On Wed, Feb 05, 2020 at 11:01:32AM +0000, Sergei Miroshnichenko wrote:
> On Fri, 2020-01-31 at 14:31 -0600, Bjorn Helgaas wrote:
> > On Wed, Jan 29, 2020 at 06:29:28PM +0300, Sergei Miroshnichenko
> > wrote:
> > > BAR allocation by BIOS/UEFI/bootloader/firmware may be
> > > non-optimal and it may even clash with the kernel's BAR
> > > assignment algorithm.
> > >
> > > For example, sometimes BIOS doesn't reserve space for SR-IOV
> > > BARs, and this bridge window can neither extend (blocked by
> > > immovable BARs) nor move (the device itself is immovable).
> > >
> > > With this patch the kernel will use its own methods of BAR
> > > allocating when possible, increasing the chances of successful
> > > boot and hotplug.
> > >
> > > Signed-off-by: Sergei Miroshnichenko <s.miroshnichenko@xxxxxxxxx>
> > > ---
> > > drivers/pci/probe.c | 16 ++++++++++++++--
> > > 1 file changed, 14 insertions(+), 2 deletions(-)
> > >
> > > diff --git a/drivers/pci/probe.c b/drivers/pci/probe.c
> > > index bb584038d5b4..f8f643dac6d1 100644
> > > --- a/drivers/pci/probe.c
> > > +++ b/drivers/pci/probe.c
> > > @@ -306,6 +306,14 @@ int __pci_read_base(struct pci_dev *dev, enum
> > > pci_bar_type type,
> > > pos, (unsigned long long)region.start);
> > > }
> > >
> > > + if (pci_can_move_bars && res->start && !(res->flags &
> > > IORESOURCE_IO)) {
> > > + pci_warn(dev, "ignore the current offset of BAR %llx-
> > > %llx\n",
> > > + l64, l64 + sz64 - 1);
> > > + res->start = 0;
> > > + res->end = sz64 - 1;
> > > + res->flags |= IORESOURCE_SIZEALIGN;
> > > + }
> > > +
> > > goto out;
> > >
> > >
> > > @@ -528,8 +536,10 @@ void pci_read_bridge_bases(struct pci_bus
> > > *child)
> > > child->resource[i] = &dev-
> > > >resource[PCI_BRIDGE_RESOURCES+i];
> > >
> > > pci_read_bridge_io(child);
> > > - pci_read_bridge_mmio(child);
> > > - pci_read_bridge_mmio_pref(child);
> > > + if (!pci_can_move_bars) {
> > > + pci_read_bridge_mmio(child);
> > > + pci_read_bridge_mmio_pref(child);
> > > + }
> >
> > I mentioned this in another response, but I'll repeat it here to
> > comment on the code directly: I don't think we should have feature
> > tests like this "!pci_can_move_bars" scattered around, and I don't
> > want basic behaviors like reading bridge windows during enumeration
> > to
> > depend on it.
> >
> > There's no obvious reason why we should ignore bridge windows if we
> > support movable BARs.
>
> That patch solves a problem which is non-fatal during boot, but is
> breaking this whole patchset when trying a PCI rescan. On a specific
> machine we have a popular i350 network card:
>
> $ lspci -tv
> -[0000:00]-...
> +-01.1-[02]--+-00.0 Intel Corporation I350 Gigabit Network
> Connection
> | +-00.1 Intel Corporation I350 Gigabit Network
> Connection
> | +-00.2 Intel Corporation I350 Gigabit Network
> Connection
> | \-00.3 Intel Corporation I350 Gigabit Network
> Connection
>
> On the "ROG STRIX Z370-F GAMING, BIOS 0612 03/01/2018" motherboard and
> vanilla kernel, not every BAR is allocated:
>
> $ dmesg -t
> ...
> pci 0000:00:01.0: PCI bridge to [bus 01]
> pci 0000:00:01.0: bridge window [mem 0xf7700000-0xf78fffff]
> pci 0000:02:00.0: BAR 7: assigned [mem 0xf7490000-0xf74affff]
> pci 0000:02:00.0: BAR 10: assigned [mem 0xf74b0000-0xf74cffff]
> pci 0000:02:00.1: BAR 7: assigned [mem 0xf74d0000-0xf74effff]
> pci 0000:02:00.1: BAR 10: no space for [mem size 0x00020000]
> pci 0000:02:00.1: BAR 10: failed to assign [mem size 0x00020000]
> pci 0000:02:00.2: BAR 7: no space for [mem size 0x00020000]
> pci 0000:02:00.2: BAR 7: failed to assign [mem size 0x00020000]
> pci 0000:02:00.2: BAR 10: no space for [mem size 0x00020000]
> pci 0000:02:00.2: BAR 10: failed to assign [mem size 0x00020000]
> pci 0000:02:00.3: BAR 7: no space for [mem size 0x00020000]
> pci 0000:02:00.3: BAR 7: failed to assign [mem size 0x00020000]
> pci 0000:02:00.3: BAR 10: no space for [mem size 0x00020000]
> pci 0000:02:00.3: BAR 10: failed to assign [mem size 0x00020000]
> pci 0000:00:01.1: PCI bridge to [bus 02]
>
> $ sudo cat /proc/iomem
> ...
