Repository navigation
Conversation
There was a problem hiding this comment.
🟡 Changes recommended
The topology changes can propagate incorrect formats across fanout connections, inflate pipeline resource accounting during repeated preparation, and accept formats incompatible with the bound input pin.
3 open issues
What changed in this PR
This PR makes SOF IPC4 format selection and propagation pin-aware, separating module configuration from the format passed along each connection.
Changes:
- Passes output-pin identity through widget preparation and isolates parameters across branching paths.
- Selects and propagates process and copier formats per pin, including secondary-pin handling.
- Validates complete pin bindings, checks ambiguous formats, and fixes process-format cleanup.
| File | Description |
|---|---|
sound/soc/sof/topology.c |
Rejects partially defined pin-binding arrays. |
sound/soc/sof/sof-audio.h |
Adds output-pin identifiers and extends the preparation interface. |
sound/soc/sof/sof-audio.c |
Prepares each connection with pin identity and independent branch parameters. |
sound/soc/sof/ipc4-topology.h |
Clarifies that parameter-change detection concerns pin 0. |
sound/soc/sof/ipc4-topology.c |
Implements pin-aware format handling, validation, and error-path cleanup. |
Review effort: Balanced
kv2019i
left a comment
There was a problem hiding this comment.
Complicated series, but commit messages look good and explain the changes. There is unfortunately lot of special logic around pin0, but that is fundamental of the firmware interface and kernel can do nothing about it. Minor comment(s) inline.
| } | ||
|
|
||
| copier_data->gtw_cfg.node_id &= ~SOF_IPC4_NODE_INDEX_MASK; | ||
| copier_data->gtw_cfg.node_id |= SOF_IPC4_NODE_INDEX(host_dma_id); |
There was a problem hiding this comment.
Not sure how to explain this simpler, but I think this commit has complexities:
""Select the copier output format among the pin 0 formats only and modify
the params for the next widget with the format of the output pin feeding
the path. The format of a pin above 0 is selected from its formats in
topology""
This seems selfcontracticting. First it is said selecting format amont pin 0 formats only, but then in next sentence, format selection is done based on other factors. Was the intention that for pin 0, only formats in pin 0 are considered?
There was a problem hiding this comment.
Yes, exactly, the pin 0 restriction applies to the copier’s initialization format. Downstream parameter propagation separately uses the format of the pin feeding that connection. I’ll clarify it.
aa84a33 to
ba430ff
Compare
|
Changes since v1:
|
There was a problem hiding this comment.
🟡 Changes recommended
The secondary copier output lookup can reject a bind-compatible mono connection by requiring the output to match the four-channel input, rather than the connected sink’s constraints.
1 open issue
3 resolved since last review
Review effort: Balanced
ba430ff to
9fe415d
Compare
|
Changes since v2:
|
9fe415d to
8d27499
Compare
|
Changes since v3:
|
8d27499 to
0752819
Compare
An SRC or ASRC can be visited again when paths reconverge or a second PCM uses a shared downstream path. Its prepared fast path currently returns without updating the pipeline parameters, leaving downstream widgets with the incoming rate instead of the configured output rate. Derive the prepared output format from the configured base format and the stored SRC sink rate or ASRC output frequency. Propagate it without reconfiguring the converter or accounting its resources again. Preserve the format, channel and rate update mask used by initial preparation. Implemented with GitHub Copilot assistance. Link: thesofproject#6030 (comment)
|
Changes since v4:
|
An SRC or ASRC can be visited again when paths reconverge or a second PCM uses a shared downstream path. Its prepared fast path currently returns without updating the pipeline parameters, leaving downstream widgets with the incoming rate instead of the configured output rate. Derive the prepared output format from the configured base format and the stored SRC sink rate or ASRC output frequency. Propagate it without reconfiguring the converter or accounting its resources again. Preserve the format, channel and rate update mask used by initial preparation. Implemented with GitHub Copilot assistance. Link: thesofproject#6030 (comment)
There was a problem hiding this comment.
🟡 Changes recommended
Two critical issues remain: valid secondary-output format conversion is rejected, and ALH allocation failure leaves an incorrect configuration length that can cause an out-of-bounds read on retry.
3 open issues
2 resolved since last review
Review effort: Balanced
sof_ipc4_widget_setup_comp_process() allocates the input and output pin
format arrays via sof_ipc4_get_audio_fmt(). When the following
sof_ipc4_widget_setup_msg() fails, e.g. because the firmware does not
provide a module matching the widget UUID, the error path jumps to the
'err' label, which frees the process struct and clears swidget->private
but skips sof_ipc4_free_audio_fmt(). The arrays are then unreachable and
are reported by kmemleak:
sof_ipc4_get_audio_fmt+0xc9/0x420 [snd_sof]
sof_ipc4_widget_setup_comp_process+0x8e/0x200 [snd_sof]
snd_sof_load_topology+0x71a/0x840 [snd_sof]
sof_pcm_probe+0x77/0x120 [snd_sof]
snd_soc_component_probe+0x21/0x60 [snd_soc_core]
...
