OSDN Git Service

bpf: Migrate release_on_unlock logic to non-owning ref semantics
authorDave Marchevsky <davemarchevsky@fb.com>
Sun, 12 Feb 2023 09:27:07 +0000 (01:27 -0800)
committerAlexei Starovoitov <ast@kernel.org>
Mon, 13 Feb 2023 21:37:37 +0000 (13:37 -0800)
This patch introduces non-owning reference semantics to the verifier,
specifically linked_list API kfunc handling. release_on_unlock logic for
refs is refactored - with small functional changes - to implement these
semantics, and bpf_list_push_{front,back} are migrated to use them.

When a list node is pushed to a list, the program still has a pointer to
the node:

  n = bpf_obj_new(typeof(*n));

  bpf_spin_lock(&l);
  bpf_list_push_back(&l, n);
  /* n still points to the just-added node */
  bpf_spin_unlock(&l);

What the verifier considers n to be after the push, and thus what can be
done with n, are changed by this patch.

Common properties both before/after this patch:
  * After push, n is only a valid reference to the node until end of
    critical section
  * After push, n cannot be pushed to any list
  * After push, the program can read the node's fields using n

Before:
  * After push, n retains the ref_obj_id which it received on
    bpf_obj_new, but the associated bpf_reference_state's
    release_on_unlock field is set to true
    * release_on_unlock field and associated logic is used to implement
      "n is only a valid ref until end of critical section"
  * After push, n cannot be written to, the node must be removed from
    the list before writing to its fields
  * After push, n is marked PTR_UNTRUSTED

After:
  * After push, n's ref is released and ref_obj_id set to 0. NON_OWN_REF
    type flag is added to reg's type, indicating that it's a non-owning
    reference.
    * NON_OWN_REF flag and logic is used to implement "n is only a
      valid ref until end of critical section"
  * n can be written to (except for special fields e.g. bpf_list_node,
    timer, ...)

Summary of specific implementation changes to achieve the above:

  * release_on_unlock field, ref_set_release_on_unlock helper, and logic
    to "release on unlock" based on that field are removed

  * The anonymous active_lock struct used by bpf_verifier_state is
    pulled out into a named struct bpf_active_lock.

  * NON_OWN_REF type flag is introduced along with verifier logic
    changes to handle non-owning refs

  * Helpers are added to use NON_OWN_REF flag to implement non-owning
    ref semantics as described above
    * invalidate_non_owning_refs - helper to clobber all non-owning refs
      matching a particular bpf_active_lock identity. Replaces
      release_on_unlock logic in process_spin_lock.
    * ref_set_non_owning - set NON_OWN_REF type flag after doing some
      sanity checking
    * ref_convert_owning_non_owning - convert owning reference w/
      specified ref_obj_id to non-owning references. Set NON_OWN_REF
      flag for each reg with that ref_obj_id and 0-out its ref_obj_id

  * Update linked_list selftests to account for minor semantic
    differences introduced by this patch
    * Writes to a release_on_unlock node ref are not allowed, while
      writes to non-owning reference pointees are. As a result the
      linked_list "write after push" failure tests are no longer scenarios
      that should fail.
    * The test##missing_lock##op and test##incorrect_lock##op
      macro-generated failure tests need to have a valid node argument in
      order to have the same error output as before. Otherwise
      verification will fail early and the expected error output won't be seen.

Signed-off-by: Dave Marchevsky <davemarchevsky@fb.com>
Link: https://lore.kernel.org/r/20230212092715.1422619-2-davemarchevsky@fb.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
include/linux/bpf.h
include/linux/bpf_verifier.h
kernel/bpf/verifier.c
tools/testing/selftests/bpf/prog_tests/linked_list.c
tools/testing/selftests/bpf/progs/linked_list.c
tools/testing/selftests/bpf/progs/linked_list_fail.c

index 4385418..8b5d0b4 100644 (file)
@@ -181,6 +181,7 @@ enum btf_field_type {
        BPF_KPTR       = BPF_KPTR_UNREF | BPF_KPTR_REF,
        BPF_LIST_HEAD  = (1 << 4),
        BPF_LIST_NODE  = (1 << 5),
+       BPF_GRAPH_NODE_OR_ROOT = BPF_LIST_NODE | BPF_LIST_HEAD,
 };
 
 struct btf_field_kptr {
@@ -576,6 +577,11 @@ enum bpf_type_flag {
        /* MEM is tagged with rcu and memory access needs rcu_read_lock protection. */
        MEM_RCU                 = BIT(13 + BPF_BASE_TYPE_BITS),
 
