgroup */
    __NR_PCG_FLAGS,
    };

    @@ -70,6 +71,9 @@ SETPCGFLAG(Migration, MIGRATION)
    CLEARPCGFLAG(Migration, MIGRATION)
    TESTPCGFLAG(Migration, MIGRATION)

    +TESTPCGFLAG(Reset, RESET)
    +SETPCGFLAG(Reset, RESET)
    +
    static inline void lock_page_cgroup(struct page_cgroup *pc)
    {
    /*
    @@ -84,16 +88,13 @@ static inline void unlock_page_cgroup(struct page_cgroup *pc)
    bit_spin_unlock(PCG_LOCK, &pc->flags);
    }

    +extern struct mem_cgroup* root_mem_cgroup;

    static inline struct mem_cgroup* pc_to_mem_cgroup(struct page_cgroup *pc)
    {
    - return pc->mem_cgroup;
    -}
    -
    -static inline void
    -pc_set_mem_cgroup(struct page_cgroup *pc, struct mem_cgroup *memcg)
    -{
    - pc->mem_cgroup = memcg;
    + if (likely(!PageCgroupReset(pc)))
    + return pc->mem_cgroup;
    + return root_mem_cgroup;
    }

    static inline void
    diff --git a/mm/memcontrol.c b/mm/memcontrol.c
    index d366b60..622fd2e 100644
    --- a/mm/memcontrol.c
    +++ b/mm/memcontrol.c
    @@ -1080,7 +1080,8 @@ struct lruvec *mem_cgroup_lru_add_list(struct zone *zone, struct page *page,
    * of pc's mem_cgroup safe.
    */
    if (!PageCgroupUsed(pc) && memcg != root_mem_cgroup) {
    - pc_set_mem_cgroup(pc, root_mem_cgroup);
    + /* this reset bit is cleared when the page is charged */
    + SetPageCgroupReset(pc);
    memcg = root_mem_cgroup;
    }

    @@ -2626,7 +2627,7 @@ static int mem_cgroup_move_account(struct page *page,
    __mem_cgroup_cancel_charge(from, nr_pages);

    /* caller should have done css_get */
    - pc_set_mem_cgroup(pc, to);
    + pc_set_mem_cgroup_and_flags(pc, to, BIT(PCG_USED) | BIT(PCG_LOCK));
    mem_cgroup_charge_statistics(to, anon, nr_pages);
    /*
    * We charges against "to" which may not have any tasks. Then, "to"
    --
    1.7.4.1



    \
     
     \ /
      Last update: 2012-03-28 13:01    [W:0.025 / U:90.224 seconds]
    ©2003-2016 Jasper Spaans. hosted at Digital OceanAdvertise on this site