Make capture work + add full-duplex (record+playback) on modern kernels

[ljadach]

Summary

This makes capture/recording work on the Akai EIE Pro - long reported broken (#1, #3, #7) - and enables simultaneous playback + capture (full duplex) on current kernels. Tested on Linux Mint 22.3 / kernel 6.17.0-29-generic: recorded a 4-channel overdub while playing back, no crashes.

The playback path is left unchanged; all changes are on the capture side plus one shared-reset serialization.

Background

Capture never produced data: arecord blocked forever and wrote a 44-byte (header-only) file. Investigating revealed a chain of issues, each hidden behind the previous one.

Root causes & fixes

1. Capture URBs were never submitted. init_cap_urbs() allocates them, but nothing ever called usb_submit_urb() on them - there was submit_init_sync_urbs() / submit_init_play_urbs() but no capture equivalent, so no completion ever fired. Added submit_init_cap_urbs() and call it from the capture trigger START (by then reset_eie() from prepare has the sync + playback streams flowing, so the device is clocked and the bulk-in endpoint is live). Resubmission in cap_urb_complete() is now gated on CAPTURE_RUNNING so the URBs drain cleanly on stop.

2. Missing URB_NO_TRANSFER_DMA_MAP on capture URBs. They use usb_alloc_coherent buffers but didn't set the flag (the playback/sync URBs do). On kernels with the vmalloc-DMA-map hardening the core then tries to DMA-map the buffer and the submit fails with -EAGAIN (xhci_hcd: rejecting DMA map of vmalloc memory). Set the flag.

3. Capture buffer overflow -> kernel panic. cap_urb_complete() writes BYTES_PER_FRAME_CAP (16, i.e. 4x __le32) per frame, but capture shared the playback hw advertising S24_3LE = 12 bytes/frame, so ALSA sized the runtime buffer at 12. Once the URBs actually flowed, the 16-byte writes overran the buffer -> memory corruption -> panic. Capture now advertises its own SNDRV_PCM_FMTBIT_S32_LE (16 bytes/frame) in eie_cpcm_open().

4. Full-duplex reset race. Opening playback and capture near-simultaneously ran reset_eie() from both *_prepare callbacks concurrently, re-submitting already-active URBs (-EBUSY, "URB submitted while active"). Serialized the if (rate != eie->rate) reset_eie() decision under devices_mutex in both prepares, so the second stream sees the device already streaming and skips the redundant reset.

5. Capture decoded to noise. The 24-bit sample was written low-aligned and unsigned, so S32 -> S16 downconversion kept only ~8 noisy bits. Sign-align the value into S32_LE (ch << 8).

Testing

• Linux Mint 22.3 (Ubuntu 24.04 base), kernel 6.17.0-29-generic, Akai EIE Pro 09e8:0010.
• Builds clean (needs the existing >6.12 vmalloc handling already in master).
• Capture-only: records 4-channel audio.
• Full duplex: played a clip out through the EIE while recording a live guitar on input 1 - clean overdub on playback, dmesg clean, no crash.

Known remaining work (honest)

• cap_urb_complete() still updates cap_buf_pos / cap_frames without the spinlock (the existing TODO); fine in light testing, should be hardened for heavy DAW load.
• Only tested at 48 kHz so far.
• Capture can reach full scale; trim the EIE input gain for headroom (or reduce the shift).

This was debugged with a lot of dmesg-driven iteration; the commit message lists the same chain. Happy to split into smaller commits if preferred.