Re: bcm43xx regression in 2.6.24 (with patch)

From: Alexey Zaytsev
Date: Tue Feb 26 2008 - 19:53:46 EST


On Wed, Feb 27, 2008 at 3:43 AM, Michael Buesch <mb@xxxxxxxxx> wrote:
>
> On Wednesday 27 February 2008 01:32:21 Alexey Zaytsev wrote:
> > On Wed, Feb 27, 2008 at 3:27 AM, Michael Buesch <mb@xxxxxxxxx> wrote:
> > >
> > > On Wednesday 27 February 2008 01:23:17 Alexey Zaytsev wrote:
> > > > On Wed, Feb 27, 2008 at 1:47 AM, John W. Linville
> > > > <linville@xxxxxxxxxxxxx> wrote:
> > > > > On Wed, Feb 27, 2008 at 01:12:32AM +0300, Alexey Zaytsev wrote:
> > > > > > On Wed, Feb 27, 2008 at 1:04 AM, Michael Buesch <mb@xxxxxxxxx> wrote:
> > > > >
> > > > >
> > > > > > > Besides that the bcm43xx driver is not broken. That's the whole reason
> > > > > > > this damn thread started at all. So it can't be broken.
> > > > > > >
> > > > > > Can't agree here. The bcm43xx driver used to work with 2.6.23 without requiring
> > > > > > any module magic.
> > > > >
> > > > > At the risk of prolonging things... :-(
> > > > >
> > > > > Isn't the fundamental problem here that the ssb driver claims the same
> > > > > PCI IDs as the bcm43xx driver? He have hit this same issue a number
> > > > > of times: 8139too vs. 8139cp, eepro vs. e100, sk98lin vs. skge,
> > > > > and I'm sure there are more. I admit that this situation is a bit
> > > > > more confusing, since the user is less likely to predict a conflict
> > > > > between bcm43xx and the ssb driver. This is especially true since
> > > > > the user isn't even selecting ssb directly, but is instead selecting
> > > > > the apparently unrelated b44.
> > > > >
> > > > > Still, the bcm43xx driver is not fundamentally damaged. This is
> > > > > fundamentally a "two drivers claiming the same PCI ID" issue, not a
> > > > > "you broke my driver" one.
> > > >
> > > > Is there any reason the ssb driver should claim the bcm43xx pci ids in
> > > > the first place? I have very little understanding what the Sonic Silicon
> > > > Backplane really is, but I see that the b44 driver claims its PCI ids
> > > > directly. I also think I understand why the b43/b43legacy drivers can't
> > > > claim the ids directly: because the driver-device matching is done not
> > > > with the pci bus methods, but with the ssb bus methods, and it would
> > > > be impossible to automatically choose the right driver for the right
> > > > device (with same ssb ids), as the first of the two drivers loaded would
> > > > succeed in probe()'ing the pci "ssb bridge" device, and not letting the
> > > > other to take control, even after moments later the ssb probe for the
> > > > non-supported ssb device would fail. (Or am I completely wrong?)
> > > >
> > > > That said, I still think that the ssb driver claims the wrong pci ids,
> > > > which is especially wrong if the the b43/b43legacy drivers are not
> > > > even built. And my patch fixes exactly this problem - the ssb driver
> > > > no more claims the broadcom pci ids, when the b43/b43legacy drivers
> > > > are not built.
> > > >
> > > > One better solution I think might be to move the b43_pci_bridge.c
> > > > code to a separate module, and let the b43/b43legacy drivers
> > > > depend on it, but as I said, I have little knowledge in the
> > > > ssb stuff, so I did it the easy way.
> > >
> > > See the comment in b43_pci_bridge.c
> > >
> > Yes, I've seen it. And this design, kind of, causes me some trouble.
>
> There are several reasons to not do a seperate module.
>
> First one being: People won't load it and complain about a regression.
> Yeah, stupid stuff like that happens all the time. That's the reason we
> SELECT the ssb code instead of using DEPENDS ON. People will otherwise not
> enable it and report regressions.
> And please don't say this won't happen. It _did_ happen when b44-PCI was
> a seperate kconfig option. People reported regressions, although there were none.
>
I mean, not to make it a separate config option, of course it should
be hidden and
autoselected, but a separate module with maybe a symbol both b43/b43legacy
could depend on.

> Second one being: It's overkill to have a seperate module for two lines of code
> and a PCI ID table.
>
> Third one being: We want the code to be as small as possible, as it must
> all run on embedded machines. In general being as small as possible should
> be the way to go.
>

If compiled statically, which I think is the right thing for embedded systems,
it should not take much additional space when separated from the ssb code.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/