lkml.org 
[lkml]   [2018]   [Nov]   [27]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCHi v2] mm: put_and_wait_on_page_locked() while page is migrated
On Mon, Nov 26, 2018 at 11:27:07AM -0800, Hugh Dickins wrote:
> Waiting on a page migration entry has used wait_on_page_locked() all
> along since 2006: but you cannot safely wait_on_page_locked() without
> holding a reference to the page, and that extra reference is enough to
> make migrate_page_move_mapping() fail with -EAGAIN, when a racing task
> faults on the entry before migrate_page_move_mapping() gets there.
>
> And that failure is retried nine times, amplifying the pain when
> trying to migrate a popular page. With a single persistent faulter,
> migration sometimes succeeds; with two or three concurrent faulters,
> success becomes much less likely (and the more the page was mapped,
> the worse the overhead of unmapping and remapping it on each try).
>
> This is especially a problem for memory offlining, where the outer
> level retries forever (or until terminated from userspace), because
> a heavy refault workload can trigger an endless loop of migration
> failures. wait_on_page_locked() is the wrong tool for the job.
>
> David Herrmann (but was he the first?) noticed this issue in 2014:
> https://marc.info/?l=linux-mm&m=140110465608116&w=2
>
> Tim Chen started a thread in August 2017 which appears relevant:
> https://marc.info/?l=linux-mm&m=150275941014915&w=2
> where Kan Liang went on to implicate __migration_entry_wait():
> https://marc.info/?l=linux-mm&m=150300268411980&w=2
> and the thread ended up with the v4.14 commits:
> 2554db916586 ("sched/wait: Break up long wake list walk")
> 11a19c7b099f ("sched/wait: Introduce wakeup boomark in wake_up_page_bit")
>
> Baoquan He reported "Memory hotplug softlock issue" 14 November 2018:
> https://marc.info/?l=linux-mm&m=154217936431300&w=2
>
> We have all assumed that it is essential to hold a page reference while
> waiting on a page lock: partly to guarantee that there is still a struct
> page when MEMORY_HOTREMOVE is configured, but also to protect against
> reuse of the struct page going to someone who then holds the page locked
> indefinitely, when the waiter can reasonably expect timely unlocking.
>
> But in fact, so long as wait_on_page_bit_common() does the put_page(),
> and is careful not to rely on struct page contents thereafter, there is
> no need to hold a reference to the page while waiting on it. That does
> mean that this case cannot go back through the loop: but that's fine for
> the page migration case, and even if used more widely, is limited by the
> "Stop walking if it's locked" optimization in wake_page_function().
>
> Add interface put_and_wait_on_page_locked() to do this, using "behavior"
> enum in place of "lock" arg to wait_on_page_bit_common() to implement it.
> No interruptible or killable variant needed yet, but they might follow:
> I have a vague notion that reporting -EINTR should take precedence over
> return from wait_on_page_bit_common() without knowing the page state,
> so arrange it accordingly - but that may be nothing but pedantic.
>
> __migration_entry_wait() still has to take a brief reference to the
> page, prior to calling put_and_wait_on_page_locked(): but now that it
> is dropped before waiting, the chance of impeding page migration is
> very much reduced. Should we perhaps disable preemption across this?
>
> shrink_page_list()'s __ClearPageLocked(): that was a surprise! This
> survived a lot of testing before that showed up. PageWaiters may have
> been set by wait_on_page_bit_common(), and the reference dropped, just
> before shrink_page_list() succeeds in freezing its last page reference:
