Skip to contentSkip to conversation

PCI: Add Extended Tag + MRRS quirk for Xeon 6

PATCH Superseded Ilpo Järvinen 6 participants 8 messages Apr 22, 2025 – Jun 2, 2025

Subsystems (2)

v2 of 3 Superseded Latest is v3 →

lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 PCI: Add Extended Tag + MRRS quirk for Xeon 6

When bifurcated to x2, Xeon 6 Root Port performance is sensitive to the configuration of Extended Tags, Max Read Request Size (MRRS), and 10-Bit Tag Requester (note: there is currently no 10-Bit Tag support in the kernel). While those can be configured to the recommended values by FW, kernel may decide to overwrite the initial values.

Unfortunately, there is no mechanism for FW to indicate OS which parts of PCIe configuration should not be altered. Thus, the only option is to add such logic into the kernel as quirks.

There is a pre-existing quirk flag to disable Extended Tags. Depending on CONFIG_PCIE_BUS_* setting, MRRS may be overwritten by what the kernel thinks is the best for performance (the largest supported value), resulting in performance degradation instead with these Root Ports. (There would have been a pre-existing quirk to disallow increasing MRRS but it is not identical to rejecting >128B MRRS.)

Add a quirk that disallows enabling Extended Tags and setting MRRS larger than 128B for devices under Xeon 6 Root Ports if the Root Port is bifurcated to x2. Reject >128B MRRS only when it is going to be written by the kernel (this assumes FW configured a good initial value for MRRS in case the kernel is not touching MRRS at all).

It was first attempted to always write MRRS when the quirk is needed (always overwrite the initial value). That turned out to be quite invasive change, however, given the complexity of the initial setup callchain and various stages returning early when they decide no changes are necessary, requiring override each. As such, the initial value for MRRS is now left into the hands of FW.

---

v2:
- Explain in changelog why FW cannot solve this on its own
- Moved the quirk under arch/x86/pci/
- Don't NULL check value from pci_find_host_bridge()
- Added comment above the quirk about the performance degradation
- Removed all setup chain 128B quirk overrides expect for MRRS write itself (assumes a sane initial value is set by FW)

 arch/x86/pci/fixup.c | 30 ++++++++++++++++++++++++++++++
 drivers/pci/pci.c    | 15 ++++++++-------
 include/linux/pci.h  |  1 +
 3 files changed, 39 insertions(+), 7 deletions(-)
arch/x86/pci/fixup.c +30 -0
index efefeb82ab61..aa9617bc4b55 100644
@@ -294,6 +294,36 @@ DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_INTEL,	PCI_DEVICE_ID_INTEL_MCH_PB1,	pcie_r
 DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_INTEL,	PCI_DEVICE_ID_INTEL_MCH_PC,	pcie_rootport_aspm_quirk);
 DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_INTEL,	PCI_DEVICE_ID_INTEL_MCH_PC1,	pcie_rootport_aspm_quirk);
 
+/*
+ * PCIe devices underneath Xeon6 PCIe Root Port bifurcated to 2x have slower
+ * performance with Extended Tags and MRRS > 128B. Workaround the performance
+ * problems by disabling Extended Tags and limiting MRRS to 128B.
+ *
+ * https://cdrdv2.intel.com/v1/dl/getContent/837176
+ */
+static void quirk_pcie2x_no_tags_no_mrrs(struct pci_dev *pdev)
+{
+	struct pci_host_bridge *bridge = pci_find_host_bridge(pdev->bus);
+	u32 linkcap;
+
+	pcie_capability_read_dword(pdev, PCI_EXP_LNKCAP, &linkcap);
+	if (FIELD_GET(PCI_EXP_LNKCAP_MLW, linkcap) != 0x2)
+		return;
+
+	bridge->no_ext_tags = 1;
+	bridge->only_128b_mrrs = 1;
+	pci_info(pdev, "Disabling Extended Tags and limiting MRRS to 128B (performance reasons due to 2x PCIe link)\n");
+}
+
+DECLARE_PCI_FIXUP_EARLY(PCI_VENDOR_ID_INTEL, 0x0db0, quirk_pcie2x_no_tags_no_mrrs);
+DECLARE_PCI_FIXUP_EARLY(PCI_VENDOR_ID_INTEL, 0x0db1, quirk_pcie2x_no_tags_no_mrrs);
+DECLARE_PCI_FIXUP_EARLY(PCI_VENDOR_ID_INTEL, 0x0db2, quirk_pcie2x_no_tags_no_mrrs);
+DECLARE_PCI_FIXUP_EARLY(PCI_VENDOR_ID_INTEL, 0x0db3, quirk_pcie2x_no_tags_no_mrrs);
+DECLARE_PCI_FIXUP_EARLY(PCI_VENDOR_ID_INTEL, 0x0db6, quirk_pcie2x_no_tags_no_mrrs);
+DECLARE_PCI_FIXUP_EARLY(PCI_VENDOR_ID_INTEL, 0x0db7, quirk_pcie2x_no_tags_no_mrrs);
+DECLARE_PCI_FIXUP_EARLY(PCI_VENDOR_ID_INTEL, 0x0db8, quirk_pcie2x_no_tags_no_mrrs);
+DECLARE_PCI_FIXUP_EARLY(PCI_VENDOR_ID_INTEL, 0x0db9, quirk_pcie2x_no_tags_no_mrrs);
+
 /*
  * Fixup to mark boot BIOS video selected by BIOS before it changes
  *
drivers/pci/pci.c +8 -7
index 4d7c9f64ea24..2ca9cb30fbd3 100644
@@ -5941,7 +5941,7 @@ EXPORT_SYMBOL(pcie_get_readrq);
 int pcie_set_readrq(struct pci_dev *dev, int rq)
 {
 	u16 v;
-	int ret;
+	int ret, max_mrrs = 4096;
 	struct pci_host_bridge *bridge = pci_find_host_bridge(dev->bus);
 
