userspace: DP: application: switch to vregion by lyakh · Pull Request #10702 · thesofproject/sof

lyakh · 2026-04-14T13:57:54Z

Switch to using vregions for private module allocations with userspace DP.

~~Currently seems to be causing a regression, therefore a draft, debugging.~~

Update: With the updated Zephyr the regression cannot be reproduced any more.

Copilot

Pull request overview

Switches userspace DP module-private allocations from a per-module k_heap to vregions, and introduces a unified sof_alloc_api handle (heap/vregion + refcount) that is threaded through buffer and module allocation paths.

Changes:

Add vregion “coherent” allocation APIs and vregion metadata query, and adjust vregion free/alignment behavior.
Replace DP module heap usage with vregion-backed allocations; propagate sof_alloc_api through module resources, buffers, and IPC buffer creation.
Update Zephyr userspace config gating to require SOF_VREGIONS for the userspace application.

Reviewed changes

Copilot reviewed 20 out of 20 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
zephyr/lib/vregion.c	Adds coherent alloc helpers, adjusts free() for uncached aliases, changes default alignment, adds vregion_mem_info().
zephyr/Kconfig	Makes `SOF_USERSPACE_APPLICATION` depend on `SOF_VREGIONS`.
src/schedule/zephyr_dp_schedule_application.c	Updates mem partitions to use cached/uncached aliases for module heap region start.
src/lib/objpool.c	Allocates objpool backing storage from vregion when available (incl. coherent).
src/ipc/ipc4/helper.c	Threads `sof_alloc_api` into IPC4 buffer creation and client refcounting.
src/ipc/ipc-helper.c	Updates `buffer_new()` signature to take `sof_alloc_api`.
src/include/sof/schedule/dp_schedule.h	Removes obsolete `dp_heap_user` struct.
src/include/sof/objpool.h	Adds optional vregion pointer to objpool head.
src/include/sof/lib/vregion.h	Moves enum definition, adds coherent alloc APIs + mem_info, adds CONFIG stubs.
src/include/sof/audio/module_adapter/module/generic.h	Replaces module resources’ heap pointer with `sof_alloc_api`.
src/include/sof/audio/component.h	Introduces `struct sof_alloc_api` (heap/vreg/client_count).
src/include/sof/audio/buffer.h	Updates buffer allocation APIs to take `sof_alloc_api`.
src/include/sof/audio/audio_buffer.h	Replaces `heap` pointer with `alloc` pointer in `sof_audio_buffer`.
src/audio/module_adapter/module_adapter_ipc4.c	Uses `alloc.heap` for module adapter IPC4 init allocations.
src/audio/module_adapter/module_adapter.c	Replaces DP per-module heap with vregion creation and vregion-backed allocations; refcounts via alloc API.
src/audio/module_adapter/module/generic.c	Routes module alloc/free paths through vregion when present; updates syscall verification checks.
src/audio/dai-zephyr.c	Switches a DMA config allocation call from rballoc() to rmalloc().
src/audio/buffers/ring_buffer.c	Uses vregion allocations for ring buffer objects/data when available; switches to alloc API.
src/audio/buffers/comp_buffer.c	Uses vregion allocations for buffer objects when available; switches to alloc API and refcount-based vregion lifetime.
src/audio/base_fw_intel.c	Makes UAOL helper(s) `__cold`, adds cold assertions, and config-guards one helper.

Comments suppressed due to low confidence (1)

src/include/sof/lib/vregion.h:11

vregion_mem_info() uses uintptr_t in its signature, but this header no longer includes a header that defines it. Add #include <stdint.h> (or equivalent) near the top so this header is self-contained and doesn’t rely on transitive includes.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Copilot · 2026-04-15T08:07:07Z

 	/* Allocate buffer memory for module */
-	void *ptr = sof_heap_alloc(res->heap, SOF_MEM_FLAG_USER | SOF_MEM_FLAG_LARGE_BUFFER,
+	void *ptr = sof_heap_alloc(res->alloc.heap, SOF_MEM_FLAG_USER | SOF_MEM_FLAG_LARGE_BUFFER,
 				   size, alignment);



mod_balloc_align() always allocates from res->alloc.heap, but when a module is backed by a vregion, alloc.heap is NULL and the allocation will fall back to the global SOF heap. Later, MOD_RES_HEAP frees go through vregion_free() when res->alloc.vreg is set, which will free a pointer that was not allocated from that vregion. Update mod_balloc_align() to allocate from the vregion when res->alloc.vreg is non-NULL (or ensure heap/vreg pairing is consistent) so allocation and free paths match.