> f7000000-f74fffff : PCI Bus 0000:02
> f7000000-f70fffff : 0000:02:00.3
> f7000000-f70fffff : igb
> f7100000-f71fffff : 0000:02:00.2
> f7100000-f71fffff : igb
> f7200000-f72fffff : 0000:02:00.1
> f7200000-f72fffff : igb
> f7300000-f73fffff : 0000:02:00.0
> f7300000-f73fffff : igb
> f7400000-f747ffff : 0000:02:00.0
> f7480000-f7483fff : 0000:02:00.3
> f7480000-f7483fff : igb
> f7484000-f7487fff : 0000:02:00.2
> f7484000-f7487fff : igb
> f7488000-f748bfff : 0000:02:00.1
> f7488000-f748bfff : igb
> f748c000-f748ffff : 0000:02:00.0
> f748c000-f748ffff : igb
> f7490000-f74affff : 0000:02:00.0
> f74b0000-f74cffff : 0000:02:00.0
> f74d0000-f74effff : 0000:02:00.1
>
> But when allowing the kernel to allocate BARs properly, the map is
> full:
>
> c8200000-c87fffff : PCI Bus 0000:02
> c8200000-c82fffff : 0000:02:00.0
> c8200000-c82fffff : igb
> c8300000-c83fffff : 0000:02:00.1
> c8300000-c83fffff : igb
> c8400000-c84fffff : 0000:02:00.2
> c8400000-c84fffff : igb
> c8500000-c85fffff : 0000:02:00.3
> c8500000-c85fffff : igb
> c8600000-c867ffff : 0000:02:00.0
> c8680000-c8683fff : 0000:02:00.0
> c8680000-c8683fff : igb
> c8684000-c86a3fff : 0000:02:00.0
> c86a4000-c86c3fff : 0000:02:00.0
> c86c4000-c86c7fff : 0000:02:00.1
> c86c4000-c86c7fff : igb
> c86c8000-c86e7fff : 0000:02:00.1
> c86e8000-c8707fff : 0000:02:00.1
> c8708000-c870bfff : 0000:02:00.2
> c8708000-c870bfff : igb
> c870c000-c872bfff : 0000:02:00.2
> c872c000-c874bfff : 0000:02:00.2
> c874c000-c874ffff : 0000:02:00.3
> c874c000-c874ffff : igb
> c8750000-c876ffff : 0000:02:00.3
> c8770000-c878ffff : 0000:02:00.3
>
> In this particular case the "repaired" BARs are not vital and are not
> used by the igb driver, but in general such behavior of BIOS can lead
> to a non-working setup.
>
> So ignoring pre-set BARs and bridge windows may be useful on its own,
> but it is also provides a working starting point required by this
> patchset, otherwise it will need to track such BARs impossible to
> assign, and don't try to assign them during a next PCI rescan.
>
> The reason I've tied this feature to the "movable BARs" flag is that I
> know at least one exception demanding a workaround - VGA. So I wanted
> to provide a flag to disable it in case of other unforeseen
> consequences, and the only feature depends on this - is movable BARs.
If we need exceptions for broken or legacy devices, we should check
for those explicitly.
If we fail to assign some BARs at boot, I think it's reasonable to try
to reassign things before drivers claim the devices. But support for
that should be in a reassignment path, not in the general enumeration
path. And, since drivers aren't involved yet, it probably doesn't
even depend on pci_can_move_bars.
> The [07/26] "PCI: hotplug: Don't allow hot-added devices to steal
> resources" patch introduces an additional step in BAR assignment:
> - try to assign every existing BAR + BARs of the hot-added device;
> - it if fails, disable BARs for the hot-added device and retry without
> them.
>
> A possible way to work-around non-working BARs could be adding more
> steps:
> - first try to assign every existing BAR + BARs not worked previously
> + "hot-added" BARs;
> - if it fails, retry without those BARs which weren't working before;
> - if it still fails, retry without "hot-added" BARs.
>
> Best regards,
> Serge
>
> > > if (dev->transparent) {
> > > pci_bus_for_each_resource(child->parent, res, i) {
> > > @@ -2945,6 +2955,8 @@ int pci_host_probe(struct pci_host_bridge
> > > *bridge)
> > > pci_bus_claim_resources(bus);
> > > } else {
> > > pci_bus_size_bridges(bus);
> > > + if (pci_can_move_bars)
> > > + pci_bus_update_realloc_range(bus);
> > > pci_bus_assign_resources(bus);
> > >
> > > list_for_each_entry(child, &bus->children, node)
> > > --
> > > 2.24.1
> > >
