mm/page_alloc: always attempt to allocate at least one page during bu…

…lk allocation

commit c572e48 upstream.

Peter Pavlisko reported the following problem on kernel bugzilla 216007.

	When I try to extract an uncompressed tar archive (2.6 milion
	files, 760.3 GiB in size) on newly created (empty) XFS file system,
	after first low tens of gigabytes extracted the process hangs in
	iowait indefinitely. One CPU core is 100% occupied with iowait,
	the other CPU core is idle (on 2-core Intel Celeron G1610T).

It was bisected to c9fa563 ("xfs: use alloc_pages_bulk_array() for
buffers") but XFS is only the messenger.  The problem is that nothing is
waking kswapd to reclaim some pages at a time the PCP lists cannot be
refilled until some reclaim happens.  The bulk allocator checks that there
are some pages in the array and the original intent was that a bulk
allocator did not necessarily need all the requested pages and it was best
to return as quickly as possible.

This was fine for the first user of the API but both NFS and XFS require
the requested number of pages be available before making progress.  Both
could be adjusted to call the page allocator directly if a bulk allocation
fails but it puts a burden on users of the API.  Adjust the semantics to
attempt at least one allocation via __alloc_pages() before returning so
kswapd is woken if necessary.

It was reported via bugzilla that the patch addressed the problem and that
the tar extraction completed successfully.  This may also address bug
215975 but has yet to be confirmed.

BugLink: https://bugzilla.kernel.org/show_bug.cgi?id=216007
BugLink: https://bugzilla.kernel.org/show_bug.cgi?id=215975
Link: https://lkml.kernel.org/r/20220526091210.GC3441@techsingularity.net
Fixes: 387ba26 ("mm/page_alloc: add a bulk page allocator")
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Cc: "Darrick J. Wong" <djwong@kernel.org>
Cc: Dave Chinner <dchinner@redhat.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Jesper Dangaard Brouer <brouer@redhat.com>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: <stable@vger.kernel.org>	[5.13+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

mm/page_alloc.c

@@ -5305,8 +5305,8 @@ unsigned long __alloc_pages_bulk(gfp_t gfp, int preferred_nid,
 		page = __rmqueue_pcplist(zone, 0, ac.migratetype, alloc_flags,
 								pcp, pcp_list);
 		if (unlikely(!page)) {
-			/* Try and get at least one page */
-			if (!nr_populated)
+			/* Try and allocate at least one page */
+			if (!nr_account)
 				goto failed_irq;
 			break;
 		}