Re: [PATCH 1/1] suspend: make sync() on suspend-to-RAM optional

[Pavel Machek <pavel@ucw.cz>]

On Sat 2015-07-18 01:54:09, Rafael J. Wysocki wrote:
> On Wednesday, July 15, 2015 02:58:22 PM Brown, Len wrote:
> > 
> > > -----Original Message-----
> > > From: Austin S Hemmelgarn [mailto:ahferroin7@gmail.com]
> > > Sent: Wednesday, July 15, 2015 10:07 AM
> > > To: Pavel Machek; Len Brown
> > > Cc: rjw@rjwysocki.net; linux-pm@vger.kernel.org; linux-
> > > kernel@vger.kernel.org; Brown, Len
> > > Subject: Re: [PATCH 1/1] suspend: make sync() on suspend-to-RAM optional
> > > 
> > > On 2015-07-15 02:43, Pavel Machek wrote:
> > > > On Tue 2015-07-14 22:24:51, Len Brown wrote:
> > > >> From: Len Brown <len.brown@intel.com>
> > > >>
> > > >> The Linux kernel suspend path has traditionally invoked sys_sync().
> > > >>
> > > >> But sys_sync() can be expensive, and some systems do not want
> > > >> to pay the cost of sys_sync() on every suspend.
> > > >
> > > > Have you measured how expensive it can be, and why it is expensive?
> > 
> > > How expensive it is can vary widely, but it pretty much boils down to
> > > how much dirty data still needs written out, and how slow the storage it
> > > needs written to is.  There's not really much that can be done in the
> > > kernel to change this, and most userspace suspend systems call sync
> > > themselves during the suspend sequence.
> > 
> > Right.
> > And now, user-space gets is no longer forced to incur that
> > delay on every suspend if they do not want it.
> > 
> > Yes, have measured this under many conditions.
> > The bottom line is that sys_sync() is rarely as fast as 1ms,
> > and is sometimes as slow as hundreds of ms.

..and failed to figure out what is going on there. Yes, sync needs to
be slow if there's a lot of data to write. But if you suspend
frequently, there should not be a lot of data to write. (And you
should not be suspending with ton of dirty data in the first place).

> > >> Why do you need CONFIG parameter?
> > 
> > So that an OS that doesn't want to change their user-space,
> > can build a kernel that does what they want by default.
> > 
> > Originally I had the config parameter remove this code entirely,
> > which would achieve the same goal.
> > But Rafael prefers the sysfs attribute always exist
> > and the config simply set the default.
> 
> Indeed.
> 
> And so I'm queuing this patch up for 4.3 (with a couple of minor fixups).

Please don't.

"OS that doesn't want to change the user-space to speed up suspend by
few milliseconds" is not a valid reason for asking about million users
one more config question. Affected users can't run mainline kernel
anyway, and will have to change their userland in non-trivial ways to
get there.

									Pavel
-- 
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html