 	if (rq < 128 || rq > 4096 || !is_power_of_2(rq))
@@ -5961,13 +5961,14 @@ int pcie_set_readrq(struct pci_dev *dev, int rq)
 
 	v = FIELD_PREP(PCI_EXP_DEVCTL_READRQ, ffs(rq) - 8);
 
-	if (bridge->no_inc_mrrs) {
-		int max_mrrs = pcie_get_readrq(dev);
+	if (bridge->no_inc_mrrs)
+		max_mrrs = pcie_get_readrq(dev);
+	if (bridge->only_128b_mrrs)
+		max_mrrs = 128;
 
-		if (rq > max_mrrs) {
-			pci_info(dev, "can't set Max_Read_Request_Size to %d; max is %d\n", rq, max_mrrs);
-			return -EINVAL;
-		}
+	if (rq > max_mrrs) {
+		pci_info(dev, "can't set Max_Read_Request_Size to %d; max is %d\n", rq, max_mrrs);
+		return -EINVAL;
 	}
 
 	ret = pcie_capability_clear_and_set_word(dev, PCI_EXP_DEVCTL,
include/linux/pci.h +1 -0
index 0e8e3fd77e96..6dc7a05f4d4b 100644
@@ -603,6 +603,7 @@ struct pci_host_bridge {
 	unsigned int	ignore_reset_delay:1;	/* For entire hierarchy */
 	unsigned int	no_ext_tags:1;		/* No Extended Tags */
 	unsigned int	no_inc_mrrs:1;		/* No Increase MRRS */
+	unsigned int	only_128b_mrrs:1;	/* Only 128B MRRS */
 	unsigned int	native_aer:1;		/* OS may use PCIe AER */
 	unsigned int	native_pcie_hotplug:1;	/* OS may use PCIe hotplug */
 	unsigned int	native_shpc_hotplug:1;	/* OS may use SHPC hotplug */

base-commit: 0af2f6be1b4281385b618cb86ad946eded089ac8

Signature · 1 line
2.39.5
#1 lore raw
Quoted 30 lines · Ilpo Järvinen

When bifurcated to x2, Xeon 6 Root Port performance is sensitive to the configuration of Extended Tags, Max Read Request Size (MRRS), and 10-Bit Tag Requester (note: there is currently no 10-Bit Tag support in the kernel). While those can be configured to the recommended values by FW, kernel may decide to overwrite the initial values.

Unfortunately, there is no mechanism for FW to indicate OS which parts of PCIe configuration should not be altered. Thus, the only option is to add such logic into the kernel as quirks.

There is a pre-existing quirk flag to disable Extended Tags. Depending on CONFIG_PCIE_BUS_* setting, MRRS may be overwritten by what the kernel thinks is the best for performance (the largest supported value), resulting in performance degradation instead with these Root Ports. (There would have been a pre-existing quirk to disallow increasing MRRS but it is not identical to rejecting >128B MRRS.)

Add a quirk that disallows enabling Extended Tags and setting MRRS larger than 128B for devices under Xeon 6 Root Ports if the Root Port is bifurcated to x2. Reject >128B MRRS only when it is going to be written by the kernel (this assumes FW configured a good initial value for MRRS in case the kernel is not touching MRRS at all).

