On Sun, Aug 19, 2001 at 09:15:55PM +0200, Seth Mos wrote:
> >
> >I noticed a later list message suggesting that a change to XFS in the 2.4.6
> >timeframe should significantly improve stability on highmem machines; I also
> >found a message suggesting that there were deadlocks elsewhere in the Linux
> >kernel on highmem machines until 2.4.7. So I had high hopes for the 2.4.9
> >patch.
>
> 2.4.9 is not available as a patch on the FTP site yet and you will need to
> fetch it from CVS.
Actually, it appears to have been on the FTP site for a couple of days. Do
you see any reason I shouldn't run it?
> >Unfortunately, though I haven't gotten the machine to hang hard yet, I'm
> >now running 2.4.9 and a simple cp -R of our CVS repository from one directory
> >on an XFS filesystem to another produces the "0 order alloc" messages.
>
> These are rather irritating but not related to XFS. If you push ext2 or
> reiserfs hard enough they will show up as well.
Hm. I can't seem to make it happen, but I'll take your word for it.
Is it too complex, or would you mind giving me a thumbnail sketch of how
the memory-allocation path differs in highmem systems that would cause
allocation to fail more frequently? I've got a decent background WRT the
Unix kernel but the guts of the Linux VM system aren't something I've looked
at before in any detail.
> >Is this actually believed to be fixed at the moment, or not? Looks like
> >not, but I'll leave this running for a few hours and see if I can get an
> >actual hang.
>
> The message is fairly harmless for now but the box should at least survive
> this. Stability on highmem is at least better then what it used to be.
My copy-copy-delete test has, indeed, been running since I sent the original
message and so far it hasn't hung, with 2.4.9 and the patch from the FTP
site. I guess I'll have to get over my nervousness about the messages. :-/
Thanks for the help!
Thor
|