mm, memory_hotplug: do not fail offlining too early

Patch series "mm, memory_hotplug: redefine memory offline retry logic", v2.

While testing memory hotplug on a large 4TB machine we have noticed that
memory offlining is just too eager to fail.  The primary reason is that
the retry logic is just too easy to give up.  We have 4 ways out of the
offline

	- we have a permanent failure (isolation or memory notifiers fail,
	  or hugetlb pages cannot be dropped)
	- userspace sends a signal
	- a hardcoded 120s timeout expires
	- page migration fails 5 times

This is way too convoluted and it doesn't scale very well.  We have seen
both temporary migration failures as well as 120s being triggered.
After removing those restrictions we were able to pass stress testing
during memory hot remove without any other negative side effects
observed.  Therefore I suggest dropping both hard coded policies.  I
couldn't have found any specific reason for them in the changelog.  I
neither didn't get any response [1] from Kamezawa.  If we need some
upper bound - e.g.  timeout based - then we should have a proper and
user defined policy for that.  In any case there should be a clear use
case when introducing it.

This patch (of 2):

Memory offlining can fail too eagerly under heavy memory pressure.

  page:ffffea22a646bd00 count:255 mapcount:252 mapping:ffff88ff926c9f38 index:0x3
  flags: 0x9855fe40010048(uptodate|active|mappedtodisk)
  page dumped because: isolation failed
  page->mem_cgroup:ffff8801cd662000
  memory offlining [mem 0x18b580000000-0x18b5ffffffff] failed

Isolation has failed here because the page is not on LRU.  Most probably
because it was on the pcp LRU cache or it has been removed from the LRU
already but it hasn't been freed yet.  In both cases the page doesn't
look non-migrable so retrying more makes sense.

__offline_pages seems rather cluttered when it comes to the retry logic.
We have 5 retries at maximum and a timeout.  We could argue whether the
timeout makes sense but failing just because of a race when somebody
isoltes a page from LRU or puts it on a pcp LRU lists is just wrong.  It
only takes it to race with a process which unmaps some pages and remove
them from the LRU list and we can fail the whole offline because of
something that is a temporary condition and actually not harmful for the
offline.

Please note that unmovable pages should be already excluded during
start_isolate_page_range.  We could argue that has_unmovable_pages is
racy and MIGRATE_MOVABLE check doesn't provide any hard guarantee either
but kernel zones (aka < ZONE_MOVABLE) will very likely detect unmovable
pages in most cases and movable zone shouldn't contain unmovable pages
at all.  Some of those pages might be pinned but not for ever because
that would be a bug on its own.  In any case the context is still
interruptible and so the userspace can easily bail out when the
operation takes too long.  This is certainly better behavior than a
hardcoded retry loop which is racy.

Fix this by removing the max retry count and only rely on the timeout
resp. interruption by a signal from the userspace.  Also retry rather
than fail when check_pages_isolated sees some !free pages because those
could be a result of the race as well.

Link: http://lkml.kernel.org/r/20170918070834.13083-2-mhocko@kernel.org
Signed-off-by: Michal Hocko <mhocko@suse.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Reza Arbab <arbab@linux.vnet.ibm.com>
Cc: Yasuaki Ishimatsu <yasu.isimatu@gmail.com>
Cc: Xishi Qiu <qiuxishi@huawei.com>
Cc: Igor Mammedov <imammedo@redhat.com>
Cc: Vitaly Kuznetsov <vkuznets@redhat.com>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/memory_hotplug.c

@@ -1594,7 +1594,7 @@ static int __ref __offline_pages(unsigned long start_pfn,
 {
 	unsigned long pfn, nr_pages, expire;
 	long offlined_pages;
-	int ret, drain, retry_max, node;
+	int ret, node;
 	unsigned long flags;
 	unsigned long valid_start, valid_end;
 	struct zone *zone;
@@ -1631,43 +1631,25 @@ static int __ref __offline_pages(unsigned long start_pfn,
 
 	pfn = start_pfn;
 	expire = jiffies + timeout;
-	drain = 0;
-	retry_max = 5;
 repeat:
 	/* start memory hot removal */
-	ret = -EAGAIN;
+	ret = -EBUSY;
 	if (time_after(jiffies, expire))
 		goto failed_removal;
 	ret = -EINTR;
 	if (signal_pending(current))
 		goto failed_removal;
-	ret = 0;
-	if (drain) {
-		lru_add_drain_all_cpuslocked();
-		cond_resched();
-		drain_all_pages(zone);
-	}
+
+	cond_resched();
+	lru_add_drain_all_cpuslocked();
+	drain_all_pages(zone);
 
 	pfn = scan_movable_pages(start_pfn, end_pfn);
 	if (pfn) { /* We have movable pages */
 		ret = do_migrate_range(pfn, end_pfn);
-		if (!ret) {
-			drain = 1;
-			goto repeat;
-		} else {
-			if (ret < 0)
-				if (--retry_max == 0)
-					goto failed_removal;
-			yield();
-			drain = 1;
-			goto repeat;
-		}
+		goto repeat;
 	}
-	/* drain all zone's lru pagevec, this is asynchronous... */
-	lru_add_drain_all_cpuslocked();
-	yield();
-	/* drain pcp pages, this is synchronous. */
-	drain_all_pages(zone);
+
 	/*
 	 * dissolve free hugepages in the memory block before doing offlining
 	 * actually in order to make hugetlbfs's object counting consistent.
@@ -1677,10 +1659,8 @@ repeat:
 		goto failed_removal;
 	/* check again */
 	offlined_pages = check_pages_isolated(start_pfn, end_pfn);
-	if (offlined_pages < 0) {
-		ret = -EBUSY;
-		goto failed_removal;
-	}
+	if (offlined_pages < 0)
+		goto repeat;
 	pr_info("Offlined Pages %ld\n", offlined_pages);
 	/* Ok, all of our target is isolated.
 	   We cannot do rollback at this point. */