Hi, Jianmin and Bjorn,
On Wed, Jun 29, 2022 at 8:33 AM Jianmin Lv <lvjianmin@xxxxxxxxxxx> wrote:
>
>
>
> On 2022/6/29 上午12:04, Bjorn Helgaas wrote:
> > On Tue, Jun 28, 2022 at 09:03:02PM +0800, Jianmin Lv wrote:
> >> On 2022/6/28 上午5:38, Bjorn Helgaas wrote:
> >>> On Fri, Jun 17, 2022 at 03:43:27PM +0800, Huacai Chen wrote:
> >>>> On LS2K/LS7A, some non-existant devices don't return 0xffffffff when
> >>>> scanning. This is a hardware flaw but we can only avoid it by software
> >>>> now.
> >>>
> >>> We should say what *does* happen if we do a config read to a device
> >>> that doesn't exit. Machine check, hang, etc?
> >>
> >> The device is a hidden device(only for debug) that should not be
> >> scanned. If scanned in a non-normal way, the machine is hang(one
> >> case in ltp pci test can trigger the issue, which is explained
> >> below).
> >
> > Reading the Vendor ID is the *normal* way to scan for a device. It
> > seems that this hardware just hangs in some cases when the device
> > doesn't exist.
> >
> >>> Generally speaking we only probe for functions > 0 if .0 is marked as
> >>> multi-function, so I guess this means 00:09.0 is marked as a
> >>> multi-function device, but config reads to 00:09.1 would fail?
> >>
> >> Yes, definitely. Actually, the 00:09.0 is a single device, so fun1(09.1)
> >> will not be scanned(e.g. the fun1 will be not scanned on pci enumeration
> >> during kernel booting).
> >>
> >> But, there is one situation: when running ltp pci test case on LS7A,
> >> the 00:08.2 is a sata controller(a valid device), and the bus number(0)
> >> and devfn(0x42) are inputted to kernel api pci_scan_slot(), which has
> >> clear note: devfn must have zero function. So, apparently, the inputted
> >> devfn's function is not zero, but 2, and then in the pci_scan_slot():
> >>
> >> for (fn = next_fn(bus, dev, 0); fn > 0; fn = next_fn(bus, dev, fn))
> >> {
> >> dev = pci_scan_single_device(bus, devfn + fn);
> >> ...
> >> }
> >>
> >> 08.2,08.3...and 09.1 will be scanned one by one, so the 09.1(fun1) is
> >> scanned.
> >
> > Does the "((bus == 0) && (device >= 9 && device <= 20) && (function > 0))"
> > test catch *all* devfns where the hang occurs? I wouldn't want to
> > only avoid the ones that LTP happens to use. If we did that, a future
> > LTP change could easily break things again. But I assume you know
> > exactly what devices are present on the root bus.
> >
>
> Yes, as you said, I'm sure that only these hidden functions(fun1 of dev
> 9 to 20) on root bus can cause issue, so this fix is enough to address it.
>
> >>>> - if (priv->data->flags & FLAG_DEV_FIX &&
> >>>> - !pci_is_root_bus(bus) && PCI_SLOT(devfn) > 0)
> >>>> + if ((priv->data->flags & FLAG_DEV_FIX) && bus->self) {
> >>>> + if (!pci_is_root_bus(bus) && (device > 0))
> >>>> + return NULL;
> >>>> + }
> >>>> +
> >>>> + /* Don't access non-existant devices */
> >>>> + if (!pdev_is_existant(busnum, device, function))
> >>>> return NULL;
> >>>
> >>> Is this a "forever" hardware bug that will never be fixed, or should
> >>> there be a flag like FLAG_DEV_FIX so we only do this on the broken
> >>> devices?
> >>
> >> No, the next new version LS7A will correct it, so maybe we can use
> >> FLAG_DEV_FIX-like to address it.
> >
> > You should add the flag now instead of waiting for the new hardware.
> > Otherwise you may not remember or notice the need to make this
> > conditional on the hardware version, you'll wonder why the fixed
> > hardware doesn't enumerate devices correctly.
> >
>
> Thanks for your suggestion, I agree that, Huacai, WDYT?
Agree, I think we can use a FLAG_DEV_HIDDEN flag.
Huacai
>
>
> > Bjorn
> >
>
