Skip to contentSkip to conversation

PCI/ASPM: Add ASPM BIOS override function

PATCH In review David E. Box 4 participants 12 messages Dec 16, 2021 – Dec 22, 2021

Subsystems (3)
lore
Share & tools
lore
#1 lore raw

PATCH v4 2/2 PCI: vmd: Override ASPM on TGL/ADL VMD devices

In reply to David E. Box

From: Michael Bottini <[email protected]>

On Tiger Lake and Alder Lake platforms, VMD controllers do not have ASPM enabled nor LTR values set by BIOS. This leads high power consumption on these platforms when VMD is enabled as reported in bugzilla [1]. Enable these features in the VMD driver using pcie_aspm_policy_override() to set the ASPM policy for the root ports.

To do this, add an additional flag in VMD features to specify devices that must have their respective policies overridden.

[1] https://bugzilla.kernel.org/show_bug.cgi?id=213717

---
V4
- Refactor vmd_enable_apsm() to exit early, making the lines shorter and more readable. Suggested by Christoph.
V3
- No changes
V2
- Use return status to print pci_info message if ASPM cannot be enabled.
- Add missing static declaration, caught by [email protected]

 drivers/pci/controller/vmd.c | 43 +++++++++++++++++++++++++++++++++---
 1 file changed, 40 insertions(+), 3 deletions(-)
drivers/pci/controller/vmd.c +40 -3
index a45e8e59d3d4..880afd450a14 100644
@@ -20,6 +20,8 @@
 
 #include <asm/irqdomain.h>
 
+#include "../pci.h"
+
 #define VMD_CFGBAR	0
 #define VMD_MEMBAR1	2
 #define VMD_MEMBAR2	4
@@ -67,6 +69,12 @@ enum vmd_features {
 	 * interrupt handling.
 	 */
 	VMD_FEAT_CAN_BYPASS_MSI_REMAP		= (1 << 4),
+
+	/*
+	 * Device must have ASPM policy overridden, as its default policy is
+	 * incorrect.
+	 */
+	VMD_FEAT_QUIRK_OVERRIDE_ASPM		= (1 << 5),
 };
 
 static DEFINE_IDA(vmd_instance_ida);
@@ -661,6 +669,30 @@ static int vmd_alloc_irqs(struct vmd_dev *vmd)
 	return 0;
 }
 
+/*
+ * Override the BIOS ASPM policy and set the LTR value for PCI storage
+ * devices on the VMD bride.
+ */
+static int vmd_enable_aspm(struct pci_dev *pdev, void *userdata)
+{
+	int features = *(int *)userdata, pos;
+
+	if (!(features & VMD_FEAT_QUIRK_OVERRIDE_ASPM) ||
+	    pdev->class != PCI_CLASS_STORAGE_EXPRESS)
+		return 0;
+
+	pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_LTR);
+	if (!pos)
+		return 0;
+
+	pci_write_config_word(pdev, pos + PCI_LTR_MAX_SNOOP_LAT, 0x1003);
+	pci_write_config_word(pdev, pos + PCI_LTR_MAX_NOSNOOP_LAT, 0x1003);
+	if (pcie_aspm_policy_override(pdev))
+		pci_info(pdev, "Unable of override ASPM policy\n");
+
+	return 0;
+}
+
 static int vmd_enable_domain(struct vmd_dev *vmd, unsigned long features)
 {
 	struct pci_sysdata *sd = &vmd->sysdata;
@@ -807,6 +839,8 @@ static int vmd_enable_domain(struct vmd_dev *vmd, unsigned long features)
 	pci_scan_child_bus(vmd->bus);
 	pci_assign_unassigned_bus_resources(vmd->bus);
 
+	pci_walk_bus(vmd->bus, vmd_enable_aspm, &features);
+
 	/*
 	 * VMD root buses are virtual and don't return true on pci_is_pcie()
 	 * and will fail pcie_bus_configure_settings() early. It can instead be
@@ -948,15 +982,18 @@ static const struct pci_device_id vmd_ids[] = {
 	{PCI_DEVICE(PCI_VENDOR_ID_INTEL, 0x467f),
 		.driver_data = VMD_FEAT_HAS_MEMBAR_SHADOW_VSCAP |
 				VMD_FEAT_HAS_BUS_RESTRICTIONS |
-				VMD_FEAT_OFFSET_FIRST_VECTOR,},
+				VMD_FEAT_OFFSET_FIRST_VECTOR |
+				VMD_FEAT_QUIRK_OVERRIDE_ASPM,},
 	{PCI_DEVICE(PCI_VENDOR_ID_INTEL, 0x4c3d),
 		.driver_data = VMD_FEAT_HAS_MEMBAR_SHADOW_VSCAP |
 				VMD_FEAT_HAS_BUS_RESTRICTIONS |
-				VMD_FEAT_OFFSET_FIRST_VECTOR,},
+				VMD_FEAT_OFFSET_FIRST_VECTOR |
+				VMD_FEAT_QUIRK_OVERRIDE_ASPM,},
 	{PCI_DEVICE(PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_VMD_9A0B),
 		.driver_data = VMD_FEAT_HAS_MEMBAR_SHADOW_VSCAP |
 				VMD_FEAT_HAS_BUS_RESTRICTIONS |
-				VMD_FEAT_OFFSET_FIRST_VECTOR,},
+				VMD_FEAT_OFFSET_FIRST_VECTOR |
+				VMD_FEAT_QUIRK_OVERRIDE_ASPM,},
 	{0,}
 };
 MODULE_DEVICE_TABLE(pci, vmd_ids);
Signature · 1 line
2.25.1
#0 lore raw

PATCH v4 1/2 PCI/ASPM: Add ASPM BIOS override function

From: Michael Bottini <[email protected]>

Devices that appear under the Intel VMD host bridge are not visible to BIOS and therefore not programmed by BIOS with ASPM settings. For these devices, it is necessary for the driver to configure ASPM. Since ASPM settings are adjustable at runtime by module parameter, use the same mechanism to allow drivers to override the default (in this case never configured) BIOS policy to ASPM_STATE_ALL. Then, reconfigure ASPM on the link. Do not override if ASPM control is disabled.

---
V4
- No changes.
V3
- Fix missing semicolon in static inline function.

V2
- Change return type to int so caller can determine if override was successful.
- Return immediately if link is not found so that lock it not unecessarily taken, suggested by [email protected].
- Don't override if aspm_disabled is true.

 drivers/pci/pci.h       |  2 ++
 drivers/pci/pcie/aspm.c | 19 +++++++++++++++++++
 2 files changed, 21 insertions(+)
drivers/pci/pci.h +2 -0
index 3d60cabde1a1..c9c55d43cd8a 100644
@@ -562,11 +562,13 @@ void pcie_aspm_init_link_state(struct pci_dev *pdev);
 void pcie_aspm_exit_link_state(struct pci_dev *pdev);
 void pcie_aspm_pm_state_change(struct pci_dev *pdev);
 void pcie_aspm_powersave_config_link(struct pci_dev *pdev);
+int pcie_aspm_policy_override(struct pci_dev *dev);
 #else
 static inline void pcie_aspm_init_link_state(struct pci_dev *pdev) { }
 static inline void pcie_aspm_exit_link_state(struct pci_dev *pdev) { }
 static inline void pcie_aspm_pm_state_change(struct pci_dev *pdev) { }
 static inline void pcie_aspm_powersave_config_link(struct pci_dev *pdev) { }
+static inline int pcie_aspm_policy_override(struct pci_dev *dev) { return -EINVAL; }
 #endif
 
 #ifdef CONFIG_PCIE_ECRC
drivers/pci/pcie/aspm.c +19 -0
index 52c74682601a..e2c61e14e724 100644
@@ -1140,6 +1140,25 @@ int pci_disable_link_state(struct pci_dev *pdev, int state)
 }
 EXPORT_SYMBOL(pci_disable_link_state);
 
+int pcie_aspm_policy_override(struct pci_dev *pdev)
+{
+	struct pcie_link_state *link = pcie_aspm_get_link(pdev);
+
+	if (!link || aspm_disabled)
+		return -EINVAL;
+
+	down_read(&pci_bus_sem);
+	mutex_lock(&aspm_lock);
+	link->aspm_default = ASPM_STATE_ALL;
+	pcie_config_aspm_link(link, policy_to_aspm_state(link));
+	pcie_set_clkpm(link, policy_to_clkpm_state(link));
+	mutex_unlock(&aspm_lock);
+	up_read(&pci_bus_sem);
+
+	return 0;
+}
+EXPORT_SYMBOL(pcie_aspm_policy_override);
+
 static int pcie_aspm_set_policy(const char *val,
 				const struct kernel_param *kp)
 {
Signature · 1 line
2.25.1
#10 lore raw
Quoted 5 lines · David E. Box

From: Michael Bottini <[email protected]>

Devices that appear under the Intel VMD host bridge are not visible to BIOS and therefore not programmed by BIOS with ASPM settings. For these devices, it is necessary for the driver to configure ASPM.

