On Tue, Nov 12, 2024 at 9:24 PM Kanchana P Sridhar
<kanchana.p.sridhar@xxxxxxxxx> wrote:
>
> This is a hotfix for a potential zpool memory leak that could result in
> the existing zswap_decompress():
>
> mutex_unlock(&acomp_ctx->mutex);
>
> if (src != acomp_ctx->buffer)
> zpool_unmap_handle(zpool, entry->handle);
>
> Releasing the lock before the conditional does not protect the integrity of
> "src", which is set earlier under the acomp_ctx mutex lock. This poses a
> risk for the conditional behaving as intended, and consequently not
> unmapping the zpool handle, which could cause a zswap zpool memory leak.
>
> This patch moves the mutex_unlock() to occur after the conditional and
> subsequent zpool_unmap_handle(). This ensures that the value of "src"
> obtained earlier, with the mutex locked, does not change.
The commit log is too complicated and incorrect. It is talking about
the stability of 'src', but that's a local variable on the stack
anyway. It doesn't need protection.
The problem is 'acomp_ctx->buffer' being reused and changed after the
mutex is released. Leading to the check not being reliable. Please
simplify this.
>
> Even though an actual memory leak was not observed, this fix seems like a
> cleaner implementation.
>
> Signed-off-by: Kanchana P Sridhar <kanchana.p.sridhar@xxxxxxxxx>
> Fixes: 9c500835f279 ("mm: zswap: fix kernel BUG in sg_init_one")
> ---
> mm/zswap.c | 3 ++-
> 1 file changed, 2 insertions(+), 1 deletion(-)
>
> diff --git a/mm/zswap.c b/mm/zswap.c
> index f6316b66fb23..58810fa8ff23 100644
> --- a/mm/zswap.c
> +++ b/mm/zswap.c
> @@ -986,10 +986,11 @@ static void zswap_decompress(struct zswap_entry *entry, struct folio *folio)
> acomp_request_set_params(acomp_ctx->req, &input, &output, entry->length, PAGE_SIZE);
> BUG_ON(crypto_wait_req(crypto_acomp_decompress(acomp_ctx->req), &acomp_ctx->wait));
> BUG_ON(acomp_ctx->req->dlen != PAGE_SIZE);
> - mutex_unlock(&acomp_ctx->mutex);
>
> if (src != acomp_ctx->buffer)
> zpool_unmap_handle(zpool, entry->handle);
Actually now that I think more about it, I think this check isn't
entirely safe, even under the lock. Is it possible that
'acomp_ctx->buffer' just happens to be equal to 'src' from a previous
decompression at the same handle? In this case, we will also
mistakenly skip the unmap.
It would be more reliable to set a boolean variable if we copy to
acomp_ctx->buffer and do the unmap, and check that flag here to check
if the unmap was done or not.
> +
> + mutex_unlock(&acomp_ctx->mutex);
> }
>
> /*********************************
>
> base-commit: 0e5bdedb39ded767bff4c6184225578595cee98c
> --
> 2.27.0
>
