Re: [PATCH] mm/zswap: invalidate old entry when store fail or !zswap_enabled

[Yosry Ahmed]

On Sun, Feb 04, 2024 at 08:34:11AM +0000, chengming.zhou@xxxxxxxxx wrote:
> From: Chengming Zhou <zhouchengming@xxxxxxxxxxxxx>
> 
> We may encounter duplicate entry in the zswap_store():
> 
> 1. swap slot that freed to per-cpu swap cache, doesn't invalidate
>    the zswap entry, then got reused. This has been fixed.
> 
> 2. !exclusive load mode, swapin folio will leave its zswap entry
>    on the tree, then swapout again. This has been removed.
> 
> 3. one folio can be dirtied again after zswap_store(), so need to
>    zswap_store() again. This should be handled correctly.
> 
> So we must invalidate the old duplicate entry before insert the
> new one, which actually doesn't have to be done at the beginning
> of zswap_store(). And this is a normal situation, we shouldn't
> WARN_ON(1) in this case, so delete it. (The WARN_ON(1) seems want
> to detect swap entry UAF problem? But not very necessary here.)
> 
> The good point is that we don't need to lock tree twice in the
> store success path.
> 
> Note we still need to invalidate the old duplicate entry in the
> store failure path, otherwise the new data in swapfile could be
> overwrite by the old data in zswap pool when lru writeback.

I think this may have been introduced by 42c06a0e8ebe ("mm: kill
frontswap"). Frontswap used to check if the page was present in
frontswap and invalidate it before calling into zswap, so it would
invalidate a previously stored page when it is dirtied and swapped out
again, even if zswap is disabled.

Johannes, does this sound correct to you? If yes, I think we need a
proper Fixes tag and a stable backport as this may cause data
corruption.

> 
> We have to do this even when !zswap_enabled since zswap can be
> disabled anytime. If the folio store success before, then got
> dirtied again but zswap disabled, we won't invalidate the old
> duplicate entry in the zswap_store(). So later lru writeback
> may overwrite the new data in swapfile.
> 
> This fix is not good, since we have to grab lock to check everytime
> even when zswap is disabled, but it's simple.

Frontswap had a bitmap that we can query locklessly to find out if there
is an outdated stored page. I think we can overcome this with the
xarray, we can do a lockless lookup first, and only take the lock if
there is an outdated entry to remove.

Meanwhile I am not sure if acquiring the lock on every swapout even with
zswap disabled is acceptable, but I think it's the simplest fix for now,
unless we revive the bitmap.

> 
> Signed-off-by: Chengming Zhou <zhouchengming@xxxxxxxxxxxxx>
> ---
>  mm/zswap.c | 33 +++++++++++++++------------------
>  1 file changed, 15 insertions(+), 18 deletions(-)
> 
> diff --git a/mm/zswap.c b/mm/zswap.c
> index cd67f7f6b302..0b7599f4116d 100644
> --- a/mm/zswap.c
> +++ b/mm/zswap.c
> @@ -1518,18 +1518,8 @@ bool zswap_store(struct folio *folio)
>  		return false;
>  
>  	if (!zswap_enabled)
> -		return false;
> +		goto check_old;
>  
> -	/*
> -	 * If this is a duplicate, it must be removed before attempting to store
> -	 * it, otherwise, if the store fails the old page won't be removed from
> -	 * the tree, and it might be written back overriding the new data.
> -	 */
> -	spin_lock(&tree->lock);
> -	entry = zswap_rb_search(&tree->rbroot, offset);
> -	if (entry)
> -		zswap_invalidate_entry(tree, entry);
> -	spin_unlock(&tree->lock);
>  	objcg = get_obj_cgroup_from_folio(folio);
>  	if (objcg && !obj_cgroup_may_zswap(objcg)) {
>  		memcg = get_mem_cgroup_from_objcg(objcg);
> @@ -1608,15 +1598,11 @@ bool zswap_store(struct folio *folio)
>  	/* map */
>  	spin_lock(&tree->lock);
>  	/*
> -	 * A duplicate entry should have been removed at the beginning of this
> -	 * function. Since the swap entry should be pinned, if a duplicate is
> -	 * found again here it means that something went wrong in the swap
> -	 * cache.
> +	 * The folio could be dirtied again, invalidate the possible old entry
> +	 * before insert this new entry.
>  	 */
> -	while (zswap_rb_insert(&tree->rbroot, entry, &dupentry) == -EEXIST) {
> -		WARN_ON(1);
> +	while (zswap_rb_insert(&tree->rbroot, entry, &dupentry) == -EEXIST)
>  		zswap_invalidate_entry(tree, dupentry);
> -	}

I always thought the loop here was confusing. We are holding the lock,
so it should be guaranteed that if we get -EEXIST once and invalidate
it, we won't find it the next time around.

This should really be a cmpxchg operation, which is simple with the
xarray. We can probably do the same with the rbtree, but perhaps it's
not worth it if the xarray change is coming soon.

For now, I think an if condition is clearer:

if (zswap_rb_insert(&tree->rbroot, entry, &dupentry) == -EEXIST) {
	zswap_invalidate_entry(tree, dupentry);
	/* Must succeed, we just removed the dup under the lock */
	WARN_ON(zswap_rb_insert(&tree->rbroot, entry, &dupentry));
}

>  	if (entry->length) {
>  		INIT_LIST_HEAD(&entry->lru);
>  		zswap_lru_add(&entry->pool->list_lru, entry);
> @@ -1638,6 +1624,17 @@ bool zswap_store(struct folio *folio)
>  reject:
>  	if (objcg)
>  		obj_cgroup_put(objcg);
> +check_old:
> +	/*
> +	 * If zswap store fail or zswap disabled, we must invalidate possible
> +	 * old entry which previously stored by this folio. Otherwise, later
> +	 * writeback could overwrite the new data in swapfile.
> +	 */
> +	spin_lock(&tree->lock);
> +	entry = zswap_rb_search(&tree->rbroot, offset);
> +	if (entry)
> +		zswap_invalidate_entry(tree, entry);
> +	spin_unlock(&tree->lock);
>  	return false;
>  
>  shrink:
> -- 
> 2.40.1
>