lkml.org 
[lkml]   [2022]   [Mar]   [22]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
SubjectRe: [GIT PULL] f2fs for 5.18
From
On 3/22/22 13:22, Linus Torvalds wrote:
> On Mon, Mar 21, 2022 at 1:39 PM Jaegeuk Kim <jaegeuk@kernel.org> wrote:
>> In this cycle, f2fs has some performance improvements for Android workloads such
>> as using read-unfair rwsems [...]
> I've pulled this, but that read-unfair rwsem code looks incredibly
> dodgy. Doing your own locking is always a bad sign, and it ahs
> traditionally come back to bite us pretty much every time. At least it
> uses real lock primitives, just in a really odd way.
>
> The whole notion of making an rwsem unfair to readers sounds really
> really odd. I mean, the whole and only _point_ of an rwsem is to
> allow concurrent readers, and traditionally if it's unfair it's unfair
> to _writers_ because that tends to be better for throughput (but
> unfairness can cause horrible latency).
>
> So it smells like there's something bad going on in f2fs.
>
> That said, I'm adding Waiman to the cc here in case he would have
> ideas at least for a cleaner interface. Our rw_semaphores are
> explicitly trying to be fair, because unfairness (the other way) was
> such a big problem.
>
> I'm wondering it the optimistic read lock stealing is bothering f2fs?

I don't believe it is the optimistic read lock stealing code that is
bothering f2fs.

AFAICS, the read-unfair rwsem code is created to resolve a potential
lock starvation problem that they found on linux-5.10.y stable tree. I
believe I have fixed that in the v5.11 kernel, see commit 2f06f702925
("locking/rwsem: Prevent potential lock starvation"). However that
commit is not in the stable tree. In fact, I have moved forward and
taken out reader optimistic spinning but added just optimistic lock
stealing instead. I believe the problem would have solved by including
that patch series in their build. I haven't gotten any response as to
whether they had tested this or not.

Apparently they prefer to upstream this stop-gap solution.

Cheers,
Longman

\
 
 \ /
  Last update: 2022-03-22 18:37    [W:1.404 / U:0.236 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site