| lists.openwall.net | lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC | |
|
Open Source and information security mailing list archives
| ||
|
Message-Id: <20101019134541.455eeaba.kamezawa.hiroyu@jp.fujitsu.com>
Date: Tue, 19 Oct 2010 13:45:41 +0900
From: KAMEZAWA Hiroyuki <kamezawa.hiroyu@...fujitsu.com>
To: KAMEZAWA Hiroyuki <kamezawa.hiroyu@...fujitsu.com>
Cc: Greg Thelen <gthelen@...gle.com>,
Andrew Morton <akpm@...ux-foundation.org>,
linux-kernel@...r.kernel.org, linux-mm@...ck.org,
containers@...ts.osdl.org, Andrea Righi <arighi@...eler.com>,
Balbir Singh <balbir@...ux.vnet.ibm.com>,
Daisuke Nishimura <nishimura@....nes.nec.co.jp>,
Minchan Kim <minchan.kim@...il.com>,
Ciju Rajan K <ciju@...ux.vnet.ibm.com>,
David Rientjes <rientjes@...gle.com>
Subject: [RFC][PATCH 2/2] memcg: move_account optimization by reduce locks
(Re: [PATCH v3 04/11] memcg: add lock to synchronize page accounting and
migration
From: KAMEZAWA Hiroyuki <kamezawa.hiroyu@...fujitsu.com>
reduce lock at account moving.
a patch "memcg: add lock to synchronize page accounting and migration" add
a new lock and make locking cost twice. This patch is for reducing the cost.
At moving charges by scanning page table, we do all jobs under pte_lock.
This means we never have race with "uncharge". Because of that,
we can remove lock_page_cgroup() in some situation.
The cost of moing 8G anon process
==
[mmotm-1013]
Before:
real 0m0.792s
user 0m0.000s
sys 0m0.780s
[dirty-limit v3 patch]
real 0m0.854s
user 0m0.000s
sys 0m0.842s
[get/put optimization ]
real 0m0.757s
user 0m0.000s
sys 0m0.746s
[this patch]
real 0m0.732s
user 0m0.000s
sys 0m0.721s
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@...fujitsu.com>
---
Documentation/cgroups/memory.txt | 23 ++++++++++++++++++++++-
mm/memcontrol.c | 29 ++++++++++++++++++++++-------
2 files changed, 44 insertions(+), 8 deletions(-)
Index: dirty_limit_new/mm/memcontrol.c
===================================================================
--- dirty_limit_new.orig/mm/memcontrol.c
+++ dirty_limit_new/mm/memcontrol.c
@@ -2386,7 +2386,6 @@ static void __mem_cgroup_move_account(st
{
VM_BUG_ON(from == to);
VM_BUG_ON(PageLRU(pc->page));
- VM_BUG_ON(!PageCgroupLocked(pc));
VM_BUG_ON(!PageCgroupUsed(pc));
VM_BUG_ON(pc->mem_cgroup != from);
@@ -2424,19 +2423,32 @@ static void __mem_cgroup_move_account(st
* __mem_cgroup_move_account()
*/
static int mem_cgroup_move_account(struct page_cgroup *pc,
- struct mem_cgroup *from, struct mem_cgroup *to, bool uncharge)
+ struct mem_cgroup *from, struct mem_cgroup *to,
+ bool uncharge, bool stable)
{
int ret = -EINVAL;
unsigned long flags;
-
- lock_page_cgroup(pc);
+ /*
+ * When stable==true, some lock (page_table_lock etc.) prevents
+ * modification of PCG_USED bit and pc->mem_cgroup never be invalid.
+ * IOW, there will be no race with charge/uncharge. From another point
+ * of view, there will be other races with codes which accesses
+ * pc->mem_cgroup under lock_page_cgroup(). Considering what
+ * pc->mem_cgroup the codes will see, they'll see old or new value and
+ * both of values will never be invalid while they holds
+ * lock_page_cgroup(). There is no probelm to skip lock_page_cgroup
+ * when we can.
+ */
+ if (!stable)
+ lock_page_cgroup(pc);
if (PageCgroupUsed(pc) && pc->mem_cgroup == from) {
move_lock_page_cgroup(pc, &flags);
__mem_cgroup_move_account(pc, from, to, uncharge);
move_unlock_page_cgroup(pc, &flags);
ret = 0;
}
- unlock_page_cgroup(pc);
+ if (!stable)
+ unlock_page_cgroup(pc);
/*
* check events
*/
@@ -2474,7 +2486,7 @@ static int mem_cgroup_move_parent(struct
if (ret || !parent)
goto put_back;
- ret = mem_cgroup_move_account(pc, child, parent, true);
+ ret = mem_cgroup_move_account(pc, child, parent, true, false);
if (ret)
mem_cgroup_cancel_charge(parent);
put_back:
@@ -5156,6 +5168,7 @@ retry:
struct page *page;
struct page_cgroup *pc;
swp_entry_t ent;
+ bool mapped = false;
if (!mc.precharge)
break;
@@ -5163,12 +5176,14 @@ retry:
type = is_target_pte_for_mc(vma, addr, ptent, &target);
switch (type) {
case MC_TARGET_PAGE:
+ mapped = true;
+ /* Fall Through */
case MC_TARGET_UNMAPPED_PAGE:
page = target.page;
if (!isolate_lru_page(page)) {
pc = lookup_page_cgroup(page);
if (!mem_cgroup_move_account(pc, mc.from,
- mc.to, false)) {
+ mc.to, false, mapped)) {
mc.precharge--;
/* we uncharge from mc.from later. */
mc.moved_charge++;
Index: dirty_limit_new/Documentation/cgroups/memory.txt
===================================================================
--- dirty_limit_new.orig/Documentation/cgroups/memory.txt
+++ dirty_limit_new/Documentation/cgroups/memory.txt
@@ -637,7 +637,28 @@ memory cgroup.
| page_mapcount(page) > 1). You must enable Swap Extension(see 2.4) to
| enable move of swap charges.
-8.3 TODO
+8.3 Implemenation Detail
+
+ At moving, we need to take care of races. At first thinking, there are
+ several sources of race when we overwrite pc->mem_cgroup.
+ - charge/uncharge
+ - file stat (dirty, writeback, etc..) accounting
+ - LRU add/remove
+
+ Against charge/uncharge, we do all "move" under pte_lock. So, if we move
+ chareges of a mapped pages, we don't need extra locks. If not mapped,
+ we need to take lock_page_cgroup.
+
+ Against file-stat accouning, we need some locks. Current implementation
+ uses 2 level locking, one is light-weight, another is heavy.
+ A light-weight scheme is to use per-cpu counter. If someone moving a charge
+ from a mem_cgroup, per-cpu "caution" counter is incremented and file-stat
+ update will use heavy lock. This heavy lock is a special lock for move_charge
+ and allow mutual execution of accessing pc->mem_cgroup.
+
+ Against LRU, we do isolate_lru_page() before move_account().
+
+8.4 TODO
- Implement madvise(2) to let users decide the vma to be moved or not to be
moved.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
Powered by blists - more mailing lists