linux-kernel - Re: threads-max observe limits

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 for Android: free password hash cracker in your pocket

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

Date:   Sun, 22 Sep 2019 16:40:26 -0500
From:   ebiederm@...ssion.com (Eric W. Biederman)
To:     Heinrich Schuchardt <xypron.glpk@....de>
Cc:     Michal Hocko <mhocko@...nel.org>,
        Andrew Morton <akpm@...ux-foundation.org>,
        LKML <linux-kernel@...r.kernel.org>
Subject: Re: threads-max observe limits

Heinrich Schuchardt <xypron.glpk@....de> writes:

> Did this patch when applied to the customer's kernel solve any problem?
>
> WebSphere MQ is a messaging application. If it hits the current limits
> of threads-max, there is a bug in the software or in the way that it has
> been set up at the customer. Instead of messing around with the kernel
> the application should be fixed.

While it is true that almost every workload will be buggy if it exceeds
1/8 of memory with just the kernel data structures for threads.  It is
not necessary true of every application.  I can easily imagine cases
up around 1/2 of memory where things could work reasonably.

Further we can exhaust all of memory much more simply in a default
configuration by malloc'ing more memory that in physically present
and zeroing it all.

Henrich, you were the one messed with the kernel by breaking a
reasonable kernel tunable.  AKA you caused a regression.  That violates
the no regression rule.

As much as possible we fix regressions so software that used to work
continues to work.  Removing footguns is not a reason to introduce a
regression.

I do agree that Michal's customer's problem sounds like it is something
else but if the kernel did not have a regression we could focus on the
real problem instead of being side tracked by the regression.

> With this patch you allow administrators to set values that will crash
> their system. And they will not even have a way to find out the limits
> which he should adhere to. So expect a lot of systems to be downed
> this way.

Nope.  The system administrator just setting a higher value whon't crash
their system.  Only using that many resources would crash the system.

Nor is a sysctl like this for discovering the physical limits of a
machine.  Which the current value is vastly inappropriate for.  As
the physical limits of many machines are much higher than 1/8 of memory.

Eric