It was first attempted to always write MRRS when the quirk is needed (always overwrite the initial value). That turned out to be quite invasive change, however, given the complexity of the initial setup callchain and various stages returning early when they decide no changes are necessary, requiring override each. As such, the initial value for MRRS is now left into the hands of FW.

Quoted 3 lines

---

v1/rfc: https://lore.kernel.org/all/[email protected]/

I suppose it's quixotic to hope for anything better than quirks that have to be added or updated for every new processor that comes along.

ACPI _HPX might be a possible way for the platform to tell us what to do here. ACPI r6.5, sec 6.2.9 says it's for hot-added devices and "Functions not configured by the platform firmware during initial system boot" (how are we supposed to determine that?) In any case, Linux does evaluate _HPX for every device in pci_configure_device().

I'm not sure _HPX really works; it's very general, and I would expect to see reports of problems if firmware really tried to use it.

Or, I guess a _DSM function would be a possible way to communicate this.

Quoted 105 lines

v2:
- Explain in changelog why FW cannot solve this on its own
- Moved the quirk under arch/x86/pci/
- Don't NULL check value from pci_find_host_bridge()
- Added comment above the quirk about the performance degradation
- Removed all setup chain 128B quirk overrides expect for MRRS write itself (assumes a sane initial value is set by FW)

 arch/x86/pci/fixup.c | 30 ++++++++++++++++++++++++++++++
 drivers/pci/pci.c    | 15 ++++++++-------
 include/linux/pci.h  |  1 +
 3 files changed, 39 insertions(+), 7 deletions(-)
arch/x86/pci/fixup.c +30 -0
index efefeb82ab61..aa9617bc4b55 100644
@@ -294,6 +294,36 @@ DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_INTEL,	PCI_DEVICE_ID_INTEL_MCH_PB1,	pcie_r
 DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_INTEL,	PCI_DEVICE_ID_INTEL_MCH_PC,	pcie_rootport_aspm_quirk);
 DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_INTEL,	PCI_DEVICE_ID_INTEL_MCH_PC1,	pcie_rootport_aspm_quirk);
 
+/*
+ * PCIe devices underneath Xeon6 PCIe Root Port bifurcated to 2x have slower
+ * performance with Extended Tags and MRRS > 128B. Workaround the performance
+ * problems by disabling Extended Tags and limiting MRRS to 128B.
+ *
+ * https://cdrdv2.intel.com/v1/dl/getContent/837176
+ */
+static void quirk_pcie2x_no_tags_no_mrrs(struct pci_dev *pdev)
+{
+	struct pci_host_bridge *bridge = pci_find_host_bridge(pdev->bus);
+	u32 linkcap;
+
+	pcie_capability_read_dword(pdev, PCI_EXP_LNKCAP, &linkcap);
+	if (FIELD_GET(PCI_EXP_LNKCAP_MLW, linkcap) != 0x2)
+		return;
+
+	bridge->no_ext_tags = 1;
+	bridge->only_128b_mrrs = 1;
+	pci_info(pdev, "Disabling Extended Tags and limiting MRRS to 128B (performance reasons due to 2x PCIe link)\n");
+}
+
+DECLARE_PCI_FIXUP_EARLY(PCI_VENDOR_ID_INTEL, 0x0db0, quirk_pcie2x_no_tags_no_mrrs);
+DECLARE_PCI_FIXUP_EARLY(PCI_VENDOR_ID_INTEL, 0x0db1, quirk_pcie2x_no_tags_no_mrrs);
+DECLARE_PCI_FIXUP_EARLY(PCI_VENDOR_ID_INTEL, 0x0db2, quirk_pcie2x_no_tags_no_mrrs);
+DECLARE_PCI_FIXUP_EARLY(PCI_VENDOR_ID_INTEL, 0x0db3, quirk_pcie2x_no_tags_no_mrrs);
+DECLARE_PCI_FIXUP_EARLY(PCI_VENDOR_ID_INTEL, 0x0db6, quirk_pcie2x_no_tags_no_mrrs);
+DECLARE_PCI_FIXUP_EARLY(PCI_VENDOR_ID_INTEL, 0x0db7, quirk_pcie2x_no_tags_no_mrrs);
+DECLARE_PCI_FIXUP_EARLY(PCI_VENDOR_ID_INTEL, 0x0db8, quirk_pcie2x_no_tags_no_mrrs);
+DECLARE_PCI_FIXUP_EARLY(PCI_VENDOR_ID_INTEL, 0x0db9, quirk_pcie2x_no_tags_no_mrrs);
+
 /*
  * Fixup to mark boot BIOS video selected by BIOS before it changes
  *
drivers/pci/pci.c +8 -7
index 4d7c9f64ea24..2ca9cb30fbd3 100644
@@ -5941,7 +5941,7 @@ EXPORT_SYMBOL(pcie_get_readrq);
 int pcie_set_readrq(struct pci_dev *dev, int rq)
 {
 	u16 v;
-	int ret;
+	int ret, max_mrrs = 4096;
 	struct pci_host_bridge *bridge = pci_find_host_bridge(dev->bus);
 
