[lkml]   [2007]   [Oct]   [30]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: pci-disable-decode-of-io-memory-during-bar-sizing.patch
Linus Torvalds wrote:
> On Tue, 30 Oct 2007, Robert Hancock wrote:
>>> You have to, anyway. Even now the MMCONFIG stuff uses CONF1 cycles for
>>> startup.
>> If it does, it's not by necessity. As soon as you read the table location out
>> of the ACPI tables you can start using it, and that shouldn't require any
>> config space accesses.
> Don't be silly. Exactly _BECAUSE_ we cannot trust the firmware, we have to
> use conf1 (which we can trust) to verify it and/or fix things up.

My point was, it's not inherently necessary in order to use MMCONFIG.
I'm not saying the checks (unreachable_devices and
pci_mmcfg_check_hostbridge) aren't useful or needed with many real
machines. However, in the event that type1 access isn't available we
just skip all those checks because we have no other option. It would
indeed be a pretty broken spec if there was no way to bootstrap with it
even under ideal conditions..

> Also, there are several devices that don't show up in the MMCFG things, or
> just otherwise get it wrong.
> So just take a look at arch/x86/pci/mmconfig-shared.c and look for
> "conf1".
> Really. Damn, I'm nervous taking any MMCFG patches that has you as an
> author, if you aren't even aware of these kinds of fundamnetal issues. You
> probably read the standards about how things are "supposed" to work, and
> then just believed them?
> Rule #1 in kernel programming: don't *ever* think that things actually
> work the way they are documented to work. The documentation is a starting
> point, nothing else.
> And please be defensive in programming. We *know* conf1 cycles work. The
> hardware has been extensively tested, and there are no firmware
> interactions. There is *zero* reasons to use MMCONF cycles for normal
> devices. Ergo: switching over to MMCONF when not needed is stupid and
> fragile.

I can't really disagree that MMCONFIG doesn't have great advantages for
most devices (though it likely is faster on a lot of platforms, which
may be significant if the device does lots of config space accesses). So
for the moment, avoiding using it except where necessary will likely
work out (except if some system does indeed puke on mixing type1 and

However, what Microsoft is doing with Vista may eventually make a
difference in the future. Many hardware vendors seem to use the testing
strategy of "test with latest Windows version. Works OK? Ship it." If
Vista decides that MMCONFIG is good to use all the time, then type1
access support is likely going to a) end up less tested and b) probably
deleted entirely in time. We've seen it before - it used to be that not
using ACPI was the safe option on most hardware with Linux. Now you
pretty much have to use it because the manufacturers only test with it
enabled. I've seen at least one board where the interrupt routing was
completely broken with ACPI off, because they obviously only tested in

To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to
More majordomo info at
Please read the FAQ at

 \ /
  Last update: 2007-10-31 01:27    [W:0.062 / U:2.980 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site