[lgirdwood]

Phase 3 of #11306 (kernel / user-space partition), first part: split the three mixed IPC source files and move their kernel halves into src/kernel/ipc. Stacked on #11421 (trace removal), which is stacked on #11420; only the last three commits are new.

The kernel half of src/X/foo.c is src/kernel/X/foo-kernel.c, and each user half says where its kernel half is. Shared helpers that used to be static are now declared in the existing IPC headers.

Patches

1. ipc/ipc-common.c → kernel/ipc/ipc-common-kernel.c:
   • Moves to the kernel half:
     • the IPC task, command completion and IDC forwarding;
     • the outgoing message queue and the kernel half of the ipc_msg_send() and ipc_msg_list_remove() system calls;
     • the setup of the user-space IPC threads and the command forwarding to them;
     • the user-space LL IPC context with ipc_get();
     • ipc_init().
   • Stays in ipc-common.c:
     • message allocation and the component lookups;
     • the user-space IPC thread function and its default dispatch. The thread entry point and its event bits are now in sof/ipc/common.h.
2. ipc/ipc4/handler-user.c → kernel/ipc/ipc4/handler-user-kernel.c:
   • Moves to the kernel half: the global and module message routers, which run in the kernel IPC thread and forward commands to the user-space IPC thread, and everything only they use:
     • the reply message;
     • the target core lookup and the multicore SET_PIPELINE_STATE sequencing;
     • the MOD_INIT_INSTANCE driver lookup and module loading;
     • the CHAIN_DMA, IPC gateway and GDB commands;
     • kernel-mode module init;
     • the base firmware LARGE_CONFIG get and set;
     • the wrappers that the routers of builds without user-space LL call directly.
   • Stays in handler-user.c: the pipeline state handling, the module config and large config processing, and the user-space IPC thread dispatch.
   • Now shared: four helpers, declared in ipc4/handler.h.
3. ipc/ipc4/helper.c → kernel/ipc/ipc4/helper-kernel.c:
   • Moves to the kernel half:
     • the component driver lookup, which reads the IMR manifest, registers loadable modules and walks the driver list;
     • comp_new_ipc4() and ipc4_user_module_load();
     • the IDC helpers ipc4_process_on_core() and ipc_comp_free_remote();
     • the chain DMA manager.
   • Stays in helper.c: pipeline and component management, binding and the DMA config parsing.
   • Includes: helper.c no longer includes the rimage manifest, platform memory map or library manager headers.

The 16 cmocka and 2 ztest unit test CMake files that build ipc-common.c also build its kernel half. No unit test builds the IPC4 files.

Some code in the user halves still calls the kernel (ipc4_process_on_core() and ipc_comp_free_remote() in builds without user-space LL, and the base firmware driver lookup). The issue tracks that as a Phase 5 item.

Functional change

None. The functions are moved, not changed. The per-function disassembly of every image (the builds listed under Testing) was compared before and after each commit. It differs only where the compiler can no longer inline or specialise a function across the split:

• ipc-common.c: no such cases. Only Xtensa alignment padding and narrow/wide encodings differ, plus the line numbers passed to the assertion handler in qemu_xtensa, which has assertions enabled.
• handler-user.c:
  • the callers of the shared helpers call them with the full argument list instead of GCC's constant-propagated clones;
  • ipc4_set_pipeline_state() only partially inlines ipc4_pipeline_id_get() (builds without user-space LL);
  • the routers' register allocation differs.
• helper.c: the callers no longer inline the shared helpers:
  • ipc_pipeline_new(), ipc_pipeline_free(), ipc4_comp_connect() and ipc4_comp_disconnect() call ipc4_process_on_core() and ipc_comp_free_remote();
  • comp_new_ipc4(), ipc4_user_module_load() and ipc4_chain_manager_create() call the module config and component list helpers.

Symbols and SYS_INIT() order are otherwise unchanged, as are every build's .config and the generated rimage configurations.

Testing

• Per commit:
  • Builds: tgl, tgl-h, mtl, lnl, ptl, wcl, nvl, imx8, imx8x, imx8m, imx8m_cm7, imx8ulp, imx95 and qemu_xtensa, and the extra CI configurations:
    • native_sim IPC3/IPC4 stub builds;
    • imx93 64-bit;
    • PTL with each userspace overlay;
    • PTL with CONFIG_SOF_USERSPACE_PROXY=n;
    • PTL and MTL with os_linux_overlay.conf;llext_relocatable.conf.
  • Comparisons: objects, LLEXT modules and firmware images (sections, symbols, SYS_INIT() order), plus the per-function disassembly comparison described above.
  • Signing: rimage runs for every Intel platform, and the generated configs are unchanged.
  • Cadence xt-clang (RI-2022.10): MTL, PTL and NVL compile.
  • ZTest: the unit tests pass, run as in CI (18 of 18 configurations).
  • Host: the host library and the cmocka unit tests build and pass for all nine defconfigs and the library config.
• At the tip:
  • 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.
• Hardware: TGL (UP Xtreme i11), sof-hda-generic-cavs25-2ch.tplg, comparing the tip of Kernel/user partition: remove the SOF DMA trace #11421 and this branch. Both boot, play 5 s, and play again after runtime suspend, with no DSP errors.

🤖 Generated with Claude Code