[lgirdwood]

Phase 1 of #11306 (kernel / user-space partition), third part: more user-space code stops including the platform memory map and the DAI, IDC and scheduling domain headers. Stacked on #11345: the first 6 commits are that PR and will drop out when it is merged.

Patches

• Cache alias helpers in rtos/cache.h: uncache_to_cache(), cache_to_uncache() and is_uncached() are defined once instead of in every platform memory header:

  • ACE uses the Zephyr sys_cache_*_ptr_get() / sys_cache_is_ptr_cached() API, as before;
  • cAVS keeps its SRAM alias arithmetic;
  • the other Zephyr platforms have no uncached alias. They keep the identity mapping, and is_uncached() still returns 0.

  The host library keeps its own definitions.

• sof/coherent.h: includes rtos/cache.h instead of the platform memory map.

• comp_dai_config(): moved from component_ext.h to a new sof/audio/component_dai.h, so component_ext.h no longer includes sof/lib/dai.h and the DMA and DAI driver headers behind it.

• CONFIG_PLATFORM_MAX_STREAMS (new): replaces the per-platform PLATFORM_MAX_STREAMS, following CONFIG_PLATFORM_MAX_CHANNELS from Phase 1 (part 2): stop user-space code depending on the platform memory map and headers #11345. The defaults equal today's values: 5 for the i.MX, MediaTek and AMD Zephyr builds, 8 for i.MX8M and qemu_xtensa, and 16 otherwise, including the host library and unit tests. Like the channel option, it has no prompt. The mixer, mux and module adapter then drop sof/platform.h.

• LL scheduler: LL_TIMER_PERIOD_US and the user LL lock helpers (user_ll_lock_sched(), user_ll_unlock_sched(), user_ll_assert_locked()) move from ll_schedule_domain.h to ll_schedule.h. Audio code then stops including the scheduling domain header, which holds the domain driver interface and pulls in the clock headers.

• component_ext.h: dropped from the 18 sources and headers that use nothing from it. It includes rtos/idc.h for its cross-core helpers, and only the pipeline, KPB, component.c, IPC and IDC code use those.

• Copier: includes sof/lib/cold.h instead of sof/lib/memory.h; __cold is all it used.

• Missing includes: three of them were hidden by the indirect includes removed here (tlv.h → rtos/string.h, intel_uaol.c → the logging header, the non-HiFi3 up/down mixer stubs → rtos/panic.h). Each gets its own patch, ahead of the patch that exposed it.

Files that still use something from the headers they no longer get indirectly now include those headers themselves.

Functional change

None. Every patch leaves the generated code unchanged.

Effect

Measured with scripts/check_kernel_includes.py on PTL, as the number of user-space sources including each header, before and after this PR (on top of #11345):

Header	Before	After
sof/lib/memory.h (platform memory map)	141	34
zephyr/drivers/dma.h, sof/lib/io.h	27	10
rtos/idc.h	26	11
Platform clock headers / rtos/clk.h	20	14
sof/platform.h	15	11
Any kernel-only header	143	45

What remains is mostly real kernel dependencies, which later parts need to replace with syscalls or move to the kernel:

• the IPC4 mailbox registers (volume peak meter, pipeline registers, IPC gateway);
• the IDC dispatchers in component_ext.h, used by the pipeline;
• the DAI/DMA headers used by the copier and dai-zephyr.h;
• the user LL lock and IRQ locking in the pipeline.

Testing

• Per commit:
  • Firmware builds: tgl, tgl-h, mtl, lnl, ptl, wcl, nvl, imx8, imx8x, imx8m, imx8m_cm7, imx8ulp, imx95 and qemu_xtensa.
  • Extra CI configurations: native_sim IPC3/IPC4 stub builds, imx93 64-bit, PTL with each userspace overlay, PTL with CONFIG_SOF_USERSPACE_PROXY=n, and PTL and MTL with os_linux_overlay.conf;llext_relocatable.conf.
  • Object comparison: every object file and LLEXT module has the same non-debug section sizes as in the previous commit. The only differences are compiler-generated local names (CSWTCH, .str.N) that are numbered differently because fewer headers are included. In the ASan fuzz builds those names are stored as strings.
  • Host: the host library and cmocka build and pass.
• Cache helpers: the patch that moves the cache alias helpers was also checked for identical section contents and relocations on tgl, ptl, imx8m, imx95 and qemu_xtensa. The only differences are platform.c, which embeds the git version, and the PTL LLEXT link addresses, which also differ between two builds of the same commit.
• At the tip:
  • all cmocka configurations, the testbench and the tools pass;
  • the platforms that already fail on main (acp_6_0, acp_7_0, acp_7_x, mt8186, mt8188, mt8195, mt8196, mt8365) have unchanged error sets.

All of the above was run on db2b388 (main including #11346) with #11345 applied.

🤖 Generated with Claude Code