On Mon, 2022-03-28 at 09:43 +0800, Xiaomeng Tong wrote:
> On Sun, 27 Mar 2022 15:20:42 +0000, Trond Myklebust wrote:
> > On Sun, 2022-03-27 at 16:02 +0800, Xiaomeng Tong wrote:
> > > The bug is here:
> > > if (!server ||
> > > server->pnfs_curr_ld->id != dev->cbd_layout_type) {
> > >
> > > The list iterator value 'server' will *always* be set and non-
> > > NULL
> > > by list_for_each_entry_rcu, so it is incorrect to assume that the
> > > iterator value will be NULL if the list is empty or no element is
> > > found (In fact, it will be a bogus pointer to an invalid struct
> > > object containing the HEAD, which is used for above check at next
> > > outer loop). Otherwise it may bypass the check in theory (iif
> > > server->pnfs_curr_ld->id == dev->cbd_layout_type, 'server' now is
> > > a bogus pointer) and lead to invalid memory access passing the
> > > check.
> > >
> > > To fix the bug, use a new variable 'iter' as the list iterator,
> > > while use the original variable 'server' as a dedicated pointer
> > > to
> > > point to the found element.
> > >
> > > Cc: stable@xxxxxxxxxxxxxxx
> > > Fixes: 1be5683b03a76 ("pnfs: CB_NOTIFY_DEVICEID")
> > > Signed-off-by: Xiaomeng Tong <xiam0nd.tong@xxxxxxxxx>
> > > ---
> > > fs/nfs/callback_proc.c | 9 +++++----
> > > 1 file changed, 5 insertions(+), 4 deletions(-)
> > >
> > > diff --git a/fs/nfs/callback_proc.c b/fs/nfs/callback_proc.c
> > > index c343666d9a42..84779785dc8d 100644
> > > --- a/fs/nfs/callback_proc.c
> > > +++ b/fs/nfs/callback_proc.c
> > > @@ -361,7 +361,7 @@ __be32 nfs4_callback_devicenotify(void *argp,
> > > void *resp,
> > > uint32_t i;
> > > __be32 res = 0;
> > > struct nfs_client *clp = cps->clp;
> > > - struct nfs_server *server = NULL;
> > > + struct nfs_server *server = NULL, *iter;
> > >
> > > if (!clp) {
> > > res = cpu_to_be32(NFS4ERR_OP_NOT_IN_SESSION);
> > > @@ -374,10 +374,11 @@ __be32 nfs4_callback_devicenotify(void
> > > *argp,
> > > void *resp,
> > > if (!server ||
> > > server->pnfs_curr_ld->id != dev-
> > > >cbd_layout_type)
> > > {
> > > rcu_read_lock();
> > > - list_for_each_entry_rcu(server, &clp-
> > > > cl_superblocks, client_link)
> > > - if (server->pnfs_curr_ld &&
> > > - server->pnfs_curr_ld->id ==
> > > dev-
> > > > cbd_layout_type) {
> > > + list_for_each_entry_rcu(iter, &clp-
> > > > cl_superblocks, client_link)
> > > + if (iter->pnfs_curr_ld &&
> > > + iter->pnfs_curr_ld->id ==
> > > dev-
> > > > cbd_layout_type) {
> > > rcu_read_unlock();
> > > + server = iter;
> >
> > Hmm... We're not holding any locks on the super block for 'iter'
> > here,
> > so nothing is preventing it from going away while we're.
> >
>
> ok, i am not a 'rcu lock' expert, i will make it hold the
> rcu_read_lock()
> if necessary.
>
> > Given that we really only want a pointer to the struct
> > pnfs_layoutdriver_type anyway, why not just convert the code to
> > save a
> > pointer to that (and do it while holding the rcu_read_lock())?
> >
>
> Maybe it's not that simple. If you only save a pointer to that and
> still
> use 'server' as the list iterator of list_for_each_entry_rcu, there
> could
> be problem.
>
> I.e., if no element found in list_for_each_entry_rcu in the first
> outer
> 'for' loop, and now 'server' is a bogus pointer to an invalid struct,
> and
> continue to go into the second outer 'for' loop, and the check below
> will
> lead to invalid memory access (server->pnfs_curr_ld->id), even can
> potentialy
> be bypassed with crafted data to make the condition false and
> mistakely run
> nfs4_delete_deviceid(server->pnfs_curr_ld, clp, &dev->cbd_dev_id);
> with bogus
> 'server'.
>
> if (!server ||
> server->pnfs_curr_ld->id != dev->cbd_layout_type) {
>
> > The struct pnfs_layoutdriver is always expected to be a statically
> > allocated structure, so it won't go away as long as the pNFS driver
> > module remains loaded.
>
Let's just do the following.
8<-----------------------------------------------