The VMD-related parts of this commit log belong in the next patch, because this patch has nothing in particular to do with VMD.

Quoted 5 lines

Since ASPM settings are adjustable at runtime by module parameter, use the same mechanism to allow drivers to override the default (in this case never configured) BIOS policy to ASPM_STATE_ALL. Then, reconfigure ASPM on the link. Do not override if ASPM control is disabled.

The module parameter ("policy") has global effect: it runs pcie_aspm_set_policy(), which assigns the global "aspm_policy" and then reconfigures all links in the system.

This is not that; it's a link-based thing that doesn't change "aspm_policy" and only affects a single link. This is more like aspm_attr_store_common() for the sysfs "l0s_aspm" and similar attributes, or the pci_disable_link_state() interface for drivers.

Quoted 49 lines

---
V4
- No changes.
V3
- Fix missing semicolon in static inline function.

V2
- Change return type to int so caller can determine if override was successful.
- Return immediately if link is not found so that lock it not unecessarily taken, suggested by [email protected].
- Don't override if aspm_disabled is true.

 drivers/pci/pci.h       |  2 ++
 drivers/pci/pcie/aspm.c | 19 +++++++++++++++++++
 2 files changed, 21 insertions(+)
drivers/pci/pci.h +2 -0
index 3d60cabde1a1..c9c55d43cd8a 100644
@@ -562,11 +562,13 @@ void pcie_aspm_init_link_state(struct pci_dev *pdev);
 void pcie_aspm_exit_link_state(struct pci_dev *pdev);
 void pcie_aspm_pm_state_change(struct pci_dev *pdev);
 void pcie_aspm_powersave_config_link(struct pci_dev *pdev);
+int pcie_aspm_policy_override(struct pci_dev *dev);
 #else
 static inline void pcie_aspm_init_link_state(struct pci_dev *pdev) { }
 static inline void pcie_aspm_exit_link_state(struct pci_dev *pdev) { }
 static inline void pcie_aspm_pm_state_change(struct pci_dev *pdev) { }
 static inline void pcie_aspm_powersave_config_link(struct pci_dev *pdev) { }
+static inline int pcie_aspm_policy_override(struct pci_dev *dev) { return -EINVAL; }
 #endif
 
 #ifdef CONFIG_PCIE_ECRC
drivers/pci/pcie/aspm.c +4 -0
index 52c74682601a..e2c61e14e724 100644
@@ -1140,6 +1140,25 @@ int pci_disable_link_state(struct pci_dev *pdev, int state)
 }
 EXPORT_SYMBOL(pci_disable_link_state);
 
+int pcie_aspm_policy_override(struct pci_dev *pdev)
+{
+	struct pcie_link_state *link = pcie_aspm_get_link(pdev);
+

drivers/pci/pcie/aspm.c int pci_disable_link_state(struct pci_dev *pdev, int state) · line 1156

Quoted 10 lines
drivers/pci/pcie/aspm.c +10 -0
+	if (!link || aspm_disabled)
+		return -EINVAL;
+
+	down_read(&pci_bus_sem);
+	mutex_lock(&aspm_lock);
+	link->aspm_default = ASPM_STATE_ALL;
+	pcie_config_aspm_link(link, policy_to_aspm_state(link));
+	pcie_set_clkpm(link, policy_to_clkpm_state(link));
+	mutex_unlock(&aspm_lock);
+	up_read(&pci_bus_sem);

This is essentially the inverse of pci_disable_link_state(). Why not name it so the connection is obvious? Probably also make the signature ("int state") similar.

Quoted 10 lines
drivers/pci/pcie/aspm.c +4 -0
+	return 0;
+}
+EXPORT_SYMBOL(pcie_aspm_policy_override);
+
 static int pcie_aspm_set_policy(const char *val,
 				const struct kernel_param *kp)
 {
Signature · 1 line
2.25.1
#2 lore raw

Re: [PATCH V4 2/2] PCI: vmd: Override ASPM on TGL/ADL VMD devices

In reply to David E. Box

[+cc Rajat for LTR max latency write]

Quoted 7 lines · David E. Box

From: Michael Bottini <[email protected]>

On Tiger Lake and Alder Lake platforms, VMD controllers do not have ASPM enabled nor LTR values set by BIOS. This leads high power consumption on these platforms when VMD is enabled as reported in bugzilla [1]. Enable these features in the VMD driver using pcie_aspm_policy_override() to set the ASPM policy for the root ports.

s/leads high/leads to high/

Does this depend on "Tiger Lake and Alder Lake platforms" specifically, or does it depend on a BIOS design choice, i.e., "don't configure ASPM or LTR for devices below a VMD"?

The subject says "override ASPM on VMD devices," but it looks like this affects the ASPM configuration of devices *below* the VMD, not of the VMD itself.

It looks like this only affects *NVMe* devices, since vmd_enable_aspm() checks for PCI_CLASS_STORAGE_EXPRESS. Why is that? Is there something special about NVMe? I'd think you would want to do this for *all* devices below a VMD.

Since it only affects PCI_CLASS_STORAGE_EXPRESS devices, I don't think it actually "sets ASPM policy for the root ports". vmd_enable_aspm() calls pcie_aspm_policy_override() on endpoints. It's true that the link ASPM state happens to be attached to the upstream end of the link, but that's an ASPM implementation detail.

This all needs to be clear in the subject and commit log.

Quoted 2 lines

To do this, add an additional flag in VMD features to specify devices that must have their respective policies overridden.

I'm not clear on why you want this to apply to only certain VMDs and not others. Do some BIOSes configure ASPM for devices below some VMDs?

Quoted 32 lines

[1] https://bugzilla.kernel.org/show_bug.cgi?id=213717

---
V4
- Refactor vmd_enable_apsm() to exit early, making the lines shorter and more readable. Suggested by Christoph.
V3
- No changes
V2
- Use return status to print pci_info message if ASPM cannot be enabled.
- Add missing static declaration, caught by [email protected]

 drivers/pci/controller/vmd.c | 43 +++++++++++++++++++++++++++++++++---
 1 file changed, 40 insertions(+), 3 deletions(-)
drivers/pci/controller/vmd.c +2 -0
index a45e8e59d3d4..880afd450a14 100644
@@ -20,6 +20,8 @@
 
 #include <asm/irqdomain.h>
 
+#include "../pci.h"
+
 #define VMD_CFGBAR	0
 #define VMD_MEMBAR1	2
 #define VMD_MEMBAR2	4

drivers/pci/controller/vmd.c enum vmd_features { · line 77

Quoted 9 lines
drivers/pci/controller/vmd.c +6 -0
@@ -67,6 +69,12 @@ enum vmd_features {
 	 * interrupt handling.
 	 */
 	VMD_FEAT_CAN_BYPASS_MSI_REMAP		= (1 << 4),
+
+	/*
+	 * Device must have ASPM policy overridden, as its default policy is
+	 * incorrect.
+	 */
+	VMD_FEAT_QUIRK_OVERRIDE_ASPM		= (1 << 5),

I think you specifically want to *enable* some ASPM link states, not just "override the default policy." "Override" tells us nothing about whether you are enabling or disabling ASPM. Applies to subject line as well.

drivers/pci/controller/vmd.c static int vmd_alloc_irqs(struct vmd_dev *vmd) · line 674

Quoted 10 lines
drivers/pci/controller/vmd.c +3 -0
 };
 
 static DEFINE_IDA(vmd_instance_ida);
@@ -661,6 +669,30 @@ static int vmd_alloc_irqs(struct vmd_dev *vmd)
 	return 0;
 }
 
+/*
+ * Override the BIOS ASPM policy and set the LTR value for PCI storage
+ * devices on the VMD bride.

I don't think there's any BIOS "policy" here. At this point BIOS is no longer involved at all, so all that's left is whatever ASPM config the BIOS did or did not do.

