soundwire: qcom: enumerate in software when hardware auto-enum fails

[nandamajay]

soundwire: qcom: enumerate in software when hardware auto-enum fails

Problem

On some QCM6490 cold boots the SoundWire sound card fails to come up
("Initialization not complete", -ETIMEDOUT). Root cause is a SoundWire
enumeration failure: the v2.0.0 controller's hardware auto-enumerator does a
one-shot DevID read at ping address 0 that breaks mid-transfer (first byte
latches, the rest return zero). The HW enumeration table is left empty, so
qcom_swrm_enumerate() assigns no device number, the WCD/WSA codecs never
initialise, and no card appears.

The controller signals exactly this via SWRM_INTERRUPT_STATUS_AUTO_ENUM_FAILED
(BIT(11)) — but that interrupt has had no case in qcom_swrm_irq_handler()
since it was introduced in commit 312355a ("soundwire: qcom: add support
for v2.0.0 controller"); it falls through to the "unknown interrupt" default and
nothing recovers.

Fix

The peripheral is awake throughout — a command-FIFO read of SCP_DEVID_0..5
(qcom_swrm_cmd_fifo_rd_cmd(), which retries and flushes the FIFO between tries)
returns the real DevID where the hardware's single-shot read did not. So on
AUTO_ENUM_FAILED, enumerate the peripheral in software over the command
FIFO:

• qcom_swrm_match_slave_at_zero() reads the DevID with the retrying command-FIFO
  primitive and matches it against the peripherals this bus declares.
• qcom_swrm_sw_program_device() picks a free number with find_first_zero_bit()
  over bus->assigned (same as the core), writes SCP_DEVNUMBER over the command
  FIFO and records the assignment the same way the hardware path does.
• The device number is assigned in the driver rather than delegating the
  address-0 program step to the generic sdw_program_device_num(), whose single
  raw re-read of address 0 would underflow on a flaky boot and mint a phantom
  slave ("Slave Entry not found") that steals the real codec's number.
• A bounded settle-retry (SWRM_SW_ENUM_ATTEMPTS × SWRM_SW_ENUM_SETTLE_MS)
  covers the few-hundred-ms a cold-boot peripheral can take to answer; a clean
  read programs on the first attempt and never sleeps. The whole fallback runs
  from the threaded IRQ handler and is bounded per bring-up by
  SWRM_SW_ENUM_RETRIES.

Also adds the previously-missing SWRM_INTERRUPT_STATUS_AUTO_ENUM_TABLE_IS_FULL
case.

Fixes: 312355a ("soundwire: qcom: add support for v2.0.0 controller")

Related issue

Fixes #1251 — the same cold-boot no-soundcard /
enumeration failure.

Validation

The fix was validated on QCM6490-IDP over repeated cold boots: card comes up on
every boot with a clean 4-device slave list (both WSA883x + both WCD937x,
all Attached and bound), no phantom node and no "Slave Entry not found",
with the capture-path (TX) WCD enumerated as well.

Note on this branch: validation was performed on a linux-next-based image.
This PR targets main (7.3-rc6), which does not carry the SCP_ADDRPAGE
shadow-cache fields; the one functional difference versus the validated tree is
that this variant omits the page-cache invalidation (a no-op on main, whose
hardware set_slave_dev_num() path also lacks it). The enumeration logic is
identical. Builds clean for arm64; checkpatch --strict reports 0/0/0.