> in such a case, unlock_page() must be used. Follow the suggestion from
> Michal Hocko, just revert a978d6f52106 ("mm: unlockless reclaim") now:
> that optimization predates PageWaiters, and won't buy much these days;
> but we can reinstate it for the !PageWaiters case if anyone notices.
>
> It does raise the question: should vmscan.c's is_page_cache_freeable()
> and __remove_mapping() now treat a PageWaiters page as if an extra
> reference were held? Perhaps, but I don't think it matters much, since
> shrink_page_list() already had to win its trylock_page(), so waiters are
> not very common there: I noticed no difference when trying the bigger
> change, and it's surely not needed while put_and_wait_on_page_locked()
> is only used for page migration.
>
> Reported-and-tested-by: Baoquan He <bhe@redhat.com>
> Signed-off-by: Hugh Dickins <hughd@google.com>
> Acked-by: Michal Hocko <mhocko@suse.com>
> Reviewed-by: Andrea Arcangeli <aarcange@redhat.com>
> ---
> include/linux/pagemap.h | 2 ++
> mm/filemap.c | 77 ++++++++++++++++++++++++++++++++++-------
> mm/huge_memory.c | 6 ++--
> mm/migrate.c | 12 +++----
> mm/vmscan.c | 10 ++----
> 5 files changed, 74 insertions(+), 33 deletions(-)
>
> diff --git a/include/linux/pagemap.h b/include/linux/pagemap.h
> index 226f96f0dee0..e2d7039af6a3 100644
> --- a/include/linux/pagemap.h
> +++ b/include/linux/pagemap.h
> @@ -537,6 +537,8 @@ static inline int wait_on_page_locked_killable(struct page *page)
> return wait_on_page_bit_killable(compound_head(page), PG_locked);
> }
>
> +extern void put_and_wait_on_page_locked(struct page *page);
> +
> /*
> * Wait for a page to complete writeback
> */
> diff --git a/mm/filemap.c b/mm/filemap.c
> index 81adec8ee02c..575e16c037ca 100644
> --- a/mm/filemap.c
> +++ b/mm/filemap.c
> @@ -981,7 +981,14 @@ static int wake_page_function(wait_queue_entry_t *wait, unsigned mode, int sync,
> if (wait_page->bit_nr != key->bit_nr)
> return 0;
>
> - /* Stop walking if it's locked */
> + /*
> + * Stop walking if it's locked.
> + * Is this safe if put_and_wait_on_page_locked() is in use?
> + * Yes: the waker must hold a reference to this page, and if PG_locked
> + * has now already been set by another task, that task must also hold
> + * a reference to the *same usage* of this page; so there is no need
> + * to walk on to wake even the put_and_wait_on_page_locked() callers.
> + */
> if (test_bit(key->bit_nr, &key->page->flags))
> return -1;
>
> @@ -1049,25 +1056,44 @@ static void wake_up_page(struct page *page, int bit)
> wake_up_page_bit(page, bit);
> }
>
> +/*
> + * A choice of three behaviors for wait_on_page_bit_common():
> + */
> +enum behavior {
> + EXCLUSIVE, /* Hold ref to page and take the bit when woken, like
> + * __lock_page() waiting on then setting PG_locked.
> + */
> + SHARED, /* Hold ref to page and check the bit when woken, like
> + * wait_on_page_writeback() waiting on PG_writeback.
> + */
> + DROP, /* Drop ref to page before wait, no check when woken,
> + * like put_and_wait_on_page_locked() on PG_locked.
> + */
> +};

Can we please make it:

/**
* enum behavior - a choice of three behaviors for wait_on_page_bit_common()
*/
enum behavior {
/**
* @EXCLUSIVE: Hold ref to page and take the bit when woken,
* like __lock_page() waiting on then setting %PG_locked.
*/
EXCLUSIVE,
/**
* @SHARED: Hold ref to page and check the bit when woken,
* like wait_on_page_writeback() waiting on %PG_writeback.
*/
SHARED,
/**
* @DROP: Drop ref to page before wait, no check when woken,
* like put_and_wait_on_page_locked() on %PG_locked.
*/
DROP,
};


--
Sincerely yours,
Mike.

\
 
 \ /
  Last update: 2018-11-27 12:00    [W:0.160 / U:3.524 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site