+       /* Used to tag PTR_TO_BTF_ID | MEM_ALLOC references which are non-owning.
+        * Currently only valid for linked-list and rbtree nodes.
+        */
+       NON_OWN_REF             = BIT(14 + BPF_BASE_TYPE_BITS),
+
        __BPF_TYPE_FLAG_MAX,
        __BPF_TYPE_LAST_FLAG    = __BPF_TYPE_FLAG_MAX - 1,
 };
index aa83de1..cf1bb1c 100644 (file)
@@ -43,6 +43,22 @@ enum bpf_reg_liveness {
        REG_LIVE_DONE = 0x8, /* liveness won't be updating this register anymore */
 };
 
+/* For every reg representing a map value or allocated object pointer,
+ * we consider the tuple of (ptr, id) for them to be unique in verifier
+ * context and conside them to not alias each other for the purposes of
+ * tracking lock state.
+ */
+struct bpf_active_lock {
+       /* This can either be reg->map_ptr or reg->btf. If ptr is NULL,
+        * there's no active lock held, and other fields have no
+        * meaning. If non-NULL, it indicates that a lock is held and
+        * id member has the reg->id of the register which can be >= 0.
+        */
+       void *ptr;
+       /* This will be reg->id */
+       u32 id;
+};
+
 struct bpf_reg_state {
        /* Ordering of fields matters.  See states_equal() */
        enum bpf_reg_type type;
@@ -226,11 +242,6 @@ struct bpf_reference_state {
         * exiting a callback function.
         */
        int callback_ref;
-       /* Mark the reference state to release the registers sharing the same id
-        * on bpf_spin_unlock (for nodes that we will lose ownership to but are
-        * safe to access inside the critical section).
-        */
-       bool release_on_unlock;
 };
 
 /* state of the program:
@@ -331,21 +342,8 @@ struct bpf_verifier_state {
        u32 branches;
        u32 insn_idx;
        u32 curframe;
-       /* For every reg representing a map value or allocated object pointer,
-        * we consider the tuple of (ptr, id) for them to be unique in verifier
-        * context and conside them to not alias each other for the purposes of
-        * tracking lock state.
-        */
-       struct {
-               /* This can either be reg->map_ptr or reg->btf. If ptr is NULL,
-                * there's no active lock held, and other fields have no
-                * meaning. If non-NULL, it indicates that a lock is held and
-                * id member has the reg->id of the register which can be >= 0.
-                */
-               void *ptr;
-               /* This will be reg->id */
-               u32 id;
-       } active_lock;
+
+       struct bpf_active_lock active_lock;
        bool speculative;
        bool active_rcu_lock;
 
index 388245e..f176bc1 100644 (file)
@@ -190,6 +190,9 @@ struct bpf_verifier_stack_elem {
 
 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx);
 static int release_reference(struct bpf_verifier_env *env, int ref_obj_id);
+static void invalidate_non_owning_refs(struct bpf_verifier_env *env);
+static int ref_set_non_owning(struct bpf_verifier_env *env,
+                             struct bpf_reg_state *reg);
 
 static bool bpf_map_ptr_poisoned(const struct bpf_insn_aux_data *aux)
 {
@@ -457,6 +460,11 @@ static bool type_is_ptr_alloc_obj(u32 type)
        return base_type(type) == PTR_TO_BTF_ID && type_flag(type) & MEM_ALLOC;
 }
 
+static bool type_is_non_owning_ref(u32 type)
+{
+       return type_is_ptr_alloc_obj(type) && type_flag(type) & NON_OWN_REF;
+}
+
 static struct btf_record *reg_btf_record(const struct bpf_reg_state *reg)
 {
        struct btf_record *rec = NULL;
@@ -1073,6 +1081,8 @@ static void print_verifier_state(struct bpf_verifier_env *env,
                                verbose_a("id=%d", reg->id);
                        if (reg->ref_obj_id)
                                verbose_a("ref_obj_id=%d", reg->ref_obj_id);
+                       if (type_is_non_owning_ref(reg->type))
+                               verbose_a("%s", "non_own_ref");
                        if (t != SCALAR_VALUE)
                                verbose_a("off=%d", reg->off);
                        if (type_is_pkt_pointer(t))
@@ -5052,7 +5062,8 @@ static int check_ptr_to_btf_access(struct bpf_verifier_env *env,
                        return -EACCES;
                }
 
