Re: [RFC][PATCH] fix swap entries is not reclaimed in proper way for mem+swap controller

[Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>]

On Tue, 21 Apr 2009 16:21:21 +0900, KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> wrote:
> maybe this patch covers almost all cases of swap-leak of memcg, Nishimura-san
> reported. There are many callers of lock_page() but lock_page() which can
> be racy with free_swap_and_cache() is not so much, I think.
> 
> Nishimura-san, How about this ?
> 
Thank you for your patch.

I've run this patch last night but unfortunately I got a BUG.

  BUG: sleeping function called from invalid context at include/linux/pagemap.h:327
  in_atomic(): 1, irqs_disabled(): 0, pid: 9230, name: page01

hmm, calling lock_page() in end_swap_bio_* seems not safe.

And, some comments are inlined.
> ==
> From: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
> 
> free_swap_and_cache(), which is called under various spin_lock,
> is designed to be best-effort function and it does nothing if the page
> is locked, under I/O. But it's ok because global lru will find the page finally.
> 
> But, when it comes to memory+swap cgroup, global LRU may not work and
> swap entry can be alive as "not used but not freed" state for very very long time
> because memory cgroup's LRU routine scans its own LRU only.
> (Such stale swap-cache is out of memcg's LRU)
> 
> Nishimura repoted such kind of swap cache makes memcg unstable and
>  - we can never free mem_cgroup object (....it's big) even if no users.
>  - OOM killer can happen because amounts of charge-to-swap is leaked.
>
And 
 
- All the swap space(swap entries) could be exhausted by these swap cache.

> This patch tries to fix the problem by adding PageCgroupStale() flag.
> If a page which is under zappped is busy swap cache, recored it as Stale.
> At the end of swap I/O, Stale flag is checked and swap and swapcache is
> freed if necessary. Page migration case is also checked.
> 
> 
> Reported-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
> Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
> 
> ---
>  include/linux/page_cgroup.h |    4 +++
>  include/linux/swap.h        |    8 ++++++
>  mm/memcontrol.c             |   51 ++++++++++++++++++++++++++++++++++++++++----
>  mm/page_io.c                |    2 +
>  mm/swapfile.c               |    1 
>  5 files changed, 62 insertions(+), 4 deletions(-)
> 
> Index: linux-2.6.30-rc2/include/linux/page_cgroup.h
> ===================================================================
> --- linux-2.6.30-rc2.orig/include/linux/page_cgroup.h
> +++ linux-2.6.30-rc2/include/linux/page_cgroup.h
> @@ -26,6 +26,7 @@ enum {
>  	PCG_LOCK,  /* page cgroup is locked */
>  	PCG_CACHE, /* charged as cache */
>  	PCG_USED, /* this object is in use. */
> +	PCG_STALE, /* may be a stale swap-cache */
>  };
>  
>  #define TESTPCGFLAG(uname, lname)			\
> @@ -46,6 +47,9 @@ TESTPCGFLAG(Cache, CACHE)
>  TESTPCGFLAG(Used, USED)
>  CLEARPCGFLAG(Used, USED)
>  
> +TESTPCGFLAG(Stale, STALE)
> +SETPCGFLAG(Stale, STALE)
> +
>  static inline int page_cgroup_nid(struct page_cgroup *pc)
>  {
>  	return page_to_nid(pc->page);
> Index: linux-2.6.30-rc2/include/linux/swap.h
> ===================================================================
> --- linux-2.6.30-rc2.orig/include/linux/swap.h
> +++ linux-2.6.30-rc2/include/linux/swap.h
> @@ -344,10 +344,18 @@ mem_cgroup_uncharge_swapcache(struct pag
>  #endif
>  #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
>  extern void mem_cgroup_uncharge_swap(swp_entry_t ent);
> +extern void mem_cgroup_mark_swapcache_stale(struct page *page);
> +extern void mem_cgroup_fixup_swapcache(struct page *page);
>  #else
>  static inline void mem_cgroup_uncharge_swap(swp_entry_t ent)
>  {
>  }
> +static void mem_cgroup_check_mark_swapcache_stale(struct page *page)
> +{
> +}
> +static void mem_cgroup_fixup_swapcache(struct page *page)
> +{
> +}
>  #endif
>  
I think they should be defined in MEM_RES_CTLR case.
Exhausting swap entries problem is not depend on MEM_RES_CTLR_SWAP.