 	if (rq < 128 || rq > 4096 || !is_power_of_2(rq))
@@ -5961,13 +5961,14 @@ int pcie_set_readrq(struct pci_dev *dev, int rq)
 
 	v = FIELD_PREP(PCI_EXP_DEVCTL_READRQ, ffs(rq) - 8);
 
-	if (bridge->no_inc_mrrs) {
-		int max_mrrs = pcie_get_readrq(dev);
+	if (bridge->no_inc_mrrs)
+		max_mrrs = pcie_get_readrq(dev);
+	if (bridge->only_128b_mrrs)
+		max_mrrs = 128;
 
-		if (rq > max_mrrs) {
-			pci_info(dev, "can't set Max_Read_Request_Size to %d; max is %d\n", rq, max_mrrs);
-			return -EINVAL;
-		}
+	if (rq > max_mrrs) {
+		pci_info(dev, "can't set Max_Read_Request_Size to %d; max is %d\n", rq, max_mrrs);
+		return -EINVAL;
 	}
 
 	ret = pcie_capability_clear_and_set_word(dev, PCI_EXP_DEVCTL,
include/linux/pci.h +1 -0
index 0e8e3fd77e96..6dc7a05f4d4b 100644
@@ -603,6 +603,7 @@ struct pci_host_bridge {
 	unsigned int	ignore_reset_delay:1;	/* For entire hierarchy */
 	unsigned int	no_ext_tags:1;		/* No Extended Tags */
 	unsigned int	no_inc_mrrs:1;		/* No Increase MRRS */
+	unsigned int	only_128b_mrrs:1;	/* Only 128B MRRS */
 	unsigned int	native_aer:1;		/* OS may use PCIe AER */
 	unsigned int	native_pcie_hotplug:1;	/* OS may use PCIe hotplug */
 	unsigned int	native_shpc_hotplug:1;	/* OS may use SHPC hotplug */

base-commit: 0af2f6be1b4281385b618cb86ad946eded089ac8

Signature · 1 line
2.39.5
#2 lore raw

Bjorn Helgaas wrote:
[..]

Quoted 11 lines

v1/rfc: https://lore.kernel.org/all/[email protected]/

I suppose it's quixotic to hope for anything better than quirks that have to be added or updated for every new processor that comes along.

ACPI _HPX might be a possible way for the platform to tell us what to do here. ACPI r6.5, sec 6.2.9 says it's for hot-added devices and "Functions not configured by the platform firmware during initial system boot" (how are we supposed to determine that?) In any case, Linux does evaluate _HPX for every device in pci_configure_device().

Quoted 5 lines

I'm not sure _HPX really works; it's very general, and I would expect to see reports of problems if firmware really tried to use it.

Or, I guess a _DSM function would be a possible way to communicate this.

Ok, I am reading this as "Maintainer asserts the quirk is unsavory, please make this be something that Linux can ask the platform firmware if it needs to apply, or make existing _HPX just work. Either of those is preferable to a new entry in the quirk table."

Acked-by on v2 1/1
#3 lore raw
In reply to Ilpo Järvinen
Quoted 52 lines · Ilpo Järvinen

When bifurcated to x2, Xeon 6 Root Port performance is sensitive to the configuration of Extended Tags, Max Read Request Size (MRRS), and 10-Bit Tag Requester (note: there is currently no 10-Bit Tag support in the kernel). While those can be configured to the recommended values by FW, kernel may decide to overwrite the initial values.

Unfortunately, there is no mechanism for FW to indicate OS which parts of PCIe configuration should not be altered. Thus, the only option is to add such logic into the kernel as quirks.

There is a pre-existing quirk flag to disable Extended Tags. Depending on CONFIG_PCIE_BUS_* setting, MRRS may be overwritten by what the kernel thinks is the best for performance (the largest supported value), resulting in performance degradation instead with these Root Ports. (There would have been a pre-existing quirk to disallow increasing MRRS but it is not identical to rejecting >128B MRRS.)