Copilot · 2026-04-15T08:07:07Z

+	 * A 1-to-1 replacement of the original heap implementation would be to
+	 * have "lifetime size" equal to 0. But (1) this is invalid for
+	 * vregion_create() and (2) we gradually move objects, that are simple
+	 * to move to the lifetime buffer. Make it 1k for the beginning.


The comment says “Make it 1k for the beginning.” but the code uses 4096 bytes for the lifetime partition. Please either update the comment or change the value so they match (otherwise it’s unclear what the intended sizing is).

Suggested change

* to move to the lifetime buffer. Make it 1k for the beginning.

* to move to the lifetime buffer. Make it 4k for the beginning.

kv2019i

Really cool to see this move forward! A couple of comments inline, please check.

wjablon1 · 2026-04-16T09:47:18Z

+	 * vregion_create() and (2) we gradually move objects, that are simple
+	 * to move to the lifetime buffer. Make it 1k for the beginning.
+	 */
+	return vregion_create(4096, buf_size - 4096);


Cosmetics: I would just declare two constants instead of magic number subtraction:
const size_t lifetime_size = 4 * 1024
const size_t interim_size = 16 * 1024

wjablon1 · 2026-04-16T09:57:13Z


-static struct k_heap *module_adapter_dp_heap_new(const struct comp_ipc_config *config,
-						 size_t *heap_size)
+static struct vregion *module_adapter_dp_heap_new(const struct comp_ipc_config *config,


module_adapter_dp_heap_new -> module_adapter_dp_vreg_new

wjablon1 · 2026-04-16T15:10:35Z

+/**
+ * @brief like vregion_alloc_align() but allocates coherent memory
+ */
+void *vregion_alloc_coherent_align(struct vregion *vr, enum vregion_mem_type type,


Wouldn't be better to have a single function for both coherent and normal allocations? Function sof_heap_alloc already does this. You could just replace enum vregion_mem_type with enum vregion_alloc_flags or something similar.

@wjablon1 you mean using flags like all other SOF allocation functions do. I thought about that. Personally I'm not a big fan of "flags"-like argument when they most often called with those flags constant. If the caller already knows which functionality it needs, why should it use a flag and have the called function check which flag was used? The caller can call the required functionality directly. And in this case the common functionality is easily enough implemented by a separate function, so I think this is a better fit. If however we decide that we want to continue using flags, we can switch too - now or later.

I agree with you in principle provided that callers know the right function at compile time, i.e. it doesn’t have to be determined by means of repeating if-else pattern. Also, in this case the flags values are already determined for the legacy allocation method. So I guess it all depends on how long we are going to maintain both solutions.

wjablon1 · 2026-04-16T15:15:05Z

@@ -170,10 +176,17 @@ static void module_adapter_mem_free(struct processing_module *mod)
 #if CONFIG_IPC_MAJOR_4
 	sof_heap_free(mod_heap, mod->priv.cfg.input_pins);


Is this intentionally omitted?

yes, this allocation hasn't been switched to the vregion so far

jsarha

Did this a bit hastily, but I did not find anything wrong. Just one question/suggestion.

jsarha · 2026-04-17T10:39:09Z

-static struct comp_buffer *buffer_alloc_struct(struct k_heap *heap,
+static struct comp_buffer *buffer_alloc_struct(struct sof_alloc_api *alloc,
 					       void *stream_addr, size_t size,
 					       uint32_t flags, bool is_shared)


Why do we need separate is_shared flag when we have defined SOF_MEM_FLAG_COHERENT in the flags field?

