diff --git a/Documentation/bpf/kfuncs.rst b/Documentation/bpf/kfuncs.rst
index 85f73e0bbd0ff..89dea6b0b0241 100644
--- a/Documentation/bpf/kfuncs.rst
+++ b/Documentation/bpf/kfuncs.rst
@@ -575,6 +575,71 @@ is also covered by this recovery. A kfunc handed an arena pointer may
 therefore access up to ``GUARD_SZ / 2`` past it without bounds-checking
 against the arena. Larger accesses must verify the range explicitly.
 
+2.9 kfunc Return Values
+-----------------------
+
+A kfunc may return a scalar, a pointer, or a small struct or union by
+value. A scalar or pointer of up to 8 bytes is returned in R0, as usual.
+
+A struct or union returned by value must be composed only of scalars
+(recursively), where a scalar is an integer or an enum; arrays of scalars are
+allowed as members. Its bytes are handed back to the program as the raw
+contents of R0 (and R2), so a pointer field would be laundered into a scalar
+and escape the verifier's pointer provenance and reference tracking. A struct
+or union with a pointer member is therefore rejected at load time, and so is
+one with a floating-point member, which the ABI may not return in R0:R2 at
+all.
+
+A kfunc may also return a value larger than 8 bytes and up to 16 bytes -- a
+scalar-only struct or union, or an ``__int128``. Such a value is returned
+in the register pair R0:R2, matching the convention LLVM uses for the BPF
+target: the first 8 bytes in R0 and the second 8 bytes in R2. A struct or
+union of 8 bytes or less is returned in R0 alone.
+
+::
+
+        struct bpf_pair { __u64 a, b; };   /* 16 bytes */
+
+        __bpf_kfunc struct bpf_pair bpf_kfunc_get_pair(void)
+        {
+                struct bpf_pair p = { .a = 1, .b = 2 };
+
+                return p;      /* p.a in R0, p.b in R2 */
+        }
+
+Returning a value in the R0:R2 pair requires the JIT to place the second
+half of the return value into R2, which not every architecture supports
+right now. A kfunc with a return value larger than 8 bytes is therefore
+rejected at load time on a JIT that does not advertise this capability (see
+``bpf_jit_supports_kfunc_ret_reg_pair()``), and such a program is never run
+by the interpreter. A return value larger than 16 bytes is not supported.
+
+The same R0:R2 convention applies to a BPF subprogram, global or static, that
+returns an ``__int128`` or a struct or union larger than 8 bytes. It is only
+used when the program is JITed, since the interpreter propagates only R0 out of
+a subprogram: without a JIT the return value stays in R0 alone, and a caller
+reading R2 is rejected for reading an uninitialized register. A global
+subprogram is verified in isolation, so its by-value struct or union return is
+restricted to scalars just like a kfunc's; a static subprogram is verified
+inline and has no such restriction. The main program is not covered: its return
+value is the program's exit code, read out of R0 alone, so a declared upper
+half is never looked at.
+
+A global subprogram must leave a scalar in *every* register of the pair, so
+both halves of the returned value have to be assigned. Leaving the upper half
+uninitialized is not merely untidy: the compiler is then free to leave R2
+holding whatever it happened to hold, which for a subprogram taking a pointer
+argument is typically that pointer. Handing the caller an unknown scalar built
+from a pointer is a leak, so the verifier rejects it with::
+
+        At subprogram exit the register R2 is not a scalar value (...)
+
+Initialize the whole return value, for example ``struct pair p = {};``, to
+avoid this. A static subprogram is exempt from the scalar-only rule: it is
+verified inline, so an unassigned R2 is simply passed back to the caller as
+uninitialized and only a caller that reads it fails. A stack pointer left in
+R2 is still rejected there, just as one in R0 is.
+
 .. _BPF_kfunc_lifecycle_expectations:
 
 3. kfunc lifecycle expectations
diff --git a/arch/arm64/net/bpf_jit_comp.c b/arch/arm64/net/bpf_jit_comp.c
index c18e005a41dbe..3aa3ea0bc30bb 100644
--- a/arch/arm64/net/bpf_jit_comp.c
+++ b/arch/arm64/net/bpf_jit_comp.c
@@ -2388,6 +2388,11 @@ bool bpf_jit_supports_kfunc_call(void)
 	return true;
 }
 
+bool bpf_jit_supports_kfunc_ret_reg_pair(void)
+{
+	return true;
+}
+
 bool bpf_jit_supports_stack_args(void)
 {
 	return true;
diff --git a/arch/riscv/net/bpf_jit_comp64.c b/arch/riscv/net/bpf_jit_comp64.c
index 74efe4b138d2a..47c7bf431ba85 100644
--- a/arch/riscv/net/bpf_jit_comp64.c
+++ b/arch/riscv/net/bpf_jit_comp64.c
@@ -2121,6 +2121,11 @@ bool bpf_jit_supports_kfunc_call(void)
 	return true;
 }
 
+bool bpf_jit_supports_kfunc_ret_reg_pair(void)
+{
+	return true;
+}
+
 bool bpf_jit_supports_ptr_xchg(void)
 {
 	return true;
diff --git a/arch/x86/net/bpf_jit_comp.c b/arch/x86/net/bpf_jit_comp.c
index 1a9fb530adc3c..48429fae06410 100644
--- a/arch/x86/net/bpf_jit_comp.c
+++ b/arch/x86/net/bpf_jit_comp.c
@@ -1689,17 +1689,12 @@ static int emit_spectre_bhb_barrier(u8 **pprog, u8 *ip,
  * arena NULL is offset 0. Return the number of emitted bytes.
  */
 static int emit_kfunc_arena_args(struct bpf_prog *bpf_prog,
-				 const struct bpf_insn *insn, u8 **pprog)
+				 const struct btf_func_model *fm, u8 **pprog)
 {
-	const struct btf_func_model *fm;
 	u8 *prog = *pprog;
 	u8 *start = prog;
 	int i;
 
-	fm = bpf_jit_find_kfunc_model(bpf_prog, insn);
-	if (!fm)
-		return -EINVAL;
-
 	for (i = 0; i < min_t(int, fm->nr_args, MAX_BPF_FUNC_REG_ARGS); i++) {
 		u8 flags = fm->arg_flags[i];
 		u32 reg = BPF_REG_1 + i;
@@ -2644,6 +2639,8 @@ st:			insn_off = insn->off;
 
