Received-SPF: pass (google.com: domain of linux-kernel+bounces-185444-linux.lists.archive=gmail.com@vger.kernel.org designates 2604:1380:45e3:2400::1 as permitted sender) client-ip=2604:1380:45e3:2400::1;
Precedence: bulk
MIME-Version: 1.0
References: <20240521104825.1060966-1-jolsa@kernel.org> <Zk0IvZU834RQ7YKp@debug.ba.rivosinc.com>
In-Reply-To: <Zk0IvZU834RQ7YKp@debug.ba.rivosinc.com>
From: Alexei Starovoitov <alexei.starovoitov@gmail.com>
Date: Tue, 21 May 2024 13:57:33 -0700
Message-ID: <CAADnVQ+2Q1992e9mRtWOavHfqKsFUxPp4f6MAAJg90TK_KTpew@mail.gmail.com>
Subject: Re: [PATCHv6 bpf-next 0/9] uprobe: uretprobe speed up
To: Deepak Gupta <debug@rivosinc.com>
Cc: Jiri Olsa <jolsa@kernel.org>, Steven Rostedt <rostedt@goodmis.org>, 
	Masami Hiramatsu <mhiramat@kernel.org>, Oleg Nesterov <oleg@redhat.com>, 
	Alexei Starovoitov <ast@kernel.org>, Daniel Borkmann <daniel@iogearbox.net>, 
	Andrii Nakryiko <andrii@kernel.org>, LKML <linux-kernel@vger.kernel.org>, 
	linux-trace-kernel <linux-trace-kernel@vger.kernel.org>, Linux API <linux-api@vger.kernel.org>, 
	linux-man <linux-man@vger.kernel.org>, X86 ML <x86@kernel.org>, bpf <bpf@vger.kernel.org>, 
	Song Liu <songliubraving@fb.com>, Yonghong Song <yhs@fb.com>, 
	John Fastabend <john.fastabend@gmail.com>, Peter Zijlstra <peterz@infradead.org>, 
	Thomas Gleixner <tglx@linutronix.de>, "Borislav Petkov (AMD)" <bp@alien8.de>, Ingo Molnar <mingo@redhat.com>, 
	Andy Lutomirski <luto@kernel.org>, "Edgecombe, Rick P" <rick.p.edgecombe@intel.com>
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable

On Tue, May 21, 2024 at 1:49=E2=80=AFPM Deepak Gupta <debug@rivosinc.com> w=
rote:
>
> On Tue, May 21, 2024 at 12:48:16PM +0200, Jiri Olsa wrote:
> >hi,
> >as part of the effort on speeding up the uprobes [0] coming with
> >return uprobe optimization by using syscall instead of the trap
> >on the uretprobe trampoline.
>
> I understand this provides an optimization on x86. I believe primary reas=
on
> is syscall is straight-line microcode and short sequence while trap deliv=
ery
> still does all the GDT / IDT and segmentation checks and it makes deliver=
y
> of the trap slow.
>
> So doing syscall improves that. Although it seems x86 is going to get rid=
 of
> that as part of FRED [1, 2]. And linux kernel support for FRED is already=
 upstream [2].
> So I am imagining x86 hardware already exists with FRED support.
>
> On other architectures, I believe trap delivery for breakpoint instructio=
n
> is same as syscall instruction.
>
> Given that x86 trap delivery is pretty much going following the suit here=
 and
> intend to make trap delivery cost similar to syscall delivery.
>
> Sorry for being buzzkill here but ...
> Is it worth introducing this syscall which otherwise has no use on other =
arches
> and x86 (and x86 kernel) has already taken steps to match trap delivery l=
atency with
> syscall latency would have similar cost?
>
> Did you do any study of this on FRED enabled x86 CPUs?

afaik CPUs with FRED do not exist on the market and it's
not clear when they will be available.
And when they finally will be on the shelves
the overhead of FRED vs int3 would still have to be measured.
int3 with FRED might still be higher than syscall with FRED.

>
> [1] - https://www.intel.com/content/www/us/en/content-details/780121/flex=
ible-return-and-event-delivery-fred-specification.html
> [2] - https://docs.kernel.org/arch/x86/x86_64/fred.html
>
> >
> >The speed up depends on instruction type that uprobe is installed
> >and depends on specific HW type, please check patch 1 for details.
> >