Re: [RFC 5/5] mm, page_alloc: disable pcplists during page isolation

From: Michal Hocko
Date: Wed Sep 09 2020 - 07:44:31 EST

On Wed 09-09-20 12:48:54, Vlastimil Babka wrote:
> Here's a version that will apply on top of next-20200908. The first 4 patches need no change.
> ----8<----
> >From 8febc17272b8e8b378e2e5ea5e76b2616f029c5b Mon Sep 17 00:00:00 2001
> From: Vlastimil Babka <vbabka@xxxxxxx>
> Date: Mon, 7 Sep 2020 17:20:39 +0200
> Subject: [PATCH] mm, page_alloc: disable pcplists during page isolation
> Page isolation can race with process freeing pages to pcplists in a way that
> a page from isolated pageblock can end up on pcplist. This can be fixed by
> repeated draining of pcplists, as done by patch "mm/memory_hotplug: drain
> per-cpu pages again during memory offline" in [1].
> David and Michal would prefer that this race was closed in a way that callers
> of page isolation don't need to care about drain. David suggested disabling
> pcplists usage completely during page isolation, instead of repeatedly draining
> them.
> To achieve this without adding special cases in alloc/free fastpath, we can use
> the same 'trick' as boot pagesets - when pcp->high is 0, any pcplist addition
> will be immediately flushed.
> The race can thus be closed by setting pcp->high to 0 and draining pcplists
> once in start_isolate_page_range(). The draining will serialize after processes
> that already disabled interrupts and read the old value of pcp->high in
> free_unref_page_commit(), and processes that have not yet disabled interrupts,
> will observe pcp->high == 0 when they are rescheduled, and skip pcplists.
> This guarantees no stray pages on pcplists in zones where isolation happens.
> We can use the variable zone->nr_isolate_pageblock (protected by zone->lock)
> to detect transitions from 0 to 1 (to change pcp->high to 0 and issue drain)
> and from 1 to 0 (to restore original pcp->high and batch values cached in
> struct zone). We have to avoid external updates to high and batch by taking
> pcp_batch_high_lock. To allow multiple isolations in parallel, change this
> lock from mutex to rwsem.
> For callers that pair start_isolate_page_range() with
> undo_isolated_page_range() properly, this is transparent. Currently that's
> alloc_contig_range(). __offline_pages() doesn't call undo_isolated_page_range()
> in the succes case, so it has to be carful to handle restoring pcp->high and batch
> and unlocking pcp_batch_high_lock.

I was hoping that it would be possible to have this completely hidden
inside start_isolate_page_range code path. If we need some sort of
disable_pcp_free/enable_pcp_free then it seems like a better fit to have
an explicit API for that (the naming would be obviously different
because we do not want to call out pcp free lists). I strongly suspect
that only the memory hotplug really cares for this hard guanrantee.
alloc_contig_range simply goes with EBUSY.

> This commit also changes drain_all_pages() to not trust reading pcp->count during
> drain for page isolation - I believe that could be racy and lead to missing some
> cpu's to drain. If others agree, this can be separated and potentially backported.
> [1]
> Suggested-by: David Hildenbrand <david@xxxxxxxxxx>
> Suggested-by: Michal Hocko <mhocko@xxxxxxxx>
> Signed-off-by: Vlastimil Babka <vbabka@xxxxxxx>
> ---
> include/linux/gfp.h | 1 +
> mm/internal.h | 4 +++
> mm/memory_hotplug.c | 55 ++++++++++++++++++++++++++++-------------
> mm/page_alloc.c | 60 +++++++++++++++++++++++++++++----------------
> mm/page_isolation.c | 45 ++++++++++++++++++++++++++++------
> 5 files changed, 119 insertions(+), 46 deletions(-)

This has turned out much larger than I would expect.
Michal Hocko