Various improvements to synchronization in osnoise/timerlat.

v6:
 - Revived the "Always check OSN_WORKLOAD" and "Split out per-cpu
   variable resets" patches, with fixes, to address minor issues raised
   by Sashiko
 - Other minor Sashiko-requseted fixes

v5:
 - Removed "Split out per-cpu variable resets" due to Sashiko feedback
   about races with NMIs and other tracing hooks; the main motivation
   for that patch was problems with osn->sampling and tlat->thread,
   both of which are now explicitly cleared.

 - Fixed some patch shuffling bugs in fd detachment that were revealed
   by dropping the reset patch

 - Added "Start all eligible cpus from hotplug workfn"

 - Removed "Always check OSN_WORKLOAD" which is no longer needed after the
   hotplug workfn patch

https://lore.kernel.org/all/[email protected]/

v4: fixed a problem with the v3 fix spotted by Sashiko

https://lore.kernel.org/all/[email protected]/

v3: make sure osn->sampling gets cleared on stop, even for non-workload
osnoise (reported by Sashiko)

https://lore.kernel.org/all/[email protected]/

v2: addresses issues raised by Sashiko, as well as a few more problems
discovered in additional testing.

https://lore.kernel.org/all/[email protected]/

v1: https://lore.kernel.org/all/[email protected]/

Crystal Wood (10):
  tracing/osnoise: Variable cleanup
  tracing/osnoise: start_kthread: Always check OSN_WORKLOAD
  tracing/osnoise: Start all eligible cpus from hotplug workfn
  tracing/osnoise: Hold a ref on osn->kthread
  tracing/osnoise: Per-cpu mutexes and fd detachment
  tracing/osnoise: Split out per-cpu variable resets
  tracing/osnoise: kthreads: Verify CPU after migrate_disable()
  tracing/osnoise: Take trace_types_lock in timerlat_fd_open
  tracing/osnoise: Print taint message for timerlat
  tracing/osnoise: Refuse timerlat fd open if tracer inactive

 Documentation/trace/timerlat-tracer.rst |   5 +-
 kernel/trace/trace_osnoise.c            | 480 ++++++++++++++----------
 2 files changed, 287 insertions(+), 198 deletions(-)

-- 
2.54.0


Reply via email to