[PATCH v3 0/7] PCI: Add common helper for 100 ms delay after link training

[PATCH v3 0/7] PCI: Add common helper for 100 ms delay after link training

[Hans Zhang]

PCIe r6.0, sec 6.6.1 (Conventional Reset) requires that for a Downstream
Port supporting Link speeds greater than 5.0 GT/s, software must wait a
minimum of 100 ms after Link training completes before sending any
Configuration Request.

Several PCIe host controller drivers currently omit this 100 ms delay
when the negotiated link speed is Gen3 (8 GT/s) or higher. Only the DWC
driver already implements it. The missing delay can lead to violations
of the PCIe specification and cause enumeration failures with high-speed
devices (e.g., NVIDIA RTX5070 GPU, PCIe 5.0 NVMe SSDs).

To fix this consistently and avoid code duplication, this series:

  1. Adds a static inline helper `pci_host_common_link_train_delay()`
     in `drivers/pci/controller/pci-host-common.h`. The helper checks
     the given max_link_speed and calls msleep(100) if speed > 5 GT/s.

  2. Converts the DWC driver to use this helper.

  3. Adds the missing 100 ms delay to the Cadence PCIe controller
     (both LGA and HPA IPs). A `max_link_speed` field is introduced
     in `struct cdns_pcie`. The j721e glue driver sets this field;
     other LGA glue drivers fall back to reading DT property
     "max-link-speed". HPA also reads the DT property.

  4. Adds the delay to the Aardvark and MediaTek Gen3 host drivers,
     and replaces the existing unconditional delay in the Renesas
     RZ/G3S driver with the conditional helper (per spec).

All changes are placed immediately after link training completes and
before any Configuration Request would be issued.

---
Our company's product is based on the HPA IP from Cadence. When connecting
to different devices, we encountered issues with the enumeration failure
when connecting to the NVIDIA RTX5070 GPU and the NVMe SSD with PCIe 5.0
interface. Our code is based on: 80dc18a0cba8d ("PCI: dwc: Ensure that
dw_pcie_wait_for_link() waits 100 ms after link up").
---
Changes since v3:
- Renamed helper from pcie_wait_after_link_train() to
  pci_host_common_link_train_delay() and moved to pci-host-common.h. (Mani)
- Reorganized Cadence patches: LGA + j721e glue in one patch, HPA separate.
- Added missing DT property reading for HPA and LGA fallback.
- Replace the existing msleep(100) call with the common helper
  pci_host_common_link_train_delay(). (Claudiu)

Changes since v2:
- Add pcie_wait_after_link_train() helper
- Reduce repetitive code comments and have each Root Port driver use the
  helper function instead.
- Increase the delay to 100ms after enabling the link-up that distinguishes
  between Cadence LGA and HPA IPs.
- Add the Aardvark, MediaTek Gen3, and Renesas RZ/G3S Root Port driver. When
  the speed is greater than GEN2, a delay of 100ms should be applied.

v2:
https://patchwork.kernel.org/project/linux-pci/cover/20260506152346.166056-1-18255117159@163.com/

v1:
https://patchwork.kernel.org/project/linux-pci/patch/20260501153553.66382-1-18255117159@163.com/
---
Hans Zhang (7):
  PCI: Add pci_host_common_link_train_delay() helper
  PCI: cadence: Add post-link delay for LGA and j721e glue driver
  PCI: cadence: HPA: Add post-link delay
  PCI: dwc: Use common pci_host_common_link_train_delay() helper
  PCI: aardvark: Add 100 ms delay after link training
  PCI: mediatek-gen3: Add 100 ms delay after link up
  PCI: rzg3s-host: Use common pci_host_common_link_train_delay() helper

 drivers/pci/controller/cadence/pci-j721e.c      |  1 +
 .../cadence/pcie-cadence-host-common.c          |  4 ++++
 .../controller/cadence/pcie-cadence-host-hpa.c  |  8 ++++++++
 .../pci/controller/cadence/pcie-cadence-host.c  |  4 ++++
 drivers/pci/controller/cadence/pcie-cadence.h   |  2 ++
 drivers/pci/controller/dwc/pcie-designware.c    |  9 ++-------
 drivers/pci/controller/pci-aardvark.c           |  5 ++++-
 drivers/pci/controller/pci-host-common.h        | 17 +++++++++++++++++
 drivers/pci/controller/pcie-mediatek-gen3.c     |  3 +++
 drivers/pci/controller/pcie-rzg3s-host.c        |  3 ++-
 10 files changed, 47 insertions(+), 9 deletions(-)


base-commit: 70390501d1944d4e5b8f7352be180fceb3a44132
-- 
2.34.1

[PATCH v3 1/7] PCI: Add pci_host_common_link_train_delay() helper

[Hans Zhang]

PCIe r6.0, sec 6.6.1 (Conventional Reset) requires that for a Downstream
Port supporting Link speeds greater than 5.0 GT/s, software must wait a
minimum of 100 ms after Link training completes before sending any
Configuration Request.

