Linux Virtualization - Apr 2016 - [PATCH v3 01/16] mm: use put_page to free page instead of putback_lru_page

On Mon, Apr 04, 2016 at 10:39:17AM +0900, Minchan Kim
wrote:
> On Fri, Apr 01, 2016 at 02:58:21PM +0200, Vlastimil Babka wrote:
> > On 03/30/2016 09:12 AM, Minchan Kim wrote:
> > >Procedure of page migration is as follows:
> > >
> > >First of all, it should isolate a page from LRU and try to
> > >migrate the page. If it is successful, it releases the page
> > >for freeing. Otherwise, it should put the page back to LRU
> > >list.
> > >
> > >For LRU pages, we have used putback_lru_page for both freeing
> > >and putback to LRU list. It's okay because put_page is aware
of
> > >LRU list so if it releases last refcount of the page, it removes
> > >the page from LRU list. However, It makes unnecessary operations
> > >(e.g., lru_cache_add, pagevec and flags operations. It would be
> > >not significant but no worth to do) and harder to support new
> > >non-lru page migration because put_page isn't aware of non-lru
> > >page's data structure.
> > >
> > >To solve the problem, we can add new hook in put_page with
> > >PageMovable flags check but it can increase overhead in
> > >hot path and needs new locking scheme to stabilize the flag check
> > >with put_page.
> > >
> > >So, this patch cleans it up to divide two semantic(ie, put and
putback).
> > >If migration is successful, use put_page instead of
putback_lru_page and
> > >use putback_lru_page only on failure. That makes code more
readable
> > >and doesn't add overhead in put_page.
> > >
> > >Comment from Vlastimil
> > >"Yeah, and compaction (perhaps also other migration users)
has to drain
> > >the lru pvec... Getting rid of this stuff is worth even by
itself."
> > >
> > >Cc: Mel Gorman <mgorman at suse.de>
> > >Cc: Hugh Dickins <hughd at google.com>
> > >Cc: Naoya Horiguchi <n-horiguchi at ah.jp.nec.com>
> > >Acked-by: Vlastimil Babka <vbabka at suse.cz>
> > >Signed-off-by: Minchan Kim <minchan at kernel.org>
> > 
> > [...]
> > 
> > >@@ -974,28 +986,28 @@ static ICE_noinline int
unmap_and_move(new_page_t get_new_page,
> > >  		list_del(&page->lru);
> > >  		dec_zone_page_state(page, NR_ISOLATED_ANON +
> > >  				page_is_file_cache(page));
> > >-		/* Soft-offlined page shouldn't go through lru cache list
*/
> > >+	}
> > >+
> > >+	/*
> > >+	 * If migration is successful, drop the reference grabbed during
> > >+	 * isolation. Otherwise, restore the page to LRU list unless we
> > >+	 * want to retry.
> > >+	 */
> > >+	if (rc == MIGRATEPAGE_SUCCESS) {
> > >+		put_page(page);
> > >  		if (reason == MR_MEMORY_FAILURE) {
> > >-			put_page(page);
> > >  			if (!test_set_page_hwpoison(page))
> > >  				num_poisoned_pages_inc();
> > >-		} else
> > >+		}
> > 
> > Hmm, I didn't notice it previously, or it's due to rebasing,
but it
> > seems that you restricted the memory failure handling (i.e. setting
> > hwpoison) to MIGRATE_SUCCESS, while previously it was done for all
> > non-EAGAIN results. I think that goes against the intention of
> > hwpoison, which is IIRC to catch and kill the poor process that
> > still uses the page?
> 
> That's why I Cc'ed Naoya Horiguchi to catch things I might make
> mistake.
> 
> Thanks for catching it, Vlastimil.
> It was my mistake. But in this chance, I looked over hwpoison code and
> I saw other places which increases num_poisoned_pages are successful
> migration, already freed page and successful invalidated page.
> IOW, they are already successful isolated page so I guess it should
> increase the count when only successful migration is done?
Yes, that's right. When exiting with migration's failure, we
shouldn't call
test_set_page_hwpoison or num_poisoned_pages_inc, so current code checking
(rc != -EAGAIN) is simply incorrect. Your change fixes the bug in memory
error handling. Great!
> And when I read memory_failure, it bails out without killing if it
> encounters HWPoisoned page so I think it's not for catching and
> kill the poor proces.
>
> > 
> > Also (but not your fault) the put_page() preceding
> > test_set_page_hwpoison(page)) IMHO deserves a comment saying which
> > pin we are releasing and which one we still have (hopefully? if I
> > read description of da1b13ccfbebe right) otherwise it looks like
> > doing something with a page that we just potentially freed.
>
> Yes, while I read the code, I had same question. I think the releasing
> refcount is for get_any_page.
As the other callers of page migration do, soft_offline_page expects the
migration source page to be freed at this put_page() (no pin remains.)
The refcount released here is from isolate_lru_page() in __soft_offline_page().
(the pin by get_any_page is released by put_hwpoison_page just after it.)

.. yes, doing something just after freeing page looks weird, but that's
how PageHWPoison flag works. IOW, many other page flags are maintained
only during one "allocate-free" life span, but PageHWPoison still does
its job beyond it.

As for commenting, this put_page() is called in any MIGRATEPAGE_SUCCESS
case (regardless of callers), so what we can say here is "we free the
source page here, bypassing LRU list" or something?

