Re: [PATCH] memcg: handle accounting race in swapin-readahead andzap_pte

[KAMEZAWA Hiroyuki]

Daisuke Nishimura wrote:
> On Fri, 15 May 2009 19:00:27 +0900, KAMEZAWA Hiroyuki
>-------------------------------------+-------------------------------------
>                                        |  trylock_page()
>                                        |    try_to_free_swap()
>                                        |      page_swapcount() -> true &
> return
>      swap_info_get()                   |
>        swap_entry_free() == 1          |
>        find_get_page() -> found        |
>        trylock_page() -> fail & return |
>                                        |    unlock_page()
>
> I don't think it happens in practice(unlock_page() would be called soon
> after
> try_to_free_swap() returns), and this patch seems to work well actually.
> I'm not sure whether we should handle this case more strictly or not, but
> I think
> it it would be better to add some comments about it at least.
>
Hmm, ok. maybe trylock in free_swap_and_cache() is the worst thing as
Andrew pointed out...

> And I have a question.
>
> If the size of swap device(or the number of used swap entries not on
> SwapCache)
> is small enough not to hit "if (memcg_swapin_buffer.nr >
> ENOUGH_LARGE_SWAPIN_BUFFER)"
> in mem_cgroup_add_swapin_buffer(), those pages in swapin buffer
> are left and unfreed by swapoff(although swap entries are freed) ?
> Isn't it better to call directly mem_cgroup_drain_swapin_buffer() at the
> end of swapoff ?
>
Hmm, maybe necessary.

> I prefer your v4(remembering only stale swap entries) to be honest,
> but I don't oppose strongly to this direction.
>
I can't believe I can handle complex race with "rememebering only stale".
I'll try to remove trylock in free_swap_and_cache...

Thank you for testing.
-Kmae

--
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/