[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20200612152921.GP19480@localhost>
Date: Fri, 12 Jun 2020 17:29:21 +0200
From: Johan Hovold <johan@...nel.org>
To: Dmitry Safonov <0x7f454c46@...il.com>
Cc: Johan Hovold <johan@...nel.org>,
Greg Kroah-Hartman <gregkh@...uxfoundation.org>,
Jiri Slaby <jslaby@...e.com>,
Andy Shevchenko <andriy.shevchenko@...ux.intel.com>,
linux-serial@...r.kernel.org, linux-kernel@...r.kernel.org
Subject: Re: [PATCH v2 2/3] serial: core: fix sysrq overhead regression
On Wed, Jun 10, 2020 at 05:24:57PM +0100, Dmitry Safonov wrote:
> Hi Johan,
>
> On 6/10/20 4:22 PM, Johan Hovold wrote:
> > Commit 8e20fc391711 ("serial_core: Move sysrq functions from header
> > file") converted the inline sysrq helpers to exported functions which
> > are now called for every received character, interrupt and break signal
> > also on systems without CONFIG_MAGIC_SYSRQ_SERIAL instead of being
> > optimised away by the compiler.
>
> The part with ifdeffing looks good to me.
>
> > Inlining these helpers again also avoids the function call overhead when
> > CONFIG_MAGIC_SYSRQ_SERIAL is enabled (e.g. when the port is not used as
> > a console).
>
> But this one, coul you add measures? (it will also help to understand if
> it's a stable material).
Interrupt processing takes 2-3% longer without the inlining with
8250_omap on a beagleboard for example.
> If one function call actually matters here, than should
> uart_insert_char() also go into header?
Good question, it actually was originally intended to be inlined as all
other per-character processing. Separate discussion though.
The point is that we don't want a rarely used debugging feature to incur
unnecessary additional overhead that can easily be avoided.
Johan
Powered by blists - more mailing lists