EMMC HS400es support on RK3399 (and probably all rockchip boards) CQE error "fix"

prahal · December 16, 2023

The mainline 5.11 patch "phy: rockchip: set pulldown for strobe line in dts" https://github.com/torvalds/linux/commit/8b5c2b45b8f0a11c9072da0f7baf9ee986d3151e (not backported to 5.10.y) seems to have broken most if not all rockchip board EMMC HS400 enhanced strobe support.

This probably affects supported rockchip boards (at least rk3399 ones, but probably most).

That is boards exhibit:

[   18.985162] mmc1: running CQE recovery
[   18.988056] ------------[ cut here ]------------
[   18.988500] mmc1: cqhci: spurious TCN for tag 12

and the filesystem ends up corrupted on write attempt.

Note that another bug about regulator core DEFERRED support (which might have produced the same issue hardware wise) was introduced in 5.10.43 (I had bisected it to https://github.com/torvalds/linux/commit/98e48cd9283dbac0e1445ee780889f10b3d1db6a "regulator: core: resolve supply for boot-on/always-on regulators").

But I was confident that it to be fixed in at least 6.1 by https://github.com/torvalds/linux/commit/8a866d527ac0441c0eb14a991fa11358b476b11d "regulator: core: Resolve supply name earlier to prevent double-init" (introduced in 6.1), still EMMC was still failing on me. Thanks to @RussianNeuroMancer telling me that not all rk339 boards had EMMC HS400es broken, I found that nanopc-t4 had https://github.com/torvalds/linux/commit/463be3cb357dab7d7e4d8dcc7c15c642e10c5bef arm64: dts: rockchip: add enable-strobe-pulldown to emmc phy on nanopi4

So the current way, from this nanopc-t4 commit, to fix EMMC HS400 on most rockchip is to add "rockchip,enable-strobe-pulldown;" to the "emmc_phy" node (at least this node alias for rk3399).

&emmc_phy {
+       rockchip,enable-strobe-pulldown;
        status = "okay";
 };

With this patch I can renable hs400es for rk3399 emmc on helios64 (it is already set for nanopc-t4 in mainline).

Details in:

I believe the https://github.com/torvalds/linux/commit/8b5c2b45b8f0a11c9072da0f7baf9ee986d3151e mainline commit was wrong in that it probably should have done the opposite (that is enable the pulldown) as most boards were hardwired so.

As was done for rk3588 boards:

"arm64: dts: rockchip: Fix eMMC Data Strobe PD on rk3588" https://github.com/torvalds/linux/commit/37f3d6108730713c411827ab4af764909f4dfc78 "

JEDEC standard JESD84-B51 defines the eMMC Data Strobe line, which is
currently used only in HS400 mode, as a device->host clock signal that
"is used only in read operation. The Data Strobe is always High-Z (not
driven by the device and pulled down by RDS) or Driven Low in write
operation, except during CRC status response." RDS is a pull-down
resistor specified in the 10K-100K ohm range. Thus per the standard, the
Data Strobe is always pulled to ground (by the eMMC and/or RDS) during
write operations.

Evidently, the eMMC host controller in the RK3588 considers an active
voltage on the eMMC-DS line during a write to be an error.

The default (i.e. hardware reset, and Rockchip BSP) behavior for the
RK3588 is to activate the eMMC-DS pin's builtin pull-down. As a result,
many RK3588 board designers do not bother adding a dedicated RDS
resistor, instead relying on the RK3588's internal bias. The current
devicetree, however, disables this bias (`pcfg_pull_none`), breaking
HS400-mode writes for boards without a dedicated RDS, but with an eMMC
chip that chooses to High-Z (instead of drive-low) the eMMC-DS line.
(The Turing RK1 is one such board.)

Fix this by changing the bias in the (common) emmc_data_strobe case to
reflect the expected hardware/BSP behavior. This is unlikely to cause
regressions elsewhere: the pull-down is only relevant for High-Z eMMCs,
and if this is redundant with a (dedicated) RDS resistor, the effective
result is only a lower resistance to ground -- where the range of
tolerance is quite high. If it does, it's better fixed in the specific
devicetrees.

Maybe one can confirm this is the case not only for rk5588 but for other rockchip boards? (about the default for hardware reset and rockchip BSP with regards to active that eMMC-DS pin's builtin pulldown if any, and board designers for other boards than rk5588 also not bothering to add a dedicated RDS resistor, instead relying on the rockchip internal bias, also if any on non rk5588)

At least two other boards disabled hs400es in mainline probbaly due to this patch disabling the internal pulldown by default "phy: rockchip: set pulldown for strobe line in dts" https://github.com/torvalds/linux/commit/8b5c2b45b8f0a11c9072da0f7baf9ee986d3151e

in 2023 in vanilla Linux:

Rock 4C+ https://github.com/torvalds/linux/commit/2bd1d2dd808c60532283e9cf05110bf1bf2f9079

Rock Pi 4 https://github.com/torvalds/linux/commit/cee572756aa2cb46e959e9797ad4b730b78a050b