Re: [PATCH 5/7] memcg: add per cgroup dirty pages accounting

From: Kamezawa Hiroyuki
Date: Thu Jul 19 2012 - 02:35:25 EST


(2012/07/11 18:32), Sha Zhengju wrote:
On 07/10/2012 05:02 AM, Greg Thelen wrote:
On Thu, Jun 28 2012, Sha Zhengju wrote:

From: Sha Zhengju<handai.szj@xxxxxxxxxx>

This patch adds memcg routines to count dirty pages, which allows memory controller
to maintain an accurate view of the amount of its dirty memory and can provide some
info for users while group's direct reclaim is working.

After Kame's commit 89c06bd5(memcg: use new logic for page stat accounting), we can
use 'struct page' flag to test page state instead of per page_cgroup flag. But memcg
has a feature to move a page from a cgroup to another one and may have race between
"move" and "page stat accounting". So in order to avoid the race we have designed a
bigger lock:

mem_cgroup_begin_update_page_stat()
modify page information -->(a)
mem_cgroup_update_page_stat() -->(b)
mem_cgroup_end_update_page_stat()

It requires (a) and (b)(dirty pages accounting) can stay close enough.

In the previous two prepare patches, we have reworked the vfs set page dirty routines
and now the interfaces are more explicit:
incrementing (2):
__set_page_dirty
__set_page_dirty_nobuffers
decrementing (2):
clear_page_dirty_for_io
cancel_dirty_page


Signed-off-by: Sha Zhengju<handai.szj@xxxxxxxxxx>
---
fs/buffer.c | 17 ++++++++++++++---
include/linux/memcontrol.h | 1 +
mm/filemap.c | 5 +++++
mm/memcontrol.c | 28 +++++++++++++++++++++-------
mm/page-writeback.c | 30 ++++++++++++++++++++++++------
mm/truncate.c | 6 ++++++
6 files changed, 71 insertions(+), 16 deletions(-)

diff --git a/fs/buffer.c b/fs/buffer.c
index 55522dd..d3714cc 100644
--- a/fs/buffer.c
+++ b/fs/buffer.c
@@ -613,11 +613,19 @@ EXPORT_SYMBOL(mark_buffer_dirty_inode);
int __set_page_dirty(struct page *page,
struct address_space *mapping, int warn)
{
+ bool locked;
+ unsigned long flags;
+ int ret = 0;
'= 0' and 'ret = 0' change (below) are redundant. My vote is to remove
'= 0' here.


Nice catch. :-)

+
if (unlikely(!mapping))
return !TestSetPageDirty(page);

- if (TestSetPageDirty(page))
- return 0;
+ mem_cgroup_begin_update_page_stat(page,&locked,&flags);
+
+ if (TestSetPageDirty(page)) {
+ ret = 0;
+ goto out;
+ }

spin_lock_irq(&mapping->tree_lock);
if (page->mapping) { /* Race with truncate? */
@@ -629,7 +637,10 @@ int __set_page_dirty(struct page *page,
spin_unlock_irq(&mapping->tree_lock);
__mark_inode_dirty(mapping->host, I_DIRTY_PAGES);

- return 1;
+ ret = 1;
+out:
+ mem_cgroup_end_update_page_stat(page,&locked,&flags);
+ return ret;
}

/*
diff --git a/include/linux/memcontrol.h b/include/linux/memcontrol.h
index 20b0f2d..ad37b59 100644
--- a/include/linux/memcontrol.h
+++ b/include/linux/memcontrol.h
@@ -38,6 +38,7 @@ enum mem_cgroup_stat_index {
MEM_CGROUP_STAT_RSS, /* # of pages charged as anon rss */
MEM_CGROUP_STAT_FILE_MAPPED, /* # of pages charged as file rss */
MEM_CGROUP_STAT_SWAP, /* # of pages, swapped out */
+ MEM_CGROUP_STAT_FILE_DIRTY, /* # of dirty pages in page cache */
MEM_CGROUP_STAT_NSTATS,
};

diff --git a/mm/filemap.c b/mm/filemap.c
index 1f19ec3..5159a49 100644
--- a/mm/filemap.c
+++ b/mm/filemap.c
@@ -140,6 +140,11 @@ void __delete_from_page_cache(struct page *page)
* having removed the page entirely.
*/
if (PageDirty(page)&& mapping_cap_account_dirty(mapping)) {
+ /*
+ * Do not change page state, so no need to use mem_cgroup_
+ * {begin, end}_update_page_stat to get lock.
+ */
+ mem_cgroup_dec_page_stat(page, MEM_CGROUP_STAT_FILE_DIRTY);
I do not understand this comment. What serializes this function and
mem_cgroup_move_account()?


The race is exist just because the two competitors share one
public variable and one reads it and the other writes it.
I thought if both sides(accounting and cgroup_move) do not
change page flag, then risks like doule-counting(see below)
will not happen.

CPU-A CPU-B
Set PG_dirty
(delay) move_lock_mem_cgroup()
if (PageDirty(page))
new_memcg->nr_dirty++
pc->mem_cgroup = new_memcg;
move_unlock_mem_cgroup()
move_lock_mem_cgroup()
memcg = pc->mem_cgroup
new_memcg->nr_dirty++


But after second thoughts, it does have problem if without lock:

CPU-A CPU-B
if (PageDirty(page)) {
move_lock_mem_cgroup()
TestClearPageDirty(page))
memcg = pc->mem_cgroup
new_memcg->nr_dirty --
move_unlock_mem_cgroup()

memcg = pc->mem_cgroup
new_memcg->nr_dirty--
}


It may occur race between clear_page_dirty() operation.
So this time I think we need the lock again...

Kame, what about your opinion...

I think Dirty bit is cleared implicitly here...So, having lock will be good.
Thanks,
-Kame



--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/