Re: bad rss-counter message in 3.14rc5

From: Hugh Dickins
Date: Tue Mar 18 2014 - 20:40:18 EST


On Tue, 11 Mar 2014, Dave Jones wrote:
> On Tue, Mar 11, 2014 at 09:36:03PM +0400, Cyrill Gorcunov wrote:
> > On Tue, Mar 11, 2014 at 01:10:45PM -0400, Dave Jones wrote:
> > > >
> > > > Dave, iirc trinity can write log file pointing which exactly syscall sequence
> > > > was passed, right? Share it too please.
> > >
> > > Hm, I may have been mistaken, and the damage was done by a previous run.
> > > I went from being able to reproduce it almost instantly to now not being able
> > > to reproduce it at all. Will keep trying.
> >
> > Sasha already gave a link to the syscalls sequence, so no rush.
>
> It'd be nice to get a more concise reproducer, his list had a little of everything in there.

I've so far failed to find any explanation for your swapops.h BUG;
but believe I have identified one cause for "Bad rss-counter"s.

My hunch is that the swapops.h BUG is "nearby", but I just cannot
fit it together (the swapops.h BUG comes when rmap cannot find all
all the migration entries it inserted earlier: it's a very useful
BUG for validating rmap).

Untested patch below: I can't quite say Reported-by, because it may
not even be one that you and Sasha have been seeing; but I'm hopeful,
remap_file_pages is in the list.

Please give this a try, preferably on 3.14-rc or earlier: I've never
seen "Bad rss-counter"s there myself (trinity uses remap_file_pages
a lot more than most of us); but have seen them on mmotm/next, so
some other trigger is coming up there, I'll worry about that once
it reaches 3.15-rc.

(Cyrill, entirely unrelated, but in preparing this patch I noticed
your soft_dirty work in install_file_pte(): which looked good at
first, until I realized that it's propagating the soft_dirty of a
pte it's about to zap completely, to the unrelated entry it's about
to insert in its place. Which seems very odd to me.)


[PATCH] mm: fix bad rss-counter if remap_file_pages raced migration

Fix some "Bad rss-counter state" reports on exit, arising from the
interaction between page migration and remap_file_pages(): zap_pte()
must count a migration entry when zapping it.

And yes, it is possible (though very unusual) to find an anon page or
swap entry in a VM_SHARED nonlinear mapping: coming from that horrid
get_user_pages(write, force) case which COWs even in a shared mapping.

Signed-off-by: Hugh Dickins <hughd@xxxxxxxxxx>
---

mm/fremap.c | 28 ++++++++++++++++++++++------
1 file changed, 22 insertions(+), 6 deletions(-)

--- 3.14-rc7/mm/fremap.c 2014-01-19 18:40:07.000000000 -0800
+++ linux/mm/fremap.c 2014-03-18 16:32:39.288612346 -0700
@@ -23,28 +23,44 @@

#include "internal.h"

+static int mm_counter(struct page *page)
+{
+ return PageAnon(page) ? MM_ANONPAGES : MM_FILEPAGES;
+}
+
static void zap_pte(struct mm_struct *mm, struct vm_area_struct *vma,
unsigned long addr, pte_t *ptep)
{
pte_t pte = *ptep;
+ struct page *page;
+ swp_entry_t entry;

if (pte_present(pte)) {
- struct page *page;
-
flush_cache_page(vma, addr, pte_pfn(pte));
pte = ptep_clear_flush(vma, addr, ptep);
page = vm_normal_page(vma, addr, pte);
if (page) {
if (pte_dirty(pte))
set_page_dirty(page);
+ update_hiwater_rss(mm);
+ dec_mm_counter(mm, mm_counter(page));
page_remove_rmap(page);
page_cache_release(page);
+ }
+ } else { /* zap_pte() is not called when pte_none() */
+ if (!pte_file(pte)) {
update_hiwater_rss(mm);
- dec_mm_counter(mm, MM_FILEPAGES);
+ entry = pte_to_swp_entry(pte);
+ if (non_swap_entry(entry)) {
+ if (is_migration_entry(entry)) {
+ page = migration_entry_to_page(entry);
+ dec_mm_counter(mm, mm_counter(page));
+ }
+ } else {
+ free_swap_and_cache(entry);
+ dec_mm_counter(mm, MM_SWAPENTS);
+ }
}
- } else {
- if (!pte_file(pte))
- free_swap_and_cache(pte_to_swp_entry(pte));
pte_clear_not_present_full(mm, addr, ptep, 0);
}
}
--
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/