[PATCH 0/1] RPi: Add Linux PCIe quirk

[PATCH 0/1] RPi: Add Linux PCIe quirk

[Jeremy Linton]

Linux has been refusing to use the generic SMC
PCIe config method, so we need to add a bit of additional
linux specific information to support a new rpi4 quirk.

This patch depends on the general XHCI/PCIe SMC set.

Jeremy Linton (1):
  Platform/RaspberryPi: Add linux quirk support

 Platform/RaspberryPi/AcpiTables/Pci.asl | 21 +++++++++++++++++++++
 1 file changed, 21 insertions(+)

-- 
2.13.7

[PATCH 1/1] Platform/RaspberryPi: Add linux quirk support

[Jeremy Linton]

Linux, for the time being has refused to support the Arm
standard SMCCC for PCIe configuration. Instead they
want to continue to maintain per device "quirks".

As the RPI isn't really ECAM this is a bit more
involved because the MADT can't really describe
the root port+config registers situation. Further
platforms which support the SMCCC shouldn't have
a MADT, so we need an additional way to tell linux
what it needs to know about this platform.

Signed-off-by: Jeremy Linton <jeremy.linton@arm.com>
---
 Platform/RaspberryPi/AcpiTables/Pci.asl | 21 +++++++++++++++++++++
 1 file changed, 21 insertions(+)

diff --git a/Platform/RaspberryPi/AcpiTables/Pci.asl b/Platform/RaspberryPi/AcpiTables/Pci.asl
index 34474f13ef..3e7fd0d5b7 100644
--- a/Platform/RaspberryPi/AcpiTables/Pci.asl
+++ b/Platform/RaspberryPi/AcpiTables/Pci.asl
@@ -123,6 +123,15 @@ DefinitionBlock (__FILE__, "SSDT", 5, "RPIFDN", "RPI4PCIE", 2)
         Name(_BBN, Zero) // PCI Base Bus Number
         Name(_CCA, 0)    // Mark the PCI noncoherent
 
+        Name (_DSD, Package () {
+          ToUUID("daffd814-6eba-4d8c-8a91-bc9bbf4aa301"),
+            Package () {
+              Package () { "linux,pcie-quirk", "bcm2711" },
+              Package () { "linux,pcie-nomsi", 1 },
+            }
+        })
+
+
         // Root Complex 0
         Device (RP0) {
          Name(_ADR, 0xF0000000)    // Dev 0, Func 0
@@ -176,6 +185,18 @@ DefinitionBlock (__FILE__, "SSDT", 5, "RPIFDN", "RPI4PCIE", 2)
               2                               // SANITIZED_PCIE_MMIO_LEN + 1
               ,,,MMI1,,TypeTranslation
             )
+
+            QWordMemory ( // Root port registers, not to be used if SMCCC is utilized
+              ResourceConsumer, ,
+              MinFixed, MaxFixed,
+              NonCacheable, ReadWrite,        // cacheable? is that right?
+              0x00000000,                     // Granularity
+              0xFD500000,                     // Root port begin
+              0xFD509FFF,                     // Root port end
+              0x00000000,                     // no translation
+              0x0000A000,                     // size
+              ,,
+            )
           }) // end Name(RBUF)
 
           // Work around ASL's inability to add in a resource definition
-- 
2.13.7

Re: [PATCH 1/1] Platform/RaspberryPi: Add linux quirk support

[Andrei Warkentin]

[-- Attachment #1: Type: text/plain, Size: 4191 bytes --]

Hi Jeremy,

Reviewed-by: Andrei Warkentin <awarkentin@vmware.com>

MADT -> MCFG

Root port registers would be NonCacheable just like the outbound mapping. (Also, with https://mantis.uefi.org/mantis/view.php?id=2220, _MEM attributes aside from  NonCacheable and Prefetchable are effectively deprecated).

Note: I really wish we had sorted out the HID/CID story for the PCIe RC... i.e. at least to make the HID custom for non-ECAM implementations and push PNP0A08/PNP0A03 in the CID... this would make linux,pcie-quirk unnecessary, using standard ACPI driver binding mechanisms to separate quirks from proper ECAM. Sadly, I think the train to do that has long left the station (with so many ACPI Arm systems out there and non-ECAM or not-quite-ECAM RCs using PNP0A08/PNP0A03 alone. ESXi, for example, ends up keying on the Table Ids (which is arguably even worse than the DT props)

Note 2: Given that DT has users in U-Boot and the BSDs (and potentially anywhere else), there's a long term hope to make DT bindings separate from Linux. To that end "linux,pcie-quirk" and "linux,pcie-nomsi" should probably be named something else (although I recognise that you're probably just wiring up something that already exists).

