Re: [PATCH 1/2] pipe: change pipe_write() to never add a zero-sized buffer

From: K Prateek Nayak
Date: Sun Feb 09 2025 - 18:39:40 EST


Hello Oleg,

On 2/10/2025 12:45 AM, Oleg Nesterov wrote:
On 02/09, Linus Torvalds wrote:

On Sun, 9 Feb 2025 at 10:45, Oleg Nesterov <oleg@xxxxxxxxxx> wrote:

Again, lets look eat_empty_buffer().

The comment says "maybe it's empty" but how/why can this happen ?

WHY DO YOU CARE?

Because it looks unclear/confusing, and I think it can confuse other
readers of this code. Especially after 1/2.

So here's the deal: either you
...
(b) you DON'T convince yourself that that is true, and you leave
eat_empty_buffer() alone.

Yes, I failed to convince myself that fs/splice.c can never add an
empty bufer. Although it seems to me it should not.

In contrast, the "eat_empty_buffer()" case just saying "if it's an
empty buffer, it doesn't satisfy my needs, so I'll just release the
empty buffer and go on".

... without wakeup_pipe_writers().

OK, nevermind, I see your point even if I do not 100% agree.

I'll send v2 without WARN_ON() and without 2/2.

Went ahead and tested that version on top of mainline with your
previous change to skip updating {a,c,m}time, here are the results:

==================================================================
Test : sched-messaging
Units : Normalized time in seconds
Interpretation: Lower is better
Statistic : AMean
==================================================================
Case: mainline + no-acm_time[pct imp](CV) patched[pct imp](CV)
1-groups 1.00 [ -0.00]( 7.19) 0.95 [ 4.90](12.39)
2-groups 1.00 [ -0.00]( 3.54) 1.02 [ -1.92]( 6.55)
4-groups 1.00 [ -0.00]( 2.78) 1.01 [ -0.85]( 2.18)
8-groups 1.00 [ -0.00]( 1.04) 0.99 [ 0.63]( 0.77)
16-groups 1.00 [ -0.00]( 1.02) 1.00 [ -0.26]( 0.98)

I don't see any regression / improvements from a performance standpoint
on my 3rd Generation EPYC system (2 x 64C/128T. boost on, C2 disabled)
Feel free to include:

Tested-by: K Prateek Nayak <kprateek.nayak@xxxxxxx>


Oleg.


--
Thanks and Regards,
Prateek