Re: Tyan 2885 ftp freeze

On Tuesday 26 April 2005 18:01, Bjarne Thomsen wrote:
> I have some new results on my Tyan 2885. See below.
>
> On Tue, 2005-04-26 at 11:05 +0200, Giuseppe Ghibò wrote:
> > Michael Riß wrote:
> > > Bjarne Thomsen wrote:
> > >> I have a Tyan 2885 dual Opteron with
> > >> a Broadcom Corp.|NetXtreme BCM5703X Gigabit Ethernet.
> > >> It is running Mandriva LE2005 x86_64 with
> > >> kernel 2.6.11-7mdksmp
> > >> The problem is, and was also with Mandrake 10.1/x86_64
> > >> that the machine freezes without any message in syslog
> > >> when I transfer about 5 Gb of small and large files by
> > >> sftp, or by ncftp from this machine to anoter machine.
> > >> It does not happen when I transfer the same data the
> > >> other way. It happens both with the tg3 and bcm5700
> > >> drivers.
> > >> I have both IDE and SCSI disks, and it happens while
> > >> transfering from both types of disks.
> > >>
> > >> How can I track this bug down? If it is a bug.
> > >> Any suggestions are welcome.

[..]

> I am not using SATA (because of problems). Instead I use SCSI disks
> and an IDE disk. For the root partition I use ext3, for the remaining
> partitions I use XFS. I should also mention that I transfer/copy to
> a relatively slow PC. This may be part of the problem.
> As suggested I installed the x86_32 version of LE2005 running the smp
> kernel. I also tried the uniprocessor kernel.
> I tried sftp and scp, and the same thing happened for both kernels.
> I then re-installed the x86_64 version of LE2005.
> I now nfs mounted a file system on the remote (slow) PC,
> and copied (cp) the same files to the NFS mounted system.
> This copy went well twice (this has never happened with sftp).
> I then tried wget via the web server. This also went well twice.
> I then returned to sftp. It crashed as usual.
>
> Conclusion: wget and NFS cp works well, but sftp, scp and ncftp
> does not.  But what does it all mean?

That's strange. NFS works in the kernel space so other code paths
get used and it's plausible that the error doesn't show up here. 
But http works in user space and should be affected, too.
It can't be sendfile(2) either because sftp and scp have to do the
encryption in user space and therefore can't use it.
Maybe memory management issues? See below.

> I can make some further tests if somebody can tell me what to do.
> Now, this is a machine that I use for work, so I am not going
> reformat /home /usr/local and /opt, but / and /usr are in principle
> open for experiments. Could some sort of debugging be turned on
> in order to locate the problem.
> I could also try to use another NIC?
>
>  -- Bjarne
>
> > I tried sending 5GB of data from one machine to another using:
> >
> > ttcp -t -s -f M -n 655360 remotehost
> >
> > (and of course the remotehost should run "ttcp -r -s -f M"
> > to receive packets). 655360 should be exactly 5GB of data.
> >
> > Rather there I get in, /var/log/kernel/warnings, this
> > kind of logs:
> >
> > kernel: mm/memory.c:97: bad pmd ffff8100425cb840(5f36387800000000)
> > kernel: mm/memory.c:97: bad pmd ffff8100425cb848(0000000000003436).
> >
> > but RAM seems OK.

I found a thread about problems of the memory management (mm) subsystem on
x86_64(AMD) systems. There has been a reorganization of mm recently to fix 
problems with those large memory machines (4-level patches), but the
problems are still not completely sorted out:

http://marc.theaimsgroup.com/?l=linux-kernel&m=111221947310457&w=2

I ran into problems myself with the 10.1-x86_64 Mandrake kernel on our 16GB
Opteron machine 
(http://qa.mandrakesoft.com/show_bug.cgi?id=13069)
and even after I fixed them somehow I had problems writing out 
large amounts of data via NFS. The mm consumed so much memory for the file 
system cache that there was no further memory available for normal programs.
The next time they did a malloc they got killed (out of memory).
Hopefully these problems get fixed soon.

Sorry Bjarne, I'm out of ideas for your problem now. :-\

cu
Michi