--
Andrei Warkentin,
Arm Enablement Architect,
Cloud Platform Business Unit, VMware
________________________________
From: Jeremy Linton <jeremy.linton@arm.com>
Sent: Thursday, August 5, 2021 7:40 PM
To: devel@edk2.groups.io <devel@edk2.groups.io>
Cc: pete@akeo.ie <pete@akeo.ie>; ardb+tianocore@kernel.org <ardb+tianocore@kernel.org>; Andrei Warkentin <awarkentin@vmware.com>; Sunny.Wang@arm.com <Sunny.Wang@arm.com>; samer.el-haj-mahmoud@arm.com <samer.el-haj-mahmoud@arm.com>; Jeremy Linton <jeremy.linton@arm.com>
Subject: [PATCH 1/1] Platform/RaspberryPi: Add linux quirk support

Linux, for the time being has refused to support the Arm
standard SMCCC for PCIe configuration. Instead they
want to continue to maintain per device "quirks".

As the RPI isn't really ECAM this is a bit more
involved because the MADT can't really describe
the root port+config registers situation. Further
platforms which support the SMCCC shouldn't have
a MADT, so we need an additional way to tell linux
what it needs to know about this platform.

Signed-off-by: Jeremy Linton <jeremy.linton@arm.com>
---
 Platform/RaspberryPi/AcpiTables/Pci.asl | 21 +++++++++++++++++++++
 1 file changed, 21 insertions(+)

diff --git a/Platform/RaspberryPi/AcpiTables/Pci.asl b/Platform/RaspberryPi/AcpiTables/Pci.asl
index 34474f13ef..3e7fd0d5b7 100644
--- a/Platform/RaspberryPi/AcpiTables/Pci.asl
+++ b/Platform/RaspberryPi/AcpiTables/Pci.asl
@@ -123,6 +123,15 @@ DefinitionBlock (__FILE__, "SSDT", 5, "RPIFDN", "RPI4PCIE", 2)
         Name(_BBN, Zero) // PCI Base Bus Number
         Name(_CCA, 0)    // Mark the PCI noncoherent

+        Name (_DSD, Package () {
+          ToUUID("daffd814-6eba-4d8c-8a91-bc9bbf4aa301"),
+            Package () {
+              Package () { "linux,pcie-quirk", "bcm2711" },
+              Package () { "linux,pcie-nomsi", 1 },
+            }
+        })
+
+
         // Root Complex 0
         Device (RP0) {
          Name(_ADR, 0xF0000000)    // Dev 0, Func 0
@@ -176,6 +185,18 @@ DefinitionBlock (__FILE__, "SSDT", 5, "RPIFDN", "RPI4PCIE", 2)
               2                               // SANITIZED_PCIE_MMIO_LEN + 1
               ,,,MMI1,,TypeTranslation
             )
+
+            QWordMemory ( // Root port registers, not to be used if SMCCC is utilized
+              ResourceConsumer, ,
+              MinFixed, MaxFixed,
+              NonCacheable, ReadWrite,        // cacheable? is that right?
+              0x00000000,                     // Granularity
+              0xFD500000,                     // Root port begin
+              0xFD509FFF,                     // Root port end
+              0x00000000,                     // no translation
+              0x0000A000,                     // size
+              ,,
+            )
           }) // end Name(RBUF)

           // Work around ASL's inability to add in a resource definition
--
2.13.7


[-- Attachment #2: Type: text/html, Size: 10155 bytes --]

Re: [PATCH 1/1] Platform/RaspberryPi: Add linux quirk support

[Jeremy Linton]

Hi,

On 8/6/21 10:47 AM, Andrei Warkentin wrote:
> Hi Jeremy,
> 
> Reviewed-by: Andrei Warkentin <awarkentin@vmware.com>
> 
> MADT -> MCFG

Yah, I have to start thinking about what I'm typing :)

> 
> Root port registers would be NonCacheable just like the outbound mapping. (Also, with https://mantis.uefi.org/mantis/view.php?id=2220, _MEM attributes aside from  NonCacheable and Prefetchable are effectively deprecated).
> 
> Note: I really wish we had sorted out the HID/CID story for the PCIe RC... i.e. at least to make the HID custom for non-ECAM implementations and push PNP0A08/PNP0A03 in the CID... this would make linux,pcie-quirk unnecessary, using standard ACPI driver binding mechanisms to separate quirks from proper ECAM. Sadly, I think the train to do that has long left the station (with so many ACPI Arm systems out there and non-ECAM or not-quite-ECAM RCs using PNP0A08/PNP0A03 alone. ESXi, for example, ends up keying on the Table Ids (which is arguably even worse than the DT props)

This isn't a bad idea (changing the HID), but I think your right that 
ship might have sailed, although it might be still be an alternative for 
the SMC.


