Re: [PATCH] [v2] TAINT_PERFORMANCE

From: Ingo Molnar
Date: Wed Aug 20 2014 - 04:12:12 EST



* Dave Hansen <dave@xxxxxxxx> wrote:

>
> From: Dave Hansen <dave.hansen@xxxxxxxxxxxxxxx>
>
> Changes from v1:
> * remove schedstats
> * add DEBUG_PAGEALLOC and SLUB_DEBUG_ON
>
> --
>
> I have more than once myself been the victim of an accidentally-
> enabled kernel config option being mistaken for a true
> performance problem.
>
> I'm sure I've also taken profiles or performance measurements
> and assumed they were real-world when really I was measuing the
> performance with an option that nobody turns on in production.

Most of these options already announce themselves in the
syslog.

> A warning like this late in boot will help remind folks when
> these kinds of things are enabled.
>
> As for the patch...
>
> I originally wanted this for CONFIG_DEBUG_VM, but I think it also
> applies to things like lockdep and slab debugging. See the patch
> for the list of offending config options. I'm open to adding
> more, but this seemed like a good list to start.

> [ 2.534574] CONFIG_LOCKDEP enabled
> [ 2.536392] Do not use this kernel for performance measurement.

This is workload dependent: for many kernel workloads this is
indeed true. For many user-space workloads it will add very
little overhead.

> [ 2.547189] CPU: 0 PID: 1 Comm: swapper/0 Not tainted 3.16.0-10473-gc8d6637-dirty #800
> [ 2.558075] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Bochs 01/01/2011
> [ 2.564483] 0000000080000000 ffff88009c70be78 ffffffff817ce318 0000000000000000
> [ 2.582505] ffffffff81dca5b6 ffff88009c70be88 ffffffff81dca5e2 ffff88009c70bef8
> [ 2.588589] ffffffff81000377 0000000000000000 0007000700000142 ffffffff81b78968
> [ 2.592638] Call Trace:
> [ 2.593762] [<ffffffff817ce318>] dump_stack+0x4e/0x68

Generating a stack dump that tells us nothing isn't really
useful.

> { TAINT_SOFTLOCKUP, 'L', ' ' },
> + { TAINT_PERFORMANCE, 'Q', ' ' },

Also this looks like a slight abuse of the taint flag: we taint
the kernel if there's a problem with it. But even for many
types of performance measurements, a debug kernel is just fine.
For other types of performance measurements, even a non-debug
kernel option can have big impact.

A better option might be to declare known performance killers
in /proc/config_debug or so, and maybe print them once at the
end of the bootup, with a 'WARNING:' or 'INFO:' prefix. That
way tooling (benchmarks, profilers, etc.) can print them, but
it's also present in the syslog, just in case.

/proc/config_debug is different from /proc/config.gz IKCONFIG,
because it would always be present when performance impacting
options are enabled. So tools would only have to check the
existence of this file, for the simplest test.

In any case I don't think it's a good idea to abuse existing
facilities just to gain attention: you'll get the extra
attention, but the abuse dillutes the utility of those only
tangentially related facilities.

Thanks,

Ingo
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/