Why only storage?

s/bride/bridge/

Quoted 5 lines
drivers/pci/controller/vmd.c +5 -0
+ */
+static int vmd_enable_aspm(struct pci_dev *pdev, void *userdata)
+{
+	int features = *(int *)userdata, pos;
+

drivers/pci/controller/vmd.c static int vmd_alloc_irqs(struct vmd_dev *vmd) · line 689

Quoted 10 lines
drivers/pci/controller/vmd.c +10 -0
+	if (!(features & VMD_FEAT_QUIRK_OVERRIDE_ASPM) ||
+	    pdev->class != PCI_CLASS_STORAGE_EXPRESS)
+		return 0;
+
+	pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_LTR);
+	if (!pos)
+		return 0;
+
+	pci_write_config_word(pdev, pos + PCI_LTR_MAX_SNOOP_LAT, 0x1003);
+	pci_write_config_word(pdev, pos + PCI_LTR_MAX_NOSNOOP_LAT, 0x1003);

1) Where did this magic 0x1003 value come from? Does that depend on the VMD device? The endpoint? The circuit design? The path between endpoint and VMD? What if there are switches in the path?

2) There exist broken devices where WORD config accesses don't work:
https://lore.kernel.org/all/[email protected]/

We might need a way to quirk config accesses to those devices, but we don't have one yet. So for now this needs to be a single DWORD write.

drivers/pci/controller/vmd.c static int vmd_alloc_irqs(struct vmd_dev *vmd) · line 691

Quoted 2 lines
drivers/pci/controller/vmd.c +2 -0
+	if (pcie_aspm_policy_override(pdev))
+		pci_info(pdev, "Unable of override ASPM policy\n");

s/Unable of/Unable to/

I think we might need a message about when we *do* override the policy. A note in dmesg might be useful for debugging. I'm worried about the LTR programming because I really don't understand how we should be doing that.

drivers/pci/controller/vmd.c static int vmd_enable_domain(struct vmd_dev *vmd, unsigned long features) · line 842

Quoted 11 lines
drivers/pci/controller/vmd.c +4 -0
+	return 0;
+}
+
 static int vmd_enable_domain(struct vmd_dev *vmd, unsigned long features)
 {
 	struct pci_sysdata *sd = &vmd->sysdata;
@@ -807,6 +839,8 @@ static int vmd_enable_domain(struct vmd_dev *vmd, unsigned long features)
 	pci_scan_child_bus(vmd->bus);
 	pci_assign_unassigned_bus_resources(vmd->bus);
 
+	pci_walk_bus(vmd->bus, vmd_enable_aspm, &features);

Do you support hotplug under VMD? This will not happen for hot-added devices.

Quoted 28 lines
drivers/pci/controller/vmd.c +6 -3
 	/*
 	 * VMD root buses are virtual and don't return true on pci_is_pcie()
 	 * and will fail pcie_bus_configure_settings() early. It can instead be
@@ -948,15 +982,18 @@ static const struct pci_device_id vmd_ids[] = {
 	{PCI_DEVICE(PCI_VENDOR_ID_INTEL, 0x467f),
 		.driver_data = VMD_FEAT_HAS_MEMBAR_SHADOW_VSCAP |
 				VMD_FEAT_HAS_BUS_RESTRICTIONS |
-				VMD_FEAT_OFFSET_FIRST_VECTOR,},
+				VMD_FEAT_OFFSET_FIRST_VECTOR |
+				VMD_FEAT_QUIRK_OVERRIDE_ASPM,},
 	{PCI_DEVICE(PCI_VENDOR_ID_INTEL, 0x4c3d),
 		.driver_data = VMD_FEAT_HAS_MEMBAR_SHADOW_VSCAP |
 				VMD_FEAT_HAS_BUS_RESTRICTIONS |
-				VMD_FEAT_OFFSET_FIRST_VECTOR,},
+				VMD_FEAT_OFFSET_FIRST_VECTOR |
+				VMD_FEAT_QUIRK_OVERRIDE_ASPM,},
 	{PCI_DEVICE(PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_VMD_9A0B),
 		.driver_data = VMD_FEAT_HAS_MEMBAR_SHADOW_VSCAP |
 				VMD_FEAT_HAS_BUS_RESTRICTIONS |
-				VMD_FEAT_OFFSET_FIRST_VECTOR,},
+				VMD_FEAT_OFFSET_FIRST_VECTOR |
+				VMD_FEAT_QUIRK_OVERRIDE_ASPM,},
 	{0,}
 };
 MODULE_DEVICE_TABLE(pci, vmd_ids);
Signature · 1 line
2.25.1
#11 lore raw
In reply to Bjorn Helgaas
Quoted 9 lines · Bjorn Helgaas
Quoted 5 lines · David E. Box

From: Michael Bottini <[email protected]>

Devices that appear under the Intel VMD host bridge are not visible to BIOS and therefore not programmed by BIOS with ASPM settings. For these devices, it is necessary for the driver to configure ASPM.

The VMD-related parts of this commit log belong in the next patch, because this patch has nothing in particular to do with VMD.

Okay, but shouldn't something be said here as to why it's needed? Maybe a more general description?

Quoted 7 lines
Quoted 6 lines

Since ASPM settings are adjustable at runtime by module parameter, use the same mechanism to allow drivers to override the default (in this case never configured) BIOS policy to ASPM_STATE_ALL. Then, reconfigure ASPM on the link. Do not override if ASPM control is disabled.

Quoted 8 lines

The module parameter ("policy") has global effect: it runs pcie_aspm_set_policy(), which assigns the global "aspm_policy" and then reconfigures all links in the system.

This is not that; it's a link-based thing that doesn't change "aspm_policy" and only affects a single link.  This is more like aspm_attr_store_common() for the sysfs "l0s_aspm" and similar attributes, or the pci_disable_link_state() interface for drivers.

Yes. I'll clean up the wording.

Quoted 56 lines
Quoted 56 lines

---
V4
 - No changes.
V3
 - Fix missing semicolon in static inline function.

V2
 - Change return type to int so caller can determine if override was successful.
 - Return immediately if link is not found so that lock it not unecessarily taken, suggested by [email protected].
 - Don't override if aspm_disabled is true.

 drivers/pci/pci.h       |  2 ++
 drivers/pci/pcie/aspm.c | 19 +++++++++++++++++++

 2 files changed, 21 insertions(+)
drivers/pci/pci.h +2 -0
index 3d60cabde1a1..c9c55d43cd8a 100644
@@ -562,11 +562,13 @@ void pcie_aspm_init_link_state(struct pci_dev *pdev);
  void pcie_aspm_exit_link_state(struct pci_dev *pdev);
  void pcie_aspm_pm_state_change(struct pci_dev *pdev);
  void pcie_aspm_powersave_config_link(struct pci_dev *pdev);
+int pcie_aspm_policy_override(struct pci_dev *dev);
  #else
  static inline void pcie_aspm_init_link_state(struct pci_dev *pdev) { }
  static inline void pcie_aspm_exit_link_state(struct pci_dev *pdev) { }
  static inline void pcie_aspm_pm_state_change(struct pci_dev *pdev) { }
  static inline void pcie_aspm_powersave_config_link(struct pci_dev *pdev) { }
+static inline int pcie_aspm_policy_override(struct pci_dev *dev) { return -EINVAL; }
  #endif


 #ifdef CONFIG_PCIE_ECRC

drivers/pci/pcie/aspm.c +0 -0
index 52c74682601a..e2c61e14e724 100644
@@ -1140,6 +1140,25 @@ int pci_disable_link_state(struct pci_dev *pdev, int state)
  }
  EXPORT_SYMBOL(pci_disable_link_state);

drivers/pci/pcie/aspm.c +10 -0
+int pcie_aspm_policy_override(struct pci_dev *pdev)
+{
+       struct pcie_link_state *link = pcie_aspm_get_link(pdev);
+
+       if (!link || aspm_disabled)
+               return -EINVAL;
+
+       down_read(&pci_bus_sem);
+       mutex_lock(&aspm_lock);
+       link->aspm_default = ASPM_STATE_ALL;
Quoted 8 lines
Quoted 4 lines

+       pcie_config_aspm_link(link, policy_to_aspm_state(link)); +       pcie_set_clkpm(link, policy_to_clkpm_state(link)); +       mutex_unlock(&aspm_lock);
+       up_read(&pci_bus_sem);

This is essentially the inverse of pci_disable_link_state().  Why not name it so the connection is obvious?  Probably also make the signature ("int state") similar.

Indeed it is the inverse. Will implement as pci_enable_link_state().

Thanks.

