[PATCH 0/4] amlogic: t7: give the MMC bus pipeline clock real consumers

[PATCH 0/4] amlogic: t7: give the MMC bus pipeline clock real consumers

[Lucas Tanure]

The previous series [2] for the intermittent boot hangs on the VIM4
marked sys_ampipe_nand as critical. But as Chuan Liu from Amlogic
explained [1], sys_ampipe_nand clocks a pipeline stage that the
hardware design inserts in the bus path between the SD/eMMC
controllers and the NIC_MATRIX fabric to help timing closure.

So the clock has real consumers: reference it from the three MMC
controller nodes and drop CLK_IS_CRITICAL instead.

The series crosses three trees; patch 4 (clk) must only land after
patches 1-3, otherwise the DMA hangs fixed by 43e1705ecab9 come back.

[1] https://lore.kernel.org/linux-amlogic/ (RFC: clk: meson: t7:
    Intermittent boot instability and memory corruption on VIM4)
[2] https://lore.kernel.org/linux-amlogic/ (arm64: amlogic: t7:
    describe the VIM4 supplies and keep the fabric clocks running)

Thanks,
Lucas

Lucas Tanure (4):
  dt-bindings: mmc: amlogic,meson-gx-mmc: document the T7 pipeline clock
  mmc: meson-gx: enable the bus pipeline clock on T7
  arm64: dts: amlogic: t7: add the pipeline clock to the MMC controllers
  clk: meson: t7: don't mark sys_ampipe_nand as critical

 .../bindings/mmc/amlogic,meson-gx-mmc.yaml    | 29 ++++++++++++++++++-
 arch/arm64/boot/dts/amlogic/amlogic-t7.dtsi   | 15 ++++++----
 drivers/clk/meson/t7-peripherals.c            | 10 +++----
 drivers/mmc/host/meson-gx-mmc.c               | 21 ++++++++++++++
 4 files changed, 63 insertions(+), 12 deletions(-)

-- 
2.55.0

[PATCH 1/4] dt-bindings: mmc: amlogic,meson-gx-mmc: document the T7 pipeline clock

[Lucas Tanure]

The SD/eMMC controllers of the T7 SoC are physically far from the
NIC_MATRIX bus fabric, so the hardware design inserts a pipeline stage
in the middle of the bus path to help timing closure. This stage is fed
by its own gate clock and, when that clock is disabled, a controller
that starts a DMA transfer can never complete it.

Allow a fourth clock entry, named "pipeline", and require it for the
amlogic,t7-mmc compatible. The other controllers of this family keep
the existing three clocks.

Assisted-by: Claude Fable 5 <noreply@anthropic.com>
Signed-off-by: Lucas Tanure <tanure@linux.com>
---
 .../bindings/mmc/amlogic,meson-gx-mmc.yaml    | 29 ++++++++++++++++++-
 1 file changed, 28 insertions(+), 1 deletion(-)

diff --git a/Documentation/devicetree/bindings/mmc/amlogic,meson-gx-mmc.yaml b/Documentation/devicetree/bindings/mmc/amlogic,meson-gx-mmc.yaml
index 976f36de2091..12342b1d78e5 100644
--- a/Documentation/devicetree/bindings/mmc/amlogic,meson-gx-mmc.yaml
+++ b/Documentation/devicetree/bindings/mmc/amlogic,meson-gx-mmc.yaml
@@ -15,6 +15,23 @@ maintainers:
 
 allOf:
   - $ref: mmc-controller.yaml#
+  - if:
+      properties:
+        compatible:
+          contains:
+            const: amlogic,t7-mmc
+    then:
+      properties:
+        clocks:
+          minItems: 4
+        clock-names:
+          minItems: 4
+    else:
+      properties:
+        clocks:
+          maxItems: 3
+        clock-names:
+          maxItems: 3
 
 properties:
   compatible:
@@ -38,13 +55,23 @@ properties:
       - description: card detect
 
   clocks:
-    maxItems: 3
+    minItems: 3
+    items:
+      - description: core clock
+      - description: clock input 0 of the internal mux
+      - description: clock input 1 of the internal mux
+      - description:
+          clock of the pipeline stage inserted in the bus path between
+          the controller and the DRAM. Without it, the controller cannot
+          complete DMA transfers.
 