Introduce a static inline helper pci_host_common_link_train_delay() that
checks the given max_link_speed (2 = 5.0 GT/s, 3 = 8.0 GT/s, etc.) and
calls msleep(100) only when the speed is greater than 5.0 GT/s.

This allows multiple host controller drivers to share the same mandatory
delay without duplicating the logic.

Signed-off-by: Hans Zhang <18255117159@163.com>
---
 drivers/pci/controller/pci-host-common.h | 17 +++++++++++++++++
 1 file changed, 17 insertions(+)

diff --git a/drivers/pci/controller/pci-host-common.h b/drivers/pci/controller/pci-host-common.h
index b5075d4bd7eb..d709f7e3e11a 100644
--- a/drivers/pci/controller/pci-host-common.h
+++ b/drivers/pci/controller/pci-host-common.h
@@ -10,6 +10,9 @@
 #ifndef _PCI_HOST_COMMON_H
 #define _PCI_HOST_COMMON_H
 
+#include <linux/delay.h>
+#include "../pci.h"
+
 struct pci_ecam_ops;
 
 int pci_host_common_probe(struct platform_device *pdev);
@@ -20,4 +23,18 @@ void pci_host_common_remove(struct platform_device *pdev);
 
 struct pci_config_window *pci_host_common_ecam_create(struct device *dev,
 	struct pci_host_bridge *bridge, const struct pci_ecam_ops *ops);
+
+/**
+ * pci_host_common_link_train_delay - Wait 100 ms if link speed > 5 GT/s
+ * @max_link_speed: the maximum link speed (2 = 5.0 GT/s, 3 = 8.0 GT/s, ...)
+ *
+ * Must be called after Link training completes and before the first
+ * Configuration Request is sent.
+ */
+static inline void pci_host_common_link_train_delay(int max_link_speed)
+{
+	if (max_link_speed > 2)
+		msleep(PCIE_RESET_CONFIG_WAIT_MS);
+}
+
 #endif
-- 
2.34.1

[PATCH v3 2/7] PCI: cadence: Add post-link delay for LGA and j721e glue driver

[Hans Zhang]

The Cadence LGA (Legacy Architecture IP) PCIe host controller currently
lacks the mandatory 100 ms delay after link training completes for speeds
> 5.0 GT/s, as required by PCIe r6.0 sec 6.6.1.

Add a 'max_link_speed' field to struct cdns_pcie. In the common host
layer function cdns_pcie_host_start_link(), after the link has been
successfully established, call pci_host_common_link_train_delay() to
insert the required delay.

For the j721e glue driver, set cdns_pcie.max_link_speed from the existing
link speed logic. For other LGA-based glue drivers (sky1, sg2042), the
common LGA host setup (pcie-cadence-host.c) provides a fallback reading
of the device tree property "max-link-speed" when available. This ensures
that the delay is not missed on those platforms once they enable the
property.

Signed-off-by: Hans Zhang <18255117159@163.com>
---
 drivers/pci/controller/cadence/pci-j721e.c                | 1 +
 drivers/pci/controller/cadence/pcie-cadence-host-common.c | 4 ++++
 drivers/pci/controller/cadence/pcie-cadence-host.c        | 4 ++++
 drivers/pci/controller/cadence/pcie-cadence.h             | 2 ++
 4 files changed, 11 insertions(+)

diff --git a/drivers/pci/controller/cadence/pci-j721e.c b/drivers/pci/controller/cadence/pci-j721e.c
index bfdfe98d5aba..ee85b8e04f5b 100644
--- a/drivers/pci/controller/cadence/pci-j721e.c
+++ b/drivers/pci/controller/cadence/pci-j721e.c
@@ -206,6 +206,7 @@ static int j721e_pcie_set_link_speed(struct j721e_pcie *pcie,
 	    (pcie_get_link_speed(link_speed) == PCI_SPEED_UNKNOWN))
 		link_speed = 2;
 