Quoted 11 lines
Quoted 11 lines

+       return 0;

drivers/pci/pcie/aspm.c +3 -0
+}
+EXPORT_SYMBOL(pcie_aspm_policy_override);
+

 static int pcie_aspm_set_policy(const char *val, const struct kernel_param *kp) {

Signature · 1 line
2.25.1
#3 lore raw

Re: [PATCH V4 2/2] PCI: vmd: Override ASPM on TGL/ADL VMD devices

In reply to Bjorn Helgaas
Quoted 11 lines · Bjorn Helgaas

[+cc Rajat for LTR max latency write]

Quoted 7 lines · David E. Box

From: Michael Bottini <[email protected]>

On Tiger Lake and Alder Lake platforms, VMD controllers do not have ASPM enabled nor LTR values set by BIOS. This leads high power consumption on these platforms when VMD is enabled as reported in bugzilla [1].  Enable these features in the VMD driver using pcie_aspm_policy_override() to set the ASPM policy for the root ports.

Quoted 5 lines

s/leads high/leads to high/

Does this depend on "Tiger Lake and Alder Lake platforms" specifically, or does it depend on a BIOS design choice, i.e., "don't configure ASPM or LTR for devices below a VMD"?

It was a BIOS design choice to have this done by the OS driver for both Tiger Lake and Alder Lake.

Quoted 4 lines

The subject says "override ASPM on VMD devices," but it looks like this affects the ASPM configuration of devices *below* the VMD, not of the VMD itself.

Yes.

Quoted 5 lines

It looks like this only affects *NVMe* devices, since vmd_enable_aspm() checks for PCI_CLASS_STORAGE_EXPRESS.  Why is that? Is there something special about NVMe?  I'd think you would want to do this for *all* devices below a VMD.

We need this for all devices under the PCIe root ports below VMD. Those will only be NVMe device AFAICS.

Quoted 8 lines

Since it only affects PCI_CLASS_STORAGE_EXPRESS devices, I don't think it actually "sets ASPM policy for the root ports".  vmd_enable_aspm() calls pcie_aspm_policy_override() on endpoints.  It's true that the link ASPM state happens to be attached to the upstream end of the link, but that's an ASPM implementation detail.

This all needs to be clear in the subject and commit log.

Ack

Quoted 7 lines
Quoted 3 lines

To do this, add an additional flag in VMD features to specify devices that must have their respective policies overridden.

I'm not clear on why you want this to apply to only certain VMDs and not others.  Do some BIOSes configure ASPM for devices below some VMDs?

Not currently. But the plan is for future devices to move back to having BIOS do the programming.

Quoted 39 lines
Quoted 39 lines

[1] https://bugzilla.kernel.org/show_bug.cgi?id=213717

---
V4
 - Refactor vmd_enable_apsm() to exit early, making the lines shorter and more readable. Suggested by Christoph.
V3
 - No changes
V2
 - Use return status to print pci_info message if ASPM cannot be enabled.
 - Add missing static declaration, caught by [email protected]

 drivers/pci/controller/vmd.c | 43 +++++++++++++++++++++++++++++++++---

 1 file changed, 40 insertions(+), 3 deletions(-)
drivers/pci/controller/vmd.c +0 -0
index a45e8e59d3d4..880afd450a14 100644
@@ -20,6 +20,8 @@


 #include <asm/irqdomain.h>

drivers/pci/controller/vmd.c +2 -0
+#include "../pci.h"
+

 #define VMD_CFGBAR     0
 #define VMD_MEMBAR1    2
 #define VMD_MEMBAR2    4

drivers/pci/controller/vmd.c +3 -0
@@ -67,6 +69,12 @@ enum vmd_features {
          * interrupt handling.
          */
         VMD_FEAT_CAN_BYPASS_MSI_REMAP           = (1 << 4),
+
+       /*
+        * Device must have ASPM policy overridden, as its default policy is
Quoted 8 lines

drivers/pci/controller/vmd.c enum vmd_features { · line 77

Quoted 3 lines
drivers/pci/controller/vmd.c +3 -0
+        * incorrect.
+        */
+       VMD_FEAT_QUIRK_OVERRIDE_ASPM            = (1 << 5),

I think you specifically want to *enable* some ASPM link states, not just "override the default policy."  "Override" tells us nothing about whether you are enabling or disabling ASPM.  Applies to subject line as well.

Ack

Quoted 16 lines
Quoted 11 lines
drivers/pci/controller/vmd.c +0 -0
  };


 static DEFINE_IDA(vmd_instance_ida);

drivers/pci/controller/vmd.c +0 -0
@@ -661,6 +669,30 @@ static int vmd_alloc_irqs(struct vmd_dev *vmd)
         return 0;
  }


+/*
+ * Override the BIOS ASPM policy and set the LTR value for PCI storage
+ * devices on the VMD bride.

I don't think there's any BIOS "policy" here.  At this point BIOS is no longer involved at all, so all that's left is whatever ASPM config the BIOS did or did not do.

Quoted 1 line

Why only storage?

Only storage devices will be on these root ports.

Quoted 14 lines

s/bride/bridge/

Quoted 11 lines

+ */

drivers/pci/controller/vmd.c +10 -0
+static int vmd_enable_aspm(struct pci_dev *pdev, void *userdata)
+{
+       int features = *(int *)userdata, pos;
+
+       if (!(features & VMD_FEAT_QUIRK_OVERRIDE_ASPM) ||
+           pdev->class != PCI_CLASS_STORAGE_EXPRESS)
+               return 0;
+
+       pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_LTR);
+       if (!pos)
Quoted 8 lines
Quoted 4 lines

+               return 0;
+
+       pci_write_config_word(pdev, pos + PCI_LTR_MAX_SNOOP_LAT, 0x1003); +       pci_write_config_word(pdev, pos + PCI_LTR_MAX_NOSNOOP_LAT, 0x1003);

1) Where did this magic 0x1003 value come from?  Does that depend on the VMD device?  The endpoint?  The circuit design?  The path between endpoint and VMD?  What if there are switches in the path?

The number comes from the BIOS team. They are tied to the SoC. I don't believe there can be switches in the path but Nirmal and Jonathan should know for sure. From what I've seen these root ports are wired directly to M.2 slots on boards that are intended for storage devices.

Quoted 6 lines

2) There exist broken devices where WORD config accesses don't work:
https://lore.kernel.org/all/[email protected]/

We might need a way to quirk config accesses to those devices, but we don't have one yet.  So for now this needs to be a single DWORD write.

Ack

Quoted 10 lines
Quoted 3 lines

+       if (pcie_aspm_policy_override(pdev))
+               pci_info(pdev, "Unable of override ASPM policy\n");

s/Unable of/Unable to/

I think we might need a message about when we *do* override the policy.  A note in dmesg might be useful for debugging.  I'm worried about the LTR programming because I really don't understand how we should be doing that.

Sure.

Quoted 15 lines
Quoted 10 lines

+       return 0;

drivers/pci/controller/vmd.c +2 -0
+}
+

 static int vmd_enable_domain(struct vmd_dev *vmd, unsigned long features) {
        struct pci_sysdata *sd = &vmd->sysdata;

drivers/pci/controller/vmd.c +0 -0
@@ -807,6 +839,8 @@ static int vmd_enable_domain(struct vmd_dev *vmd, unsigned long features)
         pci_scan_child_bus(vmd->bus);
         pci_assign_unassigned_bus_resources(vmd->bus);
Quoted 2 lines


+       pci_walk_bus(vmd->bus, vmd_enable_aspm, &features);

Do you support hotplug under VMD?  This will not happen for hot-added devices.

That I don't know. Nirmal, Jonathan?

David

Quoted 29 lines
Quoted 29 lines

        /*
         * VMD root buses are virtual and don't return true on pci_is_pcie()
         * and will fail pcie_bus_configure_settings() early. It can instead be

drivers/pci/controller/vmd.c +0 -0
@@ -948,15 +982,18 @@ static const struct pci_device_id vmd_ids[] = {
         {PCI_DEVICE(PCI_VENDOR_ID_INTEL, 0x467f),

                .driver_data = VMD_FEAT_HAS_MEMBAR_SHADOW_VSCAP | VMD_FEAT_HAS_BUS_RESTRICTIONS | -                               VMD_FEAT_OFFSET_FIRST_VECTOR,}, +                               VMD_FEAT_OFFSET_FIRST_VECTOR | +                               VMD_FEAT_QUIRK_OVERRIDE_ASPM,}, {PCI_DEVICE(PCI_VENDOR_ID_INTEL, 0x4c3d), .driver_data = VMD_FEAT_HAS_MEMBAR_SHADOW_VSCAP | VMD_FEAT_HAS_BUS_RESTRICTIONS | -                               VMD_FEAT_OFFSET_FIRST_VECTOR,}, +                               VMD_FEAT_OFFSET_FIRST_VECTOR | +                               VMD_FEAT_QUIRK_OVERRIDE_ASPM,}, {PCI_DEVICE(PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_VMD_9A0B), .driver_data = VMD_FEAT_HAS_MEMBAR_SHADOW_VSCAP | VMD_FEAT_HAS_BUS_RESTRICTIONS | -                               VMD_FEAT_OFFSET_FIRST_VECTOR,}, +                               VMD_FEAT_OFFSET_FIRST_VECTOR | +                               VMD_FEAT_QUIRK_OVERRIDE_ASPM,}, {0,}
 };
 MODULE_DEVICE_TABLE(pci, vmd_ids);

Signature · 1 line
2.25.1
#4 lore raw

Re: [PATCH V4 2/2] PCI: vmd: Override ASPM on TGL/ADL VMD devices

Quoted 14 lines · David E. Box
Quoted 13 lines · Bjorn Helgaas
Quoted 11 lines · David E. Box

From: Michael Bottini <[email protected]>

On Tiger Lake and Alder Lake platforms, VMD controllers do not have ASPM enabled nor LTR values set by BIOS. This leads high power consumption on these platforms when VMD is enabled as reported in bugzilla [1].  Enable these features in the VMD driver using pcie_aspm_policy_override() to set the ASPM policy for the root ports.
...