>  #else /* CONFIG_SWAP */
> Index: linux-2.6.30-rc2/mm/memcontrol.c
> ===================================================================
> --- linux-2.6.30-rc2.orig/mm/memcontrol.c
> +++ linux-2.6.30-rc2/mm/memcontrol.c
> @@ -1534,6 +1534,47 @@ void mem_cgroup_uncharge_swap(swp_entry_
>  	}
>  	rcu_read_unlock();
>  }
> +
> +/*
> + * free_swap_and_cache() is an best-effort function and it doesn't free
> + * swapent if the swapcache seems to be busy (ex. the page is locked.)
> + * This behavior is designed to be as is but mem+swap cgroup has to handle it.
> + * Otherwise, swp_entry seems to be leaked (for very long time)
> + */
> +void mem_cgroup_mark_swapcache_stale(struct page *page)
> +{
> +	struct page_cgroup *pc;
> +
> +	if (!PageSwapCache(page) || page_mapped(page))
> +		return;
> +
> +	pc = lookup_page_cgroup(page);
> +	lock_page_cgroup(pc);
> +
> +	/*
> +	 * This "Stale" flag will be cleared when the page is reused
> +	 * somewhere.
> +	 */
> +	if (!PageCgroupUsed(pc))
> +		SetPageCgroupStale(pc);
> +	unlock_page_cgroup(pc);
> +}
> +
> +void mem_cgroup_fixup_swapcache(struct page *page)
> +{
> +	struct page_cgroup *pc = lookup_page_cgroup(page);
> +
> +	/* Stale flag will be cleared automatically */
> +	if (unlikely(PageCgroupStale(pc))) {
> +		if (get_page_unless_zero(page)) {
> +			lock_page(page);
> +			try_to_free_swap(page);
> +			unlock_page(page);
> +			page_cache_release(page);
> +		}
> +	}
> +}
> +
>  #endif
>  
>  /*
> @@ -1604,17 +1645,19 @@ void mem_cgroup_end_migration(struct mem
>  	__mem_cgroup_commit_charge(mem, pc, ctype);
>  
>  	/*
> -	 * Both of oldpage and newpage are still under lock_page().
> -	 * Then, we don't have to care about race in radix-tree.
> -	 * But we have to be careful that this page is unmapped or not.
> +	 * oldpage is under lock_page() at migration. Then, we don't have to
> +	 * care about race in radix-tree. But we have to be careful
> +	 * that this page is unmapped or not.
>  	 *
>  	 * There is a case for !page_mapped(). At the start of
>  	 * migration, oldpage was mapped. But now, it's zapped.
>  	 * But we know *target* page is not freed/reused under us.
>  	 * mem_cgroup_uncharge_page() does all necessary checks.
>  	 */
> -	if (ctype == MEM_CGROUP_CHARGE_TYPE_MAPPED)
> +	if (ctype == MEM_CGROUP_CHARGE_TYPE_MAPPED) {
>  		mem_cgroup_uncharge_page(target);
> +		mem_cgroup_fixup_swapcache(target);
> +	}
>  }
>  
>  /*
hmm, the intention of the patch I posted(*) is not free stale SwapCache.

* http://marc.info/?l=linux-mm&m=124029196025611&w=2

If the oldpage was page_mapped() and !PageSwapCache, newpage would be
uncharge by mem_cgroup_uncharge_page() if the owner process has been exited
(because newpage is !page_mapped()).
But if the oldpage was page_mapped and PageSwapCache, newpage cannot be
uncharged by mem_cgroup_uncharge_page() even if the owner process has been exited.

Those SwapCache is put_back'ed to memcg's LRU, so it can be reclaimed
if memcg's LRU scaning(Anon) run, but my intention was to fix this behavior
for consistency.

