Message-ID: <CAF8kJuOfYMiD-aEhLa9i+oxAtasDcPhFb6__i6QRB2dGO1Lhcg@mail.gmail.com>
Date: Tue, 18 Jun 2024 02:31:58 -0700
From: Chris Li <chrisl@...nel.org>
To: "Huang, Ying" <ying.huang@...el.com>
Cc: Andrew Morton <akpm@...ux-foundation.org>, Kairui Song <kasong@...cent.com>, 
	Ryan Roberts <ryan.roberts@....com>, linux-kernel@...r.kernel.org, linux-mm@...ck.org, 
	Barry Song <baohua@...nel.org>
Subject: Re: [PATCH 0/2] mm: swap: mTHP swap allocator base on swap cluster order

On Mon, Jun 17, 2024 at 11:56 PM Huang, Ying <ying.huang@...el.com> wrote:
>
> Chris Li <chrisl@...nel.org> writes:
>
> > That is in general true with all kernel development regardless of
> > using options or not. If there is a bug in my patch, I will need to
> > debug and fix it or the patch might be reverted.
> >
> > I don't see that as a reason to take the option path or not. The
> > option just means the user taking this option will need to understand
> > the trade off and accept the defined behavior of that option.
>
> User configuration knobs are not forbidden for Linux kernel.  But we are
> more careful about them because they will introduce ABI which we need to
> maintain forever.  And they are hard to be used for users.  Optimizing
> automatically is generally the better solution.  So, I suggest you to
> think more about the automatically solution before diving into a new
> option.

I did, see my reply. Right now there are just no other options.

>
> >>
> >> >> So, I prefer the transparent methods.  Just like THP vs. hugetlbfs.
> >> >
> >> > Me too. I prefer transparent over reservation if it can achieve the
> >> > same goal. Do we have a fully transparent method spec out? How to
> >> > achieve fully transparent and also avoid fragmentation caused by mix
> >> > order allocation/free?
> >> >
> >> > Keep in mind that we are still in the early stage of the mTHP swap
> >> > development, I can have the reservation patch relatively easily. If
> >> > you come up with a better transparent method patch which can achieve
> >> > the same goal later, we can use it instead.
> >>
> >> Because we are still in the early stage, I think that we should try to
> >> improve transparent solution firstly.  Personally, what I don't like is
> >> that we don't work on the transparent solution because we have the
> >> reservation solution.
> >
> > Do you have a road map or the design for the transparent solution you can share?
> > I am interested to know what is the short term step(e.g. a month)  in
> > this transparent solution you have in mind, so we can compare the
> > different approaches. I can't reason much just by the name
> > "transparent solution" itself. Need more technical details.
> >
> > Right now we have a clear usage case we want to support, the swap
> > in/out mTHP with bigger zsmalloc buffers. We can start with the
> > limited usage case first then move to more general ones.
>
> TBH, This is what I don't like.  It appears that you refuse to think
> about the transparent (or automatic) solution.

Actually, that is not true, you make the wrong assumption about what I
have considered. I want to find out what you have in mind to compare
the near term solutions.

In my recent LSF slide I already list 3 options to address this
fragmentation problem.
>From easy to hard:
1) Assign cluster an order on allocation and remember the cluster
order. (short term).
That is this patch series
2) Buddy allocation on the swap entry (longer term)
3) Folio write out compound discontinuous swap entry. (ultimate)

I also considered 4), which I did not put into the slide, because it
is less effective than 3)
4) migrating the swap entries, which require scan page table entry.
I briefly mentioned it during the session.

3) should might qualify as your transparent solution. It is just much
harder to implement.
Even when we have 3), having some form of 1) can be beneficial as
well. (less IO count, no indirect layer of swap offset).

>
> I haven't thought about them thoroughly, but at least we may think about
>
> - promoting low order non-full cluster when we find a free high order
>   swap entries.
>
> - stealing a low order non-full cluster with low usage count for
>   high-order allocation.

Now we are talking.
These two above fall well within 2) the buddy allocators
But the buddy allocator will not be able to address all fragmentation
issues, due to the allocator not being controlled the life cycle of
the swap entry.
It will not help Barry's zsmalloc usage case much because android
likes to keep the swapfile full. I can already see that.

> - freeing more swap entries when swap devices become fragmented.

That requires a scan page table to free the swap entry, basically 4).

It is all about investment and return. 1) is relatively easy to
implement and with good improvement and return.

Chris

> >> >> >> that's really important for you, I think that it's better to design
> >> >> >> something like hugetlbfs vs core mm, that is, be separated from the
> >> >> >> normal swap subsystem as much as possible.
> >> >> >
> >> >> > I am giving hugetlbfs just to make the point using reservation, or
> >> >> > isolation of the resource to prevent mixing fragmentation existing in
> >> >> > core mm.
> >> >> > I am not suggesting copying the hugetlbfs implementation to the swap
> >> >> > system. Unlike hugetlbfs, the swap allocation is typically done from
> >> >> > the kernel, it is transparent from the application. I don't think
> >> >> > separate from the swap subsystem is a good way to go.
> >> >> >
> >> >> > This comes down to why you don't like the reservation. e.g. if we use
> >> >> > two swapfile, one swapfile is purely allocate for high order, would
> >> >> > that be better?
> >> >>
> >> >> Sorry, my words weren't accurate.  Personally, I just think that it's
> >> >> better to make reservation related code not too intrusive.
> >> >
> >> > Yes. I will try to make it not too intrusive.
> >> >
> >> >> And, before reservation, we need to consider something else firstly.
> >> >> Whether is it generally good to swap-in with swap-out order?  Should we
> >> >
> >> > When we have the reservation patch (or other means to sustain mix size
> >> > swap allocation/free), we can test it out to get more data to reason
> >> > about it.
> >> > I consider the swap in size policy an orthogonal issue.
> >>
> >> No.  I don't think so.  If you swap-out in higher order, but swap-in in
> >> lower order, you make the swap clusters fragmented.
> >
> > Sounds like that is the reason to apply swap-in the same order of the swap out.
> > In any case, my original point still stands. We need to have the
> > ability to allocate high order swap entries with reasonable success
> > rate *before* we have the option to choose which size to swap in. If
> > allocating a high order swap always fails, we will be forced to use
> > the low order one, there is no option to choose from. We can't evalute
> > "is it generally good to swap-in with swap-out order?" by actual runs.
>
> I think we don't need to fight for that.  Just prove the value of your
> patchset with reasonable use cases and normal workloads.  Data will
> persuade people.

Powered by blists - more mailing lists