Merge branch 'allow-bpf_refcount_acquire-of-mapval-obtained-via-direct-ld'

Merge branch 'allow-bpf_refcount_acquire-of-mapval-obtained-via-direct-ld'

Dave Marchevsky says:

====================
Allow bpf_refcount_acquire of mapval obtained via direct LD

Consider this BPF program:

  struct cgv_node {
    int d;
    struct bpf_refcount r;
    struct bpf_rb_node rb;
  };

  struct val_stash {
    struct cgv_node __kptr *v;
  };

  struct {
    __uint(type, BPF_MAP_TYPE_ARRAY);
    __type(key, int);
    __type(value, struct val_stash);
    __uint(max_entries, 10);
  } array_map SEC(".maps");

  long bpf_program(void *ctx)
  {
    struct val_stash *mapval;
    struct cgv_node *p;
    int idx = 0;

    mapval = bpf_map_lookup_elem(&array_map, &idx);
    if (!mapval || !mapval->v) { /* omitted */ }

    p = bpf_refcount_acquire(mapval->v); /* Verification FAILs here */

    /* Add p to some tree */
    return 0;
  }

Verification fails on the refcount_acquire:

  160: (79) r1 = *(u64 *)(r8 +8)        ; R1_w=untrusted_ptr_or_null_cgv_node(id=11,off=0,imm=0) R8_w=map_value(id=10,off=0,ks=8,vs=16,imm=0) refs=6
  161: (b7) r2 = 0                      ; R2_w=0 refs=6
  162: (85) call bpf_refcount_acquire_impl#117824
  arg#0 is neither owning or non-owning ref

The above verifier dump is actually from sched_ext's scx_flatcg [0],
which is the motivating usecase for this series' changes. Specifically,
scx_flatcg stashes a rb_node type w/ cgroup-specific info (struct
cgv_node) in a map when the cgroup is created, then later puts that
cgroup's node in a rbtree in .enqueue . Making struct cgv_node
refcounted would simplify the code a bit by virtue of allowing us to
remove the kptr_xchg's, but "later puts that cgroups node in a rbtree"
is not possible without a refcount_acquire, which suffers from the above
verification failure.

If we get rid of PTR_UNTRUSTED flag, and add MEM_ALLOC | NON_OWN_REF,
mapval->v would be a non-owning ref and verification would succeed. Due
to the most recent set of refcount changes [1], which modified
bpf_obj_drop behavior to not reuse refcounted graph node's underlying
memory until after RCU grace period, this is safe to do. Once mapval->v
has the correct flags it _is_ a non-owning reference and verification of
the motivating example will succeed.

  [0]: https://github.com/sched-ext/sched_ext/blob/52911e1040a0f94b9c426dddcc00be5364a7ae9f/tools/sched_ext/scx_flatcg.bpf.c#L275
  [1]: https://lore.kernel.org/bpf/20230821193311.3290257-1-davemarchevsky@fb.com/

Summary of patches:
  * Patch 1 fixes an issue with bpf_refcount_acquire verification
    letting MAYBE_NULL ptrs through
    * Patch 2 tests Patch 1's fix
  * Patch 3 broadens the use of "free only after RCU GP" to all
    user-allocated types
  * Patch 4 is a small nonfunctional refactoring
  * Patch 5 changes verifier to mark direct LD of stashed graph node
    kptr as non-owning ref
    * Patch 6 tests Patch 5's verifier changes

Changelog:

v1 -> v2: https://lore.kernel.org/bpf/20231025214007.2920506-1-davemarchevsky@fb.com/

Series title changed to "Allow bpf_refcount_acquire of mapval obtained via
direct LD". V1's title was mistakenly truncated.

  * Patch 5 ("bpf: Mark direct ld of stashed bpf_{rb,list}_node as non-owning ref")
    * Direct LD of percpu kptr should not have MEM_ALLOC flag (Yonghong)
  * Patch 6 ("selftests/bpf: Test bpf_refcount_acquire of node obtained via direct ld")
    * Test read from stashed value as well
====================

Link: https://lore.kernel.org/r/20231107085639.3016113-1-davemarchevsky@fb.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>