Jump to 'free_available_fmt' instead, as the other process widget error
paths after sof_ipc4_get_audio_fmt() already do.
The other IPC4 widget setup functions (pga, mixer, src, asrc) free the
audio formats in their common error label and are not affected.
Fixes: f9efae9 ("ASoC: SOF: ipc4-topology: Add support for base config extension")
Cc: [email protected]
Signed-off-by: Baoli Zhang <[email protected]>
SRC/ASRC can only change rate, there is no reason to re-set the format and channels to the same - as they are not changed by the module. Signed-off-by: Peter Ujfalusi <[email protected]>
An SRC or ASRC can be visited again when paths reconverge or a second PCM uses a shared downstream path. Its prepared fast path currently returns without updating the pipeline parameters, leaving downstream widgets with the incoming rate instead of the configured output rate. Derive the prepared output format from the configured base format and the stored SRC sink rate or ASRC output frequency. Propagate it without reconfiguring the converter or accounting its resources again. Preserve the format, channel and rate update mask used by initial preparation. Implemented with GitHub Copilot assistance. Link: thesofproject#6030 (comment)
When widget preparation fails, the error rollback clears the prepared flag only if the widget has an ipc_unprepare callback. A module that needs no cleanup callback therefore remains marked prepared even though its pipeline was never committed. Later preparation attempts can fail with stale format state. Clear the prepared flag for unused widgets during rollback regardless of whether they need an ipc_unprepare callback. Keep the use_count check to preserve widgets shared by active paths. Implemented and build-tested with GitHub Copilot assistance. Fixes: 6f2a43e ("ASoC: SOF: sof-audio: Skip unprepare for in-use widgets on error rollback") Cc: [email protected] Signed-off-by: Peter Ujfalusi <[email protected]>
A mixout shared by several playback paths is configured by the first path prepared. Further paths skip mixer preparation without checking whether their input format matches the configured mixer. With dynamic formats, a second frontend can consequently feed a 96 kHz mixin into a mixout already running at 48 kHz. The stream is accepted without rate conversion, leading to incorrect playback and delay accounting. Validate the propagated pipeline rate, channel count, valid bit depth and sample type against the configured mixer format before reusing a prepared mixer. Reject mismatches with -EINVAL without changing the active configuration. Compare the format at the mixer rather than the original frontend format so host copier format conversion remains valid. Implemented and build-tested with GitHub Copilot assistance. Fixes: 94648dc ("ASoC: SOF: ipc4-topology: Update the pipeline_params of prepared modules") Cc: [email protected] Signed-off-by: Peter Ujfalusi <[email protected]>
The pin binding array is only stored if the binding of pin 0 is defined in topology. The bindings of the other pins are assumed to be present and they are compared with strcmp() when looking for the pin a widget is connected to, which results in a NULL pointer dereference if the topology does not define the binding for all pins. If the binding of pin 0 is not defined, the parsed bindings of the other pins are leaked. The pin binding must be defined for all pins, with "NotConnected" for the unused pins as documented in struct snd_sof_widget. Reject the topology if the binding is only defined for some of the pins. Fixes: 3b3aced ("ASoC: SOF: Add support to parse pin binding array from topology") Signed-off-by: Peter Ujfalusi <[email protected]>
Move the output format selection of process modules to sof_ipc4_process_select_output_fmt(). Pass the selected input format, or NULL for modules without input formats. Keep the prepared-widget shortcut and pin 0-only parameter propagation unchanged. This prepares for separating module configuration from the output format propagated along each path. No functional change. Signed-off-by: Peter Ujfalusi <[email protected]>
…repare The ipc_prepare op of a widget is called once, before its sinks are walked, without information about which output pin the path continues on. A widget with multiple output pins, each with its own format, cannot tell which pin format to use as the input params of the next widget. The pin cannot be looked up once for the whole PCM walk either: a widget can feed several sinks of the same PCM graph via different pins, like the KPB in the Multi-WOV topology, with its two outputs reconverging at the WOV arbiter. Call ipc_prepare once per outgoing connection in the PCM widget list and pass the output pin feeding that sink: - 0 for widgets with a single output pin, - the pin bound to the sink in topology for widgets with multiple pins, - SOF_OUTPUT_PIN_UNKNOWN if the widget has no output pin bindings, - SOF_OUTPUT_PIN_NONE for widgets without a downstream connection in the PCM, which still receive one call. A connection from a widget without output pins or to a sink missing from the bindings is rejected. Virtual widgets (input, output, out_drv) are not connected in the firmware, the walk does not continue to them and they are not counted as connections. The first call initializes the widget and marks it prepared, subsequent calls only propagate the format of the given pin. When the widget fans out to multiple sinks each branch is walked with its own copy of the incoming params, so that the params propagated along one branch do not leak into the others. The source is unprepared if any of its connections fails. The IPC4 prepare callbacks are updated for the new parameter, the output pin is not used yet. Signed-off-by: Peter Ujfalusi <[email protected]>
