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;
Date:   Tue, 15 Feb 2022 13:37:40 +0000
From:   Al Viro <viro@zeniv.linux.org.uk>
To:     Miklos Szeredi <miklos@szeredi.hu>
Cc:     Xavier Roche <xavier.roche@algolia.com>,
        linux-kernel@vger.kernel.org, linux-fsdevel@vger.kernel.org,
        "Aneesh Kumar K.V" <aneesh.kumar@linux.vnet.ibm.com>
Subject: Re: race between vfs_rename and do_linkat (mv and link)
Message-ID: <YguspMvu6M6NJ1hL@zeniv-ca.linux.org.uk>
References: <20220214210708.GA2167841@xavier-xps>
 <CAJfpegvVKWHhhXwOi9jDUOJi2BnYSDxZQrp1_RRrpVjjZ3Rs2w@mail.gmail.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <CAJfpegvVKWHhhXwOi9jDUOJi2BnYSDxZQrp1_RRrpVjjZ3Rs2w@mail.gmail.com>
Sender: Al Viro <viro@ftp.linux.org.uk>
Precedence: bulk

On Tue, Feb 15, 2022 at 10:56:29AM +0100, Miklos Szeredi wrote:

> Doing "lock_rename() + lookup last components" would fix this race.

No go - thanks to the possibility of AT_SYMLINK_FOLLOW there.
Think of it - we'd need to
	* lock parents (both at the same time)
	* look up the last component of source
	* if it turns a symlink - unlock parents and repeat the entire
thing for its body, except when asked not to.
	* when we are done with the source, look the last component of
target up

... and then there is sodding -ESTALE handling, with all the elegance
that brings in.

> If this was only done on retry, then that would prevent possible
> performance regressions, at the cost of extra complexity.

Extra compared to the above, that is.  How delightful...