linux-kernel - Re: Revisiting the current set of nohz / housekeeping flags before export

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]

Message-ID: <20220211165219.GE588079@lothringen>
Date:   Fri, 11 Feb 2022 17:52:19 +0100
From:   Frederic Weisbecker <frederic@...nel.org>
To:     Paul Gortmaker <paul.gortmaker@...driver.com>
Cc:     Peter Zijlstra <peterz@...radead.org>,
        LKML <linux-kernel@...r.kernel.org>, Tejun Heo <tj@...nel.org>,
        Christoph Lameter <cl@...two.de>,
        Juri Lelli <juri.lelli@...hat.com>,
        Alex Belits <abelits@...vell.com>,
        Nitesh Lal <nilal@...hat.com>,
        Thomas Gleixner <tglx@...utronix.de>,
        Nicolas Saenz <nsaenzju@...hat.com>,
        "Paul E . McKenney" <paulmck@...nel.org>,
        Phil Auld <pauld@...hat.com>,
        Marcelo Tosatti <mtosatti@...hat.com>,
        Zefan Li <lizefan.x@...edance.com>
Subject: Re: Revisiting the current set of nohz / housekeeping flags before
 export

On Tue, Feb 08, 2022 at 11:47:49AM -0500, Paul Gortmaker wrote:
> The isolation flags [some/all?] are someday going to be exposed via cgroups.
> Frederic has a series out for review now that moves us closer to doing that:
> 
> https://lore.kernel.org/all/20220207155910.527133-1-frederic@kernel.org/
> 
> They will be user exposed tuning knobs and no longer just hidden away in
> source with a subset linked to boot args used by a small set of people.
> 
> As such, we'll need clear user-facing descriptions of each, and specifics of
> what they control/alter.  Which was the initial thought that got me here.
> 
> When the 1st group of flags was introduced (2017) they were split at a fine
> grained level, leaving the door open to re-merge some flags later if natural
> groupings arose. (see de201559df8 log)
> 
> Prior to the elevated userspace exposure they'll get via cgroups (and via
> adding a Documentation/isolation.rst or similar) it probably makes sense
> to revisit possible flag merges now, before they become cemented into API
> and thus essentially leave us stuck with the choices forever.
> 
> Open questions:
> -should HK_FLAG_SCHED be squashed into HK_FLAG_MISC ?  It isnt set and the
>  name can be misleading to new users, in that it "sounds like" the main
>  isolation flag (vs the "real" one which is essentially !HK_FLAG_DOMAIN)
> 
> -should HK_FLAG_RCU be squashed into ... ?  HK_FLAG_MISC ?  It is only used
>  for rcu_online/offline of a CPU and the name might inadvertently suggest
>  that it is somehow related to rcu_nocbs= (when it isn't).
> 
> -do we need HK_FLAG_WQ ?  Currently always or'd with DOMAIN.  Or should we
>  change to select from WQ and then fall back to DOMAIN iff WQ set is empty?
> 
> To better address this, and answer "how did we get here, and who is using
> what flags and where currently" I found myself making some notes to gather
> all that kind of information in one place.
> 
> So what follows below are my rough notes on the housekeeping flags - a
> combination of snippets of source, and commit references etc.  Maybe it
> provides others a shortcut to the overall picture w/o duplicating the work.
> 
> From here, hopefully we can decide if we are OK with the flags as they are,
> and I can take what remains and try and address the documentation I think
> we'll need, as per what I outlined at the top.
> 
> Paul.

Wow, that's quite a deep analysis and archive digging you did!

So, I don't intend to necessarily map all HK_FLAG_* to cpuset. These should be
processed one by one as we add each cpuset.isolation.$FEATURE files over time.

But we certainly need to plan ahead:


1) Unbound features (HK_FLAG_TIMER, HK_FLAG_WQ, HK_FLAG_KTHREAD)

All unbound kernel works could be gathered into a single feature/file.
In this group we have:

_ HK_FLAG_TIMER: unbound timers
_ HK_FLAG_WQ: unbound workqueues
_ HK_FLAG_KTHREAD: unbound kthreads

Now the question is: do we want to keep things finegrained and have a seperate
file for each:

    cpuset.isolation.unbound_timers
    cpuset.isolation.unbound_workqueues
    cpuset.isolation.unbound_kthreads

or do we want to group them:

    cpuset.isolation.unbound

Also note that workqueues affinity can already be controlled over sysfs. We can
still have both interfaces relying on the same cpumask though.


2) HK_FLAG_TICK

Needs its own cpuset.isolation file

3) HK_FLAG_DOMAIN

Cpuset already produces a similar behaviour with cpuset.sched_load_balance. We
just need to map isolcpus to it somehow (add a "cpuset" flag to "isolcpus=" boot
parameter?)

4) HK_FLAG_RCU

A specialization of HK_FLAG_KTHREAD, not sure if it's necessary, could be merged
into HK_FLAG_KTHREAD?

5) HK_FLAG_MANAGED_IRQ

Needs its own cpuset.isolation file

6) The leftovers: HK_FLAG_SCHED and HK_FLAG_MISC

I seem to remember HK_FLAG_SCHED was a bit messy and not well defined. I'll
check your below references. And HK_FLAG_MISC as well...

Oh and I intend to add HK_FLAG_RCU_NOCB since this also should have its own
file just in case it's ever used without nohz_full.

Thanks!