code review 5477057: runtime: madvise and SysUnused for Linux

runtime: madvise and SysUnused for Linux

SysUnused being a direct call to madvise MADV_DONTNEED.

[Sebastien Paolacci, 2011-12-10 10:37:11]

Hello golang-dev@googlegroups.com (cc: golang-dev@googlegroups.com),

I'd like you to review this change to
https://code.google.com/p/go/

[Sebastien Paolacci, 2011-12-10 10:57:32]

Note: I did cheat, the arm part is not tested (I'm currently getting illegal
instructions on tip + qemu).

Getting all archs simultenousaly pushed is desirable here, so I choose to have
an optimistic approach in the hope that it will make things easier / less work
for other people:

-> Could anyone with a working arm setup give a try, at least to ensure that it
does not break anything?

Thanks,
Sebastien

[dave_cheney.net, 2011-12-10 13:33:08]

This builds fine on arm5. Is there any way I can test this works ?

[Sebastien Paolacci, 2011-12-11 08:23:53]

Hello Dave,

Thank you for having tested that.

There's no straightforward way to assess SysUnused behavior since it's
not yet called anywhere yet, it's just a preliminary patch for the
memory scavenger. Making sure I was not breaking the build system was
my very important point here.

Depending on how much hack you want to do, you can import the 5451057
CL, replace SysFree by SysUnused (mheap.c:391), possibly tickPeriod by
2s (mheap.c:330), and strace some app (+ GOGCTRACE=1). You should
eventually see some call to madvise(...., MADV_DONTNEED) and some
memory given back to the OS. Don't spend time on that linked patch
itself, it's going to be deeply rewritten.

I hope however I can find some time to get my arm setup fixed today,
so you maybe waiting a few more hours is enough now.

Sebastien

On Sat, Dec 10, 2011 at 2:33 PM,  <dave@cheney.net> wrote:
> This builds fine on arm5. Is there any way I can test this works ?
>
> http://codereview.appspot.com/5477057/

[dave_cheney.net, 2011-12-11 08:29:51]

Sure thing. A lot of the less well used syscalls don't work properly under qemu,
please feel free to send me any ARM code you need tested. 

Sent from my iPhone

On 11/12/2011, at 19:23, Sébastien Paolacci <sebastien.paolacci@gmail.com>
wrote:

> Hello Dave,
> 
> Thank you for having tested that.
> 
> There's no straightforward way to assess SysUnused behavior since it's
> not yet called anywhere yet, it's just a preliminary patch for the
> memory scavenger. Making sure I was not breaking the build system was
> my very important point here.
> 
> Depending on how much hack you want to do, you can import the 5451057
> CL, replace SysFree by SysUnused (mheap.c:391), possibly tickPeriod by
> 2s (mheap.c:330), and strace some app (+ GOGCTRACE=1). You should
> eventually see some call to madvise(...., MADV_DONTNEED) and some
> memory given back to the OS. Don't spend time on that linked patch
> itself, it's going to be deeply rewritten.
> 
> I hope however I can find some time to get my arm setup fixed today,
> so you maybe waiting a few more hours is enough now.
> 
> Sebastien
> 
> On Sat, Dec 10, 2011 at 2:33 PM,  <dave@cheney.net> wrote:
>> This builds fine on arm5. Is there any way I can test this works ?
>> 
>> http://codereview.appspot.com/5477057/

[Sebastien Paolacci, 2011-12-11 10:35:28]

Ok. Tested, it works as expected. Thanks for helping out.

Sebastien

[rsc, 2011-12-12 21:33:15]

*** Submitted as http://code.google.com/p/go/source/detail?r=30428728ac2b ***

runtime: madvise and SysUnused for Linux

SysUnused being a direct call to madvise MADV_DONTNEED.

R=golang-dev, dave
CC=golang-dev
http://codereview.appspot.com/5477057

Committer: Russ Cox <rsc@golang.org>

[rsc, 2011-12-12 21:33:23]

LGTM

Thanks.

[dave_cheney.net, 2011-12-14 09:28:59]

Hi Sébastien,

I know that SysUnused is not used yet, but would you be interested in
someone adding support for that function in darwin? I think that might
be within my abilities.

Cheers

Dave

2011/12/13 Russ Cox <rsc@golang.org>:
> LGTM
>
> Thanks.

[Sebastien Paolacci, 2011-12-14 13:24:07]

Hi Dave,

Yes it would be great if you could do that, I don't have any short
term plan/setup for a darwin patch.

Having a SysUnused implementation is a preliminary step on the memory
scavenging path. Russ proposed to get the Unix variant done first, so
darwin would also benefit from that patch.

Sebastien

[dave_cheney.net, 2012-01-10 09:33:18]

In preparation for this work I have submitted
http://codereview.appspot.com/5533058/ which rebuilds the
defs_darwin_{amd64,386}.h files.

2011/12/15 Sébastien Paolacci <sebastien.paolacci@gmail.com>:
> Hi Dave,
>
> Yes it would be great if you could do that, I don't have any short
> term plan/setup for a darwin patch.
>
> Having a SysUnused implementation is a preliminary step on the memory
> scavenging path. Russ proposed to get the Unix variant done first, so
> darwin would also benefit from that patch.
>
> Sebastien