Skip site navigation (1)Skip section navigation (2)
Date:      Sat, 28 Feb 1998 21:29:16 -0500 (EST)
From:      Thomas David Rivers <rivers@dignus.com>
To:        hackers@FreeBSD.ORG, julian@whistle.com
Subject:   Re: The 'dave rivers' memorial panic.
Message-ID:  <199803010229.VAA01614@lakes.dignus.com>

next in thread | raw e-mail | index | archive | help
> I've gone some considerable distance towards tracking down a crash
> that seems to resemble the problem dave's been seeing.

Julian -

 As you can tell - I'm not ready to be 'memorialized' just yet - I'm
just *way* behind in mail reading (I seem to float at about 6000 messages
behind...)

 Anyway - I did a quick scan of Subject: lines and found my own name!
Kinda surprising!

> 
> Basically, a particular setup (running on several 
> pieces of hardware) is incapable of doing a full news expire,
> We've managed to simplify the case down to a reproducible test, that
> involves copying the contents of one disk to a newly newfs'd
> partition on another.
> 
> 
> The exact symptom that we see is that bits in the cylinder group
> bitmap get set by "something" after the cg has bee queued for write.
> 
> adding a test confirms that everything is alright immediatly before
> the write, but the next time the cg is accessed, there are some extra
> bits set. The changes are present on the disk.
> It's not hardware.. we've changed everything, but it's reproducible.
> with this particular setup..
> 
> <hmmmm> the more I write here the more it sounds like flaky hardware..
>     <\hmmmmm> but the patterns seen on disk do not 
> act like hardware..

 Yep - tell me about it...

> it looks like a reallocation..

 That's the path I went down for a long time; but I couldn't see
it.  Also, when I added printf()s, of course; it didn't occur.  
I wondered if something was getting reallocated because of a critical-region
issue...

> 
> some file or more likely, directory, is extended,
> and the cg summary info is never updated, though the 
> bitmaps are..

 Yes!  My reproduction goes similarly - particularly the info
being never updated... (i.e. write some trash on the disk; do
a newfs - and - whoops; the trash is still there - the 0's were
never written...)

> 
> 
> so the question is:
> 
> does anyone know of any 'covert' paths where the cg structs 
> (including bitmaps) are accessed other than in ffs_alloc.c? 
> I'd love to be able to mark the pages concerned 'read-only' when I queue
> them for write. that'd catch the other writer,, :)
> 
> anyone have any ideas on how I'd do that for a bdwrite(bp)?

 It sure does sound like my problem!   It's been some time since you
sent this.  I didn't find any responses - did you happen to get anywhere?

 Was this in 2.2.x or 3.0?    The last 3.0 I tried this with (a snap
from last summer) continued to demonstrate the problem...

	- The not-quite-memorial Dave Rivers -


To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe freebsd-hackers" in the body of the message



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?199803010229.VAA01614>