sched_ext selftest enable_cmask fails to load on 7.4-rc2: verifier rejects arena pointer cast from scx_enable_args.cmask_arena_addr

sched_ext selftest enable_cmask fails to load on 7.4-rc2: verifier rejects arena pointer cast from scx_enable_args.cmask_arena_addr

[Zhijian Han]

From: Zhijian Han <hanzhijian1991@gmail.com>

Hello,

The enable_cmask selftest (tools/testing/selftests/sched_ext/, added in
the recent merge window via d781d1b78acf "selftests/sched_ext: Check the
cmask cid-form ops.enable() receives") fails to load on v7.4-rc2 and
linux-next-20261105. The BPF program enable_cmask_enable is rejected by
the verifier at load time:

  libbpf: prog 'enable_cmask_enable': BPF program load failed: -EACCES
  94: (61) r1 = *(u32 *)(r8 +0)
  R8 invalid mem access 'scalar'
  Verification failed: Register Type Safety: Invalid dereference

The test follows the usage documented in kernel/sched/ext/internal.h:

  struct scx_enable_args {
          u64     cmask_arena_addr;
  };

  /*
   * ... The address is a plain value rather than a typed pointer ...
   * Cast it to struct scx_cmask __arena * before use.
   */

i.e. enable_cmask.bpf.c:101:

  struct scx_cmask __arena *m =
          (struct scx_cmask __arena *)args->cmask_arena_addr;

The C cast does not emit an addr_space_cast insn: in the generated
bytecode the member load (insn 1: r7 = *(u64 *)(r2 + 0)) produces
R7=scalar(), which stays scalar across the cast and the inlined
check_mask() prologue, so the first dereference m->base (insn 94) is
rejected.

The same selftest was verified by Tejun on the for-7.3-fixes branch
(per the v3 cover letter, 20260919002838.1960071-1-tj, 16-CPU VM, three
standalone runs plus the full suite); the verifier-side arena handling
that changed during the 7.4 merge window appears to no longer accept
this pattern. The struct_ops argument path (direct __arena args, e.g.
ops.set_cmask) still works — only the u64-carried address form is
affected.

Reproducer (deterministic, two consecutive runs identical):

  # kernel: defconfig + SCHED_CLASS_EXT, DEBUG_INFO_BTF, FUNCTION_TRACER
  make O=build -j$(nproc)
  cd tools/testing/selftests/sched_ext
  make LLVM=1 O=$PWD/../../../../../build OUTPUT=out/ TEST_GEN_MODS_DIR=

  # initrd with out/runner + out/build + libs + busybox + kernel .config
  # at /boot/config-$(uname -r); then:
  qemu-system-x86_64 -enable-kvm -smp 4 -m 2G -nographic \
    -kernel build/arch/x86/boot/bzImage -initrd initrd.cpio.gz \
    -append "console=ttyS0"

  /root # ./runner -t enable_cmask
  not ok 1 enable_cmask #

Tested on:
  v7.4-rc2 (7475ea8b5):          FAIL  (verifier rejects at load)
  next-20261105 (gddfe51b0a0f5): FAIL  (same rejection)
  v7.3:                          SKIP  (struct scx_enable_args not present)

Compiler versions tried: clang 18 and clang 19 — identical bytecode
around the failing insn, identical rejection, so this does not look
compiler-dependent.

The other 40 tests in the suite pass on both 7.4-rc2 and next (after
enabling CONFIG_FUNCTION_TRACER, without which the fentry-based tests
maximal/kick_immediate/reload_loop fail with -EBUSY — worth a note in
the selftest docs, but that part is config, not a bug).

Possible directions (not investigated in depth):
  - a kfunc or verifier-supported way to convert the u64 carried in
    scx_enable_args into a PTR_TO_ARENA,
  - or an __arena-typed member/alias in scx_enable_args once BTF can
    express arena pointer members (the internal.h comment already
    anticipates this),
  - or adjusting the test to obtain the cmask through a path the
    verifier accepts.

Full verifier log and QEMU boot log available on request.

Signed-off-by: Zhijian Han <hanzhijian1991@gmail.com>

[PATCH sched_ext/for-7.5] selftests/sched_ext: Fix enable_cmask load failure on older clangs

[Tejun Heo]

enable_cmask casts the u64 arena address of the task's cmask to an __arena
pointer without cast_kern(). Older clangs without
__BPF_FEATURE_ADDR_SPACE_CAST need the explicit cast, so the test fails to
load there. Add it.

Reported-by: Zhijian Han <hanzhijian1991@gmail.com>
Closes: https://lore.kernel.org/all/6af1e5f2.a438010f.130489.ab94@mx.google.com/
Signed-off-by: Tejun Heo <tj@kernel.org>
---
Applied to sched_ext/for-7.5.

 tools/testing/selftests/sched_ext/enable_cmask.bpf.c | 1 +
 1 file changed, 1 insertion(+)

diff --git a/tools/testing/selftests/sched_ext/enable_cmask.bpf.c b/tools/testing/selftests/sched_ext/enable_cmask.bpf.c
index 0068f3c7ab3c..081be5861424 100644
--- a/tools/testing/selftests/sched_ext/enable_cmask.bpf.c
+++ b/tools/testing/selftests/sched_ext/enable_cmask.bpf.c
@@ -104,6 +104,7 @@ void BPF_STRUCT_OPS(enable_cmask_enable, struct task_struct *p, struct scx_enabl
 	int ret;
 
 	asm volatile("" :: "r"(&arena));
+	cast_kern(m);
 	tctx = lookup_task_ctx(p);
 	if (!tctx)
 		return;
-- 
2.56.0