 			/* call */
 		case BPF_JMP | BPF_CALL: {
+			const struct btf_func_model *fm = NULL;
+
 			func = (u8 *) __bpf_call_base + imm32;
 			if (src_reg == BPF_PSEUDO_CALL && tail_call_reachable) {
 				LOAD_TAIL_CALL_CNT_PTR(stack_depth);
@@ -2652,7 +2649,10 @@ st:			insn_off = insn->off;
 			if (!imm32)
 				return -EINVAL;
 			if (src_reg == BPF_PSEUDO_KFUNC_CALL) {
-				err = emit_kfunc_arena_args(bpf_prog, insn, &prog);
+				fm = bpf_jit_find_kfunc_model(bpf_prog, insn);
+				if (!fm)
+					return -EINVAL;
+				err = emit_kfunc_arena_args(bpf_prog, fm, &prog);
 				if (err < 0)
 					return err;
 				ip += err;
@@ -2666,6 +2666,14 @@ st:			insn_off = insn->off;
 				return -EINVAL;
 			if (priv_frame_ptr)
 				pop_r9(&prog);
+			/*
+			 * A kfunc returning more than 8 bytes hands the second
+			 * half back in RDX (the native ABI's second return reg),
+			 * but BPF expects it in R0:R2. BPF R0 is RAX (no move
+			 * needed), while BPF R2 is RSI, so copy RDX into RSI.
+			 */
+			if (fm && fm->ret_size > 8)
+				emit_mov_reg(&prog, true, BPF_REG_2, BPF_REG_3);
 			break;
 		}
 
@@ -4156,6 +4164,11 @@ bool bpf_jit_supports_kfunc_call(void)
 	return true;
 }
 
+bool bpf_jit_supports_kfunc_ret_reg_pair(void)
+{
+	return true;
+}
+
 bool bpf_jit_supports_stack_args(void)
 {
 	return true;
diff --git a/include/linux/bpf_verifier.h b/include/linux/bpf_verifier.h
index bc2af02547fe5..938c9a9eb9d21 100644
--- a/include/linux/bpf_verifier.h
+++ b/include/linux/bpf_verifier.h
@@ -819,6 +819,8 @@ struct bpf_subprog_info {
 	bool is_async_cb: 1;
 	bool is_exception_cb: 1;
 	bool args_cached: 1;
+	/* true if the return value is passed in the R0:R2 register pair */
+	bool ret_reg_pair: 1;
 	/* true if bpf_fastcall stack region is used by functions that can't be inlined */
 	bool keep_fastcall_stack: 1;
 	bool changes_pkt_data: 1;
@@ -1055,6 +1057,11 @@ static inline struct bpf_subprog_info *subprog_info(struct bpf_verifier_env *env
 	return &env->subprog_info[subprog];
 }
 
+static inline bool bpf_ret_reg_pair(struct bpf_verifier_env *env, int subprog)
+{
+	return subprog_info(env, subprog)->ret_reg_pair;
+}
+
 struct bpf_call_summary {
 	u8 num_params;
 	bool is_void;
@@ -1460,6 +1467,8 @@ int bpf_jmp_offset(struct bpf_insn *insn);
 struct bpf_iarray *bpf_insn_successors(struct bpf_verifier_env *env, u32 idx);
 void bpf_fmt_stack_mask(char *buf, ssize_t buf_sz, u64 stack_mask);
 bool bpf_subprog_is_global(const struct bpf_verifier_env *env, int subprog);
+bool btf_type_is_scalar_struct(struct bpf_verifier_env *env, const struct btf *btf,
+			       const struct btf_type *t, int rec);
 
 int bpf_find_subprog(struct bpf_verifier_env *env, int off);
 bool bpf_is_throw_kfunc(struct bpf_insn *insn);
diff --git a/include/linux/filter.h b/include/linux/filter.h
index 4a9bc6a848f2e..6e746b0a09306 100644
--- a/include/linux/filter.h
+++ b/include/linux/filter.h
@@ -1237,6 +1237,7 @@ bool bpf_jit_inlines_helper_call(s32 imm);
 bool bpf_jit_supports_subprog_tailcalls(void);
 bool bpf_jit_supports_percpu_insn(void);
 bool bpf_jit_supports_kfunc_call(void);
+bool bpf_jit_supports_kfunc_ret_reg_pair(void);
 bool bpf_jit_supports_stack_args(void);
 bool bpf_jit_supports_arena_args(void);
 bool bpf_jit_supports_far_kfunc_call(void);
diff --git a/kernel/bpf/backtrack.c b/kernel/bpf/backtrack.c
index a2b18a9f1694c..653db80bcc47b 100644
--- a/kernel/bpf/backtrack.c
+++ b/kernel/bpf/backtrack.c
@@ -423,6 +423,10 @@ static int backtrack_insn(struct bpf_verifier_env *env, int idx, int subseq_idx,
 				 */
 				verifier_bug_if(idx + 1 != subseq_idx, env,
 						"extra insn from subprog");
+				/* global subprog always sets R0 */
+				bt_clear_reg(bt, BPF_REG_0);
+				/* and if it does not set R2, main pass would catch it */
+				bt_clear_reg(bt, BPF_REG_2);
 				/* r1-r5 are invalidated after subprog call,
 				 * so for global func call it shouldn't be set
 				 * anymore
@@ -432,8 +436,6 @@ static int backtrack_insn(struct bpf_verifier_env *env, int idx, int subseq_idx,
 						     bt_reg_mask(bt));
 					return -EFAULT;
 				}
-				/* global subprog always sets R0 */
-				bt_clear_reg(bt, BPF_REG_0);
 				return 0;
 			} else {
 				/* static subprog call instruction, which
@@ -506,6 +508,8 @@ static int backtrack_insn(struct bpf_verifier_env *env, int idx, int subseq_idx,
 				return -ENOTSUPP;
 			/* regular helper call sets R0 */
 			bt_clear_reg(bt, BPF_REG_0);
+			/* kfunc might also set R2 */
+			bt_clear_reg(bt, BPF_REG_2);
 			if (bt_reg_mask(bt) & BPF_REGMASK_ARGS) {
 				/* if backtracking was looking for registers R1-R5
 				 * they should have been found already.
@@ -520,7 +524,25 @@ static int backtrack_insn(struct bpf_verifier_env *env, int idx, int subseq_idx,
 					return -EFAULT;
 			}
 		} else if (opcode == BPF_EXIT) {
-			bool r0_precise;
+			bool from_subprog_call, r0_precise, r2_precise;
+
+			/* BPF_EXIT in subprog or callback always returns
+			 * right after the call instruction, so by checking
+			 * whether the instruction at subseq_idx-1 is subprog
+			 * call or not we can distinguish actual exit from
+			 * *subprog* from exit from *callback*. In the former
+			 * case, we need to propagate the precision of the
+			 * return registers, if necessary. In the latter we
+			 * never do that.
+			 */
+			from_subprog_call = subseq_idx - 1 >= 0 &&
+					    bpf_pseudo_call(&env->prog->insnsi[subseq_idx - 1]);
+
+			/* Sample the return registers before the callback
+			 * handling below clears R1-R5.
+			 */
+			r0_precise = from_subprog_call && bt_is_reg_set(bt, BPF_REG_0);
+			r2_precise = from_subprog_call && bt_is_reg_set(bt, BPF_REG_2);
 
