Re: [PATCH 4.4 00/11] Fix a potential infinite loop in RT futex-pi scenarios

[Joe Korty <joe.korty@concurrent-rt.com>]

On Mon, Aug 02, 2021 at 09:46:13PM +0800, Zhen Lei wrote:
> Commit 73d786bd043e "futex: Rework inconsistent rt_mutex/futex_q state"
> mentions that it could cause an infinite loop, and will fix it in the later
> patches:
> bebe5b514345f09 futex: Futex_unlock_pi() determinism
> cfafcd117da0216 futex: Rework futex_lock_pi() to use rt_mutex_*_proxy_lock()
> 
> But at the moment they're not backported. In a single-core environment, the
> probability of triggering is high.
> 
> I also backported commit b4abf91047cf ("rtmutex: Make wait_lock irq safe"),
> it fixes a potential deadlock problem. Although it hasn't actually been
> triggered in our environment at the moment.
> 
> Other patches are used to resolve conflicts or fix problems caused by new
> patches.
> 
> 
> Anna-Maria Gleixner (1):
>   rcu: Update documentation of rcu_read_unlock()
> 
> Mike Galbraith (1):
>   futex: Handle transient "ownerless" rtmutex state correctly
> 
> Peter Zijlstra (6):
>   futex: Cleanup refcounting
>   futex,rt_mutex: Introduce rt_mutex_init_waiter()
>   futex: Pull rt_mutex_futex_unlock() out from under hb->lock
>   futex: Rework futex_lock_pi() to use rt_mutex_*_proxy_lock()
>   futex: Futex_unlock_pi() determinism
>   futex,rt_mutex: Fix rt_mutex_cleanup_proxy_lock()
> 
> Thomas Gleixner (3):
>   futex: Rename free_pi_state() to put_pi_state()
>   rtmutex: Make wait_lock irq safe
>   futex: Avoid freeing an active timer
> 
>  include/linux/rcupdate.h        |   4 +-
>  kernel/futex.c                  | 245 +++++++++++++++++++++-----------
>  kernel/locking/rtmutex.c        | 185 +++++++++++++-----------
>  kernel/locking/rtmutex_common.h |   2 +-
>  4 files changed, 262 insertions(+), 174 deletions(-)


To all concerned,

I have verified that this series of patches, when applied
to 4.4.277, passes the futex-unlock-pi replicator I posted
to lkml on July 19.

  Subject: [BUG] 4.4.262: infinite loop in futex_unlock_pi (EAGAIN loop)

Acked-by: Joe Korty <joe.korty@concurrent-rt.com>