Re: [PATCH] nbd: pass nbd_sock to nbd_read_reply() instead of index

[Yu Kuai]

Hi,

在 2023/09/28 17:40, Yu Kuai 写道:
> Hi,
>
> 在 2023/09/28 17:24, Ming Lei 写道:
> > On Thu, Sep 28, 2023 at 05:06:40PM +0800, Yu Kuai wrote:
> > > Hi,
> > >
> > > 在 2023/09/28 16:57, Ming Lei 写道:
> > > > On Thu, Sep 28, 2023 at 04:55:03PM +0800, Yu Kuai wrote:
> > > > > Hi,
> > > > >
> > > > > 在 2023/09/28 15:40, Ming Lei 写道:
> > > > > > On Thu, Sep 28, 2023 at 02:03:28PM +0800, Yu Kuai wrote:
> > > > > > > Hi,
> > > > > > >
> > > > > > > 在 2023/09/28 12:05, Ming Lei 写道:
> > > > > > > > On Mon, Sep 11, 2023 at 10:33:08AM +0800, 
> > > > > > > > linan666@xxxxxxxxxxxxxxx wrote:
> > > > > > > > > From: Li Nan <linan122@xxxxxxxxxx>
> > > > > > > > >
> > > > > > > > > If a socket is processing ioctl 'NBD_SET_SOCK', config->socks 
> > > > > > > > > might be
> > > > > > > > > krealloc in nbd_add_socket(), and a garbage request is received 
> > > > > > > > > now, a UAF
> > > > > > > > > may occurs.
> > > > > > > > >
> > > > > > > > >       T1
> > > > > > > > >       nbd_ioctl
> > > > > > > > >        __nbd_ioctl
> > > > > > > > >         nbd_add_socket
> > > > > > > > >          blk_mq_freeze_queue
> > > > > > > > >                 T2
> > > > > > > > >                       recv_work
> > > > > > > > >                        nbd_read_reply
> > > > > > > > >                         sock_xmit
> > > > > > > > >          krealloc config->socks
> > > > > > > > >                    def config->socks
> > > > > > > > >
> > > > > > > > > Pass nbd_sock to nbd_read_reply(). And introduce a new function
> > > > > > > > > sock_xmit_recv(), which differs from sock_xmit only in the way 
> > > > > > > > > it get
> > > > > > > > > socket.
> > > > > > > >
> > > > > > > > I am wondering why not grab queue usage counter before calling 
> > > > > > > > nbd_read_reply()
> > > > > > > > for avoiding such issue, something like the following change:
> > > > > > > >
> > > > > > > > diff --git a/drivers/block/nbd.c b/drivers/block/nbd.c
> > > > > > > > index df1cd0f718b8..09215b605b12 100644
> > > > > > > > --- a/drivers/block/nbd.c
> > > > > > > > +++ b/drivers/block/nbd.c
> > > > > > > > @@ -837,9 +837,6 @@ static void recv_work(struct work_struct *work)
> > > > > > > >          while (1) {
> > > > > > > >              struct nbd_reply reply;
> > > > > > > > -        if (nbd_read_reply(nbd, args->index, &reply))
> > > > > > > > -            break;
> > > > > > > > -
> > > > > > > >              /*
> > > > > > > >               * Grab .q_usage_counter so request pool won't go 
> > > > > > > > away, then no
> > > > > > > >               * request use-after-free is possible during 
> > > > > > > > nbd_handle_reply().
> > > > > > > > @@ -852,6 +849,9 @@ static void recv_work(struct work_struct *work)
> > > > > > > >                  break;
> > > > > > > >              }
> > > > > > >
> > > > > > > This break how nbd works, if there is no reply yet, recv_work() will
> > > > > > > wait for reply in:
> > > > > > >
> > > > > > > nbd_read_reply
> > > > > > >     sock_xmit
> > > > > > >      sock_recvmsg
> > > > > > >
> > > > > > > After this change, recv_work() will just return if there is no io.
> > > > > >
> > > > > > OK, got it, thanks for the input.
> > > > > >
> > > > > > But I feel it isn't necessary & fragile to store one extra 
> > > > > > reference of nsock in
> > > > > > `recv_thread_args`.
> > > > > >
> > > > > > Just run a quick look, the only potential UAF on config->socks 
> > > > > > should be recv_work(),
> > > > > > so you can retrieve the `nsock` reference at the entry of 
> > > > > > recv_work(),
> > > > >
> > > > > I don't understand what you mean retrieve the 'nsock', is following 
> > > > > what
> > > > > you expected?
> > > > >
> > > > > blk_queue_enter() -> prevent concurrent with nbd_add_socket
> > > > > nsock = config->socks[args->index]
> > > > > blk_queue_exit()
> > > >
> > > > Yeah, turns out you do understand, :-)
> > >
> > > Ok, I was not sure about this blk_queue_enter(). By the way, this
> >
> > blk_queue_enter() isn't exported, but you can grab ->config_lock
> > for getting the `nsock`.
> >
> > > remind me of what you did to fix uaf of access queue->mq_hctx[] by
> > > convert the array to xarray.
> > >
> > >
> > > Maybe it's better to covert config->socks[] to xarray to fix this uaf as
> > > well?
> >
> > ->socks[idx] is needed in nbd fast path, so xarray may not be one good 
> > idea
> > since xarray_load() introduces extra load, especially ->socks[] uaf
> > should exist in recv_work() very likely. For other cases, the active
> > block request holds queue usage counter.
>
> Thanks for the explanation, grab 'config_lock' to get 'nsock' in the
> begining sounds good to me.

After reviewing some code, I found that it's wrong to grab config_lock,
because other context will grab such lock and flush_workqueue(), and
there is no gurantee that recv_work() will grab the lock first.

Will it be acceptable to export blk_queue_enter()? I can't think of
other way to retrieve the`nsock` reference at the entry of recv_work().

Thanks,
Kuai

> Kuai
>
> > Thanks,
> > Ming
> >
> > .
>
> .