+	pcie->cdns_pcie.max_link_speed = link_speed;
 	val = link_speed - 1;
 	ret = regmap_update_bits(syscon, offset, GENERATION_SEL_MASK, val);
 	if (ret)
diff --git a/drivers/pci/controller/cadence/pcie-cadence-host-common.c b/drivers/pci/controller/cadence/pcie-cadence-host-common.c
index 2b0211870f02..18e4b6c760b5 100644
--- a/drivers/pci/controller/cadence/pcie-cadence-host-common.c
+++ b/drivers/pci/controller/cadence/pcie-cadence-host-common.c
@@ -14,6 +14,7 @@
 
 #include "pcie-cadence.h"
 #include "pcie-cadence-host-common.h"
+#include "../pci-host-common.h"
 
 #define LINK_RETRAIN_TIMEOUT HZ
 
@@ -115,6 +116,9 @@ int cdns_pcie_host_start_link(struct cdns_pcie_rc *rc,
 	if (!ret && rc->quirk_retrain_flag)
 		ret = cdns_pcie_retrain(pcie, pcie_link_up);
 
+	if (!ret)
+		pci_host_common_link_train_delay(pcie->max_link_speed);
+
 	return ret;
 }
 EXPORT_SYMBOL_GPL(cdns_pcie_host_start_link);
diff --git a/drivers/pci/controller/cadence/pcie-cadence-host.c b/drivers/pci/controller/cadence/pcie-cadence-host.c
index 0bc9e6e90e0e..058e4e619654 100644
--- a/drivers/pci/controller/cadence/pcie-cadence-host.c
+++ b/drivers/pci/controller/cadence/pcie-cadence-host.c
@@ -13,6 +13,7 @@
 
 #include "pcie-cadence.h"
 #include "pcie-cadence-host-common.h"
+#include "../../pci.h"
 
 static u8 bar_aperture_mask[] = {
 	[RP_BAR0] = 0x1F,
@@ -397,6 +398,9 @@ int cdns_pcie_host_setup(struct cdns_pcie_rc *rc)
 	rc->device_id = 0xffff;
 	of_property_read_u32(np, "device-id", &rc->device_id);
 
+	if (pcie->max_link_speed < 1)
+		pcie->max_link_speed = of_pci_get_max_link_speed(np);
+
 	pcie->reg_base = devm_platform_ioremap_resource_byname(pdev, "reg");
 	if (IS_ERR(pcie->reg_base)) {
 		dev_err(dev, "missing \"reg\"\n");
diff --git a/drivers/pci/controller/cadence/pcie-cadence.h b/drivers/pci/controller/cadence/pcie-cadence.h
index 574e9cf4d003..042a4c49bb9a 100644
--- a/drivers/pci/controller/cadence/pcie-cadence.h
+++ b/drivers/pci/controller/cadence/pcie-cadence.h
@@ -86,6 +86,7 @@ struct cdns_plat_pcie_of_data {
  * @ops: Platform-specific ops to control various inputs from Cadence PCIe
  *       wrapper
  * @cdns_pcie_reg_offsets: Register bank offsets for different SoC
+ * @max_link_speed: Maximum supported link speed
  */
 struct cdns_pcie {
 	void __iomem		             *reg_base;
@@ -98,6 +99,7 @@ struct cdns_pcie {
 	struct device_link	             **link;
 	const  struct cdns_pcie_ops          *ops;
 	const  struct cdns_plat_pcie_of_data *cdns_pcie_reg_offsets;
+	int				     max_link_speed;
 };
 
 /**
-- 
2.34.1

[PATCH v3 3/7] PCI: cadence: HPA: Add post-link delay

[Hans Zhang]

The Cadence HPA (High Performance Architecture IP) specific link setup
function cdns_pcie_hpa_host_link_setup() waits for the link to come up
but does not implement the required 100 ms delay after link training
completes for speeds > 5.0 GT/s (PCIe r6.0 sec 6.6.1).

Add a call to pci_host_common_link_train_delay() immediately after the
link is confirmed to be up, using the max_link_speed field. Also, in the
HPA host setup function, read the device tree property "max-link-speed"
to initialize max_link_speed if not already set by a glue driver.

This ensures compliance for HPA-based platforms.

