[v2,0/4] tracing/user_events: Fixes and improvements for 6.4

Message

Beau Belgrave April 25, 2023, 10:51 p.m. UTC

Now that user_events is in for-next we broadened our integration of
user_events. During this integration we found a few things that can
help prevent the debugging of issues for user_events when user
processes use the ABI directly.

The most important thing found is an out of bounds fix with the
write index. If it is negative, an out of bounds access is attempted.
This bug was introduced on one of the very first user_events patches
and remained unseen for a long time. Apologies for not catching that
sooner.

We think users will expect the kernel to always clear the registered
bit when events are unregistered, even if the event is still enabled
in a kernel tracer. The user process could do this after unregistering,
but it seems appropriate for the kernel side to attempt this. We also
discussed if it makes sense for the kernel to allow user processes
to tie multiple events to the same value and bit. While this doesn't
cause any issues on the kernel side, it leads to very undefined
behavior for the user process. Depending on which event gets enabled
when, the bit will vary.

Change history

V2:
Add bracket to complex for_each line.

Add patch to ensure in all cases we only limit up to 10 attempts to
fault-in data if the user is able to cause write failures with
successful fault-in sequences.

Beau Belgrave (4):
  tracing/user_events: Ensure write index cannot be negative
  tracing/user_events: Ensure bit is cleared on unregister
  tracing/user_events: Prevent same address and bit per process
  tracing/user_events: Limit max fault-in attempts

 kernel/trace/trace_events_user.c              | 123 ++++++++++++++++--
 .../testing/selftests/user_events/abi_test.c  |   9 +-
 .../selftests/user_events/ftrace_test.c       |  14 +-
 3 files changed, 130 insertions(+), 16 deletions(-)


base-commit: 88fe1ec75fcb296579e05eaf3807da3ee83137e4