Changes in v7:

        * Rebase on top of tip/sched/core
        * Hold task_rq_lock() instead of using RCU.
        * Better document that changes to p->uclamp_ require task_rq_lock()
        * Remove smp_{wr}mp()
        * Hold the the tasklist_lock with smp_mp__after_spinlock()
        * Add patch 3 which addresses a splat I've seen while testing.
          static_branch_enable() in __setscheduler_uclamp() was causing it.
          Remove the call outside of the critical section to fix it.


*** v6 cover-letter ***

This series introduces a new sysctl_sched_uclamp_util_min_rt_default to control
at runtime the default boost value of RT tasks.

Full rationale is in patch 1 commit message.

v6 has changed the approach taken in v5 [1] and earlier by moving away from the
lazy update approach that touched the fast path to a synchronous one that is
performed when the write to the procfs entry is done.

for_each_process_thread() is used to update all existing RT tasks now. And to
handle the race with a concurrent fork() we introduce sched_post_fork() in
_do_fork() to ensure a concurrently forked RT tasks gets the right update.

To ensure the race condition is handled correctly, I wrote this small (simple!)
test program:

        https://github.com/qais-yousef/uclamp_test.git

And ran it on 4core x86 system and 8core big.LITTLE juno-r2 system.

>From juno-r2 run, 10 iterations each run:

Without sched_post_fork()

        # ./run.sh
        pid 3105 has 336 but default should be 337
        pid 13162 has 336 but default should be 337
        pid 23256 has 338 but default should be 339
        All forked RT tasks had the correct uclamp.min
        pid 10638 has 334 but default should be 335
        All forked RT tasks had the correct uclamp.min
        pid 30683 has 335 but default should be 336
        pid 8247 has 336 but default should be 337
        pid 18170 has 1024 but default should be 334
        pid 28274 has 336 but default should be 337

With sched_post_fork()

        # ./run.sh
        All forked RT tasks had the correct uclamp.min
        All forked RT tasks had the correct uclamp.min
        All forked RT tasks had the correct uclamp.min
        All forked RT tasks had the correct uclamp.min
        All forked RT tasks had the correct uclamp.min
        All forked RT tasks had the correct uclamp.min
        All forked RT tasks had the correct uclamp.min
        All forked RT tasks had the correct uclamp.min
        All forked RT tasks had the correct uclamp.min
        All forked RT tasks had the correct uclamp.min

Thanks

--
Qais Yousef

[1] https://lore.kernel.org/lkml/[email protected]/

CC: Jonathan Corbet <[email protected]>
CC: Juri Lelli <[email protected]>
CC: Vincent Guittot <[email protected]>
CC: Dietmar Eggemann <[email protected]>
CC: Steven Rostedt <[email protected]>
CC: Ben Segall <[email protected]>
CC: Mel Gorman <[email protected]>
CC: Luis Chamberlain <[email protected]>
CC: Kees Cook <[email protected]>
CC: Iurii Zaikin <[email protected]>
CC: Quentin Perret <[email protected]>
CC: Valentin Schneider <[email protected]>
CC: Patrick Bellasi <[email protected]>
CC: Pavan Kondeti <[email protected]>
CC: [email protected]
CC: [email protected]
CC: [email protected]


Qais Yousef (3):
  sched/uclamp: Add a new sysctl to control RT default boost value
  Documentation/sysctl: Document uclamp sysctl knobs
  sched/uclamp: Fix a deadlock when enabling uclamp static key

 Documentation/admin-guide/sysctl/kernel.rst |  54 +++++++
 include/linux/sched.h                       |  10 +-
 include/linux/sched/sysctl.h                |   1 +
 include/linux/sched/task.h                  |   1 +
 kernel/fork.c                               |   1 +
 kernel/sched/core.c                         | 149 ++++++++++++++++++--
 kernel/sysctl.c                             |   7 +
 7 files changed, 208 insertions(+), 15 deletions(-)

-- 
2.17.1

Reply via email to