   clock-names:
+    minItems: 3
     items:
       - const: core
       - const: clkin0
       - const: clkin1
+      - const: pipeline
 
   resets:
     maxItems: 1
-- 
2.55.0

[PATCH 2/4] mmc: meson-gx: enable the bus pipeline clock on T7

[Lucas Tanure]

On the T7 SoC, the bus path between the SD/eMMC controllers and the
NIC_MATRIX fabric goes through a pipeline stage inserted by the hardware
design to help timing closure. The stage has its own gate clock and,
when that clock is disabled, a controller that starts a DMA transfer can
never complete it, hanging the storage devices and, from there, the
whole system.

Add a dedicated match data for the amlogic,t7-mmc compatible that makes
the driver claim and enable the "pipeline" clock for as long as the
device is bound.

The clock is deliberately not optional: the hardware cannot do DMA
without it, and failing the probe with a clear error is preferable to
booting and hitting an undiagnosable DMA hang later.

Assisted-by: Claude Fable 5 <noreply@anthropic.com>
Signed-off-by: Lucas Tanure <tanure@linux.com>
---
 drivers/mmc/host/meson-gx-mmc.c | 21 +++++++++++++++++++++
 1 file changed, 21 insertions(+)

diff --git a/drivers/mmc/host/meson-gx-mmc.c b/drivers/mmc/host/meson-gx-mmc.c
index 694bb443d5f3..68566d7eccc7 100644
--- a/drivers/mmc/host/meson-gx-mmc.c
+++ b/drivers/mmc/host/meson-gx-mmc.c
@@ -52,6 +52,7 @@
 #define   CLK_RX_DELAY_MASK(h)		(h->data->rx_delay_mask)
 #define   CLK_ALWAYS_ON(h)		(h->data->always_on)
 #define   CLK_IRQ_SDIO_SLEEP(h)		(h->data->irq_sdio_sleep)
+#define   CLK_PIPELINE(h)		(h->data->has_pipeline_clk)
 
 #define SD_EMMC_DELAY 0x4
 #define SD_EMMC_ADJUST 0x8
@@ -139,6 +140,7 @@ struct meson_mmc_data {
 	unsigned int always_on;
 	unsigned int adjust;
 	unsigned int irq_sdio_sleep;
+	bool has_pipeline_clk;
 };
 
 struct sd_emmc_desc {
@@ -1204,6 +1206,15 @@ static int meson_mmc_probe(struct platform_device *pdev)
 	if (IS_ERR(core_clk))
 		return PTR_ERR(core_clk);
 
+	if (CLK_PIPELINE(host)) {
+		struct clk *pipe_clk;
+
+		pipe_clk = devm_clk_get_enabled(&pdev->dev, "pipeline");
+		if (IS_ERR(pipe_clk))
+                      return dev_err_probe(&pdev->dev, PTR_ERR(pipe_clk),
+					   "missing pipeline clock\n");
+	}
+
 	ret = meson_mmc_clk_init(host);
 	if (ret)
 		return ret;
@@ -1322,12 +1333,22 @@ static const struct meson_mmc_data meson_axg_data = {
 	.irq_sdio_sleep	= CLK_V3_IRQ_SDIO_SLEEP,
 };
 
+static const struct meson_mmc_data meson_t7_data = {
+	.tx_delay_mask	= CLK_V3_TX_DELAY_MASK,
+	.rx_delay_mask	= CLK_V3_RX_DELAY_MASK,
+	.always_on	= CLK_V3_ALWAYS_ON,
+	.adjust		= SD_EMMC_V3_ADJUST,
+	.irq_sdio_sleep	= CLK_V3_IRQ_SDIO_SLEEP,
+	.has_pipeline_clk = true,
+};
+
 static const struct of_device_id meson_mmc_of_match[] = {
 	{ .compatible = "amlogic,meson-gx-mmc",		.data = &meson_gx_data },
 	{ .compatible = "amlogic,meson-gxbb-mmc", 	.data = &meson_gx_data },
 	{ .compatible = "amlogic,meson-gxl-mmc",	.data = &meson_gx_data },
 	{ .compatible = "amlogic,meson-gxm-mmc",	.data = &meson_gx_data },
 	{ .compatible = "amlogic,meson-axg-mmc",	.data = &meson_axg_data },
+	{ .compatible = "amlogic,t7-mmc",		.data = &meson_t7_data },
 	{}
 };
 MODULE_DEVICE_TABLE(of, meson_mmc_of_match);
-- 
2.55.0

