Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
From:   Domenico Cerasuolo <cerasuolodomenico@gmail.com>
To:     sjenning@redhat.com, ddstreet@ieee.org, vitaly.wool@konsulko.com,
        minchan@kernel.org, ngupta@vflare.org, akpm@linux-foundation.org,
        linux-mm@kvack.org, linux-kernel@vger.kernel.org
Cc:     hannes@cmpxchg.org,
        Domenico Cerasuolo <cerasuolodomenico@gmail.com>
Subject: [PATCH] mm: fix zswap writeback race condition
Date:   Wed,  3 May 2023 15:21:48 +0200
Message-Id: <20230503132148.9682-1-cerasuolodomenico@gmail.com>
MIME-Version: 1.0
Content-Transfer-Encoding: 8bit
Precedence: bulk

The zswap writeback mechanism can cause a race condition resulting in
memory corruption, where a swapped out page gets swapped in with data
that was written to a different page.

The race unfolds like this:
1. a page with data A and swap offset X is stored in zswap
2. page A is removed off the LRU by zpool driver for writeback in
zswap-shrink work, data for A is mapped by zpool driver
3. user space program faults and invalidates page entry A, offset X is
considered free
4. kswapd stores page B at offset X in zswap (zswap could also be full,
if so, page B would then be IOed to X, then skip step 5.)
5. entry A is replaced by B in tree->rbroot, this doesn't affect the
local reference held by zswap-shrink work
6. zswap-shrink work writes back A at X, and frees zswap entry A
7. swapin of slot X brings A in memory instead of B

The fix:
Once the swap page cache has been allocated (case ZSWAP_SWAPCACHE_NEW),
zswap-shrink work just checks that the local zswap_entry reference is
still the same as the one in the tree. If it's not the same it means
that it's either been invalidated or replaced, in both cases the
writeback is aborted because the local entry contains stale data.

Reproducer:
I originally found this by running `stress` overnight to validate my
work on the zswap writeback mechanism, it manifested after hours on my
test machine. The key to make it happen is having zswap writebacks, so
whatever setup pumps /sys/kernel/debug/zswap/written_back_pages should
do the trick.
In order to reproduce this faster on a vm, I setup a system with ~100M
of available memory and a 500M swap file, then running
`stress --vm 1 --vm-bytes 300000000 --vm-stride 4000` makes it happen
in matter of tens of minutes. One can speed things up even more by
swinging /sys/module/zswap/parameters/max_pool_percent up and down
between, say, 20 and 1; this makes it reproduce in tens of seconds.
It's crucial to set `--vm-stride` to something other than 4096 otherwise
`stress` won't realize that memory has been corrupted because all pages
would have the same data.

Signed-off-by: Domenico Cerasuolo <cerasuolodomenico@gmail.com>
---
 mm/zswap.c | 13 +++++++++++++
 1 file changed, 13 insertions(+)

diff --git a/mm/zswap.c b/mm/zswap.c
index f6c89049cf70..d20d60266bc8 100644
--- a/mm/zswap.c
+++ b/mm/zswap.c
@@ -995,6 +995,19 @@ static int zswap_writeback_entry(struct zpool *pool, unsigned long handle)
 		goto fail;
 
 	case ZSWAP_SWAPCACHE_NEW: /* page is locked */
+		/*
+		 * if the entry in the tree has been replaced, it means that we would
+		 * be overriding the swap page with stale data, let's not do that.
+		 */
+		spin_lock(&tree->lock);
+		if (zswap_rb_search(&tree->rbroot, entry->offset) != entry) {
+			spin_unlock(&tree->lock);
+			delete_from_swap_cache(page_folio(page));
+			ret = -ENOMEM;
+			goto fail;
+		}
+		spin_unlock(&tree->lock);
+
 		/* decompress */
 		acomp_ctx = raw_cpu_ptr(entry->pool->acomp_ctx);
 		dlen = PAGE_SIZE;
-- 
2.34.1