Re: [PATCH 1/3] mm: completely disable THP by transparent_hugepage=never

[Cong Wang <amwang@redhat.com>]

于 2011年06月21日 03:21, Andrea Arcangeli 写道:
> On Mon, Jun 20, 2011 at 02:25:58PM -0400, Vivek Goyal wrote:
>> So I see some opprotunity there to save memory. But this 10kB
>> definitely sounds trivial amount to me.
>
> Agree with you and Rik. Also I already avoided the big memory waste
> (that for example isn't avoided in the ksmd and could be optimized
> away without decreasing flexibility of KSM, and ksmd surely runs on
> the kdump kernel too...) that is to make khugepaged exit and release
> kernel stack when enabled=never (either done by sysfs or at boot with
> transparent_hugepage=never) and all other structs associated with a
> (temporarily) useless kernel thread.

I agree to disable ksm in kdump kernel, thanks for pointing this out!
I will look into later, and probably send a patch for this too.

>
> The khugepaged_slab_init and mm_slot_hash_init() maybe could be
> deferred to when khugepaged starts, and be released when it shutdown
> but it makes it more tricky/racey. If you really want to optimize
> that, without preventing to ever enable THP again despite all .text
> was compiled in and ready to run. You will likely save more if you
> make ksmd exit when run=0 (which btw is a much more common config than
> enabled=never with THP). And slots hashes are allocated by ksm too so
> you could optimize those too if you want and allocate them only by the
> time ksmd starts.

The thing is that we can save ~10K by adding 3 lines of code as this
patch showed, where else in kernel can you save 10K by 3 lines of code?
(except some kfree() cases, of course) So, again, why not have it? ;)

>
> As long as it'd still possible to enable the feature again as it is
> possible now without noticing an altered behavior from userland, I'm
> not entirely against optimizing for saving ~8k of ram even if it
> increases complexity a bit (more kernel code will increase .text a bit
> though, hopefully not 8k more of .text ;).

Why do we _force_ the feature to be tunable even when user completely
don't want to disable it? Why not provide a way to let the user to decide
which is better for him?

When programming kernel, providing a mechanism rather than a policy is
what I always keep in mind, I don't know why you violate this rule here,
to be honest. :-/

Thanks.