@jsarha don't know, not my original implementation. Originally the is_shared argument was added in 4a03699

jsarha · 2026-04-17T11:26:01Z

-						 size_t *heap_size)
+static struct vregion *module_adapter_dp_heap_new(const struct comp_ipc_config *config,
+						  size_t *heap_size)
 {


Here there will be some conflicts with and connections to do with my #10603

lyakh · 2026-04-28T10:37:58Z

@lrudyX the single MTL failure seems unrelated?

lrudyX · 2026-04-28T10:56:45Z

@lrudyX the single MTL failure seems unrelated?

Yes, that one is unrelated

lyakh · 2026-04-28T13:27:33Z

the short-lived "DNM" was due to a suspected bug in the PR, which, however, turned out to be an unrelated regression, fixed in #10731

lyakh · 2026-04-28T15:46:25Z

This last "DNM" is more serious - just discovered, that this PR is causing a regression - it makes aplay -r 8000 -D hw:0,2 on nocodec fail. A fix to follow shortly.

uuid.h isn't needed in buffer.h. However just removing it from there breaks compilation because of multiple trace context definitions like DECLARE_TR_CTX(comp_tr, SOF_UUID(component_uuid), LOG_LEVEL_INFO); in files, not explicitly including uuid.h. To make a simple fix move the header to trace.h because that is anyway needed for all such lines. Signed-off-by: Guennadi Liakhovetski <guennadi.liakhovetski@linux.intel.com>

sof_dma.h is only used when actually building SOF. Put it under a proprocessor conditional to fix including the header in twister builds. Signed-off-by: Guennadi Liakhovetski <guennadi.liakhovetski@linux.intel.com>

fast-get has to choose which allocation method to use, it needs access to the full module allocation context. Signed-off-by: Guennadi Liakhovetski <guennadi.liakhovetski@linux.intel.com>

When userspace modules use fast-get to access smaller buffers, their copies should be allocated on module's vregion, if available, not on the global SOF heap. Signed-off-by: Guennadi Liakhovetski <guennadi.liakhovetski@linux.intel.com>

Modules, running in userspace while using the "application" DP implementation, are moved to vregion for all their private allocations. This has the advantage of not depending on build-time configured VMH buffers and of faster lifetime allocations. Signed-off-by: Guennadi Liakhovetski <guennadi.liakhovetski@linux.intel.com>

Copilot

Pull request overview

Copilot reviewed 25 out of 25 changed files in this pull request and generated 2 comments.

Comments suppressed due to low confidence (1)

src/include/sof/lib/vregion.h:152

In the #else /* CONFIG_SOF_VREGIONS */ stub implementation, vregion_put() frees vr when the refcount reaches 0 but still returns the (now freed) pointer. Callers in this PR treat a NULL return value as "destroyed" (e.g. if (!vregion_put(mod_vreg)) ...), so this causes both leaks and potential use-after-free. Make the stub return NULL when the refcount hits 0, matching the real implementation’s contract.

lyakh · 2026-04-30T07:51:15Z

 	bool walking;		/**< indicates if the buffer is being walked */

-	struct k_heap *heap;
+	struct mod_alloc_ctx *alloc;
 };


propose to address in a follow-up

lyakh · 2026-04-30T07:51:25Z

+	 * A 1-to-1 replacement of the original heap implementation would be to
+	 * have "lifetime size" equal to 0. But (1) this is invalid for
+	 * vregion_create() and (2) we gradually move objects, that are simple
+	 * to move to the lifetime buffer. Make it 1k for the beginning.


propose to address in a follow-up

lyakh mentioned this pull request Apr 15, 2026

zephyr: alloc: add 32K virtual heap bucket #10684

Merged

lyakh force-pushed the vreg branch 2 times, most recently from ab93282 to f84a925 Compare April 15, 2026 07:54

lyakh marked this pull request as ready for review April 15, 2026 07:58

Copilot AI review requested due to automatic review settings April 15, 2026 07:58

lyakh requested review from LaurentiuM1234, abonislawski, bardliao, dbaluta, iuliana-prodan, kv2019i, lbetlej, lgirdwood, marcinszkudlinski, mmaka1, pblaszko, plbossart, ranj063 and tmleman as code owners April 15, 2026 07:58

