Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
MIME-Version: 1.0
References: <CANn89iKUbyrJ=r2+_kK+sb2ZSSHifFZ7QkPLDpAtkJ8v4WUumA@mail.gmail.com>
 <CAHk-=whqNMUPbjCyMjyxfH_5-Xass=DrMkPT5ZTJbFrtU=qDEQ@mail.gmail.com>
 <CANn89i+bExb_P6A9ROmwqNgGdO5o8wawVZ5r3MHnz0qfhxvTtA@mail.gmail.com>
 <CAHk-=wig6VizZHtRznz7uAWa-hHWjrCNANZ9B+1G=aTWPiVH4g@mail.gmail.com>
 <CAHk-=whkci5ck5Him8Lx5ECKHEtj=bipYmOCGe8DWrrp8uDq5g@mail.gmail.com>
 <CAHk-=whtDupvWtj_ow11wU4_u=KvifTqno=5mW1VofyehjdVRA@mail.gmail.com>
 <CANn89i+u8jvfSQAQ=_JY0be56deJNhKgDWbqpDAvfm-i34qX9A@mail.gmail.com>
 <CAHk-=wh16fVwO2yZ4Fx0kyRHsNDhGddzNxfQQz2+x08=CPvk_Q@mail.gmail.com>
 <CANn89iJ3=OiZEABRQQLL6z+J-Wy8AvTJz6NPLQDOtzREiiYb4Q@mail.gmail.com> <CAHk-=whZ23EHnBG4ox9QpHFDeiCSrA2H1wrYrfyg3KP=zK5Sog@mail.gmail.com>
In-Reply-To: <CAHk-=whZ23EHnBG4ox9QpHFDeiCSrA2H1wrYrfyg3KP=zK5Sog@mail.gmail.com>
From:   Eric Dumazet <edumazet@google.com>
Date:   Fri, 26 May 2023 20:55:22 +0200
Message-ID: <CANn89iJgRLPR_53vrd2zfYiU5ejcVWACtH6h_JPnvte6eSGOLg@mail.gmail.com>
Subject: Re: x86 copy performance regression
To:     Linus Torvalds <torvalds@linux-foundation.org>
Cc:     LKML <linux-kernel@vger.kernel.org>,
        netdev <netdev@vger.kernel.org>
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
Precedence: bulk

On Fri, May 26, 2023 at 8:33=E2=80=AFPM Linus Torvalds
<torvalds@linux-foundation.org> wrote:
>
> On Fri, May 26, 2023 at 10:51=E2=80=AFAM Eric Dumazet <edumazet@google.co=
m> wrote:
> >
> > Hmmm
> >
> > [   25.532236] RIP: 0010:0xffffffffa5a85134
> > [   25.536173] Code: Unable to access opcode bytes at 0xffffffffa5a8510=
a.
>
> This was the other reason I really didn't want to use alternatives on
> the conditional branch instructions. The relocations are really not
> very natural, and we have odd rules for those things. So I suspect our
> instruction rewriting simply gets this wrong, because that's such a
> nasty pattern.
>
> I really wanted my "just hardcode the instruction bytes" to work. Not
> only did it get me the small 2-byte conditional jump, it meant that
> there was no relocation on it. But objtool really hates not
> understanding what the alternatives code does.
>
> Which is fair enough, but it's frustrating here when it only results
> in more problems.
>
> Anyway, I guess *this* avoids all issues. It creates an extra jump to
> a jump for the case where the CPU doesn't have ERMS, but I guess we
> don't really care about those CPUs anyway.
>
> And it avoids all the "alternative instructions have relocations"
> issues. And it creates all small two-byte jumps, and the "rep movsb"
> fits exactly on that same 2 bytes too. Which I guess all argues for
> this being what I should have started with.
>
> This time it *really* works.
>

Indeed, this one is working and fixes the issue for me, thanks a lot !

New numbers look similar to 6.3 ones.

Tested-by: Eric Dumazet <edumazet@google.com>

 Performance counter stats for 'taskset 02 ./tcp_mmap -H 2002:a05:6608:297:=
:':

          2,833.29 msec task-clock                       #    0.970
CPUs utilized
             1,065      context-switches                 #  375.888
/sec
                 1      cpu-migrations                   #    0.353
/sec
               128      page-faults                      #   45.177
/sec
    10,297,389,329      cycles                           #    3.634
GHz
     7,213,189,594      instructions                     #    0.70
insn per cycle
     1,220,821,121      branches                         #  430.884
M/sec
        10,430,907      branch-misses                    #    0.85% of
all branches

       2.921180547 seconds time elapsed

       0.005304000 seconds user
       2.478561000 seconds sys