To do this, add an additional flag in VMD features to specify devices that must have their respective policies overridden.

Quoted 6 lines
Quoted 3 lines

I'm not clear on why you want this to apply to only certain VMDs and not others.  Do some BIOSes configure ASPM for devices below some VMDs?

Not currently. But the plan is for future devices to move back to having BIOS do the programming.

Since this is apparently a BIOS design choice, it seems wrong to base the functionality on the Device ID instead of some signal that tells us what the BIOS is doing.

Quoted 10 lines
Quoted 8 lines
Quoted 2 lines

+ * Override the BIOS ASPM policy and set the LTR value for PCI storage
+ * devices on the VMD bride.

I don't think there's any BIOS "policy" here.  At this point BIOS is no longer involved at all, so all that's left is whatever ASPM config the BIOS did or did not do.

Why only storage?

Only storage devices will be on these root ports.

How do you know this? You say below that there's an M.2 slot, so surely the slot could contain a non-storage device? Couldn't somebody build a platform with a VMD root port connected to a regular PCIe x4 slot? Couldn't such a slot support hotplug?

It would be very unusual to hard-code topology knowledge like this into the kernel, since plug-and-play has always been a major goal of PCI.

Quoted 15 lines
Quoted 8 lines
Quoted 7 lines
drivers/pci/controller/vmd.c +7 -0
+static int vmd_enable_aspm(struct pci_dev *pdev, void *userdata)
+{
+       int features = *(int *)userdata, pos;
+
+       if (!(features & VMD_FEAT_QUIRK_OVERRIDE_ASPM) ||
+           pdev->class != PCI_CLASS_STORAGE_EXPRESS)
+               return 0;
Quoted 6 lines
Quoted 2 lines

+       pci_write_config_word(pdev, pos + PCI_LTR_MAX_SNOOP_LAT, 0x1003); +       pci_write_config_word(pdev, pos + PCI_LTR_MAX_NOSNOOP_LAT, 0x1003);

1) Where did this magic 0x1003 value come from?  Does that depend on the VMD device?  The endpoint?  The circuit design?  The path between endpoint and VMD?  What if there are switches in the path?

Quoted 5 lines

The number comes from the BIOS team. They are tied to the SoC. I don't believe there can be switches in the path but Nirmal and Jonathan should know for sure. From what I've seen these root ports are wired directly to M.2 slots on boards that are intended for storage devices.

I guess you're saying that 0x1003 is determined by the SoC. If so, I think this value should be in your .driver_data (which would mean converting it from a scalar to a struct, as many other drivers do). The current code suggests that 0x1003 is the correct value for *all* VMDs and all configurations.

I don't understand LTR well enough to be convinced that this static value would be the correct value for all possible hierarchies and devices that could appear below a VMD root port. I would love to be educated about this.

Bjorn

#5 lore raw

Re: [PATCH V4 2/2] PCI: vmd: Override ASPM on TGL/ADL VMD devices

Hi Born,

So you know this patch comes from our platform power management team to address several VMD bugs affecting power. The VMD driver developers are already cc'd but also adding Francisco from the VMD team who can maybe help clarify.

Quoted 19 lines · Bjorn Helgaas
Quoted 18 lines · David E. Box
Quoted 16 lines · Bjorn Helgaas
Quoted 11 lines · David E. Box

From: Michael Bottini <[email protected]>

On Tiger Lake and Alder Lake platforms, VMD controllers do not have ASPM enabled nor LTR values set by BIOS. This leads high power consumption on these platforms when VMD is enabled as reported in bugzilla [1].  Enable these features in the VMD driver using pcie_aspm_policy_override() to set the ASPM policy for the root ports.
...

To do this, add an additional flag in VMD features to specify devices that must have their respective policies overridden.

I'm not clear on why you want this to apply to only certain VMDs and not others.  Do some BIOSes configure ASPM for devices below some VMDs?

Quoted 6 lines
Quoted 2 lines

Not currently. But the plan is for future devices to move back to having BIOS do the programming.

Since this is apparently a BIOS design choice, it seems wrong to base the functionality on the Device ID instead of some signal that tells us what the BIOS is doing.

I don't know if there is another way to tell what the BIOS is _not_ doing. DID is tied to the SoC so it's our best method of inferring this. We know the firmware team has said they will not support programming these values for the mentioned platforms so there's no BIOS update coming to change this. But to be sure, we can add a check first to confirm that these values have not been programmed (ASPM is disabled and LTRs are 0).

Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 3 lines

+ * Override the BIOS ASPM policy and set the LTR value for PCI storage
+ * devices on the VMD bride.

I don't think there's any BIOS "policy" here.  At this point BIOS is no longer involved at all, so all that's left is whatever ASPM config the BIOS did or did not do.

Quoted 8 lines
Quoted 3 lines
Quoted 1 line

Why only storage?

Only storage devices will be on these root ports.

How do you know this?  You say below that there's an M.2 slot, so surely the slot could contain a non-storage device?  Couldn't somebody build a platform with a VMD root port connected to a regular PCIe x4 slot?

In VMD mode the root ports under the VMD bridge are used specifically to manage storage devices. So only storage devices should be attached to these ports.

Quoted 1 line

  Couldn't such a slot support hotplug?

The answer I just learned is yes. It may be possible to have a switch too. So both of those would need to be addressed though I'm note quite sure how yet.

Quoted 26 lines

It would be very unusual to hard-code topology knowledge like this into the kernel, since plug-and-play has always been a major goal of PCI.

Quoted 15 lines
Quoted 8 lines
Quoted 7 lines
drivers/pci/controller/vmd.c +7 -0
+static int vmd_enable_aspm(struct pci_dev *pdev, void *userdata)
+{
+       int features = *(int *)userdata, pos;
+
+       if (!(features & VMD_FEAT_QUIRK_OVERRIDE_ASPM) ||
+           pdev->class != PCI_CLASS_STORAGE_EXPRESS)
+               return 0;
Quoted 6 lines
Quoted 2 lines

+       pci_write_config_word(pdev, pos + PCI_LTR_MAX_SNOOP_LAT, 0x1003); +       pci_write_config_word(pdev, pos + PCI_LTR_MAX_NOSNOOP_LAT, 0x1003);

1) Where did this magic 0x1003 value come from?  Does that depend on the VMD device?  The endpoint?  The circuit design?  The path between endpoint and VMD?  What if there are switches in the path?

Quoted 5 lines

The number comes from the BIOS team. They are tied to the SoC. I don't believe there can be switches in the path but Nirmal and Jonathan should know for sure. From what I've seen these root ports are wired directly to M.2 slots on boards that are intended for storage devices.

Quoted 5 lines

I guess you're saying that 0x1003 is determined by the SoC.  If so, I think this value should be in your .driver_data (which would mean converting it from a scalar to a struct, as many other drivers do). The current code suggests that 0x1003 is the correct value for *all* VMDs and all configurations.

Okay

Quoted 5 lines

I don't understand LTR well enough to be convinced that this static value would be the correct value for all possible hierarchies and devices that could appear below a VMD root port.  I would love to be educated about this.

I am not an LTR expert either, but LTRs are "conglomerated" per spec. Switches subtract their own latency from the minimum LTR of the downstream ports and report this conglomerated value upstream. So the value will decrease as needed to account for latency in the hierarchy.

