[PATCH v3] clk: mediatek: mt7981: swap MSDC and MSDC_HCK gate parents

[PATCH v3] clk: mediatek: mt7981: swap MSDC and MSDC_HCK gate parents

[Brendon C.]

The MT7981B datasheet specifies the storage interface as "eMMC v4.5 @52
MHz 3.3V". MediaTek's SDK and U-Boot device trees give the MSDC
controller CK_TOP_EMMC_208M as its "source" clock (max-frequency 52 MHz
= 208/4) and CK_TOP_EMMC_400M as its "hclk", taken from topckgen
directly, with the infracfg gate only as "source_cg". The mainline
device tree takes both from the infracfg gates instead, and their
parents are the other way round: CLK_INFRA_MSDC_CK from emmc_400m,
CLK_INFRA_MSDC_HCK_CK from emmc_208m.

mtk-sd derives its divider, the reported actual_clock and its data
timeouts from the source rate, so with a device tree asking for 26 MHz
it programs CKDIV=4 and believes the bus runs at 25 MHz while the real
bus clock is 13 MHz, and every clock-derived timeout is sized for a bus
twice as fast as the real one.

On a GL.iNet GL-X3000 (Samsung 8GTF4R eMMC, 8-bit, 3.3 V, HS) this shows
up as msdc_track_cmd_data write data errors (host->error=0x4) on
roughly every 40 MB written, escalating under sustained multi-segment
writes to exhausted retries, "I/O error, dev mmcblk0", lost page writes
and a stopped f2fs overlay. Raw direct reads top out at 11 MB/s where
the vendor kernel reaches 23 MB/s on the same card.

Swap the two parents to match the vendor device trees. With the same
device tree the bus then runs at a real 26 MHz, reads reach 23 MB/s,
and four rounds of a 1.4 GB mixed direct/buffered write stress complete
with zero retries and zero I/O errors, matching the vendor firmware.
hclk's rate is only read for the CQE idle timer, which no MT7981 device
tree enables.

Fixes: 813c3b53b55b ("clk: mediatek: add MT7981 clock support")
Cc: stable@vger.kernel.org
Suggested-by: Shiji Yang <yangshiji66@outlook.com>
Signed-off-by: Brendon C. <bren@bren.win>
---
v3:
- also move CLK_INFRA_MSDC_HCK_CK to emmc_400m, matching the vendor
  device trees' hclk (suggested by Shiji Yang)
- correct the claim that the SDK kernel parents the gate from emmc_208m:
  its gate has mainline's parents, and its device tree takes the rates
  from topckgen directly
- v2: https://lore.kernel.org/all/20261017194308.1541507-1-bren@bren.win/

v2:
- add Fixes: and Cc: stable so the fix reaches 6.12
- v1: https://lore.kernel.org/all/20261017190729.1369302-1-bren@bren.win/

 drivers/clk/mediatek/clk-mt7981-infracfg.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/drivers/clk/mediatek/clk-mt7981-infracfg.c b/drivers/clk/mediatek/clk-mt7981-infracfg.c
--- a/drivers/clk/mediatek/clk-mt7981-infracfg.c
+++ b/drivers/clk/mediatek/clk-mt7981-infracfg.c
@@ -157,8 +157,8 @@ static const struct mtk_gate infra_clks[] = {
 	GATE_INFRA1(CLK_INFRA_SPI1_HCK_CK, "infra_spi1_hck", "infra_66m_mck",
 		    14),
 	GATE_INFRA1(CLK_INFRA_FRTC_CK, "infra_frtc", "cb_rtc_32k", 15),
-	GATE_INFRA1(CLK_INFRA_MSDC_CK, "infra_msdc", "emmc_400m", 16),
-	GATE_INFRA1(CLK_INFRA_MSDC_HCK_CK, "infra_msdc_hck", "emmc_208m", 17),
+	GATE_INFRA1(CLK_INFRA_MSDC_CK, "infra_msdc", "emmc_208m", 16),
+	GATE_INFRA1(CLK_INFRA_MSDC_HCK_CK, "infra_msdc_hck", "emmc_400m", 17),
 	GATE_INFRA1(CLK_INFRA_MSDC_133M_CK, "infra_msdc_133m", "sysaxi", 18),
 	GATE_INFRA1(CLK_INFRA_MSDC_66M_CK, "infra_msdc_66m", "sysaxi", 19),
 	GATE_INFRA1(CLK_INFRA_ADC_26M_CK, "infra_adc_26m", "infra_adc_frc", 20),

base-commit: c339f4010dea8b0d754f84ace70cd1b089878ef8
-- 
2.55.0