> Index: linux-2.6.30-rc2/mm/page_io.c
> ===================================================================
> --- linux-2.6.30-rc2.orig/mm/page_io.c
> +++ linux-2.6.30-rc2/mm/page_io.c
> @@ -68,6 +68,7 @@ static void end_swap_bio_write(struct bi
>  	}
>  	end_page_writeback(page);
>  	bio_put(bio);
> +	mem_cgroup_fixup_swapcache(page);
>  }
>  
>  void end_swap_bio_read(struct bio *bio, int err)
> @@ -87,6 +88,7 @@ void end_swap_bio_read(struct bio *bio, 
>  	}
>  	unlock_page(page);
>  	bio_put(bio);
> +	mem_cgroup_fixup_swapcache(page);
>  }
>  
>  /*
> Index: linux-2.6.30-rc2/mm/swapfile.c
> ===================================================================
> --- linux-2.6.30-rc2.orig/mm/swapfile.c
> +++ linux-2.6.30-rc2/mm/swapfile.c
> @@ -587,6 +587,7 @@ int free_swap_and_cache(swp_entry_t entr
>  		if (swap_entry_free(p, entry) == 1) {
>  			page = find_get_page(&swapper_space, entry.val);
>  			if (page && !trylock_page(page)) {
> +				mem_cgroup_mark_swapcache_stale(page);
>  				page_cache_release(page);
>  				page = NULL;
>  			}
> 

IIUC, this patch cannot handle(even if it worked as intended) cases like:

            processA                   |           processB
  -------------------------------------+-------------------------------------
    (free_swap_and_cache())            |  (read_swap_cache_async())
                                       |    swap_duplicate()
      swap_entry_free() == 1           |
      find_get_page()                  |
         -> cannot find. so            |
            PCG_STALE isn't set.       |
                                       |    add_to_swap_cache()
                                       |
                                       |    swap_readpage()
                                       |      end_swap_bio_read()
                                       |        mem_cgroup_fixup_swapcache()
                                       |          does nothing becase !PageCgroupStale

  (the page is mapped but not on swapcache)
            processA                   |           processB
  -------------------------------------+-------------------------------------
    (zap_pte_range())                  |  (shrink_page_list())
                                       |  (referenced flag is set in some reason)
      page_remove_rmap()               |
        -> uncharged(!PageSwapCache)   |
                                       |    add_to_swap()
                                       |      -> succeed
                                       |
                                       |  (not freed because referenced flag is set)

So, we should add check to shrink_page_list() anyway, IMHO.


Hmm... I tested a patch like below, but it seems to have a dead lock about swap_lock.
(This patch has unhandled corner cases yet, even if it worked.)

I'll dig and try more including another aproach..

---
diff --git a/include/linux/swap.h b/include/linux/swap.h
index 62d8143..991dd53 100644
--- a/include/linux/swap.h
+++ b/include/linux/swap.h
@@ -336,11 +336,16 @@ static inline void disable_swap_token(void)
 
 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
 extern void mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent);
+extern void mem_cgroup_fixup_swapcache(struct page *page);
 #else
 static inline void
 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent)
 {
 }
+static inline void
+mem_cgroup_fixup_swapcache(struct page *page)
+{
+}
 #endif
 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
 extern void mem_cgroup_uncharge_swap(swp_entry_t ent);
diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index ba07cab..b2a4d52 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -1568,6 +1568,19 @@ void mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent)
 		css_put(&memcg->css);
 }
 
+void mem_cgroup_fixup_swapcache(struct page *page)
+{
+	if (mem_cgroup_disabled())
+		return;
+
+	VM_BUG_ON(!PageLocked(page));
+
+	if (get_page_unless_zero(page)) {
+		try_to_free_swap(page);
+		page_cache_release(page);
+	}
+}
+
 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
 /*
  * called from swap_entry_free(). remove record in swap_cgroup and
diff --git a/mm/page_io.c b/mm/page_io.c
index 3023c47..2bafcd3 100644
--- a/mm/page_io.c
+++ b/mm/page_io.c
@@ -85,6 +85,7 @@ void end_swap_bio_read(struct bio *bio, int err)
 	} else {
 		SetPageUptodate(page);
 	}
+	mem_cgroup_fixup_swapcache(page);
 	unlock_page(page);
 	bio_put(bio);
 }