Messages in this thread | | | From | Davidlohr Bueso <> | Subject | [PATCH 0/2] lockless wake-queues | Date | Sun, 19 Apr 2015 12:17:38 -0700 |
| |
Hello,
This series is aimed at addressing some of the futex hash bucket lock hold times by introducing lockless wake-queues for futex_wake.
patch-1: introduces the lockless wake-queue machinery. patch-2: makes use of patch 1 for futexes.
Details in the individual patches.
This was suggested sometime ago by peterz, but due to it potentially causing spurious wakeups, was never given much consideration. However, nowadays, so far, I am reliably booting a 45-core box with peterz's patch to trigger spurious wakeups. While there are drivers out there that do not play nice with schedule(), they can be fixed over time -- while this is a production problem for some customers). Furthermore, after some auditing, there really aren't that many, it a lot of cases, those functions that end up calling schedule are merely wrapped in a loop, so just not clear at first sight.
Other code that could be migrated to lockless wake_q includes: s390 pfault_interrupt, sysv msg (sem already does the lockless wakeups without the ipc object lock), among others. However, futexes are the only users afaik that solve a _real_ issue.
Applies on top of Linus' tree 4.0-64fb1d0e975e. Thanks!
sched: lockless wake-queues futex: lockless wakeups
include/linux/sched.h | 30 ++++++++++++++++++++++++++++++ kernel/futex.c | 33 +++++++++++++++++---------------- kernel/sched/core.c | 49 +++++++++++++++++++++++++++++++++++++++++++++++++ 3 files changed, 96 insertions(+), 16 deletions(-)
-- 2.1.4
| |