On Fri, Nov 10, 2023 at 09:06:56AM +0800, Hou Tao wrote:
> Hi,
>
> On 11/10/2023 3:55 AM, Paul E. McKenney wrote:
> > On Thu, Nov 09, 2023 at 07:55:50AM -0800, Alexei Starovoitov wrote:
> >> On Wed, Nov 8, 2023 at 11:26 PM Hou Tao <houtao@xxxxxxxxxxxxxxx> wrote:
> >>> Hi,
> >>>
> >>> On 11/9/2023 2:36 PM, Martin KaFai Lau wrote:
> >>>> On 11/7/23 6:06 AM, Hou Tao wrote:
> >>>>> From: Hou Tao <houtao1@xxxxxxxxxx>
> >>>>>
> >>>>> bpf_map_of_map_fd_get_ptr() will convert the map fd to the pointer
> >>>>> saved in map-in-map. bpf_map_of_map_fd_put_ptr() will release the
> >>>>> pointer saved in map-in-map. These two helpers will be used by the
> >>>>> following patches to fix the use-after-free problems for map-in-map.
> >>>>>
> >>>>> Signed-off-by: Hou Tao <houtao1@xxxxxxxxxx>
> >>>>> ---
> >>>>> kernel/bpf/map_in_map.c | 51 +++++++++++++++++++++++++++++++++++++++++
> >>>>> kernel/bpf/map_in_map.h | 11 +++++++--
> >>>>> 2 files changed, 60 insertions(+), 2 deletions(-)
> >>>>>
> >>>>>
> >>> SNIP
> >>>>> +void bpf_map_of_map_fd_put_ptr(void *ptr, bool need_defer)
> >>>>> +{
> >>>>> + struct bpf_inner_map_element *element = ptr;
> >>>>> +
> >>>>> + /* Do bpf_map_put() after a RCU grace period and a tasks trace
> >>>>> + * RCU grace period, so it is certain that the bpf program which is
> >>>>> + * manipulating the map now has exited when bpf_map_put() is
> >>>>> called.
> >>>>> + */
> >>>>> + if (need_defer)
> >>>> "need_defer" should only happen from the syscall cmd? Instead of
> >>>> adding rcu_head to each element, how about
> >>>> "synchronize_rcu_mult(call_rcu, call_rcu_tasks)" here?
> >>> No. I have tried the method before, but it didn't work due to dead-lock
> >>> (will mention that in commit message in v2). The reason is that bpf
> >>> syscall program may also do map update through sys_bpf helper. Because
> >>> bpf syscall program is running with sleep-able context and has
> >>> rcu_read_lock_trace being held, so call synchronize_rcu_mult(call_rcu,
> >>> call_rcu_tasks) will lead to dead-lock.
> >> Dead-lock? why?
> >>
> >> I think it's legal to do call_rcu_tasks_trace() while inside RCU CS
> >> or RCU tasks trace CS.
> > Just confirming that this is the case. If invoking call_rcu_tasks_trace()
> > within under either rcu_read_lock() or rcu_read_lock_trace() deadlocks,
> > then there is a bug that needs fixing. ;-)
>
> The case for dead-lock is that calling synchronize_rcu_mult(call_rcu,
> call_rcu_tasks_trace) within under rcu_read_lock_trace() and I think it
> is expected. The case that calling call_rcu_tasks_trace() with
> rcu_read_lock_trace() being held is OK.
Very good, you are quite right. In this particular case, deadlock is
expected behavior.
The problem here is that synchronize_rcu_mult() doesn't just invoke its
arguments, instead, it also waits for all of the corresponding grace
periods to complete. But if you call this while under the protection of
rcu_read_lock_trace(), then synchronize_rcu_mult(call_rcu_tasks_trace)
cannot return until the corresponding rcu_read_unlock_trace() is
reached, but that rcu_read_unlock_trace() cannot be reached until after
synchronize_rcu_mult(call_rcu_tasks_trace) returns.
(I did leave out the call_rcu argument because it does not participate
in this particular deadlock.)
Thanx, Paul
