lkml.org 
[lkml]   [2010]   [Oct]   [18]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [RFC][PATCH 1/2] memcg: avoiding unnecessary get_page at move_charge
On Fri, 15 Oct 2010 17:11:09 +0900
KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> wrote:

> From: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
>
> CC'ed to Mel and Chrisotph, KOSAKi because I wanted to be double-cheked
> that I miss something important at isolate_lru_page().
>
> Checking perofrmance of memcg's move_account with perf, you may notice
> that put_page() is too much.
> #
> # Overhead Command Shared Object Symbol
> # ........ ....... ................. .....................................
> #
> 14.24% echo [kernel.kallsyms] [k] put_page
> 12.80% echo [kernel.kallsyms] [k] isolate_lru_page
> 9.67% echo [kernel.kallsyms] [k] is_target_pte_for_mc
> 8.11% echo [kernel.kallsyms] [k] ____pagevec_lru_add
> 7.22% echo [kernel.kallsyms] [k] putback_lru_page
>
> This is because mc_handle_present_pte() do get_page(). Then,
> page->count is updated 4 times.
> get_page_unless_zero() #1
> isolate_lru_page()
> putback_lru_page()
> put_page()
>
> But above is called all under pte_offset_map_lock().
> get_page_unless_zero() #1 is not necessary because we do all under a
> pte_offset_map_lock().
>
> isolate_lru_page()'s comment says
> # Restrictions:
> # (1) Must be called with an elevated refcount on the page. This is a
> # fundamentnal difference from isolate_lru_pages (which is called
> # without a stable reference).
>
> So, current implemnation does get_page_unless_zero() explicitly but
> holding pte_lock() implies a stable reference. I removed #1.
>
> Then, Performance will be
> [Before Patch]
> [root@bluextal kamezawa]# time echo 2530 > /cgroup/B/tasks
>
> real 0m0.792s
> user 0m0.000s
> sys 0m0.780s
>
> [After Patch]
> [root@bluextal kamezawa]# time echo 2257 > /cgroup/B/tasks
>
> real 0m0.694s
> user 0m0.000s
> sys 0m0.683s
>
Very nice!

Some nitpicks.

> perf's log is
> 10.82% echo [kernel.kallsyms] [k] isolate_lru_page
> 10.01% echo [kernel.kallsyms] [k] mem_cgroup_move_account
> 8.75% echo [kernel.kallsyms] [k] is_target_pte_for_mc
> 8.52% echo [kernel.kallsyms] [k] ____pagevec_lru_add
> 6.90% echo [kernel.kallsyms] [k] putback_lru_page
> 6.36% echo [kernel.kallsyms] [k] mem_cgroup_add_lru_list
> 6.22% echo [kernel.kallsyms] [k] mem_cgroup_del_lru_list
> 5.68% echo [kernel.kallsyms] [k] lookup_page_cgroup
> 5.28% echo [kernel.kallsyms] [k] __lru_cache_add
> 5.00% echo [kernel.kallsyms] [k] release_pages
> 3.79% echo [kernel.kallsyms] [k] _raw_spin_lock_irq
> 3.52% echo [kernel.kallsyms] [k] memcg_check_events
> 3.38% echo [kernel.kallsyms] [k] bit_spin_lock
> 3.25% echo [kernel.kallsyms] [k] put_page
>
> seems nice. I updated isolate_lru_page()'s comment, too.
>
> # Note: isolate_lru_page() is necessary before account move for avoinding
> memcg's LRU manipulation.
>
> Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
> ---
> mm/memcontrol.c | 63 +++++++++++++++++++++++++++++++++++---------------------
> mm/vmscan.c | 3 +-
> 2 files changed, 42 insertions(+), 24 deletions(-)
>
> Index: mmotm-1013/mm/memcontrol.c
> ===================================================================
> --- mmotm-1013.orig/mm/memcontrol.c
> +++ mmotm-1013/mm/memcontrol.c
> @@ -1169,7 +1169,6 @@ static void mem_cgroup_end_move(struct m
> * under hierarchy of moving cgroups. This is for
> * waiting at hith-memory prressure caused by "move".
> */
> -
> static bool mem_cgroup_stealed(struct mem_cgroup *mem)
> {
> VM_BUG_ON(!rcu_read_lock_held());
> @@ -4471,11 +4470,14 @@ one_by_one:
> * Returns
> * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
> * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
> - * move charge. if @target is not NULL, the page is stored in target->page
> - * with extra refcnt got(Callers should handle it).
> + * move charge and it's mapped.. if @target is not NULL, the page is
> + * stored in target->pagewithout extra refcnt.
^^ needs ' '.

> * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
> * target for charge migration. if @target is not NULL, the entry is stored
> * in target->ent.
> + * 3(MC_TARGET_UNMAPPED_PAGE): if the page corresponding to this pte is a
> + * target for move charge. if @target is not NULL, the page is stored in
> + * target->page with extra refcnt got(Callers should handle it).
> *
> * Called with pte lock held.
> */
> @@ -4486,8 +4488,9 @@ union mc_target {
>
> enum mc_target_type {
> MC_TARGET_NONE, /* not used */
> - MC_TARGET_PAGE,
> + MC_TARGET_PAGE, /* a page mapped */
> MC_TARGET_SWAP,
> + MC_TARGET_UNMAPPED_PAGE, /* a page unmapped */
> };
>
I prefer the order of "MC_TARGET_PAGE", "MC_TARGET_UNMAPPED_PAGE", and "MC_TARGET_SWAP".

> static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
> @@ -4504,9 +4507,10 @@ static struct page *mc_handle_present_pt
> } else if (!move_file())
> /* we ignore mapcount for file pages */
> return NULL;
> - if (!get_page_unless_zero(page))
> - return NULL;
> -
> + /*
> + * Because we're under pte_lock and the page is mapped,
> + * get_page() isn't necessary
> + */
> return page;
> }
>
> @@ -4570,14 +4574,18 @@ static int is_target_pte_for_mc(struct v
> struct page *page = NULL;
> struct page_cgroup *pc;
> int ret = 0;
> + bool present = true;
> swp_entry_t ent = { .val = 0 };
>
> if (pte_present(ptent))
> page = mc_handle_present_pte(vma, addr, ptent);
> - else if (is_swap_pte(ptent))
> - page = mc_handle_swap_pte(vma, addr, ptent, &ent);
> - else if (pte_none(ptent) || pte_file(ptent))
> - page = mc_handle_file_pte(vma, addr, ptent, &ent);
> + else {
> + present = false;
> + if (is_swap_pte(ptent))
> + page = mc_handle_swap_pte(vma, addr, ptent, &ent);
> + else if (pte_none(ptent) || pte_file(ptent))
> + page = mc_handle_file_pte(vma, addr, ptent, &ent);
> + }
>
> if (!page && !ent.val)
> return 0;
> @@ -4589,11 +4597,15 @@ static int is_target_pte_for_mc(struct v
> * the lock.
> */
> if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
> - ret = MC_TARGET_PAGE;
> + if (present)
> + ret = MC_TARGET_PAGE;
> + else
> + ret = MC_TARGET_UNMAPPED_PAGE;
> if (target)
> target->page = page;
> }
> - if (!ret || !target)
> + /* We got refcnt but the page is not for target */
> + if (!present && (!ret || !target))
> put_page(page);
> }
> /* There is a swap entry and a page doesn't exist or isn't charged */
> @@ -4780,19 +4792,24 @@ retry:
> type = is_target_pte_for_mc(vma, addr, ptent, &target);
> switch (type) {
> case MC_TARGET_PAGE:
> + case MC_TARGET_UNMAPPED_PAGE:
> page = target.page;
> - if (isolate_lru_page(page))
> - goto put;
> - pc = lookup_page_cgroup(page);
> - if (!mem_cgroup_move_account(pc,
> + if (!isolate_lru_page(page)) {
> + pc = lookup_page_cgroup(page);
> + if (!mem_cgroup_move_account(pc,
> mc.from, mc.to, false)) {
> - mc.precharge--;
> - /* we uncharge from mc.from later. */
> - mc.moved_charge++;
> + mc.precharge--;
> + /* we uncharge from mc.from later. */
> + mc.moved_charge++;
> + }
> + putback_lru_page(page);
> }
> - putback_lru_page(page);
> -put: /* is_target_pte_for_mc() gets the page */
> - put_page(page);
> + /*
> + * Because we holds pte_lock, we have a stable reference * to the page if mapped. If not mapped, we have an

You need a new line :)


Thanks,
Daisuke Nishimura.

> + * elevated refcnt. drop it.
> + */
> + if (type == MC_TARGET_UNMAPPED_PAGE)
> + put_page(page);
> break;
> case MC_TARGET_SWAP:
> ent = target.ent;
> Index: mmotm-1013/mm/vmscan.c
> ===================================================================
> --- mmotm-1013.orig/mm/vmscan.c
> +++ mmotm-1013/mm/vmscan.c
> @@ -1166,7 +1166,8 @@ static unsigned long clear_active_flags(
> * found will be decremented.
> *
> * Restrictions:
> - * (1) Must be called with an elevated refcount on the page. This is a
> + * (1) Must be called with an elevated refcount on the page, IOW, the
> + * caller must guarantee that there is a stable reference. This is a
> * fundamentnal difference from isolate_lru_pages (which is called
> * without a stable reference).
> * (2) the lru_lock must not be held.
>


\
 
 \ /
  Last update: 2010-10-18 06:43    [W:0.106 / U:0.100 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site