On Thu, Jan 16, 2020 at 09:25:42PM +0900, Misono Tomohiro wrote:
> > From: Stefan Hajnoczi <stefanha@redhat.com>
> > 
> > If thread A is using an inode it must not be deleted by thread B when
> > processing a FUSE_FORGET request.
> > 
> > The FUSE protocol itself already has a counter called nlookup that is
> > used in FUSE_FORGET messages.  We cannot trust this counter since the
> > untrusted client can manipulate it via FUSE_FORGET messages.
> > 
> > Introduce a new refcount to keep inodes alive for the required lifespan.
> > lo_inode_put() must be called to release a reference.  FUSE's nlookup
> > counter holds exactly one reference so that the inode stays alive as
> > long as the client still wants to remember it.
> > 
> > Note that the lo_inode->is_symlink field is moved to avoid creating a
> > hole in the struct due to struct field alignment.
> > 
> > Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
> > ---
> >  tools/virtiofsd/passthrough_ll.c | 168 ++++++++++++++++++++++++++-----
> >  1 file changed, 145 insertions(+), 23 deletions(-)
> > 
> > diff --git a/tools/virtiofsd/passthrough_ll.c b/tools/virtiofsd/passthrough_ll.c
> > index b19c9ee328..8f4ab8351c 100644
> > --- a/tools/virtiofsd/passthrough_ll.c
> > +++ b/tools/virtiofsd/passthrough_ll.c
> > @@ -99,7 +99,13 @@ struct lo_key {
> >  
> >  struct lo_inode {
> >      int fd;
> > -    bool is_symlink;
> > +
> > +    /*
> > +     * Atomic reference count for this object.  The nlookup field holds a
> > +     * reference and release it when nlookup reaches 0.
> > +     */
> > +    gint refcount;
> > +
> >      struct lo_key key;
> >  
> >      /*
> > @@ -118,6 +124,8 @@ struct lo_inode {
> >      fuse_ino_t fuse_ino;
> >      pthread_mutex_t plock_mutex;
> >      GHashTable *posix_locks; /* protected by lo_inode->plock_mutex */
> > +
> > +    bool is_symlink;
> >  };
> >  
> >  struct lo_cred {
> > @@ -473,6 +481,23 @@ static ssize_t lo_add_inode_mapping(fuse_req_t req, struct lo_inode *inode)
> >      return elem - lo_data(req)->ino_map.elems;
> >  }
> >  
> > +static void lo_inode_put(struct lo_data *lo, struct lo_inode **inodep)
> > +{
> > +    struct lo_inode *inode = *inodep;
> > +
> > +    if (!inode) {
> > +        return;
> > +    }
> > +
> > +    *inodep = NULL;
> > +
> > +    if (g_atomic_int_dec_and_test(&inode->refcount)) {
> > +        close(inode->fd);
> > +        free(inode);
> > +    }
> > +}
> > +
> > +/* Caller must release refcount using lo_inode_put() */
> >  static struct lo_inode *lo_inode(fuse_req_t req, fuse_ino_t ino)
> >  {
> >      struct lo_data *lo = lo_data(req);
> > @@ -480,6 +505,9 @@ static struct lo_inode *lo_inode(fuse_req_t req, fuse_ino_t ino)
> >  
> >      pthread_mutex_lock(&lo->mutex);
> >      elem = lo_map_get(&lo->ino_map, ino);
> > +    if (elem) {
> > +        g_atomic_int_inc(&elem->inode->refcount);
> > +    }
> >      pthread_mutex_unlock(&lo->mutex);
> >  
> >      if (!elem) {
> > @@ -489,10 +517,23 @@ static struct lo_inode *lo_inode(fuse_req_t req, fuse_ino_t ino)
> >      return elem->inode;
> >  }
> >  
> > +/*
> > + * TODO Remove this helper and force callers to hold an inode refcount until
> > + * they are done with the fd.  This will be done in a later patch to make
> > + * review easier.
> > + */
> >  static int lo_fd(fuse_req_t req, fuse_ino_t ino)
> >  {
> >      struct lo_inode *inode = lo_inode(req, ino);
> > -    return inode ? inode->fd : -1;
> > +    int fd;
> > +
> > +    if (!inode) {
> > +        return -1;
> > +    }
> > +
> > +    fd = inode->fd;
> > +    lo_inode_put(lo_data(req), &inode);
> > +    return fd;
> >  }
> >  
> >  static void lo_init(void *userdata, struct fuse_conn_info *conn)
> > @@ -547,6 +588,10 @@ static void lo_getattr(fuse_req_t req, fuse_ino_t ino,
> >      fuse_reply_attr(req, &buf, lo->timeout);
> >  }
> >  
> > +/*
> > + * Increments parent->nlookup and caller must release refcount using
> > + * lo_inode_put(&parent).
> > + */
> >  static int lo_parent_and_name(struct lo_data *lo, struct lo_inode *inode,
> >                                char path[PATH_MAX], struct lo_inode **parent)
> >  {
> > @@ -584,6 +629,7 @@ retry:
> >          p = &lo->root;
> >          pthread_mutex_lock(&lo->mutex);
> >          p->nlookup++;
> > +        g_atomic_int_inc(&p->refcount);
> >          pthread_mutex_unlock(&lo->mutex);
> >      } else {
> >          *last = '\0';
> 
> We need lo_ionde_put() in error path, right?:
> https://gitlab.com/virtio-fs/qemu/blob/virtio-fs-as-posted-2019-12-12/tools/virtiofsd/passthrough_ll.c#L680

Yes, thanks for spotting this bug!  The lo_parent_and_name() code should
look like this:

  fail_unref:
      unref_inode_lolocked(lo, p, 1);
      lo_inode_put(lo, &p);
  ...

> nit: if yes, unref_inode_lolocked() is always paired with lo_inode_put().
> So how about combine them in one function? As p->nloockup and p->refcount
> are both incremented in one place (lo_find/lo_parent_and_name) in these case,
> it seems natural for me to decrement them in one function as well.

Nice idea.  I would also drop the nlookup argument - this function will
only be used with nlookup=1.

Stefan