Copilot started reviewing on behalf of lyakh April 15, 2026 07:59 View session

Copilot AI reviewed Apr 15, 2026

View reviewed changes

kv2019i reviewed Apr 15, 2026

View reviewed changes

Comment thread src/audio/module_adapter/module_adapter.c Outdated

Comment thread zephyr/lib/vregion.c

wjablon1 reviewed Apr 16, 2026

View reviewed changes

jsarha approved these changes Apr 17, 2026

View reviewed changes

lyakh force-pushed the vreg branch 3 times, most recently from 3de5c1c to 536c442 Compare April 17, 2026 13:35

lyakh added DNM Do Not Merge tag and removed DNM Do Not Merge tag labels Apr 28, 2026

lyakh added the DNM Do Not Merge tag label Apr 28, 2026

lyakh marked this pull request as draft April 28, 2026 15:48

lyakh force-pushed the vreg branch from e6df28e to 3c46819 Compare April 29, 2026 09:49

lyakh marked this pull request as ready for review April 29, 2026 09:50

lyakh removed DNM Do Not Merge tag labels Apr 29, 2026

lyakh force-pushed the vreg branch 3 times, most recently from fd70d02 to 139f77c Compare April 29, 2026 14:44

lyakh added 5 commits April 29, 2026 19:03

dma: only include sof_dma.h when needed

c6c943e

sof_dma.h is only used when actually building SOF. Put it under a proprocessor conditional to fix including the header in twister builds. Signed-off-by: Guennadi Liakhovetski <guennadi.liakhovetski@linux.intel.com>

fast-get: switch to module allocation context

dd5ea94

fast-get has to choose which allocation method to use, it needs access to the full module allocation context. Signed-off-by: Guennadi Liakhovetski <guennadi.liakhovetski@linux.intel.com>

fast-get: add support for vregion allocations

0de748b

When userspace modules use fast-get to access smaller buffers, their copies should be allocated on module's vregion, if available, not on the global SOF heap. Signed-off-by: Guennadi Liakhovetski <guennadi.liakhovetski@linux.intel.com>

lyakh force-pushed the vreg branch from 139f77c to 9d97d23 Compare April 29, 2026 17:04

lyakh mentioned this pull request Apr 30, 2026

Memory attributes to user space dp module #10603

Open

lyakh requested a review from Copilot April 30, 2026 06:22

Copilot started reviewing on behalf of lyakh April 30, 2026 06:22 View session

Copilot AI reviewed Apr 30, 2026

View reviewed changes

kv2019i approved these changes Apr 30, 2026

View reviewed changes

kv2019i merged commit 8d75e2c into thesofproject:main Apr 30, 2026
44 of 45 checks passed

lyakh deleted the vreg branch April 30, 2026 08:18

kv2019i mentioned this pull request Apr 30, 2026

[DRAFT] userspace LL/audio test PR #10558

Draft

lyakh mentioned this pull request Apr 30, 2026

audio: (cosmetic) address comments to an earlier PR #10736

Open

	* to move to the lifetime buffer. Make it 1k for the beginning.
	* to move to the lifetime buffer. Make it 4k for the beginning.

		@@ -170,10 +176,17 @@ static void module_adapter_mem_free(struct processing_module *mod)
		#if CONFIG_IPC_MAJOR_4
		sof_heap_free(mod_heap, mod->priv.cfg.input_pins);

Conversation

lyakh commented Apr 14, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Copilot AI Apr 15, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 15, 2026

Choose a reason for hiding this comment

Uh oh!

kv2019i left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

jsarha left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

lyakh Apr 27, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

lyakh commented Apr 28, 2026

Uh oh!

lrudyX commented Apr 28, 2026

Uh oh!

lyakh commented Apr 28, 2026

Uh oh!

lyakh commented Apr 28, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants

lyakh commented Apr 14, 2026 •

edited

Loading

lyakh Apr 27, 2026 •

edited

Loading