xfs
[Top] [All Lists]

Re: XFS dying when many processes copy many files/directories

To: linux-xfs@xxxxxxxxxxx
Subject: Re: XFS dying when many processes copy many files/directories
From: Nic Doye <nic@xxxxxxxxxxx>
Date: 27 Dec 2001 12:56:00 +0000
Envelope-to: <linux-xfs@xxxxxxxxxxx>
In-reply-to: <3C2B1350.D2074E68@xxxxxxxxxxxxxxxxxx>
References: <3C2B1350.D2074E68@xxxxxxxxxxxxxxxxxx>
Sender: owner-linux-xfs@xxxxxxxxxxx
On Thu, 2001-12-27 at 12:25, cradeke wrote:

> Interesting thing...  Some people (including me) have seen this
> myserious
> kupdated runnings on smp machines only. Maybe there is something wrong
> with
> filehandling.. something like that a deadlock occures. I can reproduce
> it when I
> write to a nearly full xfs. 1% left but only on one of my partitions...
> it is
> the last one. Another partition which is the first one at the same disc
> can be
> filled up to 100%. Some people wrote that the problem went away if they
> running
> a single cpu kernel. Your simple test shows that there is a real problem
> with
> xfs which should be solved quickly. After such a kupdated amok running
> opened
> files are filled with @'s and any change or creation of new files are
> lost. I
> saw this with 2.4.2 / 2.4.5 smp-kernels and with xfs-1.0 / 1.01.

Another (possibly unrelated) data point:

Personally, all my kupdated problems disappeared with the new VM. My
problems were non-fatal but annoying. Basically kupdated enjoyed hogging
the entire machine (using about 50% of the CPU). 'Big' graphical things
(screensavers, moving windows, etc.) were annoyingly slow. Starting
large apps like Mozilla, XEmacs, Star Office took forever...

Something like this happened: slocate causes the cache to swell to the
size of the remaining memory on the machine. kupdated spent it's entire
life looking for spare bits of memory. (Allocating and freeing a hundred
meg or so made the machine responsive again).

Now this no longer happens - the computer is fast and fun. Just like in
the 2.2 days. ;-)

2 things could now happen:

1. The new VM/other kernel work has cured this bug (can't remember what
kernel version others reported).

2. The new VM/other kernel work will only trigger this bug far less
often and it'll be a bugger to track down.

nic


<Prev in Thread] Current Thread [Next in Thread>