-               if (type_is_alloc(reg->type) && !reg->ref_obj_id) {
+               if (type_is_alloc(reg->type) && !type_is_non_owning_ref(reg->type) &&
+                   !reg->ref_obj_id) {
                        verbose(env, "verifier internal error: ref_obj_id for allocated object must be non-zero\n");
                        return -EFAULT;
                }
@@ -6042,9 +6053,7 @@ static int process_spin_lock(struct bpf_verifier_env *env, int regno,
                        cur->active_lock.ptr = btf;
                cur->active_lock.id = reg->id;
        } else {
-               struct bpf_func_state *fstate = cur_func(env);
                void *ptr;
-               int i;
 
                if (map)
                        ptr = map;
@@ -6060,25 +6069,11 @@ static int process_spin_lock(struct bpf_verifier_env *env, int regno,
                        verbose(env, "bpf_spin_unlock of different lock\n");
                        return -EINVAL;
                }
-               cur->active_lock.ptr = NULL;
-               cur->active_lock.id = 0;
 
-               for (i = fstate->acquired_refs - 1; i >= 0; i--) {
-                       int err;
+               invalidate_non_owning_refs(env);
 
-                       /* Complain on error because this reference state cannot
-                        * be freed before this point, as bpf_spin_lock critical
-                        * section does not allow functions that release the
-                        * allocated object immediately.
-                        */
-                       if (!fstate->refs[i].release_on_unlock)
-                               continue;
-                       err = release_reference(env, fstate->refs[i].id);
-                       if (err) {
-                               verbose(env, "failed to release release_on_unlock reference");
-                               return err;
-                       }
-               }
+               cur->active_lock.ptr = NULL;
+               cur->active_lock.id = 0;
        }
        return 0;
 }
@@ -6546,6 +6541,23 @@ found:
        return 0;
 }
 
+static struct btf_field *
+reg_find_field_offset(const struct bpf_reg_state *reg, s32 off, u32 fields)
+{
+       struct btf_field *field;
+       struct btf_record *rec;
+
+       rec = reg_btf_record(reg);
+       if (!rec)
+               return NULL;
+
+       field = btf_record_find(rec, off, fields);
+       if (!field)
+               return NULL;
+
+       return field;
+}
+
 int check_func_arg_reg_off(struct bpf_verifier_env *env,
                           const struct bpf_reg_state *reg, int regno,
                           enum bpf_arg_type arg_type)
@@ -6567,6 +6579,18 @@ int check_func_arg_reg_off(struct bpf_verifier_env *env,
                 */
                if (arg_type_is_dynptr(arg_type) && type == PTR_TO_STACK)
                        return 0;
+
+               if ((type_is_ptr_alloc_obj(type) || type_is_non_owning_ref(type)) && reg->off) {
+                       if (reg_find_field_offset(reg, reg->off, BPF_GRAPH_NODE_OR_ROOT))
+                               return __check_ptr_off_reg(env, reg, regno, true);
+
+                       verbose(env, "R%d must have zero offset when passed to release func\n",
+                               regno);
+                       verbose(env, "No graph node or root found at R%d type:%s off:%d\n", regno,
+                               kernel_type_name(reg->btf, reg->btf_id), reg->off);
+                       return -EINVAL;
+               }
+
                /* Doing check_ptr_off_reg check for the offset will catch this
                 * because fixed_off_ok is false, but checking here allows us
                 * to give the user a better error message.
@@ -6601,6 +6625,7 @@ int check_func_arg_reg_off(struct bpf_verifier_env *env,
        case PTR_TO_BTF_ID | PTR_TRUSTED:
        case PTR_TO_BTF_ID | MEM_RCU:
        case PTR_TO_BTF_ID | MEM_ALLOC | PTR_TRUSTED:
+       case PTR_TO_BTF_ID | MEM_ALLOC | NON_OWN_REF:
                /* When referenced PTR_TO_BTF_ID is passed to release function,
                 * its fixed offset must be 0. In the other cases, fixed offset
                 * can be non-zero. This was already checked above. So pass
@@ -7363,6 +7388,17 @@ static int release_reference(struct bpf_verifier_env *env,
        return 0;
 }
 
+static void invalidate_non_owning_refs(struct bpf_verifier_env *env)
+{
+       struct bpf_func_state *unused;
+       struct bpf_reg_state *reg;
+
+       bpf_for_each_reg_in_vstate(env->cur_state, unused, reg, ({
+               if (type_is_non_owning_ref(reg->type))
+                       __mark_reg_unknown(env, reg);
+       }));
+}
+
 static void clear_caller_saved_regs(struct bpf_verifier_env *env,
                                    struct bpf_reg_state *regs)
 {
@@ -8915,38 +8951,54 @@ static int process_kf_arg_ptr_to_kptr(struct bpf_verifier_env *env,
        return 0;
 }
 