The output format lookup selects among the formats of all output pins of a module. Modules with multiple output pins need to select the format among the formats of a given pin. Add a pin parameter to sof_ipc4_init_output_audio_fmt() and sof_ipc4_is_single_format() to only consider the formats of that pin. SOF_OUTPUT_PIN_ANY considers the formats of all pins as before. Also allow selecting an output format without updating the obs of a base config, for pins which are not configured via the base config. Factor out the format comparison to sof_ipc4_is_same_fmt() and move the pin format helpers next to sof_ipc4_fmt_cfg_to_type(), so that they can also be used while the topology is parsed. All callers pass SOF_OUTPUT_PIN_ANY, no functional change. Signed-off-by: Peter Ujfalusi <[email protected]>
… single format The formats of a pin are treated as a single format if they have the same rate, channel count and valid bit depth. In that case the first format is used without matching the formats against the params. Formats which only differ in their sample type, like 32-bit integer and 32-bit float, or unsigned 8-bit and A-law/mu-law, are not the same format: the first one is used even if the params ask for the other one. Compare the sample type as well, the same way as the formats are matched against the params. Treat signed integer sample types equivalently when 32 valid bits fill a 32-bit container. Signed-off-by: Peter Ujfalusi <[email protected]>
…mats The formats of the pins of a process module are fixed when the module is initialized: pin 0 via the base config, the other pins via the base config extension or by the module itself. Unlike the copier, which can change the format of its output pins with SET_SINK_FORMAT, there is no message to change them afterwards. On a side with multiple pins the module can be initialized via a path using any of the pins. The format of the other pins can only be known if they have a single format, there are no params to select one of their formats from. Warn at topology load if a pin on a side with multiple pins has more than one distinct format. Signed-off-by: Peter Ujfalusi <[email protected]>
…ocess modules A process module with multiple output pins can be shared by paths of several PCMs. Selecting among all output formats with the FE params guesses the pin used by the path. The prepared-widget shortcut always propagates process->output_format, which describes pin 0 and may not have been set if the first path selected another pin. The module is configured with the output pin 0 format and buffer size in the base config. Other pins use their topology format via the base config extension or are configured by the module itself. Unlike a copier, a process module cannot change pin formats after initialization. Use the output pin feeding the path to separate module configuration from downstream propagation. Select the base config format only among pin 0 formats. If the path leaves on another pin, pin 0 must have a single format since the path params cannot select it. Propagate the configured format of the pin feeding the path on first and subsequent preparation. Pins without topology formats use the pin 0 format. Reject ambiguous secondary pin formats, including when building the base config extension. Modules with input formats are configured once, the pin formats are not selected again with the FE params of a subsequent path. Without output pin bindings the output pin of a sink is only allocated at bind time and the pin 0 format is propagated to it. Accept this only if all output pins are configured with the pin 0 format: pins without formats in topology use it, the others must have it as their single format. When there is no downstream connection, configure pin 0 without propagating an output format. Zero-output consumers skip output format selection entirely. Signed-off-by: Peter Ujfalusi <[email protected]>
…n > 0 The input format of a module is looked up among its pin 0 input formats with the params of the source widget in the path being prepared. This assumes that the path enters the module on input pin 0. A module with multiple input pins can be reached first via another pin, depending on the order of the widgets in the widget list. For example the ECNS module receives 16kHz/32bit/4ch from dmic16k on pin 0 and 48kHz/32bit/4ch from dmic01 on pin 1. If the path from dmic01 is walked first the 48kHz params are looked up among the pin 0 formats and the prepare fails, even though they are valid for pin 1. It only works today because the dmic16k path happens to be walked first. The base config of the module is always set from the pin 0 input format. If the params do not match any pin 0 format but match the format of another input pin, the path enters the module on that pin: use the pin 0 format for the base config. This is only possible if pin 0 has a single format, otherwise the format to configure is unknown: modules with multiple pin 0 formats must be prepared first via the path entering on pin 0. Fail the prepare with an explicit error if this is not the case. The fallback is only taken where the lookup used to fail, the selection for paths entering on pin 0 is not changed. Signed-off-by: Peter Ujfalusi <[email protected]>
