sdhci_brcmstb_init_2712() reads host->mmc->host_caps to decide whether
to force card-detect for a non-removable eMMC, or to route the CD signal
for a removable SD card. At the time this function runs from
sdhci_bcmstb_probe(), however, host->mmc->host_caps is still zero, that
field is only populated later by the MMC uclass, after the driver's
probe returns. mmc_of_parse() has already filled plat->cfg.host_caps
from the device tree by this point, so check that field instead.
Without the fix, every BCM2712 SDHCI instance takes the else branch and
writes SDIO_CFG_SD_PIN_SEL = SDIO_CFG_SD_PIN_SEL_CARD (0x02), including
the non-removable eMMC on boards such as CM5 on Home Assistant Yellow.
The SDIO_CFG block lies outside the SDHCI core's reset scope, so this
value persists across SDHCI_RESET_ALL into the next stage. On the
BCM2712, having SD_PIN_SEL set to "SD" when the Linux kernel performs
its first set_power(MMC_POWER_UP) write racily prevents the SDHCI
POWER_ON bit from latching (see [1] for the whole backstory) - the
voltage bits stick but POWER_ON drops - which wedges the first CMD0 the
full 10 s software timeout. On Home Assistant Yellow this manifested as
a ~20 s eMMC probe delay on roughly one in two Linux boots when U-Boot
was the previous stage. Booting directly from the Pi firmware (no U-Boot
in between) left SD_PIN_SEL at its default and did not exhibit the race.
Reading plat->cfg.host_caps lets init_2712 see the "non-removable"
property and take the correct branch, leaving SD_PIN_SEL untouched for
the eMMC.
[1] https://github.com/home-assistant/operating-system/pull/3700#issuecomment-
4430229511
Fixes: 10127cdbab64 ("mmc: bcmstb: Add support for bcm2712 SD controller")
Signed-off-by: Jan Čermák <sairon@sairon.cz>
Reviewed-by: Ivan T. Ivanov <iivanov@suse.de>