[PATCH 3/4] arm64: dts: amlogic: t7: add the pipeline clock to the MMC controllers

[Lucas Tanure]

The SD/eMMC controllers reach the DRAM through a pipeline stage that
the hardware design inserts in the bus path to the NIC_MATRIX fabric,
clocked by CLKID_SYS_AMPIPE_NAND. When that clock is disabled, a
controller that starts a DMA transfer can never complete it.

Reference the clock from the three MMC controller nodes so the driver
keeps it running, instead of relying on the clock being marked critical
in the clock controller.

Assisted-by: Claude Fable 5 <noreply@anthropic.com>
Signed-off-by: Lucas Tanure <tanure@linux.com>
---
 arch/arm64/boot/dts/amlogic/amlogic-t7.dtsi | 15 +++++++++------
 1 file changed, 9 insertions(+), 6 deletions(-)

diff --git a/arch/arm64/boot/dts/amlogic/amlogic-t7.dtsi b/arch/arm64/boot/dts/amlogic/amlogic-t7.dtsi
index 719e111bc3dd..8f3f1c0579a0 100644
--- a/arch/arm64/boot/dts/amlogic/amlogic-t7.dtsi
+++ b/arch/arm64/boot/dts/amlogic/amlogic-t7.dtsi
@@ -762,8 +762,9 @@ sd_emmc_a: mmc@88000 {
 				interrupts = <GIC_SPI 176 IRQ_TYPE_LEVEL_HIGH>;
 				clocks = <&clkc_periphs CLKID_SYS_SD_EMMC_A>,
 					 <&clkc_periphs CLKID_SD_EMMC_A>,
-					 <&scmi_clk CLKID_FCLK_DIV2>;
-				clock-names = "core", "clkin0", "clkin1";
+					 <&scmi_clk CLKID_FCLK_DIV2>,
+					 <&clkc_periphs CLKID_SYS_AMPIPE_NAND>;
+				clock-names = "core", "clkin0", "clkin1", "pipeline";
 				resets = <&reset RESET_SD_EMMC_A>;
 				assigned-clocks = <&clkc_periphs CLKID_SD_EMMC_A_SEL>;
 				assigned-clock-parents = <&xtal>;
@@ -776,8 +777,9 @@ sd_emmc_b: mmc@8a000 {
 				interrupts = <GIC_SPI 177 IRQ_TYPE_EDGE_RISING>;
 				clocks = <&clkc_periphs CLKID_SYS_SD_EMMC_B>,
 					 <&clkc_periphs CLKID_SD_EMMC_B>,
-					 <&scmi_clk CLKID_FCLK_DIV2>;
-				clock-names = "core", "clkin0", "clkin1";
+					 <&scmi_clk CLKID_FCLK_DIV2>,
+					 <&clkc_periphs CLKID_SYS_AMPIPE_NAND>;
+				clock-names = "core", "clkin0", "clkin1", "pipeline";
 				resets = <&reset RESET_SD_EMMC_B>;
 				assigned-clocks = <&clkc_periphs CLKID_SD_EMMC_B_SEL>;
 				assigned-clock-parents = <&xtal>;
@@ -790,8 +792,9 @@ sd_emmc_c: mmc@8c000 {
 				interrupts = <GIC_SPI 178 IRQ_TYPE_EDGE_RISING>;
 				clocks = <&clkc_periphs CLKID_SYS_SD_EMMC_C>,
 					 <&clkc_periphs CLKID_SD_EMMC_C>,
-					 <&scmi_clk CLKID_FCLK_DIV2>;
-				clock-names = "core", "clkin0", "clkin1";
+					 <&scmi_clk CLKID_FCLK_DIV2>,
+					 <&clkc_periphs CLKID_SYS_AMPIPE_NAND>;
+				clock-names = "core", "clkin0", "clkin1", "pipeline";
 				resets = <&reset RESET_SD_EMMC_C>;
 				assigned-clocks = <&clkc_periphs CLKID_SD_EMMC_C_SEL>;
 				assigned-clock-parents = <&xtal>;
-- 
2.55.0

