Re: [PATCH] afs: Fix refcount underflow from error handling race

[Marc Dionne]

On Fri, Dec 8, 2023 at 6:07 PM David Howells <dhowells@xxxxxxxxxx> wrote:
>
> If an AFS cell that has an unreachable (eg. ENETUNREACH) Volume Location
> server listed, an asynchronous probe to one of its addresses may fail
> immediately because sendmsg() returns an error.  When this happens, a
> refcount underflow can happen if certain events hit a very small window.
>
> The way this occurs is:
>
>  (1) There are two levels of "call" object, the afs_call and the
>      rxrpc_call.  Each of them can be transitioned to a "completed" state
>      in the event of success or failure.
>
>  (2) Asynchronous afs_calls are self-referential whilst they are active to
>      prevent them from evaporating when they're not being processed.  This
>      reference is disposed of when the afs_call is completed.
>
>      Note that an afs_call may only be completed once; once completed
>      completing it again will do nothing.
>
>  (3) When a call transmission is made, the app-side rxrpc code queues a Tx
>      buffer for the rxrpc I/O thread to transmit.  The I/O thread invokes
>      sendmsg() to transmit it - and in the case of failure, it transitions
>      the rxrpc_call to the completed state.
>
>  (4) When an rxrpc_call is completed, the app layer is notified.  In this
>      case, the app is kafs and it schedules a work item to process events
>      pertaining to an afs_call.
>
>  (5) When the afs_call event processor is run, it goes down through the
>      RPC-specific handler to afs_extract_data() to retrieve data from rxrpc
>      - and, in this case, it picks up the error from the rxrpc_call and
>      returns it.
>
>      The error is then propagated to the afs_call and that is completed
>      too.  At this point the self-reference is released.
>
>  (6) If the rxrpc I/O thread manages to complete the rxrpc_call within the
>      window between rxrpc_send_data() queuing the request packet and
>      checking for call completion on the way out, then
>      rxrpc_kernel_send_data() will return the error from sendmsg() to the
>      app.
>
>  (7) Then afs_make_call() will see an error and will jump to the error
>      handling path which will attempt to clean up the afs_call.
>
>  (8) The problem comes when the error handling path in afs_make_call()
>      tries to unconditionally drop an async afs_call's self-reference.
>      This self-reference, however, may already have been dropped by
>      afs_extract_data() completing the afs_call
>
>  (9) The refcount underflows when we return to afs_do_probe_vlserver() and
>      that tries to drop its reference on the afs_call.
>
> Fix this by making afs_make_call() attempt to complete the afs_call rather
> than unconditionally putting it.  That way, if afs_extract_data() manages
> to complete the call first, afs_make_call() won't do anything.
>
> The bug can be forced by making do_udp_sendmsg() return -ENETUNREACH and
> sticking an msleep() in rxrpc_send_data() after the 'success:' label.
>
> The error message looks something like:
>
>     refcount_t: underflow; use-after-free.
>     WARNING: CPU: 3 PID: 720 at lib/refcount.c:28 refcount_warn_saturate+0xba/0x110
>     ...
>     RIP: 0010:refcount_warn_saturate+0xba/0x110
>     ...
>     afs_put_call+0x1dc/0x1f0 [kafs]
>     afs_fs_get_capabilities+0x8b/0xe0 [kafs]
>     afs_fs_probe_fileserver+0x188/0x1e0 [kafs]
>     afs_lookup_server+0x3bf/0x3f0 [kafs]
>     afs_alloc_server_list+0x130/0x2e0 [kafs]
>     afs_create_volume+0x162/0x400 [kafs]
>     afs_get_tree+0x266/0x410 [kafs]
>     vfs_get_tree+0x25/0xc0
>     fc_mount+0xe/0x40
>     afs_d_automount+0x1b3/0x390 [kafs]
>     __traverse_mounts+0x8f/0x210
>     step_into+0x340/0x760
>     path_openat+0x13a/0x1260
>     do_filp_open+0xaf/0x160
>     do_sys_openat2+0xaf/0x170
>
> or something like:
>
>     refcount_t: underflow; use-after-free.
>     ...
>     RIP: 0010:refcount_warn_saturate+0x99/0xda
>     ...
>     afs_put_call+0x4a/0x175
>     afs_send_vl_probes+0x108/0x172
>     afs_select_vlserver+0xd6/0x311
>     afs_do_cell_detect_alias+0x5e/0x1e9
>     afs_cell_detect_alias+0x44/0x92
>     afs_validate_fc+0x9d/0x134
>     afs_get_tree+0x20/0x2e6
>     vfs_get_tree+0x1d/0xc9
>     fc_mount+0xe/0x33
>     afs_d_automount+0x48/0x9d
>     __traverse_mounts+0xe0/0x166
>     step_into+0x140/0x274
>     open_last_lookups+0x1c1/0x1df
>     path_openat+0x138/0x1c3
>     do_filp_open+0x55/0xb4
>     do_sys_openat2+0x6c/0xb6
>
> Fixes: 34fa47612bfe ("afs: Fix race in async call refcounting")
> Reported-by: Bill MacAllister <bill@xxxxxxxxxxxxx>
> Closes: https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=1052304
> Suggested-by: Jeffrey E Altman <jaltman@xxxxxxxxxxxx>
> Signed-off-by: David Howells <dhowells@xxxxxxxxxx>
> cc: Marc Dionne <marc.dionne@xxxxxxxxxxxx>
> cc: linux-afs@xxxxxxxxxxxxxxxxxxx
> ---
>  fs/afs/rxrpc.c |    2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/fs/afs/rxrpc.c b/fs/afs/rxrpc.c
> index ed1644e7683f..d642d06a453b 100644
> --- a/fs/afs/rxrpc.c
> +++ b/fs/afs/rxrpc.c
> @@ -424,7 +424,7 @@ void afs_make_call(struct afs_addr_cursor *ac, struct afs_call *call, gfp_t gfp)
>         if (call->async) {
>                 if (cancel_work_sync(&call->async_work))
>                         afs_put_call(call);
> -               afs_put_call(call);
> +               afs_set_call_complete(call, ret, 0);
>         }
>
>         ac->error = ret;

Reviewed-by: Marc Dionne <marc.dionne@xxxxxxxxxxxx>
Tested-by: Marc Dionne <marc.dionne@xxxxxxxxxxxx>