Subject: Re: Corruption on files read via NFS
To: Curt Sampson <firstname.lastname@example.org>
From: Manuel Bouyer <email@example.com>
Date: 01/13/1997 12:19:32
On Jan 13, Curt Sampson wrote
> I'm having an odd problem with corruption on files read via NFS.
> I have the following machines, all running a Dec. 25th -current kernel
> and userland:
> Server: Sparc IPX, 40 MB RAM, serving a 1 GB IBM disk
> Client 1: DEC Alpha Multia, 56 MB RAM
> Client 2: P100, Asus M/B, 40 MB RAM, NCR SCSI controller and drives (no IDE),
> 3c905 Ethernet
> Client 3: 486/40 16 MB RAM, IDE hard drive (Connor 400 MB) (no SCSI),
> 3c509 Etherent
> wdc0 at isa0 port 0x1f0-0x1f7 irq 14
> wd0 at wdc0 drive 0: <Conner Peripherals 420MB - CFS420A>
> wd0: 406MB, 826 cyl, 16 head, 63 sec, 512 bytes/sec
> wd0: using 16-sector 16-bit pio transfers, chs addressing
> For some reason client 3 is getting corruption when reading files
> via NFS.
Just a question here :
Is the server running SunOS os Slowlaris ? If so, did you turn on udp
checksum on it ?
> The corruption is exactly the same every time; I can do
> a cksum on the file ten times in a row and get the same (wrong)
> checksum every time. If I look at the file on the server, it's
> fine. In the client, it appears as if a block of the file has been
> replaced by binary garbage--a lot of nulls with a few other things
> chucked in.
The fact that it is always the same may be due to the cache, i.e. the file
isn't reloaded from the server.
> This machine ran 1.1 just fine when I used it before, although at
> that time it was a SCSI machine (Adaptec 1542 and Quantum Empire).
> I'm wondering if perhaps these problems might not be related in
> some way to using IDE on this machine; I rarely run NetBSD on IDE
> systems. Any thoughts?
There may be a problem with IDE and 3c509: this adapter has a really small
fifo, and may be loose packets if interrupts are disabled for long times
(as does the IDE drivers). I remember having very poor perfs with a 3c509
by ftp from a remote server to an IDE drive, and 'normal' perfs writing
to an scsi drive on the same machine.
Manuel Bouyer, Ecole Nationale Superieure de Techniques Avancees, Paris