[Devel] [PATCH rh7] mm/tswap: fix lockup in tswap_evict_page()

[Andrey Ryabinin]

Surprisingly after the upstream commit 7c00bafee87c
("mm/swap: free swap slots in batch"), backported
by redhat, SWAP_HAS_CACHE in swap_info->swap_map[offset]
doesn't mean that swap entry has a page in swap cache anymore.

Now we may set SWAP_HAS_CACHE even if we know that page is not cached.
Commit 7c00bafee87c changed swapcache_free() so it will set SWAP_HAS_CACHE:
-       p->swap_map[offset] = usage;
+       p->swap_map[offset] = usage ? : SWAP_HAS_CACHE;

And it will be cleared only by swapcache_free_entries(), but swapcache_free_entries()
call may be delayed indefinitely, until the 'swp_slots' per-cpu cache
is full.

It's not that obvious, how to fix this mess. For now, I think, the safest
course of action would be a workaround in the tswap, similar to what we
do in __read_swap_cache_async() - bail out on zero swap_count. Apperently
it works for __read_swap_cache_async(), so it should work for tswap.

https://jira.sw.ru/browse/PSBM-86344
Signed-off-by: Andrey Ryabinin <aryabinin at virtuozzo.com>
---
 mm/tswap.c | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/mm/tswap.c b/mm/tswap.c
index b7a990e8cd8d..7b02c1142db2 100644
--- a/mm/tswap.c
+++ b/mm/tswap.c
@@ -183,6 +183,11 @@ static int tswap_evict_page(struct page *page)
 		goto out;
 	}
 
+	if (!__swp_swapcount(entry) && swap_slot_cache_enabled) {
+		err = -ENOENT;
+		goto out;
+	}
+
 	err = swapcache_prepare(entry);
 	if (err == -EEXIST) {
 		cond_resched();
-- 
2.16.4