 			/* Backtracking to a nested function call, 'idx' is a part of
 			 * the inner frame 'subseq_idx' is a part of the outer frame.
@@ -533,30 +555,22 @@ static int backtrack_insn(struct bpf_verifier_env *env, int idx, int subseq_idx,
 			if (subseq_idx >= 0 && bpf_calls_callback(env, subseq_idx))
 				for (i = BPF_REG_1; i <= BPF_REG_5; i++)
 					bt_clear_reg(bt, i);
+
+			bt_clear_reg(bt, BPF_REG_0);
+			bt_clear_reg(bt, BPF_REG_2);
 			if (bt_reg_mask(bt) & BPF_REGMASK_ARGS) {
 				verifier_bug(env, "backtracking exit unexpected regs %x",
 					     bt_reg_mask(bt));
 				return -EFAULT;
 			}
 
-			/* BPF_EXIT in subprog or callback always returns
-			 * right after the call instruction, so by checking
-			 * whether the instruction at subseq_idx-1 is subprog
-			 * call or not we can distinguish actual exit from
-			 * *subprog* from exit from *callback*. In the former
-			 * case, we need to propagate r0 precision, if
-			 * necessary. In the former we never do that.
-			 */
-			r0_precise = subseq_idx - 1 >= 0 &&
-				     bpf_pseudo_call(&env->prog->insnsi[subseq_idx - 1]) &&
-				     bt_is_reg_set(bt, BPF_REG_0);
-
-			bt_clear_reg(bt, BPF_REG_0);
 			if (bt_subprog_enter(bt))
 				return -EFAULT;
 
 			if (r0_precise)
 				bt_set_reg(bt, BPF_REG_0);
+			if (r2_precise)
+				bt_set_reg(bt, BPF_REG_2);
 			/* r6-r9 and stack slots will stay set in caller frame
 			 * bitmasks until we return back from callee(s)
 			 */
diff --git a/kernel/bpf/btf.c b/kernel/bpf/btf.c
index 5b9d767895c91..ba83fa1d52a38 100644
--- a/kernel/bpf/btf.c
+++ b/kernel/bpf/btf.c
@@ -7591,7 +7591,7 @@ int btf_distill_func_proto(struct bpf_verifier_log *log,
 		return -EINVAL;
 	}
 	ret = __get_type_size(btf, func->type, &t);
-	if (ret < 0 || btf_type_is_struct(t)) {
+	if (ret < 0 || ret > 16) {
 		bpf_log(log,
 			"The function %s return type %s is unsupported.\n",
 			tname, btf_type_str(t));
@@ -7684,6 +7684,12 @@ static int btf_check_func_type_match(struct bpf_verifier_log *log,
 			btf_type_str(t2), fn2);
 		return -EINVAL;
 	}
+	if (btf_type_has_size(t1) && (t1->size > 8 || t2->size > 8)) {
+		bpf_log(log,
+			"Return type of %s() has size %u while %s() has size %u, and a size above 8 bytes cannot be replaced\n",
+			fn1, t1->size, fn2, t2->size);
+		return -EINVAL;
+	}
 
 	for (i = 0; i < nargs1; i++) {
 		t1 = btf_type_skip_modifiers(btf1, args1[i].type, NULL);
@@ -7964,7 +7970,7 @@ static int btf_scan_type_tags(struct bpf_verifier_env *env,
 
 /* Check whether the type is a valid return type. */
 static int btf_validate_return_type(struct bpf_verifier_env *env, struct btf *btf,
-		const struct btf_type *t, int subprog)
+		const struct btf_type *t, int subprog, bool is_global)
 {
 	u32 tags = 0;
 	int err;
@@ -7987,6 +7993,19 @@ static int btf_validate_return_type(struct bpf_verifier_env *env, struct btf *bt
 	if (btf_type_is_void(t) || btf_type_is_int(t) || btf_is_any_enum(t))
 		return 0;
 
+	if (btf_type_is_struct(t) && t->size <= 16) {
+		/*
+		 * A global function's caller models the return as an opaque
+		 * scalar pair, so it may only return scalars by value. A local
+		 * function is verified inline, so a pointer field stays tracked
+		 * and needs no such restriction.
+		 */
+		bool local_func = subprog && !is_global;
+
+		if (local_func || btf_type_is_scalar_struct(env, btf, t, 0))
+			return 0;
+	}
+
 	return -EOPNOTSUPP;
 }
 
@@ -8074,12 +8093,12 @@ int btf_prepare_func_args(struct bpf_verifier_env *env, int subprog)
 		return -EINVAL;
 	}
 
-	err = btf_validate_return_type(env, btf, t, subprog);
+	err = btf_validate_return_type(env, btf, t, subprog, is_global);
 	if (err) {
 		if (is_global) {
 			bpf_log(log,
-				"Global function %s() return value not void or scalar. "
-				"Only those are supported.\n",
+				"Global function %s() has unsupported return type. "
+				"Only void, scalar, or a scalar-only struct/union up to 16 bytes is supported.\n",
 				tname);
 		}
 		return err;
diff --git a/kernel/bpf/core.c b/kernel/bpf/core.c
index d55e737ed75ad..5db77d7915dfc 100644
--- a/kernel/bpf/core.c
+++ b/kernel/bpf/core.c
@@ -3287,6 +3287,11 @@ bool __weak bpf_jit_supports_kfunc_call(void)
 	return false;
 }
 
+bool __weak bpf_jit_supports_kfunc_ret_reg_pair(void)
+{
+	return false;
+}
+
 bool __weak bpf_jit_supports_stack_args(void)
 {
 	return false;
diff --git a/kernel/bpf/liveness.c b/kernel/bpf/liveness.c
index 74fc4b3f80d6e..71f998c6eb888 100644
--- a/kernel/bpf/liveness.c
+++ b/kernel/bpf/liveness.c
@@ -2060,7 +2060,8 @@ static inline u16 mask_hi(u32 m) { return (u16)(m >> 16); }
 /* Compute info->{use,def} fields for the instruction */
 static void compute_insn_live_regs(struct bpf_verifier_env *env,
 				   struct bpf_insn *insn,
-				   struct insn_live_regs *info)
+				   struct insn_live_regs *info,
+				   bool ret_reg_pair)
 {
 	struct bpf_call_summary cs;
 	const u8 class = BPF_CLASS(insn->code);
@@ -2072,6 +2073,7 @@ static void compute_insn_live_regs(struct bpf_verifier_env *env,
 	const u32 src32 = mask_lo(src);
 	const u32 dst32 = mask_lo(dst);
 	const u32 r0  = reg64_mask(0);
+	const u32 r2  = reg64_mask(BPF_REG_2);
 	u32 def = 0;
 	u32 use = U32_MAX;
 
