Received-SPF: pass (google.com: domain of linux-nfs-owner@vger.kernel.org designates 2620:137:e000::3:7 as permitted sender) client-ip=2620:137:e000::3:7;
Date:   Tue, 26 Sep 2023 22:25:15 +0100
From:   Al Viro <viro@zeniv.linux.org.uk>
To:     Christoph Hellwig <hch@lst.de>
Cc:     Christian Brauner <brauner@kernel.org>,
        Heiko Carstens <hca@linux.ibm.com>,
        Vasily Gorbik <gor@linux.ibm.com>,
        Alexander Gordeev <agordeev@linux.ibm.com>,
        Fenghua Yu <fenghua.yu@intel.com>,
        Reinette Chatre <reinette.chatre@intel.com>,
        Miquel Raynal <miquel.raynal@bootlin.com>,
        Richard Weinberger <richard@nod.at>,
        Vignesh Raghavendra <vigneshr@ti.com>,
        Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com>,
        Tejun Heo <tj@kernel.org>,
        Trond Myklebust <trond.myklebust@hammerspace.com>,
        Anna Schumaker <anna@kernel.org>,
        Kees Cook <keescook@chromium.org>,
        Damien Le Moal <dlemoal@kernel.org>,
        Naohiro Aota <naohiro.aota@wdc.com>,
        Greg Kroah-Hartman <gregkh@linuxfoundation.org>,
        linux-usb@vger.kernel.org, linux-kernel@vger.kernel.org,
        linux-s390@vger.kernel.org, linux-rdma@vger.kernel.org,
        linux-nfs@vger.kernel.org, linux-hardening@vger.kernel.org,
        cgroups@vger.kernel.org
Subject: Re: [PATCH 03/19] fs: release anon dev_t in deactivate_locked_super
Message-ID: <20230926212515.GN800259@ZenIV>
References: <20230913111013.77623-1-hch@lst.de>
 <20230913111013.77623-4-hch@lst.de>
 <20230913232712.GC800259@ZenIV>
 <20230926093834.GB13806@lst.de>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20230926093834.GB13806@lst.de>
Sender: Al Viro <viro@ftp.linux.org.uk>
Precedence: bulk

On Tue, Sep 26, 2023 at 11:38:34AM +0200, Christoph Hellwig wrote:

> How?
> 
> Old sequence before his patch:
> 
> 	deactivate_locked_super()
> 	  -> kill_anon_super()
> 	    -> generic_shutdown_super()
> 	    -> kill_super_notify()
> 	    -> free_anon_bdev()
> 	  -> kill_super_notify()
> 
> New sequence with this patch:
> 
> 	deactivate_locked_super()
> 	  -> generic_shutdown_super()
> 	    -> kill_super_notify()
> 	    -> free_anon_bdev()
> 

Before your patch: foo_kill_super() calls kill_anon_super(),
which calls kill_super_notify(), which removes the sucker from
the list, then frees ->s_fs_info.  After your patch:
removal from the lists happens via the call of kill_super_notify()
*after* both of your methods had been called, while freeing
->s_fs_info happens from the method call.  IOW, you've restored
the situation prior to "super: ensure valid info".  The whole
point of that commit had been to make sure that we have nothing
in the lists with ->s_fs_info pointing to a freed object.

It's not about free_anon_bdev(); that part is fine - it's the
"we can drop the weird second call site of kill_super_notify()"
thing that is broken.

Al, still slogging through the rcu pathwalk races in the methods...
The latest catch: nfs_set_verifier() can get called on a dentry
that had just been seen to have positive parent, but is not
pinned down.
	grab ->d_lock; OK, we know that dentry won't get freed under us
	fetch ->d_parent->d_inode
	pass that to nfs_verify_change_attribute()
... which assumes that inode it's been given is not NULL.  Normally it
would've been - ->d_lock stabilizes ->d_parent, and negative dentries
obviously have no children.  Except that we might've been just hit
by dentry_kill() due to eviction on memory pressure, got ->d_lock
right after that and proceeded to play with ->d_parent, just as
that parent is going through dentry_kill() from the same eviction on
memory pressure...  If it gets to dentry_unlink_inode() before we get to
fetching ->d_parent->d_inode, nfs_verify_change_attribute(NULL, whatever)
is going to oops...