Re: [PATCH] fuse: virtiofs: fix illegal inode address access in fuse_release_end

[Zhihao Cheng]

在 2026/5/16 5:07, Darrick J. Wong 写道:
> [add fuse-devel]
>
> On Thu, May 14, 2026 at 08:41:02PM +0800, Zhihao Cheng wrote:
> > When a submount (e.g. virtiofs with fc->auto_submounts=true) is umounted
> > while an async RELEASE request is still pending, an illegal inode address
> > accessing occurs in fuse_release_end()->iput().
> >
> > Trigger process:
> >   1. virtiofs has a submount; user opens and closes a file under it
> >   2. Close calls fuse_file_put() with sync=false, sending RELEASE
> >      asynchronously
> >   3. fuse_release_end() is scheduled to run later via igrab() holding
> >      inode ref
> >   4. File is freed, mount/dentry refcounts are released
> >   5. User umounts the submount; fuse connection detects remaining
> >      superblock and does NOT flush the connection's requests
> >   6. generic_shutdown_super() destroys the superblock and poisons busy
> >      inodes' inode->i_sb = VFS_PTR_POISON
> >   7. Later, fuse_request_end() calls fuse_release_end() which does
> >      iput(inode)
> >   8. iput() accesses inode->i_sb->s_op at the poisoned address, crash!
> >
> > There are two solutions to fix it:
> >   Solution A: Hold path reference in fuse_file_put, and put path
> >   synchronously, which could reintroduce the issue fixed by commit
> >   5a18ec176c934 ("fuse: fix hang of single threaded fuseblk filesystem").
> >
> >   Solution B (chosen): Use synchronous RELEASE for auto_submounts (which
> >   is only supported by virtiofs). The virtiofsd(fuse daemon) and virtiofs
> >   won't work together under one same kernel instance, so the problem fixed
> >   by commit 26e5c67deb2e ("fuse: fix livelock in synchronous file put from
> >   fuseblk workers") won't be brought back in virtiofs.
>
> Does the weird AIO "kernel breaks if you close the fd before all the IOs
> complete" behavior happen on virtiofsd?

Hi Darrick,
IMHO, the problem fixed by commit 26e5c67deb2e1f42a9("fuse: fix livelock 
in synchronous file put from fuseblk workers") happens as following process:
   fuse user              fuse daemon
fd = open(file)
io_submit_one
 req->ki_filp = fget(iocb->aio_fildes)
close(fd) // file->f_ref = 1, file won't be released
                        process FUSE_WRITE req
                        write(/dev/fuse, reply)
                         fuse_request_end
                          fuse_aio_complete_req
                           aio_complete_rw
                            fput(iocb->ki_filp)
                             task_work_add
                         task_work_run
                          __fput
                           fuse_release
                            __fuse_simple_request
                             request_wait_answer

For viriofs, the user(virtiofs) runs in kernel A(guest), the 
daemon(virtiofsd) runs kernel B(host), and the fuse_request_end is 
called by virtio_fs_requests_done_work -> virtio_fs_complete_req_work -> 
virtio_fs_request_complete, which is under the kworker context(kernel A, 
guest), so it won't block the daemon thread(kernel B, host).

> --D
>
> > Fetch a reproducer in https://bugzilla.kernel.org/show_bug.cgi?id=221519.
> >
> > Fixes: 26e5c67deb2e ("fuse: fix livelock in synchronous file put from fuseblk workers")
> > Signed-off-by: Zhihao Cheng <chengzhihao1@xxxxxxxxxx>
> > ---
> >   fs/fuse/file.c | 19 +++++++++++++------
> >   1 file changed, 13 insertions(+), 6 deletions(-)
> >
> > diff --git a/fs/fuse/file.c b/fs/fuse/file.c
> > index c59452d60b8d..a6192b96d861 100644
> > --- a/fs/fuse/file.c
> > +++ b/fs/fuse/file.c
> > @@ -375,13 +375,20 @@ void fuse_file_release(struct inode *inode, struct fuse_file *ff,
> >   	 * synchronous RELEASE is allowed (and desirable) in this case
> >   	 * because the server can be trusted not to screw up.
> >   	 *
> > -	 * Always use the asynchronous file put because the current thread
> > -	 * might be the fuse server.  This can happen if a process starts some
> > -	 * aio and closes the fd before the aio completes.  Since aio takes its
> > -	 * own ref to the file, the IO completion has to drop the ref, which is
> > -	 * how the fuse server can end up closing its clients' files.
> > +	 * For auto_submounts (e.g. virtiofs), always use synchronous
> > +	 * release to avoid illegal inode address access when umount
> > +	 * happens before async release completes. The async release
> > +	 * holds inode reference via igrab(), but umount can shutdown
> > +	 * superblock and poison inode->i_sb before release ends,
> > +	 * causing crash in fuse_release_end()->iput(). Otherwise,
> > +	 * always use the asynchronous file put because the current
> > +	 * thread might be the fuse server. This can happen if a
> > +	 * process starts some aio and closes the fd before the aio
> > +	 * completes. Since aio takes its own ref to the file, the IO
> > +	 * completion has to drop the ref, which is how the fuse server
> > +	 * can end up closing its clients' files.
> >   	 */
> > -	fuse_file_put(ff, false);
> > +	fuse_file_put(ff, ff->fm->fc->auto_submounts);
> >   }
> >   
> >   void fuse_release_common(struct file *file, bool isdir)
> > --
> > 2.52.0
>
> .