On Tue, 14 Nov 2006, Roland Dreier wrote:
>
> Huh? The device can't generate any legacy interrupts once MSI is
> enabled. As the PCI spec says:
PLEASE.
The spec is just so much toilet paper. The ONLY thing that matters is what
real hardware does. So please please please PLEASE don't start quoting
specs as a way to "prove your point". It is totally meaningless.
The fact is, we _know_ there are broken devices out there. Not just wrt
MSI - pretty much every single sentence of the PCI spec is probably broken
by _some_ device or subsystem. For example, I would expect that the DMA
ordering rule is likely broken by the big SGI boxes, because they do
memory coherency traffic separately from the IO traffic.
This should be the #1 rule for every kernel programmer: read the
documentation, but don't actually assume it's complete, or that it is
always true.
One of the big reasons user-level programming is so much simpler than
kernel programming is that you can basically "think" about your
algorithms. You seldom have cases where you literally just need to test
out whether all hardware really works that way.
The bottom line: MSI looks great on paper. I think we might make it a rule
that we can enable it by default on bridges we trust (which would seem to
be mainly just Intel). But even then, I think it makes sense for
per-driver rules, because
- some drivers may care more than others (sound, for example, certainly
doesn't really care at all), apart from irq routing issues, and right
now those are _worse_ with MSI due to bugs.
- we definitely have the potential for per-driver bugs (eg the "disable
INTx explicitly" kind of situation), so even if MSI works on a system
level, I think we all know that it doesn't always work on a device
level, and the safe thing tends to be to keep it disabled unless you
have some really overriding concern to use it.
End result: I think that at least for 2.6.19, and at least for the HDA
sound driver, keeping it disabled by default is the right choice. We
should probably _also_ make "pci_msi_supported()" just return an error
(probably by just clearing "pci_msi_enable") for any non-intel host bridge
for now.
I don't dispute AT ALL that we should try to enable things more in the
future. I'm just saying that we're better off being careful, and realizing
that documentation and "official rules" seldom match reality.
(People seem to trust documentation, because for _most_ devices, the
documentation is correct. But if it's wrong for 1% of all devices, you
should still realize that IT IS WRONG. You can't quote "the standard" to
prove a point - a single exception is enough to show that a standard is
incomplete)
Linus
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
[Index of Archives]
[Kernel Newbies]
[Netfilter]
[Bugtraq]
[Photo]
[Stuff]
[Gimp]
[Yosemite News]
[MIPS Linux]
[ARM Linux]
[Linux Security]
[Linux RAID]
[Video 4 Linux]
[Linux for the blind]
[Linux Resources]
