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:   ZiyangZhang <ZiyangZhang@linux.alibaba.com>
To:     ming.lei@redhat.com, axboe@kernel.dk
Cc:     xiaoguang.wang@linux.alibaba.com, linux-block@vger.kernel.org,
        linux-kernel@vger.kernel.org, joseph.qi@linux.alibaba.com,
        ZiyangZhang <ZiyangZhang@linux.alibaba.com>
Subject: [RFC PATCH V2 2/6] ublk_drv: refactor ublk_cancel_queue()
Date:   Wed, 31 Aug 2022 23:51:32 +0800
Message-Id: <20220831155136.23434-3-ZiyangZhang@linux.alibaba.com>
In-Reply-To: <20220831155136.23434-1-ZiyangZhang@linux.alibaba.com>
References: <20220831155136.23434-1-ZiyangZhang@linux.alibaba.com>
MIME-Version: 1.0
Content-Transfer-Encoding: 8bit
Precedence: bulk

Assume only a few FETCH_REQ ioucmds are sent to ublk_drv, then the
ubq_daemon exits, We have to call io_uring_cmd_done() for all ioucmds
received so that io_uring ctx will not leak.

ublk_cancel_queue() may be called before START_DEV or after STOP_DEV,
we decrease ubq->nr_io_ready and clear UBLK_IO_FLAG_ACTIVE so that we
won't call io_uring_cmd_done() twice for one ioucmd to avoid UAF. Also
clearing UBLK_IO_FLAG_ACTIVE makes the code more reasonable.

Signed-off-by: ZiyangZhang <ZiyangZhang@linux.alibaba.com>
---
 drivers/block/ublk_drv.c | 23 ++++++++++++++++++++---
 1 file changed, 20 insertions(+), 3 deletions(-)

diff --git a/drivers/block/ublk_drv.c b/drivers/block/ublk_drv.c
index c39b67d7133d..0c6db0978ed0 100644
--- a/drivers/block/ublk_drv.c
+++ b/drivers/block/ublk_drv.c
@@ -963,22 +963,39 @@ static inline bool ublk_queue_ready(struct ublk_queue *ubq)
 	return ubq->nr_io_ready == ubq->q_depth;
 }
 
+/* If ublk_cancel_queue() is called before sending START_DEV(), ->mutex
+ * provides protection on above update.
+ *
+ * If ublk_cancel_queue() is called after sending START_DEV(), disk is
+ * deleted first, UBLK_IO_RES_ABORT is returned so that any new io
+ * command can't be issued to driver, so updating on io flags and
+ * nr_io_ready is safe here.
+ *
+ * Also ->nr_io_ready is guaranteed to become zero after ublk_cance_queue()
+ * returns since request queue is either frozen or not present in both two
+ * cases.
+ */
 static void ublk_cancel_queue(struct ublk_queue *ubq)
 {
 	int i;
 
-	if (!ublk_queue_ready(ubq))
+	if (!ubq->nr_io_ready)
 		return;
 
 	for (i = 0; i < ubq->q_depth; i++) {
 		struct ublk_io *io = &ubq->ios[i];
 
-		if (io->flags & UBLK_IO_FLAG_ACTIVE)
+		if (io->flags & UBLK_IO_FLAG_ACTIVE) {
+			pr_devel("%s: done old cmd: qid %d tag %d\n",
+					__func__, ubq->q_id, i);
 			io_uring_cmd_done(io->cmd, UBLK_IO_RES_ABORT, 0);
+			io->flags &= ~UBLK_IO_FLAG_ACTIVE;
+			ubq->nr_io_ready--;
+		}
 	}
 
 	/* all io commands are canceled */
-	ubq->nr_io_ready = 0;
+	WARN_ON_ONCE(ubq->nr_io_ready);
 }
 
 /* Cancel all pending commands, must be called after del_gendisk() returns */
-- 
2.27.0