PCI 5.0 Version 1.0, Section 6.18 (pg 611)

Thanks

David

Quoted 2 lines

Bjorn

#9 lore raw

Re: [PATCH V4 2/2] PCI: vmd: Override ASPM on TGL/ADL VMD devices

In reply to David E. Box

Hello,

Quoted 3 lines · David E. Box
Quoted 1 line · Bjorn Helgaas

[+cc Rajat for LTR max latency write]

Thanks!

Quoted 153 lines
Quoted 10 lines · David E. Box
Quoted 7 lines

From: Michael Bottini <[email protected]>

On Tiger Lake and Alder Lake platforms, VMD controllers do not have ASPM enabled nor LTR values set by BIOS. This leads high power consumption on these platforms when VMD is enabled as reported in bugzilla [1]. Enable these features in the VMD driver using pcie_aspm_policy_override() to set the ASPM policy for the root ports.

Quoted 5 lines

s/leads high/leads to high/

Does this depend on "Tiger Lake and Alder Lake platforms" specifically, or does it depend on a BIOS design choice, i.e., "don't configure ASPM or LTR for devices below a VMD"?

It was a BIOS design choice to have this done by the OS driver for both Tiger Lake and Alder Lake.

Quoted 4 lines

The subject says "override ASPM on VMD devices," but it looks like this affects the ASPM configuration of devices *below* the VMD, not of the VMD itself.

Yes.

Quoted 5 lines

It looks like this only affects *NVMe* devices, since vmd_enable_aspm() checks for PCI_CLASS_STORAGE_EXPRESS. Why is that? Is there something special about NVMe? I'd think you would want to do this for *all* devices below a VMD.

We need this for all devices under the PCIe root ports below VMD. Those will only be NVMe device AFAICS.

Quoted 8 lines

Since it only affects PCI_CLASS_STORAGE_EXPRESS devices, I don't think it actually "sets ASPM policy for the root ports". vmd_enable_aspm() calls pcie_aspm_policy_override() on endpoints. It's true that the link ASPM state happens to be attached to the upstream end of the link, but that's an ASPM implementation detail.

This all needs to be clear in the subject and commit log.

Ack

Quoted 7 lines
Quoted 3 lines

To do this, add an additional flag in VMD features to specify devices that must have their respective policies overridden.

I'm not clear on why you want this to apply to only certain VMDs and not others. Do some BIOSes configure ASPM for devices below some VMDs?

Not currently. But the plan is for future devices to move back to having BIOS do the programming.

Quoted 39 lines
Quoted 39 lines

[1] https://bugzilla.kernel.org/show_bug.cgi?id=213717

---
V4
- Refactor vmd_enable_apsm() to exit early, making the lines shorter and more readable. Suggested by Christoph.
V3
- No changes
V2
- Use return status to print pci_info message if ASPM cannot be enabled.
- Add missing static declaration, caught by [email protected]

 drivers/pci/controller/vmd.c | 43 +++++++++++++++++++++++++++++++++---
 1 file changed, 40 insertions(+), 3 deletions(-)
drivers/pci/controller/vmd.c +5 -0
index a45e8e59d3d4..880afd450a14 100644
@@ -20,6 +20,8 @@
 
 #include <asm/irqdomain.h>
 
+#include "../pci.h"
+
 #define VMD_CFGBAR     0
 #define VMD_MEMBAR1    2
 #define VMD_MEMBAR2    4
@@ -67,6 +69,12 @@ enum vmd_features {
         * interrupt handling.
         */
        VMD_FEAT_CAN_BYPASS_MSI_REMAP           = (1 << 4),
+
+       /*
+        * Device must have ASPM policy overridden, as its default policy is
Quoted 8 lines

drivers/pci/controller/vmd.c enum vmd_features { · line 77

Quoted 3 lines
drivers/pci/controller/vmd.c +3 -0
+        * incorrect.
+        */
+       VMD_FEAT_QUIRK_OVERRIDE_ASPM            = (1 << 5),

I think you specifically want to *enable* some ASPM link states, not just "override the default policy." "Override" tells us nothing about whether you are enabling or disabling ASPM. Applies to subject line as well.

Ack

Quoted 16 lines

drivers/pci/controller/vmd.c static int vmd_alloc_irqs(struct vmd_dev *vmd) · line 674

Quoted 11 lines
drivers/pci/controller/vmd.c +3 -0
 };
 
 static DEFINE_IDA(vmd_instance_ida);
@@ -661,6 +669,30 @@ static int vmd_alloc_irqs(struct vmd_dev *vmd)
        return 0;
 }
 
+/*
+ * Override the BIOS ASPM policy and set the LTR value for PCI storage
+ * devices on the VMD bride.

I don't think there's any BIOS "policy" here. At this point BIOS is no longer involved at all, so all that's left is whatever ASPM config the BIOS did or did not do.

Quoted 1 line

Why only storage?

Only storage devices will be on these root ports.

Quoted 14 lines

s/bride/bridge/

Quoted 11 lines
drivers/pci/controller/vmd.c +11 -0
+ */
+static int vmd_enable_aspm(struct pci_dev *pdev, void *userdata)
+{
+       int features = *(int *)userdata, pos;
+
+       if (!(features & VMD_FEAT_QUIRK_OVERRIDE_ASPM) ||
+           pdev->class != PCI_CLASS_STORAGE_EXPRESS)
+               return 0;
+
+       pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_LTR);
+       if (!pos)
Quoted 8 lines

drivers/pci/controller/vmd.c static int vmd_alloc_irqs(struct vmd_dev *vmd) · line 689

Quoted 4 lines
drivers/pci/controller/vmd.c +4 -0
+               return 0;
+
+       pci_write_config_word(pdev, pos + PCI_LTR_MAX_SNOOP_LAT, 0x1003);
+       pci_write_config_word(pdev, pos + PCI_LTR_MAX_NOSNOOP_LAT, 0x1003);

1) Where did this magic 0x1003 value come from? Does that depend on the VMD device? The endpoint? The circuit design? The path between endpoint and VMD? What if there are switches in the path?

The number comes from the BIOS team. They are tied to the SoC. I don't believe there can be switches in the path but Nirmal and Jonathan should know for sure. From what I've seen these root ports are wired directly to M.2 slots on boards that are intended for storage devices.

Quoted 1 line
Quoted 8 lines
Quoted 5 lines

2) There exist broken devices where WORD config accesses don't work:
https://lore.kernel.org/all/[email protected]/

We might need a way to quirk config accesses to those devices, but we don't have one yet. So for now this needs to be a single DWORD write.

Ack

Sideband: Yes, I'm trying to work out such a quirk mechanism, but for now, yes, please do that as a single DWORD write access.

Thanks,

Rajat

Quoted 64 lines
Quoted 10 lines

drivers/pci/controller/vmd.c static int vmd_alloc_irqs(struct vmd_dev *vmd) · line 691

Quoted 3 lines
drivers/pci/controller/vmd.c +2 -0
+       if (pcie_aspm_policy_override(pdev))
+               pci_info(pdev, "Unable of override ASPM policy\n");

s/Unable of/Unable to/

I think we might need a message about when we *do* override the policy. A note in dmesg might be useful for debugging. I'm worried about the LTR programming because I really don't understand how we should be doing that.

Sure.

Quoted 11 lines
Quoted 11 lines
drivers/pci/controller/vmd.c +3 -0
+       return 0;
+}
+
 static int vmd_enable_domain(struct vmd_dev *vmd, unsigned long features)
 {
        struct pci_sysdata *sd = &vmd->sysdata;
@@ -807,6 +839,8 @@ static int vmd_enable_domain(struct vmd_dev *vmd, unsigned long features)
        pci_scan_child_bus(vmd->bus);
        pci_assign_unassigned_bus_resources(vmd->bus);
 
Quoted 4 lines

drivers/pci/controller/vmd.c static int vmd_enable_domain(struct vmd_dev *vmd, unsigned long features) · line 842

Quoted 1 line
drivers/pci/controller/vmd.c +1 -0
+       pci_walk_bus(vmd->bus, vmd_enable_aspm, &features);

Do you support hotplug under VMD? This will not happen for hot-added devices.

That I don't know. Nirmal, Jonathan?

