[PATCH v2 00/11] soundwire: amd: add ACP7B/7E/7F SoundWire manager support
From: Vijendar Mukunda
Date: Tue Nov 17 2026 - 03:06:24 EST
This is v2 of the AMD ACP7B/7E/7F SoundWire manager support series.
It incorporates review feedback from v1 and rebases on the updated fix
and refactor series. These revisions represent a significant hardware
evolution compared to the ACP7.0/7.1/7.2 generation:
- Up to four SoundWire manager instances (ACP_SDW0 through ACP_SDW3),
compared to two in earlier generations.
- Banked registers on the manager side for clock frequency and frame
shape parameters, enabling seamless transitions without bus
re-enumeration.
- Dynamic clock scaling with both fixed frame shape and dynamic frame
shape selection, enabling efficient bandwidth use across a wider
range of bus clock frequencies.
- Extended sample rate and bit-depth support (48/96/192 kHz,
S16/S24/S32) and a broader set of supported bus clock frequencies
compared to ACP7.0/7.1/7.2.
- Up to 22 DAIs per manager with 11 TX and 11 RX data ports and
8 TX and 8 RX DMA channels, compared to fewer ports and DMAs
in earlier generations.
- SSP synchronisation tag propagation per bus transfer, required by
the ACP7B/7E/7F hardware and absent in older revisions.
- Shared per-platform wake enable control across all manager instances
with software serialisation to prevent races between concurrent
managers during clock stop.
- Separate wake and bus-event IRQ paths, giving cleaner interrupt
handling compared to earlier generations.
- Per-revision IO pad pin assignments with hardware-required restore
order on D0 resume after power-off (ACP7B, ACP7E, and ACP7F each
have distinct physical SoundWire clock and data pin mappings).
- Custom bandwidth allocation with dynamic frame-shape selection
across candidate clock frequencies.
The new backend amd_acp7x_manager.c plugs into the hw_ops dispatch
framework introduced by the refactor series. It handles the full
driver lifecycle: IRQ handling and slave status processing, port
configuration, DAI PCM capability registration, host wake
enable/disable, clock-stop suspend and resume, and IO pad management.
Multi-manager synchronisation (GSYNC) is planned for a future series
once the required design work is complete; the existing implementation
is structured to accommodate it incrementally.
Two patches to generic_bandwidth_allocation.c are included: one
fixes the hwidth computation for the skipped block-offset slot
(a correctness fix relevant to any user of the generic algorithm),
and one exports the bandwidth helper functions needed by the new
ACP7x compute_params callback.
The interrupt_enabled guard used by ACP6.3/7.0/7.1/7.2 to prevent
new work from being queued during teardown is mirrored in the ACP7x
enable/disable and IRQ thread paths. The corresponding ASoC-side
patches that guard interrupt_enabled checks in the ACP7B/7E/7F IRQ
handler are not included in this series and will be posted separately.
Changes since v1:
- Rebase on AMD SoundWire fix series v6 (23 patches) and refactor
series v4 (11 patches) in place of fix v5 (25 patches) and refactor
v3 (11 patches); all context lines and base commit SHAs updated
throughout.
- Patch 2: treat IMM status cleanup timeout as fatal: replace
dev_warn() (non-fatal, residual bit comment) with dev_err() and
return ret.
- Patch 4: add missing return -EINVAL after dev_err() for unexpected
acp_rev in register_dais(); use local dev variable in the error
message consistently.
- Patch 6: extend amd_sdw_clock_stop() wake_en_mask dispatch to cover
the full set of supported revisions (ACP63/70/71/72/7B/7E/7F) with
a default returning -EINVAL; required because refactor v4 simplified
the existing code to a plain unconditional write, so this patch must
rebuild the complete dispatch table.
- Patch 7: revise pad_fully_torn_down comment to be concise and
accurate for the ACP7x context; rephrase interrupt_enabled comment
to describe what the guard prevents more precisely.
- Patch 11: use WRITE_ONCE()/READ_ONCE() for all interrupt_enabled
stores and loads; move the interrupt_enabled guard to the top of
amd_acp7x_sdw_irq_thread() (before register reads) so the full
IRQ thread body is gated, not just the schedule_work() call.
- Link to v1:
https://lore.kernel.org/linux-sound/20261108213520.3570680-1-Vijendar.Mukunda@xxxxxxx/
Dependencies:
This series applies on top of the AMD SoundWire fix series (v6,
23 patches) and the AMD SoundWire refactor series (v4, 11 patches):
https://lore.kernel.org/linux-sound/20261117070906.803139-1-Vijendar.Mukunda@xxxxxxx
https://lore.kernel.org/linux-sound/20261117071927.830373-1-Vijendar.Mukunda@xxxxxxx
Vijendar Mukunda (11):
soundwire: amd: add ACP7B/7E/7F register definitions and revision IDs
soundwire: amd: add ACP7B/7E/7F hardware backend
soundwire: amd: extend ctl_word_prep to set SSP_TAG for ACP7B/7E/7F
soundwire: amd: add ACP7B/7E/7F DAI PCM capabilities in
register_dais()
soundwire: amd: enable host wake support for ACP7B/7E/7F
soundwire: amd: add ACP7B/7E/7F clock-stop support
soundwire: amd: add IO pad enable/disable for ACP7B/7E/7F D0/D3
sequence
soundwire: generic_bandwidth_allocation: fix hwidth for skipped block
offset
soundwire: generic_bandwidth_allocation: export bandwidth helper
functions
soundwire: amd: add ACP7x generic bandwidth allocation support
soundwire: amd: add interrupt_enabled synchronization for ACP7B/7E/7F
drivers/soundwire/Makefile | 2 +-
drivers/soundwire/amd_acp7x_manager.c | 1162 +++++++++++++++++
drivers/soundwire/amd_init.c | 365 +++++-
drivers/soundwire/amd_init.h | 13 +
drivers/soundwire/amd_manager.h | 297 +++++
drivers/soundwire/amd_manager_common.c | 146 ++-
drivers/soundwire/bus.h | 10 +
.../soundwire/generic_bandwidth_allocation.c | 129 +-
include/linux/soundwire/sdw_amd.h | 17 +
9 files changed, 2124 insertions(+), 17 deletions(-)
create mode 100644 drivers/soundwire/amd_acp7x_manager.c
--
2.53.0