@@ -2191,7 +2193,7 @@ static void compute_insn_live_regs(struct bpf_verifier_env *env,
 			break;
 		case BPF_EXIT:
 			def = 0;
-			use = r0;
+			use = ret_reg_pair ? (r0 | r2) : r0;
 			break;
 		case BPF_CALL:
 			def = ALL_CALLER_SAVED_REGS;
@@ -2228,8 +2230,8 @@ int bpf_compute_live_registers(struct bpf_verifier_env *env)
 	struct insn_live_regs *state;
 	int insn_cnt = env->prog->len;
 	u64 pos, insn_pos;
-	int err = 0, i, j;
-	bool changed;
+	int err = 0, i, j, subprog, start, end;
+	bool changed, ret_reg_pair;
 
 	/* Use the following algorithm:
 	 * - define the following:
@@ -2256,8 +2258,14 @@ int bpf_compute_live_registers(struct bpf_verifier_env *env)
 		goto out;
 	}
 
-	for (i = 0; i < insn_cnt; ++i)
-		compute_insn_live_regs(env, &insns[i], &state[i]);
+	for (subprog = 0; subprog < env->subprog_cnt; subprog++) {
+		start = env->subprog_info[subprog].start;
+		end = env->subprog_info[subprog + 1].start;
+		ret_reg_pair = bpf_ret_reg_pair(env, subprog);
+
+		for (i = start; i < end; ++i)
+			compute_insn_live_regs(env, &insns[i], &state[i], ret_reg_pair);
+	}
 
 	/* Forward pass: resolve stack access through FP-derived pointers */
 	err = bpf_compute_subprog_arg_access(env);
diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
index d17f14b35b79f..edc1b4a19f64b 100644
--- a/kernel/bpf/verifier.c
+++ b/kernel/bpf/verifier.c
@@ -385,27 +385,75 @@ bool bpf_subprog_is_global(const struct bpf_verifier_env *env, int subprog)
 	return aux && aux[subprog].linkage == BTF_FUNC_GLOBAL;
 }
 
-static bool subprog_returns_void(struct bpf_verifier_env *env, int subprog)
+static const struct btf_type *subprog_ret_type(struct bpf_verifier_env *env, int subprog)
 {
-	const struct btf_type *type, *func, *func_proto;
+	const struct btf_type *func, *func_proto;
 	const struct btf *btf = env->prog->aux->btf;
 	u32 btf_id;
 
+	if (!btf || !env->prog->aux->func_info)
+		return NULL;
+
 	btf_id = env->prog->aux->func_info[subprog].type_id;
 
+	/* Both already validated by prepare_btf_func() at prog load. */
 	func = btf_type_by_id(btf, btf_id);
-	if (verifier_bug_if(!func, env, "btf_id %u not found", btf_id))
-		return false;
-
 	func_proto = btf_type_by_id(btf, func->type);
-	if (!func_proto)
-		return false;
 
-	type = btf_type_skip_modifiers(btf, func_proto->type, NULL);
-	if (!type)
-		return false;
+	return btf_type_skip_modifiers(btf, func_proto->type, NULL);
+}
+
+static bool subprog_returns_void(struct bpf_verifier_env *env, int subprog)
+{
+	const struct btf_type *type = subprog_ret_type(env, subprog);
+
+	return type && btf_type_is_void(type);
+}
+
+static u32 ret_regs_cnt(u32 size)
+{
+	return size > 8 && size <= 16 ? 2 : 1;
+}
+
+/* Registers holding a function return value, in order. See ret_regs_cnt(). */
+static const int ret_regs[] = { BPF_REG_0, BPF_REG_2 };
+
+static int bpf_compute_subprog_ret_regs(struct bpf_verifier_env *env)
+{
+	const struct btf *btf = env->prog->aux->btf;
+	const struct btf_type *type;
+	int subprog;
+	u32 size;
 
-	return btf_type_is_void(type);
+	if (!env->prog->jit_requested || bpf_prog_is_offloaded(env->prog->aux))
+		return 0;
+
+	/*
+	 * Skip the main program: its return value is the program's exit code,
+	 * read out of R0, so it never uses the register pair. An extension does
+	 * have a real prototype for subprog 0, but bpf_check_attach_target()
+	 * refuses to replace a function returning more than 8 bytes.
+	 */
+	for (subprog = 1; subprog < env->subprog_cnt; subprog++) {
+		type = subprog_ret_type(env, subprog);
+		/*
+		 * This runs before btf_validate_return_type(), so apply the same
+		 * type filter here. BTF that no compiler would emit can declare
+		 * a return type that validation rejects, an array or a float
+		 * say, which btf_resolve_size() would happily size at 16 bytes.
+		 */
+		if (!type || !(btf_type_is_struct(type) || btf_type_is_scalar(type)))
+			continue;
+		if (verifier_bug_if(IS_ERR(btf_resolve_size(btf, type, &size)), env,
+				    "cannot size return type of subprog %d", subprog))
+			return -EFAULT;
+		if (ret_regs_cnt(size) > 1) {
+			subprog_info(env, subprog)->ret_reg_pair = true;
+			env->prog->jit_required = 1;
+		}
+	}
+
+	return 0;
 }
 
 const char *bpf_subprog_name(const struct bpf_verifier_env *env, int subprog)
@@ -2877,6 +2925,18 @@ int bpf_add_kfunc_call(struct bpf_verifier_env *env, u32 func_id, u16 offset)
 	err = btf_distill_func_proto(&env->log, kfunc.btf, kfunc.proto, kfunc.name, &func_model);
 	if (err)
 		return err;
+	if (func_model.ret_size > 8) {
+		if (kfunc.flags && (*kfunc.flags & KF_FASTCALL)) {
+			verbose(env, "kfunc %s with >8-byte return is not supported with KF_FASTCALL\n",
+				kfunc.name);
+			return -EOPNOTSUPP;
+		}
+		if (!bpf_jit_supports_kfunc_ret_reg_pair()) {
+			verbose(env, "kfunc %s with >8-byte return is not supported by JIT\n",
+				kfunc.name);
+			return -EOPNOTSUPP;
+		}
+	}
 
 	memset(&meta, 0, sizeof(meta));
 	meta.btf = kfunc.btf;
@@ -9855,6 +9915,7 @@ static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
 	u16 callee_incoming, stack_arg_cnt;
 	struct bpf_func_state *caller;
 	int err, subprog, target_insn;
+	u32 i, nregs;
 
 	target_insn = *insn_idx + insn->imm + 1;
 	subprog = bpf_find_subprog(env, target_insn);
@@ -9910,9 +9971,14 @@ static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
 		clear_caller_saved_regs(env, caller->regs);
 		invalidate_outgoing_stack_args(env, cur_func(env));
 