> 
> Note 2: Given that DT has users in U-Boot and the BSDs (and potentially anywhere else), there's a long term hope to make DT bindings separate from Linux. To that end "linux,pcie-quirk" and "linux,pcie-nomsi" should probably be named something else (although I recognise that you're probably just wiring up something that already exists).

Well I want this one to be linux specific, because hopefully everyone 
else realizes that the SMC is a better plan and ignores the property. 
Its not set anywhere so I can potentially change it. The linux posting 
is here:

https://lkml.org/lkml/2021/8/5/1109

(where I apparently typed MCFG instead of MADT correctly, well at least 
most of the time :) .

But I don't have a problem dropping the linux bit, although I would like 
that part to be generic so future SMC+quirked platforms don't have to 
modify the quirk detection code. Although lets see what Lorenzo/etc say 
about what i'm doing in that patch. Thats why I left this patch off the 
other set.



> 
> --
> Andrei Warkentin,
> Arm Enablement Architect,
> Cloud Platform Business Unit, VMware
> ________________________________
> From: Jeremy Linton <jeremy.linton@arm.com>
> Sent: Thursday, August 5, 2021 7:40 PM
> To: devel@edk2.groups.io <devel@edk2.groups.io>
> Cc: pete@akeo.ie <pete@akeo.ie>; ardb+tianocore@kernel.org <ardb+tianocore@kernel.org>; Andrei Warkentin <awarkentin@vmware.com>; Sunny.Wang@arm.com <Sunny.Wang@arm.com>; samer.el-haj-mahmoud@arm.com <samer.el-haj-mahmoud@arm.com>; Jeremy Linton <jeremy.linton@arm.com>
> Subject: [PATCH 1/1] Platform/RaspberryPi: Add linux quirk support
> 
> Linux, for the time being has refused to support the Arm
> standard SMCCC for PCIe configuration. Instead they
> want to continue to maintain per device "quirks".
> 
> As the RPI isn't really ECAM this is a bit more
> involved because the MADT can't really describe
> the root port+config registers situation. Further
> platforms which support the SMCCC shouldn't have
> a MADT, so we need an additional way to tell linux
> what it needs to know about this platform.
> 
> Signed-off-by: Jeremy Linton <jeremy.linton@arm.com>
> ---
>   Platform/RaspberryPi/AcpiTables/Pci.asl | 21 +++++++++++++++++++++
>   1 file changed, 21 insertions(+)
> 
> diff --git a/Platform/RaspberryPi/AcpiTables/Pci.asl b/Platform/RaspberryPi/AcpiTables/Pci.asl
> index 34474f13ef..3e7fd0d5b7 100644
> --- a/Platform/RaspberryPi/AcpiTables/Pci.asl
> +++ b/Platform/RaspberryPi/AcpiTables/Pci.asl
> @@ -123,6 +123,15 @@ DefinitionBlock (__FILE__, "SSDT", 5, "RPIFDN", "RPI4PCIE", 2)
>           Name(_BBN, Zero) // PCI Base Bus Number
>           Name(_CCA, 0)    // Mark the PCI noncoherent
> 
> +        Name (_DSD, Package () {
> +          ToUUID("daffd814-6eba-4d8c-8a91-bc9bbf4aa301"),
> +            Package () {
> +              Package () { "linux,pcie-quirk", "bcm2711" },
> +              Package () { "linux,pcie-nomsi", 1 },
> +            }
> +        })
> +
> +
>           // Root Complex 0
>           Device (RP0) {
>            Name(_ADR, 0xF0000000)    // Dev 0, Func 0
> @@ -176,6 +185,18 @@ DefinitionBlock (__FILE__, "SSDT", 5, "RPIFDN", "RPI4PCIE", 2)
>                 2                               // SANITIZED_PCIE_MMIO_LEN + 1
>                 ,,,MMI1,,TypeTranslation
>               )
> +
> +            QWordMemory ( // Root port registers, not to be used if SMCCC is utilized
> +              ResourceConsumer, ,
> +              MinFixed, MaxFixed,
> +              NonCacheable, ReadWrite,        // cacheable? is that right?
> +              0x00000000,                     // Granularity
> +              0xFD500000,                     // Root port begin
> +              0xFD509FFF,                     // Root port end
> +              0x00000000,                     // no translation
> +              0x0000A000,                     // size
> +              ,,
> +            )
>             }) // end Name(RBUF)
> 
>             // Work around ASL's inability to add in a resource definition
> --
> 2.13.7
> 
>