Add a quirk that disallows enabling Extended Tags and setting MRRS larger than 128B for devices under Xeon 6 Root Ports if the Root Port is bifurcated to x2. Reject >128B MRRS only when it is going to be written by the kernel (this assumes FW configured a good initial value for MRRS in case the kernel is not touching MRRS at all).

It was first attempted to always write MRRS when the quirk is needed (always overwrite the initial value). That turned out to be quite invasive change, however, given the complexity of the initial setup callchain and various stages returning early when they decide no changes are necessary, requiring override each. As such, the initial value for MRRS is now left into the hands of FW.

---

v2:
- Explain in changelog why FW cannot solve this on its own
- Moved the quirk under arch/x86/pci/
- Don't NULL check value from pci_find_host_bridge()
- Added comment above the quirk about the performance degradation
- Removed all setup chain 128B quirk overrides expect for MRRS write itself (assumes a sane initial value is set by FW)

 arch/x86/pci/fixup.c | 30 ++++++++++++++++++++++++++++++
 drivers/pci/pci.c    | 15 ++++++++-------
 include/linux/pci.h  |  1 +
 3 files changed, 39 insertions(+), 7 deletions(-)
arch/x86/pci/fixup.c +0 -0
index efefeb82ab61..aa9617bc4b55 100644

arch/x86/pci/fixup.c DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_MCH_PB1, pcie_r · line 300

Quoted 7 lines
arch/x86/pci/fixup.c +4 -0
@@ -294,6 +294,36 @@ DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_INTEL,	PCI_DEVICE_ID_INTEL_MCH_PB1,	pcie_r
 DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_INTEL,	PCI_DEVICE_ID_INTEL_MCH_PC,	pcie_rootport_aspm_quirk);
 DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_INTEL,	PCI_DEVICE_ID_INTEL_MCH_PC1,	pcie_rootport_aspm_quirk);
 
+/*
+ * PCIe devices underneath Xeon6 PCIe Root Port bifurcated to 2x have slower
+ * performance with Extended Tags and MRRS > 128B. Workaround the performance
+ * problems by disabling Extended Tags and limiting MRRS to 128B.

No objections to your fix, just an obligatory:

s/Workaround
/Work around

:)

With that, FWIIW:

Thanks,

Ingo

#4 lore raw
In reply to Ilpo Järvinen
Quoted 4 lines · Ilpo Järvinen

When bifurcated to x2, Xeon 6 Root Port performance is sensitive to the configuration of Extended Tags, Max Read Request Size (MRRS), and 10-Bit Tag Requester (note: there is currently no 10-Bit Tag support in the kernel).

[...]

Quoted 5 lines

Add a quirk that disallows enabling Extended Tags and setting MRRS larger than 128B for devices under Xeon 6 Root Ports if the Root Port is bifurcated to x2. Reject >128B MRRS only when it is going to be written by the kernel (this assumes FW configured a good initial value for MRRS in case the kernel is not touching MRRS at all).

I note that there's the existing quirk_brcm_5719_limit_mrrs(), which limits MRRS to 2048 on certain revisions of Broadcom Ethernet adapters. This became necessary to work around an internal FIFO problem, see commit 2c55a3d08ade ("tg3: Scale back code that modifies MRRS") and commit 0b471506712d ("tg3: Recode PCI MRRS adjustment as a PCI quirk").

The quirk works by overriding the MRRS which was originally set on enumeration by pcie_bus_configure_settings(). The overriding happens at enable time, i.e. when a driver starts to makes use of the device:

do_pci_enable_device()
pci_host_bridge_enable_device()
pcibios_enable_device()
pci_fixup_device()

    quirk_brcm_5719_limit_mrrs()

Now if you look further above in do_pci_enable_device(), there's a call to pci_host_bridge_enable_device(), which invokes the ->enable_device() callback in struct pci_host_bridge. Currently there's only a single host brige driver implementing that callback, controller/dwc/pci-imx6.c.

One option would be to set that callback on the host bridge if a Granite Rapids Root Port is found. And then enforce the mrrs limit in the callback. That approach may be more acceptable upstream than adding a custom "only_128b_mrrs" bit to struct pci_host_bridge.

Another option would be to amend x86's pcibios_enable_device() to check whether there's a Granite Rapids Root Port above the device and enforce the mrrs limit if so.

The only downside I see is that the Broadcom quirk will run afterwards and increase the MRRS again. But it's highly unlikely that one of these old Broadcom chips is used on a present-day Granite Rapids server, so it may not be a problem in practice. And the worst thing that can happen is suboptimal performance.