-static int ref_set_release_on_unlock(struct bpf_verifier_env *env, u32 ref_obj_id)
+static int ref_set_non_owning(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
 {
-       struct bpf_func_state *state = cur_func(env);
+       struct bpf_verifier_state *state = env->cur_state;
+
+       if (!state->active_lock.ptr) {
+               verbose(env, "verifier internal error: ref_set_non_owning w/o active lock\n");
+               return -EFAULT;
+       }
+
+       if (type_flag(reg->type) & NON_OWN_REF) {
+               verbose(env, "verifier internal error: NON_OWN_REF already set\n");
+               return -EFAULT;
+       }
+
+       reg->type |= NON_OWN_REF;
+       return 0;
+}
+
+static int ref_convert_owning_non_owning(struct bpf_verifier_env *env, u32 ref_obj_id)
+{
+       struct bpf_func_state *state, *unused;
        struct bpf_reg_state *reg;
        int i;
 
-       /* bpf_spin_lock only allows calling list_push and list_pop, no BPF
-        * subprogs, no global functions. This means that the references would
-        * not be released inside the critical section but they may be added to
-        * the reference state, and the acquired_refs are never copied out for a
-        * different frame as BPF to BPF calls don't work in bpf_spin_lock
-        * critical sections.
-        */
+       state = cur_func(env);
+
        if (!ref_obj_id) {
-               verbose(env, "verifier internal error: ref_obj_id is zero for release_on_unlock\n");
+               verbose(env, "verifier internal error: ref_obj_id is zero for "
+                            "owning -> non-owning conversion\n");
                return -EFAULT;
        }
+
        for (i = 0; i < state->acquired_refs; i++) {
-               if (state->refs[i].id == ref_obj_id) {
-                       if (state->refs[i].release_on_unlock) {
-                               verbose(env, "verifier internal error: expected false release_on_unlock");
-                               return -EFAULT;
+               if (state->refs[i].id != ref_obj_id)
+                       continue;
+
+               /* Clear ref_obj_id here so release_reference doesn't clobber
+                * the whole reg
+                */
+               bpf_for_each_reg_in_vstate(env->cur_state, unused, reg, ({
+                       if (reg->ref_obj_id == ref_obj_id) {
+                               reg->ref_obj_id = 0;
+                               ref_set_non_owning(env, reg);
                        }
-                       state->refs[i].release_on_unlock = true;
-                       /* Now mark everyone sharing same ref_obj_id as untrusted */
-                       bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({
-                               if (reg->ref_obj_id == ref_obj_id)
-                                       reg->type |= PTR_UNTRUSTED;
-                       }));
-                       return 0;
-               }
+               }));
+               return 0;
        }
+
        verbose(env, "verifier internal error: ref state missing for ref_obj_id\n");
        return -EFAULT;
 }
@@ -9081,7 +9133,6 @@ static int process_kf_arg_ptr_to_list_node(struct bpf_verifier_env *env,
 {
        const struct btf_type *et, *t;
        struct btf_field *field;
-       struct btf_record *rec;
        u32 list_node_off;
 
        if (meta->btf != btf_vmlinux ||
@@ -9098,9 +9149,8 @@ static int process_kf_arg_ptr_to_list_node(struct bpf_verifier_env *env,
                return -EINVAL;
        }
 
-       rec = reg_btf_record(reg);
        list_node_off = reg->off + reg->var_off.value;
-       field = btf_record_find(rec, list_node_off, BPF_LIST_NODE);
+       field = reg_find_field_offset(reg, list_node_off, BPF_LIST_NODE);
        if (!field || field->offset != list_node_off) {
                verbose(env, "bpf_list_node not found at offset=%u\n", list_node_off);
                return -EINVAL;
@@ -9126,8 +9176,8 @@ static int process_kf_arg_ptr_to_list_node(struct bpf_verifier_env *env,
                        btf_name_by_offset(field->graph_root.btf, et->name_off));
                return -EINVAL;
        }
-       /* Set arg#1 for expiration after unlock */
-       return ref_set_release_on_unlock(env, reg->ref_obj_id);
+
+       return 0;
 }
 
