linux-kernel - Re: [RFC PATCH] migrate_pages: Never block waiting for the page lock

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <87edohvpzk.fsf@yhuang6-desk2.ccr.corp.intel.com>
Date:   Tue, 18 Apr 2023 11:17:03 +0800
From:   "Huang, Ying" <ying.huang@...el.com>
To:     Doug Anderson <dianders@...omium.org>
Cc:     Andrew Morton <akpm@...ux-foundation.org>,
        Yu Zhao <yuzhao@...gle.com>, linux-kernel@...r.kernel.org,
        linux-mm@...ck.org
Subject: Re: [RFC PATCH] migrate_pages: Never block waiting for the page lock

Doug Anderson <dianders@...omium.org> writes:

> Hi,
>
> On Sun, Apr 16, 2023 at 6:15 PM Huang, Ying <ying.huang@...el.com> wrote:
>>
>> Doug Anderson <dianders@...omium.org> writes:
>>
>> > Hi,
>> >
>> > On Thu, Apr 13, 2023 at 8:10 PM Huang, Ying <ying.huang@...el.com> wrote:
>> >>
>> >> Douglas Anderson <dianders@...omium.org> writes:
>> >>
>> >> > Currently when we try to do page migration and we're in "synchronous"
>> >> > mode (and not doing direct compaction) then we'll wait an infinite
>> >> > amount of time for a page lock. This does not appear to be a great
>> >> > idea.
>> >> >
>> >> > One issue can be seen when I put a device under extreme memory
>> >> > pressure. I took a sc7180-trogdor Chromebook (4GB RAM, 8GB zram
>> >> > swap). I ran the browser along with Android (which runs from a
>> >> > loopback mounted 128K block-size squashfs "disk"). I then manually ran
>> >> > the mmm_donut memory pressure tool [1]. The system is completely
>> >> > unusable both with and without this patch since there are 8 processes
>> >> > completely thrashing memory, but it was still interesting to look at
>> >> > how migration was behaving. I put some timing code in and I could see
>> >> > that we sometimes waited over 25 seconds (in the context of
>> >> > kcompactd0) for a page lock to become available. Although the 25
>> >> > seconds was the high mark, it was easy to see tens, hundreds, or
>> >> > thousands of milliseconds spent waiting on the lock.
>> >> >
>> >> > Instead of waiting, if I bailed out right away (as this patch does), I
>> >> > could see kcompactd0 move forward to successfully to migrate other
>> >> > pages instead. This seems like a better use of kcompactd's time.
>> >> >
>> >> > Thus, even though this didn't make the system any more usable in my
>> >> > absurd test case, it still seemed to make migration behave better and
>> >> > that feels like a win. It also makes the code simpler since we have
>> >> > one fewer special case.
>> >>
>> >> TBH, the test case is too extreme for me.
>> >
>> > That's fair. That being said, I guess the point I was trying to make
>> > is that waiting for this lock could take an unbounded amount of time.
>> > Other parts of the system sometimes hold a page lock and then do a
>> > blocking operation. At least in the case of kcompactd there are better
>> > uses of its time than waiting for any given page.
>> >
>> >> And, we have multiple "sync" mode to deal with latency requirement, for
>> >> example, we use MIGRATE_SYNC_LIGHT for compaction to avoid too long
>> >> latency.  If you have latency requirement for some users, you may
>> >> consider to add new "sync" mode.
>> >
>> > Sure. kcompactd_do_work() is currently using MIGRATE_SYNC_LIGHT. I
>> > guess my first thought would be to avoid adding a new mode and make
>> > MIGRATE_SYNC_LIGHT not block here. Then anyone that truly needs to
>> > wait for all the pages to be migrated can use the heavier sync modes.
>> > It seems to me like the current users of MIGRATE_SYNC_LIGHT would not
>> > want to block for an unbounded amount of time here. What do you think?
>>
>> It appears that you can just use MIGRATE_ASYNC if you think the correct
>> behavior is "NOT block at all".  I found that there are more
>> fine-grained controls on this in compaction code, please take a look at
>> "enum compact_priority" and its comments.
>
> Actually, the more I think about it the more I think the right answer
> is to keep kcompactd as using MIGRATE_SYNC_LIGHT and make
> MIGRATE_SYNC_LIGHT not block on the folio lock.

Then, what is the difference between MIGRATE_SYNC_LIGHT and
MIGRATE_ASYNC?

> kcompactd can accept some blocking but we don't want long / unbounded
> blocking. Reading the comments for MIGRATE_SYNC_LIGHT, this also seems
> like it fits pretty well. MIGRATE_SYNC_LIGHT says that the stall time
> of writepage() is too much. It's entirely plausible that someone else
> holding the lock is doing something as slow as writepage() and thus
> waiting on the lock can be just as bad for latency.

IIUC, during writepage(), the page/folio will be unlocked.

But, during page reading, the page/folio will be locked.  I don't really
understand why we can wait for page reading but cannot wait for page
writeback.

Best Regards,
Huang, Ying

> I'll try to send out a v2 with this approach today and we can see what
> people think.
>
> -Doug