James Washer wrote:
>
> I kinda suspected the order didn't matter ( by testing it both ways ;-)
>
> One concern I still have though, is the requirements for saving the dump
> before enabling swap. I cannot say for sure that this is an issue with
> linux, but on other unices I've used, particularly on multi proc boxes with
> LOTS of filesystems, there will be enough (parallel) fsck activity after a
> crash, to cause swapping... In other words, many systems add swap BEFORE
> fscking the disks ( i believe rh71 does this) ... and of course you need to
> fsck /var/log/dump, before you can mount it, and therefore before you can
> save the dump... Catch-22
This has been a problem for years in Unices. If you don't save the
dump to a dedicated dump partition, you have the potential to lose it
on the way up. The reason for moving the swapon is to make sure that
Linux doesn't prove too agressive (which it has been in the past) in using
swap when it may not need to.
> Of course, there are LOTS of workarounds.. like keeping /var/log/dump
> unmounted during normal operations, so that it is CLEAN after a crash and
> can be mounted, o r dumping to a secondary ( or unused ) swap partition,
> etc etc...
Right. All these are options and are usable by any customer.
> I only bring this up here in the spirit of robustness, as an issue we need
> to be aware of.
Yes, we're very aware of this problem ... :)
While at SGI, this was an issue we ran into all the time (before XFS, when
efs could use up enough memory to touch swap). We didn't see it that often,
but it did occur on occasion.
One potential item would be to start writing at the _end_ of swap and then
moving towards the head of the partition. It's ugly and wierd, but it's
an option. It doesn't buy you a whole lot if you really, really have to
swap a lot.
Most systems these days have sufficient memory during the boot-up process
to avoid this problem. Those systems that don't are normally large enough
to run a log-volume based filesystem. Those caught in between are normally
going to be okay, but there is the very uncommon case where they will be
caught by this.
--Matt
> - jim
>
> "Matt D. Robinson" <yakker@xxxxxxxxxxxxxx>@oss.sgi.com on 10/26/2001
> 12:41:30 AM
>
> Sent by: owner-lkcd@xxxxxxxxxxx
>
> To: James Washer/Beaverton/IBM@IBMUS
> cc: Tony Dziedzic <Tony.Dziedzic@xxxxxxxxxxxx>, lkcd@xxxxxxxxxxx
> Subject: Re: Which first? config or save
>
> 'lkcd config' followed by 'lkcd save' was the order required
> a long time ago, due to the way the dump parameters were set.
> /proc/sys/vmdump/??* were used in the save case.
>
> Today, that isn't the case. When 5.0 comes out, dump devices
> will have independent methods of retrieval based on their
> configuration. As long as that data doesn't change between
> configuration and retrieval, it won't be a problem.
>
> I can remove that sentence. It's nice to see someone
> re-validating the documentation based on the current
> set of code. :) My biggest concern is re-writing everything
> now, as with 5.0 it's just going to change again to allow
> for multiple dump devices. Either way, it should be done.
>
> The most important thing is to configure/save the dump before
> swap space is enabled, otherwise you could potentially write
> over the dump if you're using swap space for storage of a dump.
>
> Thanks, guys.
>
> --Matt
>
> James Washer wrote:
> >
> > Tony,
> >
> > Are you sure that "Save relies on configuration parameters set by
> config"??
> >
> > I've looked ( briefly ) at the dump save (lcrash) and don't see that it
> > requires
> > anything from the config step.
> >
> > What did I miss?
> >
> > - jim
> >
> > "Tony Dziedzic" <Tony.Dziedzic@xxxxxxxxxxxx>@oss.sgi.com on 10/23/2001
> > 11:11:51 AM
> >
> > Sent by: owner-lkcd@xxxxxxxxxxx
> >
> > To: James Washer/Beaverton/IBM@IBMUS, <lkcd@xxxxxxxxxxx>
> > cc:
> > Subject: RE: Which first? config or save
> >
> > Correct order is in the sysinit patch: config first, then save. (Save
> > relies on configuration parameters set by config.)
> >
> > Tony Dziedzic
> > Storigen Systems, Inc.
> >
> > > -----Original Message-----
> > > From: James Washer [mailto:washer@xxxxxxxxxx]
> > > Sent: Tuesday, October 23, 2001 1:15 PM
> > > To: lkcd@xxxxxxxxxxx
> > > Subject: Which first? config or save
> > >
> > >
> > > The patch files for sysinit place '/sbin/lkcd config' before
> > > '/sbin/lkcd
> > > save',
> > > however README.lkcd implies the opposite order.
> > >
> > > Which is correct?
> > >
> > >
|