Date: Mon, 10 Jan 2000 00:17:56 +0100 (MET) From: Gerard Roudier <groudier@club-internet.fr> To: "David O'Brien" <obrien@FreeBSD.org> Cc: cvs-committers@FreeBSD.org, cvs-all@FreeBSD.org Subject: Re: cvs commit: src/sys/i386/conf GENERIC LINT Message-ID: <Pine.LNX.3.95.1000109234058.464A-100000@localhost> In-Reply-To: <20000109122649.A43599@dragon.nuxi.com>
next in thread | previous in thread | raw e-mail | index | archive | help
On Sun, 9 Jan 2000, David O'Brien wrote:
> On Mon, Jan 10, 2000 at 01:38:00AM +0800, Peter Wemm wrote:
> > Note that I said "most" and not "all". Paul (now at Yahoo) and one ot=
her
> > person at Yahoo, as well as a few folks who have posted to the lists, h=
ave
> > reported that the change stops the fxp_add_rfabuf() fault. I think Pau=
l was
> > having different problems at Hotmail (ie: the fxp device stopping).
>=20
> ----- Forwarded message from Paul Saab <ps@yahoo-inc.com> -----
> Date: Sun, 9 Jan 2000 11:24:31 -0800
>=20
> What happed at Hotmail was we saw an large decrease in the number
> of panics that peter is talking about. Basically they went away
> for the most part and only occasionally show up. Ie: the machines
> stay up for weeks/months instead of hours/days.
>=20
> I also know that this is the same case at Yahoo. The test programs
> they had to easily reproduce the "Weird Shit(tm)" panics stopped
> after they started using the sym driver.
> .snip..
Just a thought:
In my opinion, the sym driver is probably better with avoiding to trigger
hardware bugs (on the chip itself and on shared resources as PCI ones as
well). I wanted it to use just the features that are needed and no more.
The LOAD/STORE SCRIPTS instructions helped a _lot_. For example:
- it does not require the chip to master itself.
- it does not write to the on-chip RAM from SCRIPTS (it only loads=20
SCRIPTS at initialization from the C code)
- it does not use MEMORY MOVE. This one got complex since the core=20
supports cache line based transactions.
On the other hand, (still just examples)
- it does perform dummy (or relies on existing) PCI reads, to ensure
ordering of posted transactions when it expects such ordering.
- On some chips I also get rid of complex features. For example, the
prefetch unit is disabled on the 896/895A/C1010 since all SCRIPTS are=
=20
executed from on-chip RAM for those ones (8K on-chip RAM).
You may colour me paranoid :), but being only able to deal with software
bugs, I do want to avoid to trigger hardware bugs as possible, but not by
killing performances and user convenience, obviously.=20
G=E9rard.
To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe cvs-all" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?Pine.LNX.3.95.1000109234058.464A-100000>