linux-kernel - Re: [PATCH] locking: Remove rt_rwlock_is

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [day] [month] [year] [list]

Date:   Mon, 13 Sep 2021 13:20:32 +0200
From:   Peter Zijlstra <peterz@...radead.org>
To:     Sebastian Andrzej Siewior <bigeasy@...utronix.de>
Cc:     linux-kernel@...r.kernel.org, Ingo Molnar <mingo@...hat.com>,
        Will Deacon <will@...nel.org>,
        Waiman Long <longman@...hat.com>,
        Boqun Feng <boqun.feng@...il.com>,
        Thomas Gleixner <tglx@...utronix.de>
Subject: Re: [PATCH] locking: Remove rt_rwlock_is_contended()

On Fri, Sep 10, 2021 at 06:37:04PM +0200, Sebastian Andrzej Siewior wrote:
> On 2021-09-10 18:16:14 [+0200], Peter Zijlstra wrote:
> > On Tue, Sep 07, 2021 at 12:34:58PM +0200, Sebastian Andrzej Siewior wrote:

> Yes. I got arguments against it after sleeping :)

Sleep is magical :-)

> > AFAICT the _is_contended() can still use useful even with preemption,
> > the typicla use case is a long lock-holder deciding to drop the lock in
> > order to let someone else in. That still works with preemptible locks,
> > no?
> 
> Sure. We can do that. Then we should look into:
> - fixing rwsem_is_contended() for the writer. The writer always observes
>   true even with no waiter around.

Right, that function does look somewhat dodgy. I'm thinking the current
function returns true if there's more than a single reader present (or a
writer) present, which is not the same.

I suppose it shoud return something like:

  for a writer: rt_mutex_is_contended(&rwb->rtmutex);
  for a reader: rt_mutex_is_locked(&rwb->rtmutex);

However, given the below arguments,,,

> - checking the top waiter list vs priority of the lock owner/current. If
>   the current lock owner has the highest priority then the unlock+lock
>   is probably pointless as he regains the lock.
>   For the spin_lock() case, if the owner is SCHED_OTHER and the waiter
>   is SCHED_OTHER then unlock+lock will give the lock to the previous
>   owner due to rt_mutex_steal() working in his favour. Unless there is a
>   preemption.

That is a good argument against all this; I had not considered that.

> - reader checking for contention is probably pointless. It works with a
>   pending writer and one reader since a second reader will hold-off the
>   writer from acquiring the lock. Also if the reader does unlock+lock
>   then writer might not be quick enough.

Should be fixable with a handoff, but yeah.

OK, I suppose the safe and easy option is to never report contention as
per your latest patch, and if/when someone complains about it, they can
sort through these issues :-)