Messages in this thread | | | Date | Sat, 13 Dec 2008 19:38:35 +0900 (JST) | Subject | Re: [BUGFIX][PATCH mmotm] memcg fix swap accounting leak (v2) | From | "KAMEZAWA Hiroyuki" <> |
| |
Hugh Dickins said: > On Sat, 13 Dec 2008, KAMEZAWA Hiroyuki wrote: >> --- mmotm-2.6.28-Dec12.orig/mm/memory.c >> +++ mmotm-2.6.28-Dec12/mm/memory.c >> >> - mem_cgroup_commit_charge_swapin(page, ptr); >> inc_mm_counter(mm, anon_rss); >> pte = mk_pte(page, vma->vm_page_prot); >> if (write_access && reuse_swap_page(page)) { >> pte = maybe_mkwrite(pte_mkdirty(pte), vma); >> write_access = 0; >> } >> - >> flush_icache_page(vma, page); >> set_pte_at(mm, address, page_table, pte); >> page_add_anon_rmap(page, vma, address); >> + /* It's better to call commit-charge after rmap is established */ >> + mem_cgroup_commit_charge_swapin(page, ptr); >> >> swap_free(entry); >> if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || >> PageMlocked(page)) > > That ordering is back to how it was before I adjusted it > for reuse_swap_page()'s delete_from_swap_cache(), isn't it? > > So I don't understand how you've fixed the bug I hit (not an > accounting imbalance but an oops or BUG, I forget) with this > ordering, without making some other change elsewhere. > Ah, this is a fix for the new bug by this order. == try_charge() commit_charge() reuse_swap_page() -> delete_from_swapcache() -> uncharge_swapcache(). increase mapcount here. == Because ucharge_swapcache() assumes following a. if mapcount==0, this swap cache is of no use and will be discarded. b. if mapcount >0, this swap cache is in use. A charge commited by commit_charge() is discarded by reuse_swap_page().
By delaying commit (means checking flag of page_cgroup). == try_charge() reuse_swap_page() commit_charge() == the leak of charge doesn't happen. (reuse_swap_page() may drop page from swap-cache, but it's no probelm to commit. But as you say, this has swp_entry==0 bug.)
> mem_cgroup_commit_charge_swapin calls swap_cgroup_record with > bogus swp_entry_t 0, which appears to belong to swp_offset 0 of > swp_type 0, but the ctrl->map for type 0 may have been freed > ages ago (we do always start from 0, but maybe we swapped on > type 1 and swapped off type 0 meanwhile). I'm guessing that > by looking at the code, not by retesting it, so I may have the > details wrong; but I didn't reorder your code just for fun. > > Perhaps your restored ordering works if you check PageSwapCache > in mem_cgroup_commit_charge_swapin or check 0 in swap_cgroup_record, > but I don't see that in yesterday's mmotm, nor in this patch. > Ahhhh, sorry. ok, swp_entry==0 is valid...Sigh... I'll revisit this and check how commit_charge() works. I think checking PageSwapCache() is enough but if not, do somehing other. (Maybe Nishimura's suggestion to pass swp_entry directly to commit_charge() is one way.)
> (And I should admit, I've not even attempted to follow your > accounting justification: I'll leave that to you memcg guys.) > Sorry for complication ;(
> An alternative could be not to clear page->private when deleting > from swap cache, that's only done for tidiness and to force notice > of races like this; but I'd want a much stronger reason to change that. > It seems that it will add another complex or unexpected behavior.. I think I can do something workaround.
> Or am I making this up? As I say, I've not tested it this time around. > I'll ask you if I found I can't do anything ;(
Thank you for pointing out! I'll revisit this on Monday.
-Kame
| |