On Fri, May 31, 2019 at 11:12:16AM -0600, Logan Gunthorpe wrote:
> One odd quirk of PLX switches is that their upstream bridge port has
> 256K of space allocated behind its BAR0 (most other bridge
> implementations do not report any BAR space).
Somewhat unusual, but completely legal, of course.
If a bridge has memory BARs, AFAIK it is impossible to enable a memory
window without also enabling the BARs, so if we want to use the bridge
at all, we *must* allocate space for its BARs, just like for any other
device.
> The lspci for such device
> looks like:
>
> 04:00.0 PCI bridge: PLX Technology, Inc. PEX 8724 24-Lane, 6-Port PCI
> Express Gen 3 (8 GT/s) Switch, 19 x 19mm FCBGA (rev ca)
> (prog-if 00 [Normal decode])
> Physical Slot: 1
> Flags: bus master, fast devsel, latency 0, IRQ 30, NUMA node 0
> Memory at 90a00000 (32-bit, non-prefetchable) [size=256K]
> Bus: primary=04, secondary=05, subordinate=0a, sec-latency=0
> I/O behind bridge: 00002000-00003fff
> Memory behind bridge: 90000000-909fffff
> Prefetchable memory behind bridge: 0000380000800000-0000380000bfffff
> Kernel driver in use: pcieport
>
> It's not clear what the purpose of the memory at 0x90a00000 is, and
> currently the kernel never actually uses it for anything. In most cases,
> it's safely ignored and does not cause a problem.
>
> However, when the kernel assigns the resource addresses (with the
> pci=realloc command line parameter, for example) it can inadvertently
> disable the struct resource corresponding to the bar. When this happens,
> lspci will report this memory as ignored:
>
> Region 0: Memory at <ignored> (32-bit, non-prefetchable) [size=256K]
>
> This is because the kernel reports a zero start address and zero flags
> in the corresponding sysfs resource file and in /proc/bus/pci/devices.
> Investigation with 'lspci -x', however shows the bios-assigned address
> will still be programmed in the device's BAR registers.
Ugh, yep. It took me a while to trace through this, but "lspci -v" by
default shows the kernel view of addresses, i.e., the pdev->resource[]
values, which it gets from the sysfs "resource" file (resource_show())
or "/proc/bus/pci/devices" (show_device()).
But "lspci -x" shows the config space values (I think "lspci -bv"
should also show these) from the sysfs "config" file
(pci_read_config()).
It's definitely a kernel bug that we lost track of what's in config
space.
> In many cases, this still isn't a problem. Nothing uses the memory,
> so nothing is affected. However, a big problem shows up when an IOMMU
> is in use: the IOMMU will not reserve this space in the IOVA because the
> kernel no longer thinks the range is valid. (See
> dmar_init_reserved_ranges() for the Intel implementation of this.)
>
> Without the proper reserved range, we have a situation where a DMA
> mapping may occasionally allocate an IOVA which the PCI bus will actually
> route to a BAR in the PLX switch. This will result in some random DMA
> writes not actually writing to the RAM they are supposed to, or random
> DMA reads returning all FFs from the PLX BAR when it's supposed to have
> read from RAM.
>
> The problem is caused in pci_assign_unassigned_root_bus_resources().
> When any resource from a bridge device fails to get assigned, the code
> sets the resource's flags to zero. This makes sense for bridge resources,
> as they will be re-enabled later, but for regular BARs, it disables them
> permanently. To fix the problem, we only set the flags to zero for
> bridge resources and treat any other resources like non-bridge devices.
>
> Reported-by: Kit Chow <kchow@xxxxxxxxxx>
> Fixes: da7822e5ad71 ("PCI: update bridge resources to get more big ranges when allocating space (again)")
> Signed-off-by: Logan Gunthorpe <logang@xxxxxxxxxxxx>
> Cc: Bjorn Helgaas <bhelgaas@xxxxxxxxxx>
> Cc: Yinghai Lu <yinghai@xxxxxxxxxx>
> ---
> drivers/pci/setup-bus.c | 7 ++++++-
> 1 file changed, 6 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/pci/setup-bus.c b/drivers/pci/setup-bus.c
> index 0eb40924169b..7adbd4bedd16 100644
> --- a/drivers/pci/setup-bus.c
> +++ b/drivers/pci/setup-bus.c
> @@ -1784,11 +1784,16 @@ void pci_assign_unassigned_root_bus_resources(struct pci_bus *bus)
> /* restore size and flags */
> list_for_each_entry(fail_res, &fail_head, list) {
> struct resource *res = fail_res->res;
> + int idx;
>
> res->start = fail_res->start;
> res->end = fail_res->end;
> res->flags = fail_res->flags;
> - if (fail_res->dev->subordinate)
> +
> + idx = res - &fail_res->dev->resource[0];
> + if (fail_res->dev->subordinate &&
> + idx >= PCI_BRIDGE_RESOURCES &&
> + idx <= PCI_BRIDGE_RESOURCE_END)
> res->flags = 0;
In my ideal world we wouldn't zap the flags of any resource. I think
we should derive the flags from the device's config space *once*
during enumeration and remember them for the life of the device.
This patch preserves res->flags for bridge BARs just like for any
other device, so I think this is definitely a step in the right
direction.
I'm not sure the "dev->subordinate" test is really correct, though.
I think the original intent of this code was to clear res->flags for
bridge windows under the assumptions that (a) we can identify bridges
by "dev->subordinate" being non-zero, and (b) bridges only have
windows and didn't have BARs.
This patch fixes assumption (b), but I think (a) is false, and we
should fix it as well. One can imagine a bridge device without a
subordinate bus (maybe we ran out of bus numbers), so I don't think we
should test dev->subordinate.
We could test something like pci_is_bridge(), although testing for idx
being in the PCI_BRIDGE_RESOURCES range should be sufficient because I
don't think we use those resource for anything other than windows.
> }
> free_list(&fail_head);
> --
> 2.20.1
>