[PATCH 4/4] clk: meson: t7: don't mark sys_ampipe_nand as critical

[Lucas Tanure]

sys_ampipe_nand clocks the pipeline stage inserted in the bus path
between the SD/eMMC controllers and the NIC_MATRIX fabric. It does have
identifiable consumers - the three MMC controllers - so marking it
critical was the wrong tool: the clock should be referenced from the
consumer nodes and claimed by the mmc driver instead.

Drop the CLK_IS_CRITICAL flag and update the comment accordingly.
sys_am2axi0..2 remain critical as they clock the AXI DMA bus itself and
have no single identifiable consumer.

This depends on the MMC controllers actually referencing the clock:
the dt-bindings, mmc driver and t7 DTS changes adding the "pipeline"
clock must land before this one, otherwise the boot hangs and memory
corruption this clock caused are reintroduced.

Fixes: 43e1705ecab9 ("clk: meson: t7: keep the memory fabric clocks running")
Assisted-by: Claude Fable 5 <noreply@anthropic.com>
Signed-off-by: Lucas Tanure <tanure@linux.com>
---
 drivers/clk/meson/t7-peripherals.c | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)

diff --git a/drivers/clk/meson/t7-peripherals.c b/drivers/clk/meson/t7-peripherals.c
index 55d09e78593b..2f429ea86796 100644
--- a/drivers/clk/meson/t7-peripherals.c
+++ b/drivers/clk/meson/t7-peripherals.c
@@ -945,13 +945,13 @@ static T7_SYS_PCLK(sys_aucpu,		SYS_CLK_EN0_REG0, 14,	0);
 static T7_SYS_PCLK(sys_cec,		SYS_CLK_EN0_REG0, 16,	0);
 static T7_SYS_PCLK(sys_gdc,		SYS_CLK_EN0_REG0, 17,	0);
 static T7_SYS_PCLK(sys_deswarp,		SYS_CLK_EN0_REG0, 18,	0);
+static T7_SYS_PCLK(sys_ampipe_nand,	SYS_CLK_EN0_REG0, 19,	0);
+static T7_SYS_PCLK(sys_ampipe_eth,	SYS_CLK_EN0_REG0, 20,	0);
 /*
- * NOTE: sys_ampipe_nand and sys_am2axi0..2 provide the clock to the AXI bus
- * used for DMA between the peripherals and the DRAM. After the clocks are
- * disabled, a device that starts a transfer cannot complete it.
+ * NOTE: sys_am2axi0..2 provide the clock to the AXI bus used for DMA between
+ * the peripherals and the DRAM. After the clocks are disabled, a device that
+ * starts a transfer cannot complete it.
  */
-static T7_SYS_PCLK(sys_ampipe_nand,	SYS_CLK_EN0_REG0, 19,	CLK_IS_CRITICAL);
-static T7_SYS_PCLK(sys_ampipe_eth,	SYS_CLK_EN0_REG0, 20,	0);
 static T7_SYS_PCLK(sys_am2axi0,		SYS_CLK_EN0_REG0, 21,	CLK_IS_CRITICAL);
 static T7_SYS_PCLK(sys_am2axi1,		SYS_CLK_EN0_REG0, 22,	CLK_IS_CRITICAL);
 static T7_SYS_PCLK(sys_am2axi2,		SYS_CLK_EN0_REG0, 23,	CLK_IS_CRITICAL);
-- 
2.55.0

