Skip site navigation (1)Skip section navigation (2)
Date:      Tue, 23 Feb 1999 21:25:24 +0000 (GMT)
From:      Terry Lambert <tlambert@primenet.com>
To:        dfr@nlsystems.com (Doug Rabson)
Cc:        julian@whistle.com, Don.Lewis@tsc.tdk.com, tlambert@primenet.com, dillon@apollo.backplane.com, freebsd-hackers@FreeBSD.ORG
Subject:   Re: Panic in FFS/4.0 as of yesterday
Message-ID:  <199902232125.OAA13140@usr06.primenet.com>
In-Reply-To: <Pine.BSF.4.05.9902230847020.60339-100000@herring.nlsystems.com> from "Doug Rabson" at Feb 23, 99 08:50:57 am

next in thread | previous in thread | raw e-mail | index | archive | help
> > softupdates already "kinda" doesn this..
> > it queues data writes at one point in the future and directory writes
> > at a different point in the future. I believe that data writes must be
> > completed before inode writes which must be completed before directory
> > writes. If they are not the the dependencies will FORCE that ordering.
> > 
> > The reason to preschedule the different actions is to make it all happen
> > in the right order anyhow, so that the dependency tracking is a big NOP.
> 
> I think softupdates will be less affected by this but there can still be
> problems with latency.  The time for a simple directory read (not a
> softupdate controlled operation) can be delayed significantly since it
> gets queued behind all the rest of the async i/o.  In Matt's test, I saw
> about 5Mb queued at one point which translates to a latency of over
> 0.5sec, assuming the drive throughput is about 8Mb/sec.

Soft updates is still vulnerable to this, though directory operations
are promoted forward, they are not grouped forward, so one set of file
I/O can find itself in front of a set of (unrelated) directory I/O.

I think the dangers are lessened by "clocking it forward", however;
the worst interference is a process with itself, after queueing a
tremendous amount of normal I/O before engaging in directory
operations.


					Terry Lambert
					terry@lambert.org
---
Any opinions in this posting are my own and not those of my present
or previous employers.


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?199902232125.OAA13140>