…nput process modules The pipeline params are updated with the output format of a process module based on the changed_params mask, which is evaluated between the input and output formats of the module. When the path enters a multi-input module on an input pin other than pin 0 the pipeline params carry the format of that pin while the output format is derived from the pin 0 format of the module. The params not flagged as changed would be propagated from the wrong pin. This only worked because the changed_params is evaluated over the formats of all pins and the format of the other input pin usually differs from the output format. Update all params with the output format for modules with multiple input pins. Signed-off-by: Peter Ujfalusi <[email protected]>
…s only The changed_params mask is evaluated by comparing all input formats with all output formats of a module regardless of the pin they belong to. For modules with multiple pins this results in params flagged as changed which are not changed by the module on its data path. For example the google-rtc-aec has the 2 channel echo reference on input pin 1 and with 4 channel microphones it has a 4 channel format on input pin 0 and on the output, but the channels are flagged as changed. Only compare the formats of input pin 0 and output pin 0. The params of a path leaving a multi-output module on other pins or entering a multi-input module are updated with the full output format, they are not affected by the mask. Signed-off-by: Peter Ujfalusi <[email protected]>
The copier is initialized with the format of output pin 0, but its output format lookup currently searches formats from all output pins. A format declared for another pin can therefore be used for pin 0. Restrict the lookup to pin 0. For module copiers, select at the input rate: use a fixed channel count or sample representation when the pin has only one, otherwise preserve the corresponding input parameter. Require the resulting combination to be declared. Keep the endpoint-specific selection for gateway copiers. Signed-off-by: Peter Ujfalusi <[email protected]>
…ailure The DMA configuration TLV size is temporarily added to the gateway configuration length before allocating the IPC payload. If allocation fails, the original length is not restored. On retry, the TLV size is counted again and the copier configuration blob can be read beyond its allocation. Defer updating the gateway configuration length until the IPC payload is ready to be copied, after the fallible preparation steps. Keep the existing restoration after the payload is assembled. Fixes: a0659f8 ("ASoC: SOF: ipc4-topology: add DMA config TLV to IPC data") Signed-off-by: Peter Ujfalusi <[email protected]>
Copier preparation always propagates the output pin 0 format to the next widget, even when the path is fed by another output pin. Use the connected output pin to select and propagate its format. When that pin has no declared formats, or its identity is not known until bind, continue propagating the pin 0 format. Skip propagation when there is no downstream connection. Move ALH group allocation after the parameter propagation, so that a failure during preparation does not leave an allocated group behind. Signed-off-by: Peter Ujfalusi <[email protected]>
The sink determines the format configured for a copier output pin above pin 0. Check that the sink format is declared for that pin when the topology provides a format list. Also check that it matches the output format propagated during prepare. This prevents the sink from being prepared with one format while the copier is configured with a different one at bind time. Signed-off-by: Peter Ujfalusi <[email protected]>
0752819 to
3317909
Compare
|
Changes since v5:
|
There was a problem hiding this comment.
🔵 Needs a closer look
Firmware-facing pin-format propagation across branching and reconvergent pipelines needs SOF maintainer validation of end-to-end capture, as Multi-WOV reaches RUNNING but frame delivery remains unverified.
No open issues 🎉
3 resolved since last review
Review effort: Balanced


Separate module format configuration from the format propagated along each connection. Previously, format lookup could consider formats belonging to unrelated pins, while already-prepared modules propagated the pin 0 format regardless of the output feeding the path.
This series:
Passes the connection’s output pin to widget preparation.
Preserves independent parameters across branching paths, including Multi-WOV branches that reconverge.
Configures process modules using pin 0 and propagates the configured format of the actual output pin.
Handles copier secondary outputs separately, preserving bind-time format configuration.
Supports paths entering modules through secondary input pins.
Restricts parameter-change detection to pin 0 formats and includes sample type in format-equivalence checks.
Validates complete pin-binding arrays and warns about ambiguous process-module pin formats.
Also includes the upstream process-widget pin-format leak fix.
Validation
Focused W=1 builds and strict checkpatch checks.
DUT: ECNS capture and silent speaker/deepbuffer playback passed.
Multi-WOV PCM12 prepares and reaches RUNNING without kernel warnings/errors; frame delivery remains unverified because capture stayed in the listening state.