Thanks,

Lukas

#5 lore raw
In reply to Ilpo Järvinen
Quoted 63 lines · Ilpo Järvinen

When bifurcated to x2, Xeon 6 Root Port performance is sensitive to the configuration of Extended Tags, Max Read Request Size (MRRS), and 10-Bit Tag Requester (note: there is currently no 10-Bit Tag support in the kernel). While those can be configured to the recommended values by FW, kernel may decide to overwrite the initial values.

Unfortunately, there is no mechanism for FW to indicate OS which parts of PCIe configuration should not be altered. Thus, the only option is to add such logic into the kernel as quirks.

There is a pre-existing quirk flag to disable Extended Tags. Depending on CONFIG_PCIE_BUS_* setting, MRRS may be overwritten by what the kernel thinks is the best for performance (the largest supported value), resulting in performance degradation instead with these Root Ports. (There would have been a pre-existing quirk to disallow increasing MRRS but it is not identical to rejecting >128B MRRS.)

Add a quirk that disallows enabling Extended Tags and setting MRRS larger than 128B for devices under Xeon 6 Root Ports if the Root Port is bifurcated to x2. Reject >128B MRRS only when it is going to be written by the kernel (this assumes FW configured a good initial value for MRRS in case the kernel is not touching MRRS at all).

It was first attempted to always write MRRS when the quirk is needed (always overwrite the initial value). That turned out to be quite invasive change, however, given the complexity of the initial setup callchain and various stages returning early when they decide no changes are necessary, requiring override each. As such, the initial value for MRRS is now left into the hands of FW.

---

v2:
- Explain in changelog why FW cannot solve this on its own
- Moved the quirk under arch/x86/pci/
- Don't NULL check value from pci_find_host_bridge()
- Added comment above the quirk about the performance degradation
- Removed all setup chain 128B quirk overrides expect for MRRS write itself (assumes a sane initial value is set by FW)

 arch/x86/pci/fixup.c | 30 ++++++++++++++++++++++++++++++
 drivers/pci/pci.c    | 15 ++++++++-------
 include/linux/pci.h  |  1 +
 3 files changed, 39 insertions(+), 7 deletions(-)
arch/x86/pci/fixup.c +8 -0
index efefeb82ab61..aa9617bc4b55 100644
@@ -294,6 +294,36 @@ DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_INTEL,	PCI_DEVICE_ID_INTEL_MCH_PB1,	pcie_r
 DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_INTEL,	PCI_DEVICE_ID_INTEL_MCH_PC,	pcie_rootport_aspm_quirk);
 DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_INTEL,	PCI_DEVICE_ID_INTEL_MCH_PC1,	pcie_rootport_aspm_quirk);
 
+/*
+ * PCIe devices underneath Xeon6 PCIe Root Port bifurcated to 2x have slower
+ * performance with Extended Tags and MRRS > 128B. Workaround the performance
+ * problems by disabling Extended Tags and limiting MRRS to 128B.
+ *
+ * https://cdrdv2.intel.com/v1/dl/getContent/837176
+ */
+static void quirk_pcie2x_no_tags_no_mrrs(struct pci_dev *pdev)

arch/x86/pci/fixup.c DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_MCH_PB1, pcie_r · line 314

Quoted 10 lines
arch/x86/pci/fixup.c +10 -0
+{
+	struct pci_host_bridge *bridge = pci_find_host_bridge(pdev->bus);
+	u32 linkcap;
+
+	pcie_capability_read_dword(pdev, PCI_EXP_LNKCAP, &linkcap);
+	if (FIELD_GET(PCI_EXP_LNKCAP_MLW, linkcap) != 0x2)
+		return;
+
+	bridge->no_ext_tags = 1;
+	bridge->only_128b_mrrs = 1;

My 2 cents here. Wouldn't it work if you hardcode MRRS to 128 in PCI_EXP_DEVCTL here and then set pci_host_bridge::no_inc_mrrs to 1? This would avoid introducing an extra flag and also serve the same purpose.

- Mani

Signature · 1 line
மணிவண்ணன் சதாசிவம்
#6 lore raw
Quoted 70 lines · Manivannan Sadhasivam
Quoted 69 lines · Ilpo Järvinen

When bifurcated to x2, Xeon 6 Root Port performance is sensitive to the configuration of Extended Tags, Max Read Request Size (MRRS), and 10-Bit Tag Requester (note: there is currently no 10-Bit Tag support in the kernel). While those can be configured to the recommended values by FW, kernel may decide to overwrite the initial values.