 static int check_kfunc_args(struct bpf_verifier_env *env, struct bpf_kfunc_call_arg_meta *meta)
@@ -9406,11 +9456,11 @@ static int check_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
                            int *insn_idx_p)
 {
        const struct btf_type *t, *func, *func_proto, *ptr_type;
+       u32 i, nargs, func_id, ptr_type_id, release_ref_obj_id;
        struct bpf_reg_state *regs = cur_regs(env);
        const char *func_name, *ptr_type_name;
        bool sleepable, rcu_lock, rcu_unlock;
        struct bpf_kfunc_call_arg_meta meta;
-       u32 i, nargs, func_id, ptr_type_id;
        int err, insn_idx = *insn_idx_p;
        const struct btf_param *args;
        const struct btf_type *ret_t;
@@ -9505,6 +9555,24 @@ static int check_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
                }
        }
 
+       if (meta.func_id == special_kfunc_list[KF_bpf_list_push_front] ||
+           meta.func_id == special_kfunc_list[KF_bpf_list_push_back]) {
+               release_ref_obj_id = regs[BPF_REG_2].ref_obj_id;
+               err = ref_convert_owning_non_owning(env, release_ref_obj_id);
+               if (err) {
+                       verbose(env, "kfunc %s#%d conversion of owning ref to non-owning failed\n",
+                               func_name, func_id);
+                       return err;
+               }
+
+               err = release_reference(env, release_ref_obj_id);
+               if (err) {
+                       verbose(env, "kfunc %s#%d reference has not been acquired before\n",
+                               func_name, func_id);
+                       return err;
+               }
+       }
+
        for (i = 0; i < CALLER_SAVED_REGS; i++)
                mark_reg_not_init(env, regs, caller_saved[i]);
 
@@ -11825,8 +11893,10 @@ static void mark_ptr_or_null_reg(struct bpf_func_state *state,
                 */
                if (WARN_ON_ONCE(reg->smin_value || reg->smax_value || !tnum_equals_const(reg->var_off, 0)))
                        return;
-               if (reg->type != (PTR_TO_BTF_ID | MEM_ALLOC | PTR_MAYBE_NULL) && WARN_ON_ONCE(reg->off))
+               if (!(type_is_ptr_alloc_obj(reg->type) || type_is_non_owning_ref(reg->type)) &&
+                   WARN_ON_ONCE(reg->off))
                        return;
+
                if (is_null) {
                        reg->type = SCALAR_VALUE;
                        /* We don't need id and ref_obj_id from this point
index 9a7d4c4..2592b8a 100644 (file)
@@ -78,8 +78,6 @@ static struct {
        { "direct_write_head", "direct access to bpf_list_head is disallowed" },
        { "direct_read_node", "direct access to bpf_list_node is disallowed" },
        { "direct_write_node", "direct access to bpf_list_node is disallowed" },
-       { "write_after_push_front", "only read is supported" },
-       { "write_after_push_back", "only read is supported" },
        { "use_after_unlock_push_front", "invalid mem access 'scalar'" },
        { "use_after_unlock_push_back", "invalid mem access 'scalar'" },
        { "double_push_front", "arg#1 expected pointer to allocated object" },
index 4ad88da..4fa4a9b 100644 (file)
@@ -260,7 +260,7 @@ int test_list_push_pop_multiple(struct bpf_spin_lock *lock, struct bpf_list_head
 {
        int ret;
 
-       ret = list_push_pop_multiple(lock ,head, false);
+       ret = list_push_pop_multiple(lockhead, false);
        if (ret)
                return ret;
        return list_push_pop_multiple(lock, head, true);
index 1d90172..69cdc07 100644 (file)
                return 0;                                   \
        }
 
-CHECK(kptr, push_front, &f->head);
-CHECK(kptr, push_back, &f->head);
 CHECK(kptr, pop_front, &f->head);
 CHECK(kptr, pop_back, &f->head);
 
-CHECK(global, push_front, &ghead);
-CHECK(global, push_back, &ghead);
 CHECK(global, pop_front, &ghead);
 CHECK(global, pop_back, &ghead);
 
-CHECK(map, push_front, &v->head);
-CHECK(map, push_back, &v->head);
 CHECK(map, pop_front, &v->head);
 CHECK(map, pop_back, &v->head);
 
-CHECK(inner_map, push_front, &iv->head);
-CHECK(inner_map, push_back, &iv->head);
 CHECK(inner_map, pop_front, &iv->head);
 CHECK(inner_map, pop_back, &iv->head);
 
