[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Message-ID: <CACRpkdZjRpcA92FARSXLNHdSBLxaNwvG6N2ZJPg8NK8cprh7UQ@mail.gmail.com>
Date: Sat, 2 Nov 2024 00:01:00 +0100
From: Linus Walleij <linus.walleij@...aro.org>
To: "Russell King (Oracle)" <linux@...linux.org.uk>
Cc: Mark Rutland <mark.rutland@....com>, Jinjie Ruan <ruanjinjie@...wei.com>, catalin.marinas@....com,
will@...nel.org, oleg@...hat.com, tglx@...utronix.de, peterz@...radead.org,
luto@...nel.org, kees@...nel.org, wad@...omium.org, rostedt@...dmis.org,
arnd@...db.de, ardb@...nel.org, broonie@...nel.org,
rick.p.edgecombe@...el.com, leobras@...hat.com, linux-kernel@...r.kernel.org,
linux-arm-kernel@...ts.infradead.org
Subject: Re: [PATCH v3 0/3] arm64: entry: Convert to generic entry
On Tue, Oct 22, 2024 at 3:59 PM Russell King (Oracle)
<linux@...linux.org.uk> wrote:
> So, I think it's important to always quantify what the impact is of
> any major change to the kernel entry/exit paths - it's all part of
> properly evaluating whether the change makes sense.
>
> If switching to generic code causes a significant degredation in
> performance, then that needs to be investigated and fixed - not
> waited for until someone pops up and eventually says "hey, that
> change to use the generic code resulted in our systems becoming
> much less performant!" six or twelve months down the road.
I agree, I did some measurements for the ARM32 patch which is
in RFC v2. I ran:
perf bench syscall all
This issues 10 million getppid calls and measures how much time
that takes (on a system with some background IRQ noise). I issued
this three times (which takes some time to execute) before and
after the patch set, then compared the best performance before the
patch with the worst performance after the patch.
On ARM32 generic entry costs some 3-4% of performance overhead
(increased execution time).
By chance I discovered that PAN (privileged access never,
CPU_SW_DOMAIN_PAN) costs around 400% increased syscall
execution time overhead (two orders of magnitude more than generic
entry) Debian apparently turns PAN off, maybe by mistake, maybe
for this reason. I was a bit surprised by it, I don't know if I made
some error in the measurement.
Yours,
Linus Walleij
Powered by blists - more mailing lists