Re: [PATCH] PCI: layerscape: Add the SRIOV support in host side

From: Robin Murphy
Date: Tue Dec 03 2019 - 10:20:32 EST

On 03/12/2019 11:51 am, Marc Zyngier wrote:
On 2019-12-03 01:42, Xiaowei Bao wrote:
-----Original Message-----
From: Marc Zyngier <maz@xxxxxxxxxxxxxxx>
Sent: 2019å12æ2æ 20:48
To: Xiaowei Bao <xiaowei.bao@xxxxxxx>
Cc: robh+dt@xxxxxxxxxx; frowand.list@xxxxxxxxx; M.h. Lian
<minghuan.lian@xxxxxxx>; Mingkai Hu <mingkai.hu@xxxxxxx>; Roy Zang
<roy.zang@xxxxxxx>; lorenzo.pieralisi@xxxxxxx; andrew.murray@xxxxxxx;
bhelgaas@xxxxxxxxxx; devicetree@xxxxxxxxxxxxxxx;
linux-kernel@xxxxxxxxxxxxxxx; linux-pci@xxxxxxxxxxxxxxx;
linux-arm-kernel@xxxxxxxxxxxxxxxxxxx; Z.q. Hou <zhiqiang.hou@xxxxxxx>
Subject: Re: [PATCH] PCI: layerscape: Add the SRIOV support in host side

On 2019-12-02 10:45, Xiaowei Bao wrote:
> GIC get the map relations of devid and stream id from the msi-map
> property of DTS, our platform add this property in u-boot base on the
> PCIe device in the bus, but if enable the vf device in kernel, the vf
> device msi-map will not set, so the vf device can't work, this patch
> purpose is that manage the stream id and device id map relations
> dynamically in kernel, and make the new PCIe device work in kernel.
>
> Signed-off-by: Xiaowei Bao <xiaowei.bao@xxxxxxx>
> ---
>Â drivers/of/irq.cÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ |Â 9 +++
>Â drivers/pci/controller/dwc/pci-layerscape.c | 94
> +++++++++++++++++++++++++++++
>Â drivers/pci/probe.cÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ |Â 6 ++
>Â drivers/pci/remove.cÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ |Â 6 ++
>Â 4 files changed, 115 insertions(+)
>
> diff --git a/drivers/of/irq.c b/drivers/of/irq.c index
> a296eaf..791e609 100644
> --- a/drivers/of/irq.c
> +++ b/drivers/of/irq.c
> @@ -576,6 +576,11 @@ void __init of_irq_init(const struct of_device_id
> *matches)
>ÂÂÂÂÂ }
>Â }
>
> +u32 __weak ls_pcie_streamid_fix(struct device *dev, u32 rid) {
> +ÂÂÂ return rid;
> +}
> +
>Â static u32 __of_msi_map_rid(struct device *dev, struct device_node
> **np,
>ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ u32 rid_in)
>Â {
> @@ -590,6 +595,10 @@ static u32 __of_msi_map_rid(struct device *dev,
> struct device_node **np,
>ÂÂÂÂÂÂÂÂÂ if (!of_map_rid(parent_dev->of_node, rid_in, "msi-map",
>ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ "msi-map-mask", np, &rid_out))
>ÂÂÂÂÂÂÂÂÂÂÂÂÂ break;
> +
> +ÂÂÂ if (rid_out == rid_in)
> +ÂÂÂÂÂÂÂ rid_out = ls_pcie_streamid_fix(parent_dev, rid_in);

Over my dead body. Get your firmware to properly program the LUT so that it
presents the ITS with a reasonable topology. There is absolutely no way this
kind of change makes it into the kernel.

Sorry for this, I know it is not reasonable, but I have no other way,
as I know, ARM
get the mapping of stream ID to request ID from the msi-map property
of DTS, if
add a new device which need the stream ID and try to get it from the
msi-map of DTS,
it will failed and not work, yes? So could you give me a better
advice to fix this issue,
I would really appreciate any comments or suggestions, thanks a lot.

Why can't firmware expose an msi-map/msi-map-mask that has a large
enough range to ensure mapping of VFs? What are the limitations of
the LUT that would prevent this from being configured before the
kernel boots?

Furthermore, note that this attempt isn't doing anything for the SMMU Stream IDs, so the moment anyone tries to assign those VFs they're still going to go bang anyway. Any firmware-based fixup for ID mappings, config space addresses, etc. needs to be SR-IOV-aware and account for all *possible* BDFs.

On LS2085 at least, IIRC you can configure a single LUT entry to just translate the Bus:Device identifier and pass some or all of the Function bits straight through as the LSBs of the Stream ID, so I don't believe the relatively limited number of LUT registers should be too much of an issue. For example, last time I hacked on that I apparently had it set up statically like this:

&pcie3 {
/* Squash 8:5:3 BDF down to 2:2:3 */
msi-map-mask = <0x031f>;
msi-map = <0x000 &its 0x00 0x20>,
<0x100 &its 0x20 0x20>,
<0x200 &its 0x40 0x20>,
<0x300 &its 0x60 0x20>;
};

Robin.