[lkml]   [2009]   [May]   [25]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
Subject[RFC][PATCH] memcg: fix swap account (26/May)[0/5]

As Nishimura reported, there is a race at handling swap cache.

Typical cases are following (from Nishimura's mail)

== Type-1 ==
If some pages of processA has been swapped out, it calls free_swap_and_cache().
And if at the same time, processB is calling read_swap_cache_async() about
a swap entry *that is used by processA*, a race like below can happen.

processA | processB
(free_swap_and_cache()) | (read_swap_cache_async())
| swap_duplicate()
| __set_page_locked()
| add_to_swap_cache()
swap_entry_free() == 0 |
find_get_page() -> found |
try_lock_page() -> fail & return |
| lru_cache_add_anon()
| doesn't link this page to memcg's
| LRU, because of !PageCgroupUsed.

This type of leak can be avoided by setting /proc/sys/vm/page-cluster to 0.

== Type-2 ==
Assume processA is exiting and pte points to a page(!PageSwapCache).
And processB is trying reclaim the page.

processA | processB
(page_remove_rmap()) | (shrink_page_list())
mem_cgroup_uncharge_page() |
->uncharged because it's not |
PageSwapCache yet. |
So, both mem/memsw.usage |
are decremented. |
| add_to_swap() -> added to swap cache.

If this page goes thorough without being freed for some reason, this page
doesn't goes back to memcg's LRU because of !PageCgroupUsed.

This patch is a trial for fixing above problems by fixing memcg's swap account logic.
But this requires some amount of changes in swap.

Comaparing with my previous post (22/May)
I think this one is much easier to read...

[1/5] change interface of swap_duplicate()/swap_free()
Adds an function swapcache_prepare() and swapcache_free().

[2/5] add SWAP_HAS_CACHE flag to swap_map
Add SWAP_HAS_CACHE flag to swap_map array for knowing an information that
"there is an only swap cache and swap has no reference"
without calling find_get_page().

[3/5] Count the number of swap-cache-only swaps
After repeating swap-in/out, there are tons of cache-only swaps.
(via a mapped swapcache under vm_swap_full()==false)
This patch counts the number of entry and show it in debug information.
(for example, sysrq-m)

[4/5] fix memcg's swap accounting.
change the memcg's swap accounting logic to see # of references to swap.

[5/5] experimental garbage collection for cache-only swaps.
reclaim swap enty which is not used.

patch [4/5] is for type-1
patch [5/5] is for type-2 and sanity of swaps control...

Thank you for all helps. Any comments are welcome.


 \ /
  Last update: 2009-05-26 05:17    [W:0.122 / U:1.280 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site