Signed-off-by: Hans Zhang <18255117159@163.com>
---
 drivers/pci/controller/cadence/pcie-cadence-host-hpa.c | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/drivers/pci/controller/cadence/pcie-cadence-host-hpa.c b/drivers/pci/controller/cadence/pcie-cadence-host-hpa.c
index 0f540bed58e8..8ef58ed01daa 100644
--- a/drivers/pci/controller/cadence/pcie-cadence-host-hpa.c
+++ b/drivers/pci/controller/cadence/pcie-cadence-host-hpa.c
@@ -15,6 +15,8 @@
 
 #include "pcie-cadence.h"
 #include "pcie-cadence-host-common.h"
+#include "../pci-host-common.h"
+#include "../../pci.h"
 
 static u8 bar_aperture_mask[] = {
 	[RP_BAR0] = 0x3F,
@@ -304,6 +306,8 @@ int cdns_pcie_hpa_host_link_setup(struct cdns_pcie_rc *rc)
 	ret = cdns_pcie_host_wait_for_link(pcie, cdns_pcie_hpa_link_up);
 	if (ret)
 		dev_dbg(dev, "PCIe link never came up\n");
+	else
+		pci_host_common_link_train_delay(pcie->max_link_speed);
 
 	return ret;
 }
@@ -313,6 +317,7 @@ int cdns_pcie_hpa_host_setup(struct cdns_pcie_rc *rc)
 {
 	struct device *dev = rc->pcie.dev;
 	struct platform_device *pdev = to_platform_device(dev);
+	struct device_node *np = dev->of_node;
 	struct pci_host_bridge *bridge;
 	enum   cdns_pcie_rp_bar bar;
 	struct cdns_pcie *pcie;
@@ -343,6 +348,9 @@ int cdns_pcie_hpa_host_setup(struct cdns_pcie_rc *rc)
 		rc->cfg_res = res;
 	}
 
+	if (pcie->max_link_speed < 1)
+		pcie->max_link_speed = of_pci_get_max_link_speed(np);
+
 	/* Put EROM Bar aperture to 0 */
 	cdns_pcie_hpa_writel(pcie, REG_BANK_IP_CFG_CTRL_REG, CDNS_PCIE_EROM, 0x0);
 
-- 
2.34.1

[PATCH v3 4/7] PCI: dwc: Use common pci_host_common_link_train_delay() helper

[Hans Zhang]

The DWC driver already implements the 100 ms delay required by PCIe
r6.0 sec 6.6.1 by checking pci->max_link_speed and calling msleep(100).

Replace the open-coded msleep() with the new common helper
pci_host_common_link_train_delay() to reduce code duplication and
improve maintainability. No functional change intended.

Signed-off-by: Hans Zhang <18255117159@163.com>
---
 drivers/pci/controller/dwc/pcie-designware.c | 9 ++-------
 1 file changed, 2 insertions(+), 7 deletions(-)

diff --git a/drivers/pci/controller/dwc/pcie-designware.c b/drivers/pci/controller/dwc/pcie-designware.c
index c11cf61b8319..7021d21bb601 100644
--- a/drivers/pci/controller/dwc/pcie-designware.c
+++ b/drivers/pci/controller/dwc/pcie-designware.c
@@ -22,6 +22,7 @@
 #include <linux/sizes.h>
 #include <linux/types.h>
 
+#include "../pci-host-common.h"
 #include "../../pci.h"
 #include "pcie-designware.h"
 
@@ -799,13 +800,7 @@ int dw_pcie_wait_for_link(struct dw_pcie *pci)
 		return -ETIMEDOUT;
 	}
 
-	/*
-	 * As per PCIe r6.0, sec 6.6.1, a Downstream Port that supports Link
-	 * speeds greater than 5.0 GT/s, software must wait a minimum of 100 ms
-	 * after Link training completes before sending a Configuration Request.
-	 */
-	if (pci->max_link_speed > 2)
-		msleep(PCIE_RESET_CONFIG_WAIT_MS);
+	pci_host_common_link_train_delay(pci->max_link_speed);
 
 	offset = dw_pcie_find_capability(pci, PCI_CAP_ID_EXP);
 	val = dw_pcie_readw_dbi(pci, offset + PCI_EXP_LNKSTA);
-- 
2.34.1

Re: [PATCH v3 4/7] PCI: dwc: Use common pci_host_common_link_train_delay() helper

[Krzysztof Wilczyński]

Hello,

> -	/*
> -	 * As per PCIe r6.0, sec 6.6.1, a Downstream Port that supports Link
> -	 * speeds greater than 5.0 GT/s, software must wait a minimum of 100 ms
> -	 * after Link training completes before sending a Configuration Request.
> -	 */
> -	if (pci->max_link_speed > 2)
> -		msleep(PCIE_RESET_CONFIG_WAIT_MS);
> +	pci_host_common_link_train_delay(pci->max_link_speed);

This comment could move to the helper you added.

Thank you!

	Krzysztof

[PATCH v3 5/7] PCI: aardvark: Add 100 ms delay after link training

[Hans Zhang]

The Aardvark PCIe controller driver waits for the link to come up but
does not implement the mandatory 100 ms delay after link training
completes for speeds greater than 5.0 GT/s (PCIe r6.0 sec 6.6.1).