-		/* All non-void global functions return a 64-bit SCALAR_VALUE. */
+		/*
+		 * A non-void global function returns a 64-bit SCALAR_VALUE in
+		 * R0, or a >8 byte SCALAR_VALUE in the R0:R2 register pair.
+		 */
 		if (!returns_void) {
-			mark_reg_unknown(env, caller->regs, BPF_REG_0);
+			nregs = bpf_ret_reg_pair(env, subprog) ? 2 : 1;
+			for (i = 0; i < nregs; i++)
+				mark_reg_unknown(env, caller->regs, ret_regs[i]);
 			bpf_diag_mod_end(env);
 		}
 
@@ -10272,11 +10338,15 @@ static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
 	struct bpf_func_state *caller, *callee;
 	struct bpf_reg_state *r0;
 	bool in_callback_fn;
+	u32 i, nregs;
 	int err;
 
 	callee = state->frame[state->curframe];
 	r0 = &callee->regs[BPF_REG_0];
-	if (r0->type == PTR_TO_STACK) {
+	nregs = bpf_ret_reg_pair(env, callee->subprogno) ? 2 : 1;
+	for (i = 0; i < nregs; i++) {
+		if (callee->regs[ret_regs[i]].type != PTR_TO_STACK)
+			continue;
 		/* technically it's ok to return caller's stack pointer
 		 * (or caller's caller's pointer) back to the caller,
 		 * since these pointers are valid. Only current stack
@@ -10311,9 +10381,13 @@ static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
 			return -EFAULT;
 		}
 	} else {
-		/* return to the caller whatever r0 had in the callee */
+		/*
+		 * return to the caller whatever the callee had in the
+		 * return register(s)
+		 */
 		bpf_diag_mod_begin(env, &caller->regs[BPF_REG_0], r0, BPF_DIAG_MOD_WRITE);
-		caller->regs[BPF_REG_0] = *r0;
+		for (i = 0; i < nregs; i++)
+			caller->regs[ret_regs[i]] = callee->regs[ret_regs[i]];
 		bpf_diag_mod_end(env);
 	}
 
@@ -11248,6 +11322,19 @@ static int check_helper_call(struct bpf_verifier_env *env, struct bpf_insn *insn
 	return 0;
 }
 
+/*
+ * Mark the register(s) holding a @size byte kfunc return value as unknown
+ * scalars. Both halves of a register pair are treated the same way.
+ */
+static void mark_kfunc_ret_regs(struct bpf_verifier_env *env,
+				struct bpf_reg_state *regs, u32 size)
+{
+	u32 i, nregs = ret_regs_cnt(size);
+
+	for (i = 0; i < nregs; i++)
+		mark_reg_unknown(env, regs, ret_regs[i]);
+}
+
 static bool is_kfunc_acquire(struct bpf_call_arg_meta *meta)
 {
 	return meta->kfunc_flags & KF_ACQUIRE;
@@ -11515,9 +11602,9 @@ static bool is_kfunc_arg_implicit(const struct bpf_call_arg_meta *meta, u32 arg_
 }
 
 /* Returns true if struct is composed of scalars, 4 levels of nesting allowed */
-static bool __btf_type_is_scalar_struct(struct bpf_verifier_env *env,
-					const struct btf *btf,
-					const struct btf_type *t, int rec)
+bool btf_type_is_scalar_struct(struct bpf_verifier_env *env,
+			       const struct btf *btf,
+			       const struct btf_type *t, int rec)
 {
 	const struct btf_type *member_type;
 	const struct btf_member *member;
@@ -11535,7 +11622,7 @@ static bool __btf_type_is_scalar_struct(struct bpf_verifier_env *env,
 				verbose(env, "max struct nesting depth exceeded\n");
 				return false;
 			}
-			if (!__btf_type_is_scalar_struct(env, btf, member_type, rec + 1))
+			if (!btf_type_is_scalar_struct(env, btf, member_type, rec + 1))
 				return false;
 			continue;
 		}
@@ -11934,7 +12021,7 @@ get_kfunc_arg_type(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
 		 (is_kfunc_arg_mem_size(meta->btf, &args[arg + 1]) ||
 		  is_kfunc_arg_const_mem_size(meta->btf, &args[arg + 1]))) {
 		if (!btf_type_is_void(ref_t) && !btf_type_is_scalar(ref_t) &&
-		    !__btf_type_is_scalar_struct(env, meta->btf, ref_t, 0)) {
+		    !btf_type_is_scalar_struct(env, meta->btf, ref_t, 0)) {
 			verbose(env, "%s pointer type %s %s must point to void, scalar, or struct with scalar\n",
 				reg_arg_name(env, argno), btf_type_str(ref_t), ref_tname);
 			return -EINVAL;
@@ -11950,7 +12037,7 @@ get_kfunc_arg_type(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
 		 * scalars. The access size is derived from the pointed-to BTF type.
 		 */
 		if (!btf_type_is_scalar(ref_t) &&
-		    !__btf_type_is_scalar_struct(env, meta->btf, ref_t, 0)) {
+		    !btf_type_is_scalar_struct(env, meta->btf, ref_t, 0)) {
 			verbose(env, "%s pointer type %s %s must point to scalar, or struct with scalar\n",
 				reg_arg_name(env, argno), btf_type_str(ref_t), ref_tname);
 			return -EINVAL;
@@ -13006,7 +13093,7 @@ static int check_kfunc_args(struct bpf_verifier_env *env, struct bpf_call_arg_me
 				break;
 			}
 
-			if (!__btf_type_is_scalar_struct(env, meta->btf, ref_t, 0)) {
+			if (!btf_type_is_scalar_struct(env, meta->btf, ref_t, 0)) {
 				enum bpf_reg_type reg2btf_type = lookup_reg2btf_ids(ref_id);
 				const char *expected_type;
 
