spi: uclass: skip cs lookup if bus not probed

[jiez]

spi_find_bus_and_cs() uses uclass_find_device_by_seq() which locates the bus without probing it. Calling spi_find_chip_select() on an unprobed bus causes drivers that validate chip selects using private data (initialized only at probe time) to spuriously reject valid CS numbers. Since no SPI device can exist on an unprobed bus, return -ENODEV early instead.

U-Boot SPL 2025.10 (May 04 2026 - 12:40:02 -0400)
ADI Boot Mode: 0x1 (QSPI Master)
Trying to boot from BOOTROM


U-Boot 2025.10 (May 04 2026 - 12:40:02 -0400)

CPU:   ADSP ADSP-SC598-0.0 (QSPI Master boot)
Model: ADI SC598-SOM-EZKIT
DRAM:  224 MiB
ADP558x Detected: Rev 4, Rev ID 4
Core:  139 devices, 23 uclasses, devicetree: embed
MMC:   mmc@310C7000: 0
Loading Environment from nowhere... OK
In:    serial@0x31003000
Out:   serial@0x31003000
Err:   serial@0x31003000
Net:   eth0: eth0
=> sf probe
adi_spi3 spi2: invalid chipselect 1
adi_spi3 spi2: Invalid cs 1 (err=-22)
SF: Detected is25lp01g with page size 256 Bytes, erase size 4 KiB, total 128 MiB
=>

[CalebEthridgeADI]

Tested on a Rev E sc598.

[ozan956]

spi_find_bus_and_cs() is documented as a lookup helper that finds the bus and slave without activating either side, but this change now returns -ENODEV because the bus is not active yet.

Also the commit message is misleading. A bound SPI child device can already exist on an unprobed bus: the SPI uclass scans child nodes at bind time via .post_bind = dm_scan_fdt_dev, and child SPI plat is populated in spi_child_post_bind(). So this patch can hide a valid, already-bound slave device simply because the parent bus is still inactive.

It also does not fix the underlying issue generally. The real problem here is that spi_find_chip_select() may call ->cs_info() on an unprobed bus, but this patch only papers over one caller, spi_find_bus_and_cs(). Other paths such as spi_cs_info() / spi_cs_is_valid() still go through the same unprobed-bus flow.

If the goal is to address cs_info() relying on probe-time private data, I would expect the fix either in spi_find_chip_select() or in the driver-side cs_info() handling, not as a behavioral change in spi_find_bus_and_cs().

[jiez]

spi_find_bus_and_cs() is documented as a lookup helper that finds the bus and slave without activating either side, but this change now returns -ENODEV because the bus is not active yet.

Also the commit message is kinda misleading. A bound SPI child device can already exist on an unprobed bus. The SPI uclass scans child nodes at bind time via .post_bind = dm_scan_fdt_dev, and child SPI plat is populated in spi_child_post_bind(). So this patch can hide a valid, already-bound slave device simply because the parent bus is still inactive.

It also does not fix the underlying issue generally. The real problem here is that spi_find_chip_select() may call ->cs_info() on an unprobed bus, but this patch only papers over one caller, spi_find_bus_and_cs(). Other paths such as spi_cs_info() / spi_cs_is_valid() still go through the same unprobed-bus flow.

If the goal is to address cs_info() relying on probe-time private data, I would expect the fix either in spi_find_chip_select() or in the driver-side cs_info() handling, not as a behavioral change in spi_find_bus_and_cs().

Thanks for your review!

I agree with your comments. I pushed a new commit, which moves cs validation from spi_find_chip_select() to spi_child_pre_probe(), i.e. move cs validation to probe phase. spi_find_chip_select() now does not validate cs.

[pamolloy]

I haven't had a chance to read this in detail, but it looks like there are still issues. See review-inline.txt from Claude Sonnet 4.6 including Ozan's comment. Given that review and looking at other implementations in U-Boot and Linux Claude suggested the following:

  For adi_spi3.c:
  static int adi_spi_cs_info(struct udevice *bus, uint cs,
                             struct spi_cs_info *info)
  {
        // Hardware supports CS 1-7 (1-indexed)
        // This matches the default in DT: dev_read_u32_default(bus, "num-cs", 7)
        if (cs == 0 || cs > MAX_CTRL_CS) {
                return -EINVAL;
        }

        return 0;
  }