The driver already maintains a 'link_gen' field that holds the negotiated
link speed. Use it together with pci_host_common_link_train_delay() to
insert the required delay immediately after confirming that the link
is up.

Signed-off-by: Hans Zhang <18255117159@163.com>
---
 drivers/pci/controller/pci-aardvark.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/drivers/pci/controller/pci-aardvark.c b/drivers/pci/controller/pci-aardvark.c
index e34bea1ff0ac..fd9c7d53e8a7 100644
--- a/drivers/pci/controller/pci-aardvark.c
+++ b/drivers/pci/controller/pci-aardvark.c
@@ -26,6 +26,7 @@
 #include <linux/of_address.h>
 #include <linux/of_pci.h>
 
+#include "pci-host-common.h"
 #include "../pci.h"
 #include "../pci-bridge-emul.h"
 
@@ -350,8 +351,10 @@ static int advk_pcie_wait_for_link(struct advk_pcie *pcie)
 
 	/* check if the link is up or not */
 	for (retries = 0; retries < LINK_WAIT_MAX_RETRIES; retries++) {
-		if (advk_pcie_link_up(pcie))
+		if (advk_pcie_link_up(pcie)) {
+			pci_host_common_link_train_delay(pcie->link_gen);
 			return 0;
+		}
 
 		usleep_range(LINK_WAIT_USLEEP_MIN, LINK_WAIT_USLEEP_MAX);
 	}
-- 
2.34.1

[PATCH v3 6/7] PCI: mediatek-gen3: Add 100 ms delay after link up

[Hans Zhang]

The MediaTek Gen3 PCIe host driver lacks the required 100 ms delay after
link training completes for speeds > 5.0 GT/s, as specified in PCIe r6.0
sec 6.6.1.

The driver already stores max_link_speed (from the device tree). After
mtk_pcie_startup_port() successfully brings up the link, call
pci_host_common_link_train_delay() to comply with the specification.

Signed-off-by: Hans Zhang <18255117159@163.com>
---
 drivers/pci/controller/pcie-mediatek-gen3.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/drivers/pci/controller/pcie-mediatek-gen3.c b/drivers/pci/controller/pcie-mediatek-gen3.c
index b0accd828589..5abddec4e9be 100644
--- a/drivers/pci/controller/pcie-mediatek-gen3.c
+++ b/drivers/pci/controller/pcie-mediatek-gen3.c
@@ -30,6 +30,7 @@
 #include <linux/regmap.h>
 #include <linux/reset.h>
 
+#include "pci-host-common.h"
 #include "../pci.h"
 
 #define PCIE_BASE_CFG_REG		0x14
@@ -570,6 +571,8 @@ static int mtk_pcie_startup_port(struct mtk_gen3_pcie *pcie)
 		goto err_power_down_device;
 	}
 
+	pci_host_common_link_train_delay(pcie->max_link_speed);
+
 	return 0;
 
 err_power_down_device:
-- 
2.34.1

[PATCH v3 7/7] PCI: rzg3s-host: Use common pci_host_common_link_train_delay() helper

[Hans Zhang]

Replace the unconditional msleep(100) with the common helper
pci_host_common_link_train_delay(). The helper only waits when
max_link_speed > 2, as required by PCIe r6.0 sec 6.6.1.

This avoids unnecessary delay for Gen1/Gen2 links while retaining
the mandatory 100 ms for higher speeds.

Signed-off-by: Hans Zhang <18255117159@163.com>
---
 drivers/pci/controller/pcie-rzg3s-host.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/drivers/pci/controller/pcie-rzg3s-host.c b/drivers/pci/controller/pcie-rzg3s-host.c
index d86e7516dcc2..66f687304c1c 100644
--- a/drivers/pci/controller/pcie-rzg3s-host.c
+++ b/drivers/pci/controller/pcie-rzg3s-host.c
@@ -35,6 +35,7 @@
 #include <linux/slab.h>
 #include <linux/units.h>
 
+#include "pci-host-common.h"
 #include "../pci.h"
 
 /* AXI registers */
@@ -1663,7 +1664,7 @@ rzg3s_pcie_host_setup(struct rzg3s_pcie_host *host,
 	if (ret)
 		dev_info(dev, "Failed to set max link speed\n");
 
-	msleep(PCIE_RESET_CONFIG_WAIT_MS);
+	pci_host_common_link_train_delay(host->max_link_speed);
 
 	return 0;
 
-- 
2.34.1