On Tue, May 8, 2012 at 10:02 AM, Bjorn Helgaas <bhelgaas@xxxxxxxxxx> wrote: > On Tue, May 8, 2012 at 12:43 AM, Andreas Herrmann > <andreas.herrmann3@xxxxxxx> wrote: >> On Mon, May 07, 2012 at 09:44:16AM -0700, Bjorn Helgaas wrote: >>> On Mon, May 7, 2012 at 12:35 AM, Andreas Herrmann >>> <andreas.herrmann3@xxxxxxx> wrote: >>> > On Fri, May 04, 2012 at 10:35:05AM -0600, Bjorn Helgaas wrote: >>> >> On Fri, May 4, 2012 at 7:03 AM, Andreas Herrmann >>> >> <andreas.herrmann3@xxxxxxx> wrote: >>> >> > On Wed, May 02, 2012 at 11:33:17AM -0600, Bjorn Helgaas wrote: >>> >> >> On Fri, Apr 27, 2012 at 8:36 AM, Andreas Herrmann >>> >> >> <andreas.herrmann3@xxxxxxx> wrote: >>> >> >> > >>> >> >> > Once upon a time this function was overloaded with quirky stuff to fix >>> >> >> > resource detection on systems w/ _CRS defects (seems that some Sun and >>> >> >> > HP systems were affected). >>> >> >> > >>> >> >> > See commit 30a18d6c3f1e774de656ebd8ff219d53e2ba4029 >>> >> >> > (x86: multi pci root bus with different io resource range, on 64-bit) >>> >> >> > >>> >> >> > Restore the old function and thus decouple it from the quirk that is >>> >> >> > CPU family specific (e.g. it won't work on AMD family 15h CPUs). BTW, >>> >> >> > I assume that the _CRS stuff is working on current systems. >>> >> >> > >>> >> >> > This is required to properly initilize the numa_node information of >>> >> >> > existing PCI busses and associated devices. >>> >> >> >>> >> >> I applied some of Yinghai's patches that also touch this area. Can >>> >> >> you refresh these so they apply on top of my "next" branch >>> >> >> (git://git.kernel.org/pub/scm/linux/kernel/git/helgaas/pci.git next)? >>> >> > >>> >> > Arrgh, will adapt my patch and resend it (asap). >>> >> > >>> >> >> Can you also be more specific about what these patches fix? >>> >> > >>> >> >> My understanding is that amd_bus.c (1) sets NUMA info with >>> >> >> set_mp_bus_to_node() and (2) figures out MMIO and I/O port apertures, >>> >> >> which are only used when blind probing and when ignoring _CRS. >>> >> >> >>> >> >> It seems like the main change in this patch is that we skip (2) >>> >> >> completely when family >= 0x11, and I don't understand what that could >>> >> >> fix. >>> >> > >>> >> > The patch restores a very old function that was used to detect the >>> >> > nearest node for a PCI bus, so yes it's used to do (1). IMHO this >>> >> > function was totally screwed up with Yinghai's code to do (2). It >>> >> > seems that Sun has (had?) some systems where (2) was req'd. I don't >>> >> > care about this part. But I'd like to do (1) on all AMD CPU NUMA >>> >> > systems. >>> >> >>> >> Thanks for the explanation. But I'm afraid I'm still confused. >>> >> >>> >> First, it sounds like you're trying to change the way we do part (1), >>> >> i.e., the set_mp_bus_to_node() calls, but I think the effect of your >>> >> patch is to stop doing part (2) in some cases. >>> >> >>> >> Second, I am pretty sure that the current early_fill_mp_bus_info() >>> >> (before your patch) does the exact same set_mp_bus_to_node() calls as >>> >> your early_fill_mp_bus_to_node() does. >>> > >>> > >>> > I want to do (1) on all AMD CPUs that might be used in NUMA systems. >>> > >>> > What's done for (2) is very specific to certain AMD CPU families -- >>> > some of the register accesses are wrong/incomplete for newer AMD >>> > CPUs. Furhtermore _CRS should provide the required info. I really >>> > don't want to extend all the quirky stuff in (2) for future AMD CPUs. >>> >>> I'm all in favor of limiting part (2) to older AMD CPUs. I certainly >>> don't want to maintain it for future CPUs. >>> >>> >> Finally, on all systems with ACPI, the set_mp_bus_to_node() call in >>> >> pci_acpi_scan_root() should be doing what you need. In fact, that >>> >> call happens later, so it should be overwriting the information filled >>> >> in by amd_bus.c. If there's something wrong in this ACPI path, the >>> >> most likely cause is a BIOS defect, such as a missing _PXM method on >>> >> the PNP0A03/0A08 host bridge device. >>> > >>> > Good point. I'll check what's wrong in this ACPI path. >>> >>> I hope you find something, especially if it's a bug in the Linux code >>> that interprets the NUMA info. Then we could fix that and limit both >>> parts to older CPUs. >> >> Simply, there is no _PXM object for the host bridge devices. At least >> on the systems that I checked. >> >> I'll try to find out whether this is sort of "common BIOS practice" on >> AMD boxes and how to avoid that in the future. > > _PXM can also be attached to any parent of the host bridge, since > devices default to the domain of their parents. It looks like > acpi_get_pxm() should already handle that correctly, so I assume these > systems just don't have any _PXM anywhere in the path between the host > bridge and the root. > > If these are just old machines with BIOS bugs, I guess I'm OK with > doing a Linux fix along the lines of your patch. What I don't like is > just silently covering up BIOS bugs in new platforms by keeping this > CPU-specific code when we have a perfectly good generic mechanism for > doing proximity. That's a maintenance problem, as you pointed out for > the aperture code (part (2)). I think we still need to do something here, don't we? I'm expecting we'll end up with at least two patches: one to keep us from looking at MSRs on future CPUs where they might be different from current CPUs, and another to work around BIOS defects (missing _PXM methods) on some systems. I'm guessing the second should be somehow limited to certain CPU families and possibly BIOS dates. Bjorn -- To unsubscribe from this list: send the line "unsubscribe linux-pci" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
