Re: [PATCH v4 2/6] nbd: convert to use blk_mq_find_and_get_req()

On Wed, Sep 08, 2021 at 04:29:57PM +0800, yukuai (C) wrote:

On 2021/09/08 16:00, Ming Lei wrote:

quoted

On Wed, Sep 08, 2021 at 03:37:06PM +0800, yukuai (C) wrote:

quoted

On 2021/09/08 15:30, Ming Lei wrote:

quoted

quoted

+put_req:
+	if (req)
+		blk_mq_put_rq_ref(req);
   	return ret ? ERR_PTR(ret) : cmd;

After the request's refcnt is dropped, it can be freed immediately, then
one stale command is returned to caller.

Hi, Ming

It's right this patch leave this problem. Please take a look at patch 3
and patch 4, the problem should be fixed with these patches.

Not see it is addressed in patch 3 & 4, and it is one obvious fault in
patch 2, please fix it from beginning by moving the refcnt drop
into recv_work().

Hi, Ming

With patch 3 & 4:

if nbd_read_stat() return a valid cmd, then the refcnt should not drop
to 0 before blk_mq_complete_request() in recv_work().

The valid cmd won't be timed out just between nbd_read_stat() and
calling blk_mq_complete_request()?

Yeah, the issue is addressed by patch 4, then please put 2 after
3 & 4, and suggest to add comment why request ref won't drop to zero
in recv_work().

if nbd_read_stat() failed, it won't be a problem if the request is freed
immediately when refcnt is dropped in nbd_read_stat().

That's why I said that the problem will be fixed.

BTW, if we move the refcnt drop into recv_work, we can only do that if
nbd_read_stat() return a valid cmd. If we get a valid rq and failed in
the following checkings in nbd_read_stat(), it's better to drop the
refcnt in nbd_read_stat().

The usual pattern is to drop the refcnt on when the object isn't used
any more, so it is perfectly fine to hold the ref until that time.


Thanks,
Ming