Hi Michal,

On Fri, Sep 12, 2014 at 07:18:09PM +0200, Michal Hocko wrote:

On Fri 12-09-14 19:26:58, Vladimir Davydov wrote:

quoted

memory.kmem.tcp.limit_in_bytes works as the system-wide tcp_mem sysctl,
but per memory cgroup. While the existence of the latter is justified
(it prevents the system from becoming unusable due to uncontrolled tcp
buffers growth) the reason why we need such a knob in containers isn't
clear to me.

Parallels was the primary driver for this change. I haven't heard of
anybody using the feature other than Parallels. I also remember there
was a strong push for this feature before it was merged besides there
were some complains at the time. I do not remember details (and I am
one half way gone for the weekend now) so I do not have pointers to
discussions.

I would love to get rid of the code and I am pretty sure that networking
people would love this go even more. I didn't plan to provide kmem.tcp.*
knobs for the cgroups v2 interface but getting rid of it altogether
sounds even better. I am just not sure whether some additional users
grown over time.
Nevertheless I am really curious. What has changed that Parallels is not
interested in kmem.tcp anymore?

In our product (OpenVZ) we have home-bred counters for many types of
resources, but we stopped setting limits for most of them, including tcp
buffers accounting, long time ago, and our customers don't set them
either. In the next product release we are going to drop them all and
use only mem, anon+swap, and kmem limits.

I don't know what was the reason to push this stuff, because I wasn't in
it at that time. From what I read from comments to the patches I found
it was something like the first step towards kmem accounting. However,
if we had fully functioning kmem accounting there would be no point in
this.

[...]

Anyway, more than welcome
Acked-by: Michal Hocko <redacted>

Thank you.

In case we happened to grow more users, which I hope hasn't happened, we
would need to keep this around at least with the legacy cgroups API.

The whole CONFIG_MEMCG_KMEM is marked as DON'T ENABLE IT, BECAUSE IT
DOESN'T WORK (kudos to you). That's why I think we could probably close
our eye to wailing users if any.

Thanks,
Vladimir

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>