Unfortunately, there is no mechanism for FW to indicate OS which parts of PCIe configuration should not be altered. Thus, the only option is to add such logic into the kernel as quirks.

There is a pre-existing quirk flag to disable Extended Tags. Depending on CONFIG_PCIE_BUS_* setting, MRRS may be overwritten by what the kernel thinks is the best for performance (the largest supported value), resulting in performance degradation instead with these Root Ports. (There would have been a pre-existing quirk to disallow increasing MRRS but it is not identical to rejecting >128B MRRS.)

Add a quirk that disallows enabling Extended Tags and setting MRRS larger than 128B for devices under Xeon 6 Root Ports if the Root Port is bifurcated to x2. Reject >128B MRRS only when it is going to be written by the kernel (this assumes FW configured a good initial value for MRRS in case the kernel is not touching MRRS at all).

It was first attempted to always write MRRS when the quirk is needed (always overwrite the initial value). That turned out to be quite invasive change, however, given the complexity of the initial setup callchain and various stages returning early when they decide no changes are necessary, requiring override each. As such, the initial value for MRRS is now left into the hands of FW.

---

v2:
- Explain in changelog why FW cannot solve this on its own
- Moved the quirk under arch/x86/pci/
- Don't NULL check value from pci_find_host_bridge()
- Added comment above the quirk about the performance degradation
- Removed all setup chain 128B quirk overrides expect for MRRS write itself (assumes a sane initial value is set by FW)

 arch/x86/pci/fixup.c | 30 ++++++++++++++++++++++++++++++
 drivers/pci/pci.c    | 15 ++++++++-------
 include/linux/pci.h  |  1 +
 3 files changed, 39 insertions(+), 7 deletions(-)
arch/x86/pci/fixup.c +14 -0
index efefeb82ab61..aa9617bc4b55 100644
@@ -294,6 +294,36 @@ DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_INTEL,	PCI_DEVICE_ID_INTEL_MCH_PB1,	pcie_r
 DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_INTEL,	PCI_DEVICE_ID_INTEL_MCH_PC,	pcie_rootport_aspm_quirk);
 DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_INTEL,	PCI_DEVICE_ID_INTEL_MCH_PC1,	pcie_rootport_aspm_quirk);
 
+/*
+ * PCIe devices underneath Xeon6 PCIe Root Port bifurcated to 2x have slower
+ * performance with Extended Tags and MRRS > 128B. Workaround the performance
+ * problems by disabling Extended Tags and limiting MRRS to 128B.
+ *
+ * https://cdrdv2.intel.com/v1/dl/getContent/837176
+ */
+static void quirk_pcie2x_no_tags_no_mrrs(struct pci_dev *pdev)
+{
+	struct pci_host_bridge *bridge = pci_find_host_bridge(pdev->bus);
+	u32 linkcap;
+
+	pcie_capability_read_dword(pdev, PCI_EXP_LNKCAP, &linkcap);
+	if (FIELD_GET(PCI_EXP_LNKCAP_MLW, linkcap) != 0x2)
Quoted 8 lines

arch/x86/pci/fixup.c DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_MCH_PB1, pcie_r · line 314

Quoted 4 lines
arch/x86/pci/fixup.c +4 -0
+		return;
+
+	bridge->no_ext_tags = 1;
+	bridge->only_128b_mrrs = 1;

My 2 cents here. Wouldn't it work if you hardcode MRRS to 128 in PCI_EXP_DEVCTL here and then set pci_host_bridge::no_inc_mrrs to 1? This would avoid introducing an extra flag and also serve the same purpose.

Hi Mani,

Thanks for the suggestion but it won't work because this is the Root Port. The devices underneath it need this setting so we cannot set them to 128B reliable here (is there anything that guarantees those devices have been enumerated at this point?).

I've v3 already prepared which uses the enable device hook as suggested by Lukas. I'll send it soon.

Signature · 1 line
 i.
#7 lore raw
Quoted 81 lines · Ilpo Järvinen
Quoted 70 lines · Manivannan Sadhasivam
Quoted 69 lines · Ilpo Järvinen

When bifurcated to x2, Xeon 6 Root Port performance is sensitive to the configuration of Extended Tags, Max Read Request Size (MRRS), and 10-Bit Tag Requester (note: there is currently no 10-Bit Tag support in the kernel). While those can be configured to the recommended values by FW, kernel may decide to overwrite the initial values.

Unfortunately, there is no mechanism for FW to indicate OS which parts of PCIe configuration should not be altered. Thus, the only option is to add such logic into the kernel as quirks.