@@ -13544,7 +13631,7 @@ static int check_special_kfunc(struct bpf_verifier_env *env, struct bpf_call_arg
 
 		struct_meta = btf_find_struct_meta(ret_btf, ret_btf_id);
 		if (is_bpf_percpu_obj_new_kfunc(meta->func_id)) {
-			if (!__btf_type_is_scalar_struct(env, ret_btf, ret_t, 0)) {
+			if (!btf_type_is_scalar_struct(env, ret_btf, ret_t, 0)) {
 				verbose(env, "bpf_percpu_obj_new type ID argument must be of a struct of scalars\n");
 				return -EINVAL;
 			}
@@ -13912,10 +13999,25 @@ static int check_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
 	}
 
 	if (btf_type_is_scalar(t)) {
-		mark_reg_unknown(env, regs, BPF_REG_0);
+		mark_kfunc_ret_regs(env, regs, t->size);
 		if (meta.btf == btf_vmlinux && (meta.func_id == special_kfunc_list[KF_bpf_res_spin_lock] ||
 		    meta.func_id == special_kfunc_list[KF_bpf_res_spin_lock_irqsave]))
 			__mark_reg_const_zero(env, &regs[BPF_REG_0]);
+	} else if (btf_type_is_struct(t)) {
+		/*
+		 * The returned struct comes back as raw register bits modeled
+		 * as an unknown scalar, so it must contain only scalars:
+		 * otherwise a pointer field would be laundered into a scalar
+		 * and escape provenance and reference tracking.
+		 */
+		if (!btf_type_is_scalar_struct(env, desc_btf, t, 0)) {
+			verbose(env,
+				"kernel function %s returns %s %s that is not composed of scalars\n",
+				func_name, btf_type_str(t),
+				btf_name_by_offset(desc_btf, t->name_off));
+			return -EINVAL;
+		}
+		mark_kfunc_ret_regs(env, regs, t->size);
 	} else if (btf_type_is_ptr(t)) {
 		ptr_type = btf_type_skip_modifiers(desc_btf, t->type, &ptr_type_id);
 		err = check_special_kfunc(env, &meta, regs, insn_aux, ptr_type, desc_btf);
@@ -17451,37 +17553,54 @@ static int check_return_code(struct bpf_verifier_env *env, int regno, const char
 	return 0;
 }
 
-static int check_global_subprog_return_code(struct bpf_verifier_env *env)
+static int check_global_ret_scalar_reg(struct bpf_verifier_env *env, u32 regno)
 {
-	struct bpf_reg_state *reg = reg_state(env, BPF_REG_0);
-	struct bpf_func_state *cur_frame = cur_func(env);
+	struct bpf_reg_state *reg;
 	int err;
 
-	if (subprog_returns_void(env, cur_frame->subprogno))
-		return 0;
-
-	err = check_reg_arg(env, BPF_REG_0, SRC_OP);
+	err = check_reg_arg(env, regno, SRC_OP);
 	if (err)
 		return err;
 
 	/* Pointers to arena are safe to pass between subprograms. */
-	if (is_arena_reg(env, BPF_REG_0))
+	if (is_arena_reg(env, regno))
 		return 0;
 
-	if (is_pointer_value(env, BPF_REG_0)) {
-		verbose(env, "R%d leaks addr as return value\n", BPF_REG_0);
+	if (is_pointer_value(env, regno)) {
+		verbose(env, "R%d leaks addr as return value\n", regno);
 		return -EACCES;
 	}
 
+	reg = reg_state(env, regno);
 	if (reg->type != SCALAR_VALUE) {
-		verbose(env, "At subprogram exit the register R0 is not a scalar value (%s)\n",
-			reg_type_str(env, reg->type));
+		verbose(env, "At subprogram exit the register R%d is not a scalar value (%s)\n",
+			regno, reg_type_str(env, reg->type));
 		return -EINVAL;
 	}
 
 	return 0;
 }
 
+static int check_global_subprog_return_code(struct bpf_verifier_env *env)
+{
+	struct bpf_func_state *cur_frame = cur_func(env);
+	u32 subprog = cur_frame->subprogno;
+	u32 i, nregs;
+	int err;
+
+	if (subprog_returns_void(env, subprog))
+		return 0;
+
+	nregs = bpf_ret_reg_pair(env, subprog) ? 2 : 1;
+	for (i = 0; i < nregs; i++) {
+		err = check_global_ret_scalar_reg(env, ret_regs[i]);
+		if (err)
+			return err;
+	}
+
+	return 0;
+}
+
 /* Bitmask with 1s for all caller saved registers */
 #define ALL_CALLER_SAVED_REGS ((1u << CALLER_SAVED_REGS) - 1)
 
@@ -21153,6 +21272,11 @@ int bpf_check(struct bpf_prog **prog, union bpf_attr *attr, bpfptr_t uattr,
 	if (ret < 0)
 		goto skip_full_check;
 
+	/* must precede the first bpf_ret_reg_pair() user below */
+	ret = bpf_compute_subprog_ret_regs(env);
+	if (ret < 0)
+		goto skip_full_check;
+
 	ret = bpf_compute_live_registers(env);
 	if (ret < 0)
 		goto skip_full_check;
diff --git a/tools/testing/selftests/bpf/prog_tests/aggregate_ret.c b/tools/testing/selftests/bpf/prog_tests/aggregate_ret.c
new file mode 100644
index 0000000000000..e0b94ed10f941
--- /dev/null
+++ b/tools/testing/selftests/bpf/prog_tests/aggregate_ret.c
@@ -0,0 +1,11 @@
+// SPDX-License-Identifier: GPL-2.0
+/* Copyright (c) 2026 Meta Platforms, Inc. and affiliates. */
+#include <test_progs.h>
+#include "aggregate_ret_func.skel.h"
+#include "aggregate_ret_kfunc.skel.h"
+
+void test_aggregate_ret(void)
+{
+	RUN_TESTS(aggregate_ret_func);
+	RUN_TESTS(aggregate_ret_kfunc);
+}
diff --git a/tools/testing/selftests/bpf/prog_tests/fexit_bpf2bpf.c b/tools/testing/selftests/bpf/prog_tests/fexit_bpf2bpf.c
index 2523c07a16c65..6c438df380bb1 100644
--- a/tools/testing/selftests/bpf/prog_tests/fexit_bpf2bpf.c
+++ b/tools/testing/selftests/bpf/prog_tests/fexit_bpf2bpf.c
@@ -441,6 +441,20 @@ static void test_func_replace_int_with_void(void)
 				     " doesn't match type INT of global_func2()");
 }
 