Thanks,
Naoya Horiguchi

On 04/04/2016 06:45 AM, Naoya Horiguchi wrote:
> On Mon, Apr 04, 2016 at 10:39:17AM +0900, Minchan Kim wrote:
>> Thanks for catching it, Vlastimil.
>> It was my mistake. But in this chance, I looked over hwpoison code and
>> I saw other places which increases num_poisoned_pages are successful
>> migration, already freed page and successful invalidated page.
>> IOW, they are already successful isolated page so I guess it should
>> increase the count when only successful migration is done?
> 
> Yes, that's right. When exiting with migration's failure, we
shouldn't call
> test_set_page_hwpoison or num_poisoned_pages_inc, so current code checking
> (rc != -EAGAIN) is simply incorrect. Your change fixes the bug in memory
> error handling. Great!
Ah, I see, soft onlining works differently than I thought.
>> And when I read memory_failure, it bails out without killing if it
>> encounters HWPoisoned page so I think it's not for catching and
>> kill the poor proces.
>>
>>>
>>> Also (but not your fault) the put_page() preceding
>>> test_set_page_hwpoison(page)) IMHO deserves a comment saying which
>>> pin we are releasing and which one we still have (hopefully? if I
>>> read description of da1b13ccfbebe right) otherwise it looks like
>>> doing something with a page that we just potentially freed.
>>
>> Yes, while I read the code, I had same question. I think the releasing
>> refcount is for get_any_page.
> 
> As the other callers of page migration do, soft_offline_page expects the
> migration source page to be freed at this put_page() (no pin remains.)
> The refcount released here is from isolate_lru_page() in
__soft_offline_page().
> (the pin by get_any_page is released by put_hwpoison_page just after it.)
> 
> .. yes, doing something just after freeing page looks weird, but that's
> how PageHWPoison flag works. IOW, many other page flags are maintained
> only during one "allocate-free" life span, but PageHWPoison still
does
> its job beyond it.
But what prevents the page from being allocated again between put_page()
and test_set_page_hwpoison()? In that case we would be marking page
poisoned while still in use, which is the same as marking it while still
in use after a failed migration?

(Also, which part prevents pages with PageHWPoison to be allocated
again, anyway? I can't find it and test_set_page_hwpoison() doesn't
remove from buddy freelists).

Thanks.
> As for commenting, this put_page() is called in any MIGRATEPAGE_SUCCESS
> case (regardless of callers), so what we can say here is "we free the
> source page here, bypassing LRU list" or something?
> 
> Thanks,
> Naoya Horiguchi
>

On Mon, Apr 04, 2016 at 04:46:31PM +0200, Vlastimil Babka
wrote:
> On 04/04/2016 06:45 AM, Naoya Horiguchi wrote:
> > On Mon, Apr 04, 2016 at 10:39:17AM +0900, Minchan Kim wrote:
...
> >>>
> >>> Also (but not your fault) the put_page() preceding
> >>> test_set_page_hwpoison(page)) IMHO deserves a comment saying
which
> >>> pin we are releasing and which one we still have (hopefully?
if I
> >>> read description of da1b13ccfbebe right) otherwise it looks
like
> >>> doing something with a page that we just potentially freed.
> >>
> >> Yes, while I read the code, I had same question. I think the
releasing
> >> refcount is for get_any_page.
> > 
> > As the other callers of page migration do, soft_offline_page expects
the
> > migration source page to be freed at this put_page() (no pin remains.)
> > The refcount released here is from isolate_lru_page() in
__soft_offline_page().
> > (the pin by get_any_page is released by put_hwpoison_page just after
it.)
> > 
> > .. yes, doing something just after freeing page looks weird, but
that's
> > how PageHWPoison flag works. IOW, many other page flags are maintained
> > only during one "allocate-free" life span, but PageHWPoison
still does
> > its job beyond it.
> 
> But what prevents the page from being allocated again between put_page()
> and test_set_page_hwpoison()? In that case we would be marking page
> poisoned while still in use, which is the same as marking it while still
> in use after a failed migration?
Actually nothing prevents that race. But I think that the result of the race
is that the error page can be reused for allocation, which results in killing
processes at page fault time. Soft offline is kind of mild/precautious thing
(for correctable errors that don't require immediate handling), so killing
processes looks to me an overkill. And marking hwpoison means that we can no
longer do retry from userspace.

And another practical thing is the race with unpoison_memory() as described
in commit da1b13ccfbebe. unpoison_memory() properly works only for properly
poisoned pages, so doing unpoison for in-use hwpoisoned pages is fragile.
That's why I'd like to avoid setting PageHWPoison for in-use pages if
possible.
> (Also, which part prevents pages with PageHWPoison to be allocated
> again, anyway? I can't find it and test_set_page_hwpoison() doesn't
> remove from buddy freelists).
check_new_page() in mm/page_alloc.c should prevent reallocation of PageHWPoison.
As you pointed out, memory error handler doens't remove it from buddy
freelists.


BTW, it might be a bit off-topic, but recently I felt that check_new_page()
might be improvable, because when check_new_page() returns 1, the whole buddy
block (not only the bad page) seems to be leaked from buddy freelist.
For example, if thp (order 9) is requested, and PageHWPoison (or any other
types of bad pages) is found in an order 9 block, all 512 page are discarded.
Unpoison can't bring it back to buddy.
So, some code to split buddy block including bad page (and recovering code from
unpoison) might be helpful, although that's another story ...

Thanks,
Naoya Horiguchi