There is a pre-existing quirk flag to disable Extended Tags. Depending on CONFIG_PCIE_BUS_* setting, MRRS may be overwritten by what the kernel thinks is the best for performance (the largest supported value), resulting in performance degradation instead with these Root Ports. (There would have been a pre-existing quirk to disallow increasing MRRS but it is not identical to rejecting >128B MRRS.)

Add a quirk that disallows enabling Extended Tags and setting MRRS larger than 128B for devices under Xeon 6 Root Ports if the Root Port is bifurcated to x2. Reject >128B MRRS only when it is going to be written by the kernel (this assumes FW configured a good initial value for MRRS in case the kernel is not touching MRRS at all).

It was first attempted to always write MRRS when the quirk is needed (always overwrite the initial value). That turned out to be quite invasive change, however, given the complexity of the initial setup callchain and various stages returning early when they decide no changes are necessary, requiring override each. As such, the initial value for MRRS is now left into the hands of FW.

---

v2:
- Explain in changelog why FW cannot solve this on its own
- Moved the quirk under arch/x86/pci/
- Don't NULL check value from pci_find_host_bridge()
- Added comment above the quirk about the performance degradation
- Removed all setup chain 128B quirk overrides expect for MRRS write itself (assumes a sane initial value is set by FW)

 arch/x86/pci/fixup.c | 30 ++++++++++++++++++++++++++++++
 drivers/pci/pci.c    | 15 ++++++++-------
 include/linux/pci.h  |  1 +
 3 files changed, 39 insertions(+), 7 deletions(-)
arch/x86/pci/fixup.c +14 -0
index efefeb82ab61..aa9617bc4b55 100644
@@ -294,6 +294,36 @@ DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_INTEL,	PCI_DEVICE_ID_INTEL_MCH_PB1,	pcie_r
 DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_INTEL,	PCI_DEVICE_ID_INTEL_MCH_PC,	pcie_rootport_aspm_quirk);
 DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_INTEL,	PCI_DEVICE_ID_INTEL_MCH_PC1,	pcie_rootport_aspm_quirk);
 
+/*
+ * PCIe devices underneath Xeon6 PCIe Root Port bifurcated to 2x have slower
+ * performance with Extended Tags and MRRS > 128B. Workaround the performance
+ * problems by disabling Extended Tags and limiting MRRS to 128B.
+ *
+ * https://cdrdv2.intel.com/v1/dl/getContent/837176
+ */
+static void quirk_pcie2x_no_tags_no_mrrs(struct pci_dev *pdev)
+{
+	struct pci_host_bridge *bridge = pci_find_host_bridge(pdev->bus);
+	u32 linkcap;
+
+	pcie_capability_read_dword(pdev, PCI_EXP_LNKCAP, &linkcap);
+	if (FIELD_GET(PCI_EXP_LNKCAP_MLW, linkcap) != 0x2)
Quoted 8 lines

arch/x86/pci/fixup.c DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_MCH_PB1, pcie_r · line 314

Quoted 4 lines
arch/x86/pci/fixup.c +4 -0
+		return;
+
+	bridge->no_ext_tags = 1;
+	bridge->only_128b_mrrs = 1;

My 2 cents here. Wouldn't it work if you hardcode MRRS to 128 in PCI_EXP_DEVCTL here and then set pci_host_bridge::no_inc_mrrs to 1? This would avoid introducing an extra flag and also serve the same purpose.

Quoted 7 lines

Hi Mani,

Thanks for the suggestion but it won't work because this is the Root Port. The devices underneath it need this setting so we cannot set them to 128B reliable here (is there anything that guarantees those devices have been enumerated at this point?).

I was under assumption that the kernel would honor the Root Port MRRS value while setting the device MRRS, but I was wrong. Kernel just takes the MPS value (but that considers the Root Port's MPS) for MRRS and scales it down if the hardware doesn't support it.

So yes, setting the Root Port's MRRS would have no effect in limiting the device MRRS. And I believe, Xeon6 Root Ports doesn't suffer from same 128B MRRS limitation for MPS, otherwise, you could just set Root Port MPS to 128B and it will make sure that both MPS and MRRS of endpoint devices will be capped to this value.

Quoted 3 lines

I've v3 already prepared which uses the enable device hook as suggested by Lukas. I'll send it soon.

The callback is supposed to be used for performing enablement steps required by the *Host Bridge* devices, not Root Ports. So I do have a concern to use it to address the Root Port quirks. That's why I thought of limiting it to the PCI quirks.

But feel free to submit the patch. Let's see what Bjorn and others feel about it.

- Mani

Signature · 1 line
மணிவண்ணன் சதாசிவம்