libceph: fix potential use-after-free on linger ping and resends
authorIlya Dryomov <idryomov@gmail.com>
Sat, 14 May 2022 10:16:47 +0000 (12:16 +0200)
committerIlya Dryomov <idryomov@gmail.com>
Wed, 18 May 2022 19:21:05 +0000 (21:21 +0200)
commit75dbb685f4e8786c33ddef8279bab0eadfb0731f
treeec20a2187adbec36beb60b438ba0e5bb5ec3bcbe
parent42226c989789d8da4af1de0c31070c96726d990c
libceph: fix potential use-after-free on linger ping and resends

request_reinit() is not only ugly as the comment rightfully suggests,
but also unsafe.  Even though it is called with osdc->lock held for
write in all cases, resetting the OSD request refcount can still race
with handle_reply() and result in use-after-free.  Taking linger ping
as an example:

    handle_timeout thread                     handle_reply thread

                                              down_read(&osdc->lock)
                                              req = lookup_request(...)
                                              ...
                                              finish_request(req)  # unregisters
                                              up_read(&osdc->lock)
                                              __complete_request(req)
                                                linger_ping_cb(req)

      # req->r_kref == 2 because handle_reply still holds its ref

    down_write(&osdc->lock)
    send_linger_ping(lreq)
      req = lreq->ping_req  # same req
      # cancel_linger_request is NOT
      # called - handle_reply already
      # unregistered
      request_reinit(req)
        WARN_ON(req->r_kref != 1)  # fires
        request_init(req)
          kref_init(req->r_kref)

                   # req->r_kref == 1 after kref_init

                                              ceph_osdc_put_request(req)
                                                kref_put(req->r_kref)

            # req->r_kref == 0 after kref_put, req is freed

        <further req initialization/use> !!!

This happens because send_linger_ping() always (re)uses the same OSD
request for watch ping requests, relying on cancel_linger_request() to
unregister it from the OSD client and rip its messages out from the
messenger.  send_linger() does the same for watch/notify registration
and watch reconnect requests.  Unfortunately cancel_request() doesn't
guarantee that after it returns the OSD client would be completely done
with the OSD request -- a ref could still be held and the callback (if
specified) could still be invoked too.

The original motivation for request_reinit() was inability to deal with
allocation failures in send_linger() and send_linger_ping().  Switching
to using osdc->req_mempool (currently only used by CephFS) respects that
and allows us to get rid of request_reinit().

Cc: stable@vger.kernel.org
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Xiubo Li <xiubli@redhat.com>
Acked-by: Jeff Layton <jlayton@kernel.org>
include/linux/ceph/osd_client.h
net/ceph/osd_client.c