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

From: Robin Murphy
Date: Wed Dec 04 2019 - 06:59:19 EST

On 2019-12-04 4:34 am, Xiaowei Bao wrote:


-----Original Message-----
From: Robin Murphy <robin.murphy@xxxxxxx>
Sent: 2019å12æ3æ 23:20
To: Marc Zyngier <maz@xxxxxxxxxx>; Xiaowei Bao <xiaowei.bao@xxxxxxx>
Cc: Roy Zang <roy.zang@xxxxxxx>; lorenzo.pieralisi@xxxxxxx;
devicetree@xxxxxxxxxxxxxxx; linux-pci@xxxxxxxxxxxxxxx; Z.q. Hou
<zhiqiang.hou@xxxxxxx>; linux-kernel@xxxxxxxxxxxxxxx; M.h. Lian
<minghuan.lian@xxxxxxx>; robh+dt@xxxxxxxxxx;
linux-arm-kernel@xxxxxxxxxxxxxxxxxxx; bhelgaas@xxxxxxxxxx;
andrew.murray@xxxxxxx; frowand.list@xxxxxxxxx; Mingkai Hu
<mingkai.hu@xxxxxxx>
Subject: Re: [PATCH] PCI: layerscape: Add the SRIOV support in host side

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>
---
 drivers/of/irq.c | 9 +++
 drivers/pci/controller/dwc/pci-layerscape.c | 94
+++++++++++++++++++++++++++++
 drivers/pci/probe.c | 6 ++
 drivers/pci/remove.c | 6 ++
 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 }
 }

+u32 __weak ls_pcie_streamid_fix(struct device *dev, u32 rid) {
+ÂÂÂ return rid;
+}
+
 static u32 __of_msi_map_rid(struct device *dev, struct
device_node **np,
ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ u32 rid_in)
 {
@@ -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?

Thanks for your comments, yes, this is the root cause, we only have 16 stream
IDs for PCIe domain, this is the hardware limitation, if there have enough stream
IDs, we can expose an msi-map/msi-map-mask for all PCIe devices in system,
unfortunately, the stream IDs is not enough, I think other ARM vendor have same
issue that they don't have enough stream IDs.

Some SMMUv2 configurations may have an uncomfortably limited number of context banks, but they almost always have more than enough stream ID bits. Your ICID allocation policy is most certainly an issue unique to Layerscape platforms.

Furthermore, that argument doesn't make a whole lot of sense anyway - if you don't have enough stream IDs for all possible VFs at boot time, then you still won't have enough later, so pretending to support SR-IOV, only for things to start subtly going wrong if the user has too many endpoints active at once, isn't going to cut it.

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>;
};

Thanks Robin, this is a effective way, but we only have total 16 stream IDs for PCIe domain,
and only assign 4 stream IDs for each PCIe controller if the board have 4 PCIe controllers,
this is the root cause, I submitted this patch to dynamically manage these stream IDs,
so that it looks like each PCIe controller has 16 stream IDs. I can dynamically allocate and
release these stream IDs based on the PCIe devices in the current system. If use your method,
we support up to 4 PCIe devices(2 PFs and 2 VFs), it will not achieve our purpose.

Sure, that was just an example to illustrate that you don't need a separate msi-map entry (and corresponding LUT entry) for each individual PCI RID - that dates from before U-Boot had ICID support, so I had hacks all over various kernel drivers to set them arbitrarily when I was playing with the SMMU.

Realistically, at this point your options are a) reserve more ICIDs for PCIe and allocate them in a way that accounts for the present endpoints' SR-IOV capabilities, or b) don't expose SR-IOV functionality at all on the root complex if it can't be guaranteed to work properly.

Robin.