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>
75dbb685f4