David

Quoted 29 lines
Quoted 29 lines
drivers/pci/controller/vmd.c +6 -4
        /*
         * VMD root buses are virtual and don't return true on pci_is_pcie()
         * and will fail pcie_bus_configure_settings() early. It can instead be
@@ -948,15 +982,18 @@ static const struct pci_device_id vmd_ids[] = {
        {PCI_DEVICE(PCI_VENDOR_ID_INTEL, 0x467f),
                .driver_data = VMD_FEAT_HAS_MEMBAR_SHADOW_VSCAP |
                                VMD_FEAT_HAS_BUS_RESTRICTIONS |
-                               VMD_FEAT_OFFSET_FIRST_VECTOR,},
+                               VMD_FEAT_OFFSET_FIRST_VECTOR |
+                               VMD_FEAT_QUIRK_OVERRIDE_ASPM,},
        {PCI_DEVICE(PCI_VENDOR_ID_INTEL, 0x4c3d),
                .driver_data = VMD_FEAT_HAS_MEMBAR_SHADOW_VSCAP |
                                VMD_FEAT_HAS_BUS_RESTRICTIONS |
-                               VMD_FEAT_OFFSET_FIRST_VECTOR,},
+                               VMD_FEAT_OFFSET_FIRST_VECTOR |
+                               VMD_FEAT_QUIRK_OVERRIDE_ASPM,},
        {PCI_DEVICE(PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_VMD_9A0B),
                .driver_data = VMD_FEAT_HAS_MEMBAR_SHADOW_VSCAP |
                                VMD_FEAT_HAS_BUS_RESTRICTIONS |
-                               VMD_FEAT_OFFSET_FIRST_VECTOR,},
+                               VMD_FEAT_OFFSET_FIRST_VECTOR |
+                               VMD_FEAT_QUIRK_OVERRIDE_ASPM,},
        {0,}
 };
 MODULE_DEVICE_TABLE(pci, vmd_ids);
--

2.25.1

#6 lore raw

Re: [PATCH V4 2/2] PCI: vmd: Override ASPM on TGL/ADL VMD devices

In reply to David E. Box

Hello,

Quoted 30 lines · David E. Box

Hi Born,

So you know this patch comes from our platform power management team to address several VMD bugs affecting power. The VMD driver developers are already cc'd but also adding Francisco from the VMD team who can maybe help clarify.

Quoted 22 lines · Bjorn Helgaas
Quoted 20 lines · David E. Box
Quoted 10 lines · Bjorn Helgaas
Quoted 8 lines · David E. Box

From: Michael Bottini <[email protected]>

On Tiger Lake and Alder Lake platforms, VMD controllers do not have ASPM enabled nor LTR values set by BIOS. This leads high power consumption on these platforms when VMD is enabled as reported in bugzilla [1]. Enable these features in the VMD driver using pcie_aspm_policy_override() to set the ASPM policy for the root ports.
...

Quoted 6 lines
Quoted 2 lines

To do this, add an additional flag in VMD features to specify devices that must have their respective policies overridden.

I'm not clear on why you want this to apply to only certain VMDs and not others. Do some BIOSes configure ASPM for devices below some VMDs?

Not currently. But the plan is for future devices to move back to having BIOS do the programming.

Quoted 8 lines
Quoted 3 lines

Since this is apparently a BIOS design choice, it seems wrong to base the functionality on the Device ID instead of some signal that tells us what the BIOS is doing.

I don't know if there is another way to tell what the BIOS is _not_ doing. DID is tied to the SoC so it's our best method of inferring this. We know the firmware team has said they will not support programming these values for the mentioned platforms so there's no BIOS update coming to change this.

I'd like to mention here that ChromeOS also runs linux kernel on TGL and ADL platforms, and we may run a different BIOS altogether, so yes, we'd greatly appreciate if you do not assume BIOS setting from device ID / SOC ID, if the setting is *not tied to device ID / SOC ID*. Perhaps you use ACPI or device trees to convey what BIOS has or hasn't configured?

Quoted 2 lines

But to be sure, we can add a check first to confirm that these values have not been programmed (ASPM is disabled and LTRs are 0).

Yes, reading the ASPM configuration from hardware sounds OK to me.

Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 4 lines

+ * Override the BIOS ASPM policy and set the LTR value for PCI storage
+ * devices on the VMD bride.

I don't think there's any BIOS "policy" here. At this point BIOS is no longer involved at all, so all that's left is whatever ASPM config the BIOS did or did not do.

Should the Policy override really take place if aspm_policy = POLICY_DEFAULT which seems to suggest we leave it to whatever BIOS left it at?

Perhaps for this override to take place, aspm_policy needs to be the policy needs to be POLICY_POWERSAVE?

#define POLICY_DEFAULT 0        /* BIOS default setting */
#define POLICY_PERFORMANCE 1    /* high performance */
#define POLICY_POWERSAVE 2      /* high power saving */

#define POLICY_POWER_SUPERSAVE 3 /* possibly even more power saving */

Quoted 5 lines
Quoted 5 lines
Quoted 4 lines
Quoted 2 lines

Why only storage?

Only storage devices will be on these root ports.

Quoted 7 lines
Quoted 4 lines

How do you know this? You say below that there's an M.2 slot, so surely the slot could contain a non-storage device? Couldn't somebody build a platform with a VMD root port connected to a regular PCIe x4 slot?

In VMD mode the root ports under the VMD bridge are used specifically to manage storage devices. So only storage devices should be attached to these ports.

Bjorn: from https://www.intel.com/content/www/us/en/architecture-and-technology/intel-volume-management-device-overview.html
"Intel® Volume Management Device (Intel® VMD) is specifically designed
for enterprise-grade management of NVMe SSDs connected to Intel® Xeon®
CPUs."

I think what David means is that any other device other than an NVME SSD showing up under VMD might be an error / bad device.

Quoted 46 lines
Quoted 2 lines

Couldn't such a slot support hotplug?

The answer I just learned is yes. It may be possible to have a switch too. So both of those would need to be addressed though I'm note quite sure how yet.

Quoted 26 lines

It would be very unusual to hard-code topology knowledge like this into the kernel, since plug-and-play has always been a major goal of PCI.

Quoted 15 lines
Quoted 8 lines
Quoted 7 lines
drivers/pci/controller/vmd.c +7 -0
+static int vmd_enable_aspm(struct pci_dev *pdev, void *userdata)
+{
+       int features = *(int *)userdata, pos;
+
+       if (!(features & VMD_FEAT_QUIRK_OVERRIDE_ASPM) ||
+           pdev->class != PCI_CLASS_STORAGE_EXPRESS)
+               return 0;
Quoted 6 lines

drivers/pci/controller/vmd.c static int vmd_alloc_irqs(struct vmd_dev *vmd) · line 689

Quoted 2 lines
drivers/pci/controller/vmd.c +2 -0
+       pci_write_config_word(pdev, pos + PCI_LTR_MAX_SNOOP_LAT, 0x1003);
+       pci_write_config_word(pdev, pos + PCI_LTR_MAX_NOSNOOP_LAT, 0x1003);

1) Where did this magic 0x1003 value come from? Does that depend on the VMD device? The endpoint? The circuit design? The path between endpoint and VMD? What if there are switches in the path?

Quoted 5 lines

The number comes from the BIOS team. They are tied to the SoC. I don't believe there can be switches in the path but Nirmal and Jonathan should know for sure. From what I've seen these root ports are wired directly to M.2 slots on boards that are intended for storage devices.

Quoted 5 lines

I guess you're saying that 0x1003 is determined by the SoC. If so, I think this value should be in your .driver_data (which would mean converting it from a scalar to a struct, as many other drivers do). The current code suggests that 0x1003 is the correct value for *all* VMDs and all configurations.

Okay

Quoted 5 lines

I don't understand LTR well enough to be convinced that this static value would be the correct value for all possible hierarchies and devices that could appear below a VMD root port. I would love to be educated about this.

Quoted 5 lines

I am not an LTR expert either, but LTRs are "conglomerated" per spec. Switches subtract their own latency from the minimum LTR of the downstream ports and report this conglomerated value upstream. So the value will decrease as needed to account for latency in the hierarchy.

PCI 5.0 Version 1.0, Section 6.18 (pg 611)

As described here and as much as I understand, I think these LTR values (0x1003 here) depend on the platform topology. So what works for David's platform may not work for another one that may have Switched for e.g., or for perhaps for another NVME that is hotplugged later (and has different characteristics than what is on board). Technically tying the LTR values to a SOC or VID/DID sounds incorrect to me, and instead should be tied to the platform. Ideally I *thiink* they should come from ACPI or device tree. However, I wouldn't protest putting them in driver_data for now given that there aren't any other users. We may have to find a solution once there are more platforms that begin to use this driver. I think this is a bigger problem that needs to be addressed w.r.t. ASPM in general.

Thanks & Best Regards,

Rajat

Quoted 9 lines

Thanks

David

Quoted 2 lines

Bjorn

#7 lore raw

Re: [PATCH V4 2/2] PCI: vmd: Override ASPM on TGL/ADL VMD devices

Hi,

Quoted 50 lines · Rajat Jain

Hello,

Quoted 38 lines · David E. Box

Hi Born,

So you know this patch comes from our platform power management team to address several VMD bugs
affecting power. The VMD driver developers are already cc'd but also adding Francisco from the VMD
team who can maybe help clarify.

