Messages in this thread | ![/](/images/icornerl.gif) | | From | Alexander Shishkin <> | Subject | [PATCH 0/7] perf: Untangle aux refcounting | Date | Thu, 3 Dec 2015 12:32:35 +0200 |
| |
Hi Peter,
As we discussed, here's a patchset that tweaks aux refcounting so that freeing stuff in NMI context is no longer possible. Also, as discussed, I tried to generalize the unlock context - task_function_call - lock context - check stuff - repeat sequence into a helper and converted the existing offenders to use that instead.
I'm not very happy with the way perf_mmap_close turned out, but I figured it's an unmap path and it's at least a start. At some point, a better way of doing things to events on rb::event_list may have to be devised.
For bisectability purposes, it *may* make sense to bring the last two patches forward as they should work just fine with or without the perf_event_stop() trick in mmap_close.
I may have gone slightly over the top with the verbal explanations around what's happening, but at least we won't have to dig up old irc conversations next time it blows up (which it won't do).
Alexander Shishkin (7): perf: Refuse to begin aux transaction after aux_mmap_count drops perf: Generalize task_function_call()ers perf: Add a helper to stop running events perf: Free aux pages in unmap path perf: Document aux api usage perf/x86/intel/pt: Move transaction start/stop to pmu start/stop callbacks perf/x86/intel/bts: Move transaction start/stop to start/stop callbacks
arch/x86/kernel/cpu/perf_event_intel_bts.c | 105 +++++------ arch/x86/kernel/cpu/perf_event_intel_pt.c | 85 +++------ include/linux/perf_event.h | 1 + kernel/events/core.c | 287 ++++++++++++++++------------- kernel/events/internal.h | 1 - kernel/events/ring_buffer.c | 52 +++--- 6 files changed, 262 insertions(+), 269 deletions(-)
-- 2.6.2
| ![\](/images/icornerr.gif) |