On 2021/08/09 14:28, Ming Lei wrote:

On Mon, Aug 09, 2021 at 11:09:27AM +0800, Yu Kuai wrote:

quoted

blk_mq_tag_to_rq() might return freed request, use
blk_mq_get_rq_by_tag() instead.

Signed-off-by: Yu Kuai <redacted>
---
  drivers/block/nbd.c | 11 ++++++-----
  1 file changed, 6 insertions(+), 5 deletions(-)

diff --git a/drivers/block/nbd.c b/drivers/block/nbd.c
index c38317979f74..9e56975a8eee 100644
--- a/drivers/block/nbd.c
+++ b/drivers/block/nbd.c

@@ -713,11 +713,10 @@ static struct nbd_cmd *nbd_read_stat(struct nbd_device *nbd, int index)
  	tag = nbd_handle_to_tag(handle);
  	hwq = blk_mq_unique_tag_to_hwq(tag);
  	if (hwq < nbd->tag_set.nr_hw_queues)
-		req = blk_mq_tag_to_rq(nbd->tag_set.tags[hwq],
-				       blk_mq_unique_tag_to_tag(tag));
-	if (!req || !blk_mq_request_started(req)) {
-		dev_err(disk_to_dev(nbd->disk), "Unexpected reply (%d) %p\n",
-			tag, req);
+		req = blk_mq_get_rq_by_tag(nbd->tag_set.tags[hwq],
+					   blk_mq_unique_tag_to_tag(tag));
+	if (!req) {
+		dev_err(disk_to_dev(nbd->disk), "Unexpected reply %d\n", tag);
  		return ERR_PTR(-ENOENT);
  	}
  	trace_nbd_header_received(req, handle);

@@ -779,6 +778,8 @@ static struct nbd_cmd *nbd_read_stat(struct nbd_device *nbd, int index)
  	}
  out:
  	trace_nbd_payload_received(req, handle);
+	if (req)
+		blk_mq_put_rq_ref(req);
  	mutex_unlock(&cmd->lock);
  	return ret ? ERR_PTR(ret) : cmd;

After blk_mq_put_rq_ref() returns, this request may have been freed,
so the returned 'cmd' may have been freed too.

As I replied in your another thread, it is driver's responsibility to
cover race between normal completion and timeout/error handling, that
means the caller of blk_mq_tag_to_rq need to make sure that the request
represented by the passed 'tag' can't be freed.

Hi, Ming

There are two problems here in nbd, both reported by our syzkaller.

The first is that blk_mq_tag_to_rq() returned a freed request, which is
because tags->static_rq[] is freed without clearing tags->rq[].
Syzkaller log shows that a reply package is sent to client without
the client's request package. And this patch is trying to solve this
problem.

The second is that flush_end_io() decrement it's refcount to -1. I guess
this is because nbd_clear_que concurrent with normal completion /
timeout / error handling, and somehow trigger the problem.
However I'm still trying to understand the logic in nbd. I tried to
add a check in flush_end_io() to fix this, as you replied, it's
driver's responsibility to fix the problem.

It seems that the two problems are not related. And this patch is ok
to fix the first problem. Any suggestions?

Thanks
Kuai

I'd suggest to understand why nbd_read_stat()/blk_mq_tag_to_rq() may return
one freed request first, who frees the request and how when calling
blk_mq_tag_to_rq() before figuring out solutions.