[erikwb]

Every time the sound card binds again, each TAS2764 adds another handler for itself on its fault IRQ. This frees the IRQ when the component is removed.

What happens

tas2764_codec_probe() requests the fault IRQ with devm_request_threaded_irq() on the I2C device. The I2C device stays bound when the card is torn down, so the handler outlives the component. The next bind probes the component again and requests the IRQ again, and tas2764_codec_remove() never frees it.

On the J713 the four amps share one active-low fault line. After two rebinds of the speaker card (for example rmmod snd_soc_macaudio; modprobe snd_soc_macaudio, or reloading snd-soc-aop with #759), /proc/interrupts listed twelve tas2764 handlers on it. Each fault interrupt runs every one of them, and each reads six latch registers over I2C, so faults are handled later with every rebind. That includes the over-temperature and over-current group shutdown.

The change

• Free the IRQ in the remove: tas2764_codec_remove() frees it with devm_free_irq() when the probe got it. That's tracked, because an amp outside an SDZ group carries on without the IRQ if the request fails.
• Request it last: the request moves to the end of the probe, after the steps that can fail, so a probe that fails after it can't leave a handler behind either. The interrupt masks are still written right after the reset, as before.

Testing (J713)

	Handlers on the fault line
Without this commit, after two card rebinds	12
With it: fresh module	4
With it: after two macaudio rebinds and an snd-soc-aop reload	4

• The IRQ still works: after the speaker stream closed, the amps reported the stopped bus clock 19 times and each report was handled. Nothing was logged about unhandled interrupts or a disabled line.
• Checks:
  • checkpatch --strict is clean;
  • the module builds with W=1 on integration/aurora with no warnings.