exporting INVARIANTS easily

[Warner Losh]

> On Sep 7, 2015, at 4:48 PM, Pawel Jakub Dawidek <pjd at FreeBSD.org> wrote:
> 
> On Mon, Sep 07, 2015 at 04:18:25PM -0600, Warner Losh wrote:
>> On Mon, Sep 7, 2015 at 3:29 PM, Pawel Jakub Dawidek <pjd at freebsd.org> wrote:
>> 
>>> On Wed, Aug 26, 2015 at 11:20:06AM -0600, Warner Losh wrote:
>>>> Greetings,
>>>> 
>>>> Many of the performance eating features are exported via some kind of
>>> sysctl, usually
>>>> patterned after the case of witness as debug.foo. INVARIANTS isn’t one
>>> of those
>>>> features.
>>>> 
>>>> https://reviews.freebsd.org/D3488
>>>> 
>>>> implements debug.invariants. Please comment.
>>>> 
>>>> I’d thought about adding it to the kern.features sysctl, but thought
>>> better of it since it
>>>> isn’t a facility that people can use.
>>>> 
>>>> If you include the kernel config in the kernel, you can get this
>>> information via
>>>>      config -x | grep INVARIANTS
>>>> but not all kernels do that. This is more robust.
>>>> 
>>>> I also know that you can load some modules compiled INVARIANTS when the
>>> base
>>>> kernel isn’t built that way and this won’t reflect that. There’s no good
>>> want to include
>>>> that information and is an uncommon use case.
>>>> 
>>>> Our use case? We have a raft of test machines. Most run without
>>> INVARIANTS since
>>>> we want to characterize the performance of the release under test. Some
>>> are running
>>>> INVARIANTS since we want to test the robustness as well, even at the
>>> expense of
>>>> some performance. To ensure we don’t accidentally include INVARIANTS
>>> systems
>>>> in the performance number, we’ve adding a key to an internal database
>>> that’s driven
>>>> off this sysctl.
>>>> 
>>>> Comments?
>>> 
>>> As long as the ultimate goal is to have INVARIANTS in GENERIC I'm all
>>> for it! I use to run even production machines with INVARIANTS, which was
>>> helpful to catch VirtualBox's kernel memory corruption, but we've moved
>>> to GENERIC since I wanted to use freebsd-update. Having INVARIANTS in
>>> GENERIC would be great.
>> 
>> 
>> That's not my goal. And I doubt my employer would run it in their
>> kernel config because it costs too much in bandwidth when we're
>> streaming all that video people binge watch.
> 
> Do you have some numbers handy? I'd be interested in seeing how much
> performance drop there is when you have INVARIANTS compiled in, but
> debug.invariants set to 0.

I don’t have any direct numbers, but the belief is that we’ll see between 10-20%
drop in our peak throughput with an INVARIANTS kernel.

Warner

-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 842 bytes
Desc: Message signed with OpenPGP using GPGMail
URL: <http://lists.freebsd.org/pipermail/freebsd-arch/attachments/20150907/1af74551/attachment.bin>