 #undef CHECK
 
+#define CHECK(test, op, hexpr, nexpr)                                  \
+       SEC("?tc")                                                      \
+       int test##_missing_lock_##op(void *ctx)                         \
+       {                                                               \
+               INIT;                                                   \
+               void (*p)(void *, void *) = (void *)&bpf_list_##op;     \
+               p(hexpr, nexpr);                                        \
+               return 0;                                               \
+       }
+
+CHECK(kptr, push_front, &f->head, b);
+CHECK(kptr, push_back, &f->head, b);
+
+CHECK(global, push_front, &ghead, f);
+CHECK(global, push_back, &ghead, f);
+
+CHECK(map, push_front, &v->head, f);
+CHECK(map, push_back, &v->head, f);
+
+CHECK(inner_map, push_front, &iv->head, f);
+CHECK(inner_map, push_back, &iv->head, f);
+
+#undef CHECK
+
 #define CHECK(test, op, lexpr, hexpr)                       \
        SEC("?tc")                                          \
        int test##_incorrect_lock_##op(void *ctx)           \
@@ -108,13 +124,49 @@ CHECK(inner_map, pop_back, &iv->head);
        CHECK(inner_map_global, op, &iv->lock, &ghead);        \
        CHECK(inner_map_map, op, &iv->lock, &v->head);
 
-CHECK_OP(push_front);
-CHECK_OP(push_back);
 CHECK_OP(pop_front);
 CHECK_OP(pop_back);
 
 #undef CHECK
 #undef CHECK_OP
+
+#define CHECK(test, op, lexpr, hexpr, nexpr)                           \
+       SEC("?tc")                                                      \
+       int test##_incorrect_lock_##op(void *ctx)                       \
+       {                                                               \
+               INIT;                                                   \
+               void (*p)(void *, void*) = (void *)&bpf_list_##op;      \
+               bpf_spin_lock(lexpr);                                   \
+               p(hexpr, nexpr);                                        \
+               return 0;                                               \
+       }
+
+#define CHECK_OP(op)                                                   \
+       CHECK(kptr_kptr, op, &f1->lock, &f2->head, b);                  \
+       CHECK(kptr_global, op, &f1->lock, &ghead, f);                   \
+       CHECK(kptr_map, op, &f1->lock, &v->head, f);                    \
+       CHECK(kptr_inner_map, op, &f1->lock, &iv->head, f);             \
+                                                                       \
+       CHECK(global_global, op, &glock2, &ghead, f);                   \
+       CHECK(global_kptr, op, &glock, &f1->head, b);                   \
+       CHECK(global_map, op, &glock, &v->head, f);                     \
+       CHECK(global_inner_map, op, &glock, &iv->head, f);              \
+                                                                       \
+       CHECK(map_map, op, &v->lock, &v2->head, f);                     \
+       CHECK(map_kptr, op, &v->lock, &f2->head, b);                    \
+       CHECK(map_global, op, &v->lock, &ghead, f);                     \
+       CHECK(map_inner_map, op, &v->lock, &iv->head, f);               \
+                                                                       \
+       CHECK(inner_map_inner_map, op, &iv->lock, &iv2->head, f);       \
+       CHECK(inner_map_kptr, op, &iv->lock, &f2->head, b);             \
+       CHECK(inner_map_global, op, &iv->lock, &ghead, f);              \
+       CHECK(inner_map_map, op, &iv->lock, &v->head, f);
+
+CHECK_OP(push_front);
+CHECK_OP(push_back);
+
+#undef CHECK
+#undef CHECK_OP
 #undef INIT
 
 SEC("?kprobe/xyz")
@@ -304,34 +356,6 @@ int direct_write_node(void *ctx)
 }
 
 static __always_inline
-int write_after_op(void (*push_op)(void *head, void *node))
-{
-       struct foo *f;
-
-       f = bpf_obj_new(typeof(*f));
-       if (!f)
-               return 0;
-       bpf_spin_lock(&glock);
-       push_op(&ghead, &f->node);
-       f->data = 42;
-       bpf_spin_unlock(&glock);
-
-       return 0;
-}
-
-SEC("?tc")
-int write_after_push_front(void *ctx)
-{
-       return write_after_op((void *)bpf_list_push_front);
-}
-
-SEC("?tc")
-int write_after_push_back(void *ctx)
-{
-       return write_after_op((void *)bpf_list_push_back);
-}
-
-static __always_inline
 int use_after_unlock(void (*op)(void *head, void *node))
 {
        struct foo *f;