Re: [PATCH 4/4] clk: meson: t7: don't mark sys_ampipe_nand as critical

[Jerome Brunet]

On jeu. 03 sept. 2026 at 23:59, Lucas Tanure <tanure@linux.com> wrote:

> sys_ampipe_nand clocks the pipeline stage inserted in the bus path
> between the SD/eMMC controllers and the NIC_MATRIX fabric. It does have
> identifiable consumers - the three MMC controllers - so marking it
> critical was the wrong tool: the clock should be referenced from the
> consumer nodes and claimed by the mmc driver instead.
>
> Drop the CLK_IS_CRITICAL flag and update the comment accordingly.
> sys_am2axi0..2 remain critical as they clock the AXI DMA bus itself and
> have no single identifiable consumer.
>
> This depends on the MMC controllers actually referencing the clock:
> the dt-bindings, mmc driver and t7 DTS changes adding the "pipeline"
> clock must land before this one, otherwise the boot hangs and memory
> corruption this clock caused are reintroduced.

This information typically goes bellow the '---' because we will not
keep it for posterity. Please resend when your dependencies have been satisfied. 

>
> Fixes: 43e1705ecab9 ("clk: meson: t7: keep the memory fabric clocks running")

Thinking of it, we probably do not want stable to pick this up since
your other change are improvements that will no be backported. You can
drop the Fixes

> Assisted-by: Claude Fable 5 <noreply@anthropic.com>

The Assisted-by should not have an email, especially a no reply one.
check https://docs.kernel.org/process/coding-assistants.html

> Signed-off-by: Lucas Tanure <tanure@linux.com>
> ---
>  drivers/clk/meson/t7-peripherals.c | 10 +++++-----
>  1 file changed, 5 insertions(+), 5 deletions(-)
>
> diff --git a/drivers/clk/meson/t7-peripherals.c b/drivers/clk/meson/t7-peripherals.c
> index 55d09e78593b..2f429ea86796 100644
> --- a/drivers/clk/meson/t7-peripherals.c
> +++ b/drivers/clk/meson/t7-peripherals.c
> @@ -945,13 +945,13 @@ static T7_SYS_PCLK(sys_aucpu,		SYS_CLK_EN0_REG0, 14,	0);
>  static T7_SYS_PCLK(sys_cec,		SYS_CLK_EN0_REG0, 16,	0);
>  static T7_SYS_PCLK(sys_gdc,		SYS_CLK_EN0_REG0, 17,	0);
>  static T7_SYS_PCLK(sys_deswarp,		SYS_CLK_EN0_REG0, 18,	0);
> +static T7_SYS_PCLK(sys_ampipe_nand,	SYS_CLK_EN0_REG0, 19,	0);
> +static T7_SYS_PCLK(sys_ampipe_eth,	SYS_CLK_EN0_REG0, 20,	0);
>  /*
> - * NOTE: sys_ampipe_nand and sys_am2axi0..2 provide the clock to the AXI bus
> - * used for DMA between the peripherals and the DRAM. After the clocks are
> - * disabled, a device that starts a transfer cannot complete it.
> + * NOTE: sys_am2axi0..2 provide the clock to the AXI bus used for DMA between
> + * the peripherals and the DRAM. After the clocks are disabled, a device that
> + * starts a transfer cannot complete it.
>   */
> -static T7_SYS_PCLK(sys_ampipe_nand,	SYS_CLK_EN0_REG0, 19,	CLK_IS_CRITICAL);
> -static T7_SYS_PCLK(sys_ampipe_eth,	SYS_CLK_EN0_REG0, 20,	0);
>  static T7_SYS_PCLK(sys_am2axi0,		SYS_CLK_EN0_REG0, 21,	CLK_IS_CRITICAL);
>  static T7_SYS_PCLK(sys_am2axi1,		SYS_CLK_EN0_REG0, 22,	CLK_IS_CRITICAL);
>  static T7_SYS_PCLK(sys_am2axi2,		SYS_CLK_EN0_REG0, 23,	CLK_IS_CRITICAL);
> -- 
> 2.55.0
>

-- 
Jerome