Re: [PATCH] [5/48] Suspend2 2.1.9.8 for 2.6.12: 350-workthreads.patch

[Nigel Cunningham <ncunningham@cyclades.com>]

Hi.

On Tue, 2005-07-12 at 21:15, Pavel Machek wrote:
> Hi!
> 
> > > > > Again, why do you think you need this?
> > > > 
> > > > 1. If something should be wrong with the freezer, it forms part of a
> > > > safety net that stops your data on disk being trashed.
> > > 
> > > > 2. Separating out threads doing syncing from threads submitting I/O
> > > > makes the refrigerator much more reliable, even under extreme load.
> > > 
> > > This seems to be red herring. Sometimes sync took way too long (like
> > > hours) with older kernels and reiserfs, but I believe that has been
> > > fixed. If not, we need to fix it, anyway; no need to work around it in
> > > suspend2.
> > 
> > Are you considering races such as the case where one process is
> > submitting I/O (say dd) while another is trying to sync? Even if sync
> > does return sooner (presumably because it only syncs writes submitted
> > before the sync), there will still be dirty data after the sync
> > completes. And if we start another sync thread, it will suffer the same
> > problem. The only solution is to stop I/O being submitted, then sync.
> > But having stopped I/O being submitted, we need to still have the
> > threads the process the I/O (possibly userspace ones) unfrozen. Hence we
> > need to differentiate 'syncthreads'.
> 
> OTOH: this is only critical for "niceness", not for
> correctness. Calling sync() before suspend is simply nice thing to do,
> but it is not required in any way. If someone is doing long dd, tough,
> they are going to loose some data if wakeup fails. It is no worse than
> sudden poweroff.

How can you say it's only required for niceness one minute, then admit
it might result in data loss the next?

Regards,

Nigel
-- 
Evolution.
Enumerate the requirements.
Consider the interdependencies.
Calculate the probabilities.
Be amazed that people believe it happened.