Probing devices by their less-specific "compatible" bindings (here: brcmnand)

[Rafał Miłecki]

Hi, I just spent few hours debugging hidden hw lockup and I need to
consult driver core code behaviour.

I have a BCM4908 SoC based board with a NAND controller on it.


### Hardware binding

Hardware details:
arch/arm64/boot/dts/broadcom/bcmbca/bcm4908.dtsi

Relevant part:
nand-controller at 1800 {
	compatible = "brcm,nand-bcm63138", "brcm,brcmnand-v7.1", "brcm,brcmnand";
	reg = <0x1800 0x600>, <0x2000 0x10>;
	reg-names = "nand", "nand-int-base";
}:

Above binding is based on the documentation:
Documentation/devicetree/bindings/mtd/brcm,brcmnand.yaml


### Linux drivers

Linux has separated drivers for few Broadcom's NAND controller bindings:

1. drivers/mtd/nand/raw/brcmnand/bcm63138_nand.c for:
brcm,nand-bcm63138

2. drivers/mtd/nand/raw/brcmnand/brcmnand.c for:
brcm,brcmnand-v2.1
brcm,brcmnand-v2.2
brcm,brcmnand-v4.0
brcm,brcmnand-v5.0
brcm,brcmnand-v6.0
brcm,brcmnand-v6.1
brcm,brcmnand-v6.2
brcm,brcmnand-v7.0
brcm,brcmnand-v7.1
brcm,brcmnand-v7.2
brcm,brcmnand-v7.3

3. drivers/mtd/nand/raw/brcmnand/brcmstb_nand.c for:
brcm,brcmnand


### Problem

As first Linux probes my hardware using the "brcm,nand-bcm63138"
compatibility string driver bcm63138_nand.c. That's good.

It that fails however (.probe() returns an error) then Linux core starts
probing using drivers for less specific bindings.

In my case probing with the "brcm,brcmnand" string driver brcmstb_nand.c
results in ignoring SoC specific bits and causes a hardware lockup. Hw
isn't initialized properly and writel_relaxed(0x00000009, base + 0x04)
just make it hang.

That obviously isn't an acceptable behavior for me. So I'm wondering
what's going on wrong here.

Should Linux avoid probing with less-specific compatible strings?
Or should I not claim hw to be "brcm,brcmnand" compatible if it REQUIRES
SoC-specific handling?

An extra note: that fallback probing happens even with .probe()
returning -EPROBE_DEFER. This actually smells fishy for me on the Linux
core part.
I'm not an expect but I think core should wait for actual error without
trying less-specific compatible strings & drivers.