Quoted 22 lines · Bjorn Helgaas
Quoted 21 lines · David E. Box
Quoted 19 lines · Bjorn Helgaas
Quoted 14 lines · David E. Box

From: Michael Bottini <[email protected]>

On Tiger Lake and Alder Lake platforms, VMD controllers do not have ASPM
enabled nor LTR values set by BIOS. This leads high power consumption on
these platforms when VMD is enabled as reported in bugzilla [1]. Enable
these features in the VMD driver using pcie_aspm_policy_override() to set
the ASPM policy for the root ports.
...
To do this, add an additional flag in VMD features to specify devices that must have their respective policies overridden.

I'm not clear on why you want this to apply to only certain VMDs and not others. Do some BIOSes configure ASPM for devices below some VMDs?

Quoted 6 lines
Quoted 2 lines

Not currently. But the plan is for future devices to move back to having BIOS do the programming.

Since this is apparently a BIOS design choice, it seems wrong to base the functionality on the Device ID instead of some signal that tells us what the BIOS is doing.

Quoted 7 lines

I don't know if there is another way to tell what the BIOS is _not_ doing. DID is tied to the SoC so
it's our best method of inferring this. We know the firmware team has said they will not support
programming these values for the mentioned platforms so there's no BIOS update coming to change
this.

Quoted 6 lines

I'd like to mention here that ChromeOS also runs linux kernel on TGL and ADL platforms, and we may run a different BIOS altogether, so yes, we'd greatly appreciate if you do not assume BIOS setting from device ID / SOC ID, if the setting is *not tied to device ID / SOC ID*. Perhaps you use ACPI or device trees to convey what BIOS has or hasn't configured?

The reason why BIOS is not programming these values is because when VMD is enabled the ports are not visible to BIOS. This would apply to any BIOS including those used by ChromeOS.

And because BIOS doesn't see these ports ...

Quoted 11 lines
Quoted 4 lines

But to be sure, we can add a check first to confirm that these values have not been programmed
(ASPM is disabled and LTRs are 0).

Yes, reading the ASPM configuration from hardware sounds OK to me.

Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 3 lines

+ * Override the BIOS ASPM policy and set the LTR value for PCI storage
+ * devices on the VMD bride.

Quoted 7 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines

I don't think there's any BIOS "policy" here. At this point BIOS is no longer involved at all, so all that's left is whatever ASPM config the BIOS did or did not do.

Should the Policy override really take place if aspm_policy = POLICY_DEFAULT which seems to suggest we leave it to whatever BIOS left it at?

... we cannot rely on POLICY_DEFAULT for their ASPM setting. This leaves the devices with ASPM disabled.

That said, we do honor the PCI aspm_disabled flag which is tied to both the FADT bit and module parameter. So we only enable it if OS control is allowed.

Quoted 8 lines

Perhaps for this override to take place, aspm_policy needs to be the policy needs to be POLICY_POWERSAVE?

#define POLICY_DEFAULT 0        /* BIOS default setting */
#define POLICY_PERFORMANCE 1    /* high performance */
#define POLICY_POWERSAVE 2      /* high power saving */

#define POLICY_POWER_SUPERSAVE 3 /* possibly even more power saving */

We do have to set a default in place of the missing BIOS configuration. We chose to do ASPM_STATE_ALL, allowing all ASPM settings supported by the endpoint.

Quoted 15 lines
Quoted 6 lines
Quoted 6 lines
Quoted 5 lines
Quoted 3 lines

Why only storage?

Only storage devices will be on these root ports.

Quoted 8 lines
Quoted 4 lines

How do you know this? You say below that there's an M.2 slot, so surely the slot could contain a non-storage device? Couldn't somebody build a platform with a VMD root port connected to a regular PCIe x4 slot?

In VMD mode the root ports under the VMD bridge are used specifically to manage storage devices. So
only storage devices should be attached to these ports.

Quoted 8 lines

Bjorn: from
https://www.intel.com/content/www/us/en/architecture-and-technology/intel-volume-management-device-overview.html
"Intel® Volume Management Device (Intel® VMD) is specifically designed for enterprise-grade management of NVMe SSDs connected to Intel® Xeon® CPUs."

I think what David means is that any other device other than an NVME SSD showing up under VMD might be an error / bad device.

I don't know if that would happen. It certainly would not be used as intended.

Quoted 56 lines
Quoted 47 lines
Quoted 3 lines

Couldn't such a slot support hotplug?

The answer I just learned is yes. It may be possible to have a switch too. So both of those would
need to be addressed though I'm note quite sure how yet.

Quoted 26 lines

It would be very unusual to hard-code topology knowledge like this into the kernel, since plug-and-play has always been a major goal of PCI.

Quoted 16 lines
Quoted 7 lines
Quoted 7 lines
drivers/pci/controller/vmd.c +7 -0
+static int vmd_enable_aspm(struct pci_dev *pdev, void *userdata)
+{
+       int features = *(int *)userdata, pos;
+
+       if (!(features & VMD_FEAT_QUIRK_OVERRIDE_ASPM) ||
+           pdev->class != PCI_CLASS_STORAGE_EXPRESS)
+               return 0;
Quoted 8 lines
Quoted 4 lines

+ pci_write_config_word(pdev, pos + PCI_LTR_MAX_SNOOP_LAT,
0x1003);
+ pci_write_config_word(pdev, pos + PCI_LTR_MAX_NOSNOOP_LAT,
0x1003);

1) Where did this magic 0x1003 value come from? Does that depend on the VMD device? The endpoint? The circuit design? The path between endpoint and VMD? What if there are switches in the path?

Quoted 5 lines

The number comes from the BIOS team. They are tied to the SoC. I don't believe there can be switches in the path but Nirmal and Jonathan should know for sure. From what I've seen these root ports are wired directly to M.2 slots on boards that are intended for storage devices.

Quoted 5 lines

I guess you're saying that 0x1003 is determined by the SoC. If so, I think this value should be in your .driver_data (which would mean converting it from a scalar to a struct, as many other drivers do). The current code suggests that 0x1003 is the correct value for *all* VMDs and all configurations.

Okay

Quoted 4 lines

I don't understand LTR well enough to be convinced that this static value would be the correct value for all possible hierarchies and devices that could appear below a VMD root port. I would love to be educated about this.

Quoted 8 lines

I am not an LTR expert either, but LTRs are "conglomerated" per spec. Switches subtract their own
latency from the minimum LTR of the downstream ports and report this conglomerated value upstream.
So the value will decrease as needed to account for latency in the hierarchy.

PCI 5.0 Version 1.0, Section 6.18 (pg 611)

Quoted 12 lines

As described here and as much as I understand, I think these LTR values (0x1003 here) depend on the platform topology. So what works for David's platform may not work for another one that may have Switched for e.g., or for perhaps for another NVME that is hotplugged later (and has different characteristics than what is on board). Technically tying the LTR values to a SOC or VID/DID sounds incorrect to me, and instead should be tied to the platform. Ideally I *thiink* they should come from ACPI or device tree. However, I wouldn't protest putting them in driver_data for now given that there aren't any other users. We may have to find a solution once there are more platforms that begin to use this driver. I think this is a bigger problem that needs to be addressed w.r.t. ASPM in general.

I'll seek more clarification from firmware team on the LTR. But again the fact that switches can reduce the reported LTR seems to allow for at least some decoupling of the endpoint LTR value from the topology.

David

Quoted 11 lines

Thanks & Best Regards,

Rajat

Quoted 5 lines

Thanks

David

Quoted 1 line

Bjorn

#8 lore raw

Re: [PATCH V4 2/2] PCI: vmd: Override ASPM on TGL/ADL VMD devices

Quoted 4 lines · David E. Box

The reason why BIOS is not programming these values is because when VMD is enabled the ports are not visible to BIOS. This would apply to any BIOS including those used by ChromeOS.
And because BIOS doesn't see these ports ...

Isn't VMD enabled by the BIOS? The bios should be able see them the same way as Linux does. But given that the whole point of VMD is to hide these ports from Windows it obviously doesn't. For Linux and thus for ChromeOS VMD is completely pointless, so give that the ChromeOS people aren't as dumb as some people at Intel I'm pretty sure they won't enable it and just use the ports as normal one and avoid this whole self-inflicted pain.

If only Intel could give the OS a way to disable VMD at runtime and discover the actual ports, then we would not have this whole nightmare of having to support it.