+static void test_func_replace_ret_pair(void)
+{
+	const char *msg = "Return type of new_agg_ret_target_func() has size 8 "
+			  "while agg_ret_target_func() has size 16";
+
+	/*
+	 * An extension cannot replace a function whose return value comes back
+	 * in the R0:R2 pair: only R0 is checked at the extension's exit, so it
+	 * would leave R2 stale for the target's callers.
+	 */
+	test_obj_load_failure_common("freplace_ret_pair.bpf.o",
+				     "./aggregate_ret_target.bpf.o", msg);
+}
+
 static int find_prog_btf_id(const char *name, __u32 attach_prog_fd)
 {
 	struct bpf_prog_info info = {};
@@ -660,6 +674,8 @@ void serial_test_fexit_bpf2bpf(void)
 		test_func_replace_progmap();
 	if (test__start_subtest("freplace_int_with_void"))
 		test_func_replace_int_with_void();
+	if (test__start_subtest("freplace_ret_pair"))
+		test_func_replace_ret_pair();
 	if (test__start_subtest("freplace_void"))
 		test_func_replace_void();
 	if (test__start_subtest("sleepable_fentry_to_xdp"))
diff --git a/tools/testing/selftests/bpf/prog_tests/verifier.c b/tools/testing/selftests/bpf/prog_tests/verifier.c
index 64ac49ad67e63..f7f94ccebce27 100644
--- a/tools/testing/selftests/bpf/prog_tests/verifier.c
+++ b/tools/testing/selftests/bpf/prog_tests/verifier.c
@@ -5,6 +5,7 @@
 #include "arena_kfunc.skel.h"
 #include "arena_kfunc_jit.skel.h"
 #include "cap_helpers.h"
+#include "verifier_aggregate_ret.skel.h"
 #include "verifier_align.skel.h"
 #include "verifier_and.skel.h"
 #include "verifier_arena.skel.h"
@@ -170,6 +171,7 @@ void test_arena_kfunc(void)                   { RUN_TESTS(arena_kfunc); }
 
