tracepoint: Fix static call function vs data state mismatch
authorMathieu Desnoyers <mathieu.desnoyers@efficios.com>
Thu, 5 Aug 2021 13:27:16 +0000 (09:27 -0400)
committerSteven Rostedt (VMware) <rostedt@goodmis.org>
Thu, 5 Aug 2021 19:42:08 +0000 (15:42 -0400)
commit231264d6927f6740af36855a622d0e240be9d94c
treea26710d11fd2c955f63083d50421ce9c0fac1a41
parentf7ec4121256393e1d03274acdca73eb18958f27e
tracepoint: Fix static call function vs data state mismatch

On a 1->0->1 callbacks transition, there is an issue with the new
callback using the old callback's data.

Considering __DO_TRACE_CALL:

        do {                                                            \
                struct tracepoint_func *it_func_ptr;                    \
                void *__data;                                           \
                it_func_ptr =                                           \
                        rcu_dereference_raw((&__tracepoint_##name)->funcs); \
                if (it_func_ptr) {                                      \
                        __data = (it_func_ptr)->data;                   \

----> [ delayed here on one CPU (e.g. vcpu preempted by the host) ]

                        static_call(tp_func_##name)(__data, args);      \
                }                                                       \
        } while (0)

It has loaded the tp->funcs of the old callback, so it will try to use the old
data. This can be fixed by adding a RCU sync anywhere in the 1->0->1
transition chain.

On a N->2->1 transition, we need an rcu-sync because you may have a
sequence of 3->2->1 (or 1->2->1) where the element 0 data is unchanged
between 2->1, but was changed from 3->2 (or from 1->2), which may be
observed by the static call. This can be fixed by adding an
unconditional RCU sync in transition 2->1.

Note, this fixes a correctness issue at the cost of adding a tremendous
performance regression to the disabling of tracepoints.

Before this commit:

  # trace-cmd start -e all
  # time trace-cmd start -p nop

  real 0m0.778s
  user 0m0.000s
  sys 0m0.061s

After this commit:

  # trace-cmd start -e all
  # time trace-cmd start -p nop

  real 0m10.593s
  user 0m0.017s
  sys 0m0.259s

A follow up fix will introduce a more lightweight scheme based on RCU
get_state and cond_sync, that will return the performance back to what it
was. As both this change and the lightweight versions are complex on their
own, for bisecting any issues that this may cause, they are kept as two
separate changes.

Link: https://lkml.kernel.org/r/20210805132717.23813-3-mathieu.desnoyers@efficios.com
Link: https://lore.kernel.org/io-uring/4ebea8f0-58c9-e571-fd30-0ce4f6f09c70@samba.org/
Cc: stable@vger.kernel.org
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: "Paul E. McKenney" <paulmck@kernel.org>
Cc: Stefan Metzmacher <metze@samba.org>
Fixes: d25e37d89dd2 ("tracepoint: Optimize using static_call()")
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
kernel/tracepoint.c