 void test_arena_kfunc_jit(void)               { RUN_TESTS(arena_kfunc_jit); }
 
+void test_verifier_aggregate_ret(void)        { RUN_TESTS(verifier_aggregate_ret); }
 void test_verifier_align(void)                { RUN(verifier_align); }
 void test_verifier_and(void)                  { RUN(verifier_and); }
 void test_verifier_arena(void)                { RUN(verifier_arena); }
diff --git a/tools/testing/selftests/bpf/progs/aggregate_ret_func.c b/tools/testing/selftests/bpf/progs/aggregate_ret_func.c
new file mode 100644
index 0000000000000..e35baaa10fea2
--- /dev/null
+++ b/tools/testing/selftests/bpf/progs/aggregate_ret_func.c
@@ -0,0 +1,235 @@
+// SPDX-License-Identifier: GPL-2.0
+/* Copyright (c) 2026 Meta Platforms, Inc. and affiliates. */
+#include <linux/bpf.h>
+#include <bpf/bpf_helpers.h>
+#include "bpf_misc.h"
+
+typedef unsigned __int128 u128;
+
+__naked u128 global_agg_good(void)
+{
+	asm volatile (
+	"r0 = 0x1234;"	/* low 64 bits */
+	"r2 = 0x5678;"	/* high 64 bits */
+	"exit;"
+	);
+}
+
+__naked u128 global_agg_bad(void)
+{
+	asm volatile (
+	"r0 = 0;"
+	"exit;"
+	);
+}
+
+__naked u128 global_agg_bad_ptr(void)
+{
+	asm volatile (
+	"r0 = 0;"
+	"r2 = r10;"
+	"exit;"
+	);
+}
+
+SEC("tc")
+__failure __msg("R2 !read_ok")
+__naked int aggregate_ret_global_fail(void)
+{
+	asm volatile (
+	"call %[global_agg_bad];"
+	"r0 = r2;"
+	"exit;"
+	:
+	: __imm(global_agg_bad)
+	: __clobber_all);
+}
+
+SEC("tc")
+__failure __msg("At subprogram exit the register R2 is not a scalar value")
+__naked int aggregate_ret_global_ptr_fail(void)
+{
+	asm volatile (
+	"call %[global_agg_bad_ptr];"
+	"r0 = r2;"
+	"exit;"
+	:
+	: __imm(global_agg_bad_ptr)
+	: __clobber_all);
+}
+
+static __naked __noinline u128 static_agg_bad_ptr(void)
+{
+	asm volatile (
+	"r0 = 0;"
+	"r2 = r10;"	/* stack pointer placed in the second return register */
+	"exit;"
+	);
+}
+
+/*
+ * R2 is a return register once the subprogram returns a pair, so a stack
+ * pointer left in it is rejected at the callee's exit exactly as one in R0
+ * is: the callee frame is gone by the time the caller could use it.
+ */
+SEC("tc")
+__failure __msg("cannot return stack pointer to the caller")
+__naked int aggregate_ret_static_ptr_fail(void)
+{
+	asm volatile (
+	"call %[static_agg_bad_ptr];"
+	"r0 = 0;"
+	"exit;"
+	:
+	: __imm(static_agg_bad_ptr)
+	: __clobber_all);
+}
+
+static __naked __noinline u128 static_agg_no_r2(void)
+{
+	asm volatile (
+	"r0 = 0;"
+	"exit;"
+	);
+}
+
+SEC("tc")
+__failure __msg("R2 !read_ok")
+__naked int aggregate_ret_static_uninit_fail(void)
+{
+	asm volatile (
+	"call %[static_agg_no_r2];"
+	"r0 = r2;"
+	"exit;"
+	:
+	: __imm(static_agg_no_r2)
+	: __clobber_all);
+}
+
+static __naked __noinline u128 static_agg_precise(void)
+{
+	asm volatile (
+	"r0 = 0;"
+	"r2 = 4;"	/* second half; its value is made precise below */
+	"exit;"
+	);
+}
+
+SEC("tc")
+__load_if_JITed()
+__success __retval(0)
+__log_level(2)
+__msg("mark_precise: frame0: last_idx 5 first_idx 0 subseq_idx -1")
+__msg("mark_precise: frame0: regs=r6 stack= before 4: (07) r1 += -8")
+__msg("mark_precise: frame0: regs=r6 stack= before 3: (bf) r1 = r10")
+__msg("mark_precise: frame0: regs=r6 stack= before 2: (57) r6 &= 7")
+__msg("mark_precise: frame0: regs=r6 stack= before 1: (bf) r6 = r2")
+__msg("mark_precise: frame0: regs=r2 stack= before 12: (95) exit")
+__msg("mark_precise: frame1: regs=r2 stack= before 11: (b7) r2 = 4")
+__naked int aggregate_ret_static_precise(void)
+{
+	asm volatile (
+	"call %[static_agg_precise];"
+	"r6 = r2;"		/* derived from the aggregate's second half */
+	"r6 &= 7;"		/* keep it in [0, 7] to index the stack */
+	"r1 = r10;"
+	"r1 += -8;"
+	"r1 += r6;"		/* ptr += scalar marks r6 (hence R2) precise */
+	"r0 = 0;"
+	"*(u8 *)(r1 + 0) = r0;"
+	"r0 = 0;"
+	"exit;"
+	:
+	: __imm(static_agg_precise)
+	: __clobber_all);
+}
+
+SEC("tc")
+__load_if_JITed()
+__success __retval(0)
+__log_level(2)
+__msg("mark_precise: frame0: last_idx 5 first_idx 0 subseq_idx -1")
+__msg("mark_precise: frame0: regs=r6 stack= before 4: (07) r1 += -8")
+__msg("mark_precise: frame0: regs=r6 stack= before 3: (bf) r1 = r10")
+__msg("mark_precise: frame0: regs=r6 stack= before 2: (57) r6 &= 7")
+__msg("mark_precise: frame0: regs=r6 stack= before 1: (bf) r6 = r2")
+__msg("mark_precise: frame0: regs=r2 stack= before 0: (85) call pc+9")
+__naked int aggregate_ret_global_precise(void)
+{
+	asm volatile (
+	"call %[global_agg_good];"
+	"r6 = r2;"		/* derived from the aggregate's second half */
+	"r6 &= 7;"		/* keep it in [0, 7] to index the stack */
+	"r1 = r10;"
+	"r1 += -8;"
+	"r1 += r6;"		/* ptr += scalar marks r6 (hence R2) precise */
+	"r0 = 0;"
+	"*(u8 *)(r1 + 0) = r0;"
+	"r0 = 0;"
+	"exit;"
+	:
+	: __imm(global_agg_good)
+	: __clobber_all);
+}
+
+#if defined(__clang_major__) && __clang_major__ >= 23
+
+/* A by-value struct that smuggles a pointer, which must be rejected. */
+struct with_ptr {
+	void *p;
+	__u64 x;
+};
+
+/* A by-value union that smuggles a pointer, which must be rejected too. */
+union upair_with_ptr {
+	void *p;
+	__u64 halves[2];
+};
+
+__naked struct with_ptr global_ret_struct_ptr(void)
+{
+	asm volatile (
+	"r0 = 0;"
+	"r2 = 0;"
+	"exit;"
+	);
+}
+
+SEC("tc")
+__failure __msg("Global function global_ret_struct_ptr() has unsupported return type")
+__naked int aggregate_ret_global_struct_ptr_fail(void)
+{
+	asm volatile (
+	"call %[global_ret_struct_ptr];"
+	"r0 = 0;"
+	"exit;"
+	:
+	: __imm(global_ret_struct_ptr)
+	: __clobber_all);
+}
+
+__naked union upair_with_ptr global_ret_union_ptr(void)
+{
+	asm volatile (
+	"r0 = 0;"
+	"r2 = 0;"
+	"exit;"
+	);
+}
+
+SEC("tc")
+__failure __msg("Global function global_ret_union_ptr() has unsupported return type")
+__naked int aggregate_ret_global_union_ptr_fail(void)
+{
+	asm volatile (
+	"call %[global_ret_union_ptr];"
+	"r0 = 0;"
+	"exit;"
+	:
+	: __imm(global_ret_union_ptr)
+	: __clobber_all);
+}
+
+#endif
+
+char _license[] SEC("license") = "GPL";
diff --git a/tools/testing/selftests/bpf/progs/aggregate_ret_kfunc.c b/tools/testing/selftests/bpf/progs/aggregate_ret_kfunc.c
new file mode 100644
index 0000000000000..c23b4beb1773c
--- /dev/null
+++ b/tools/testing/selftests/bpf/progs/aggregate_ret_kfunc.c
@@ -0,0 +1,122 @@
+// SPDX-License-Identifier: GPL-2.0
+/* Copyright (c) 2026 Meta Platforms, Inc. and affiliates. */
+#include <vmlinux.h>
+#include <bpf/bpf_helpers.h>
+#include "bpf_misc.h"
+#include "../test_kmods/bpf_testmod_kfunc.h"
+
+/*
+ * Reference kfunc addresses to force those BTF to be emitted. Taking the address
+ * (rather than calling) avoids any dependence on the compiler lowering an
+ * __int128 or struct return value, which the BPF backend only supports from
+ * LLVM 23 on.
+ */
+void __kfunc_btf_root(void)
+{
+	asm volatile (""
+	:
+	: "r"(&bpf_kfunc_call_test_i128),
+	  "r"(&bpf_kfunc_call_test_ret_fastcall),
+	  "r"(&bpf_kfunc_call_test_ret_ptr),
+	  "r"(&bpf_kfunc_call_test_ret_ii),
+	  "r"(&bpf_kfunc_call_test_ret_big));
+}
+
+SEC("tc")
+__arch_x86_64 __arch_arm64
+__load_if_JITed()
+__success __retval(0)
+__log_level(2)
+__msg("mark_precise: frame0: last_idx 7 first_idx 0 subseq_idx -1")
+__msg("mark_precise: frame0: regs=r6 stack= before 6: (07) r1 += -8")
+__msg("mark_precise: frame0: regs=r6 stack= before 5: (bf) r1 = r10")
+__msg("mark_precise: frame0: regs=r6 stack= before 4: (57) r6 &= 7")
+__msg("mark_precise: frame0: regs=r6 stack= before 3: (bf) r6 = r2")
+__msg("mark_precise: frame0: regs=r2 stack= before 2: (85) call bpf_kfunc_call_test_i128")
+__naked int aggregate_ret_kfunc_precise(void)
+{
+	asm volatile (
+	"r1 = 1;"
+	"r2 = 2;"
+	"call %[bpf_kfunc_call_test_i128];"
+	"r6 = r2;"		/* second return half */
+	"r6 &= 7;"		/* keep it in [0, 7] to index the stack */
+	"r1 = r10;"
+	"r1 += -8;"
+	"r1 += r6;"		/* ptr += scalar marks r6 (hence R2) precise */
+	"r0 = 0;"
+	"*(u8 *)(r1 + 0) = r0;"
+	"r0 = 0;"
+	"exit;"
+	:
+	: __imm(bpf_kfunc_call_test_i128)
+	: __clobber_all);
+}
+
+SEC("tc")
+__arch_x86_64 __arch_arm64
+__failure __msg("kfunc bpf_kfunc_call_test_ret_fastcall with >8-byte return is not supported with KF_FASTCALL")
+__naked int aggregate_ret_kfunc_fastcall_fail(void)
+{
+	asm volatile (
+	"r1 = 1;"
+	"r2 = 2;"
+	"call %[bpf_kfunc_call_test_ret_fastcall];"
+	"r0 = 0;"
+	"exit;"
+	:
+	: __imm(bpf_kfunc_call_test_ret_fastcall)
+	: __clobber_all);
+}
+
+SEC("tc")
+__arch_x86_64 __arch_arm64
+__failure __msg("is not composed of scalars")
+__naked int aggregate_ret_kfunc_ptr_fail(void)
+{
+	asm volatile (
+	"r1 = 0;"
+	"call %[bpf_kfunc_call_test_ret_ptr];"
+	"r0 = 0;"
+	"exit;"
+	:
+	: __imm(bpf_kfunc_call_test_ret_ptr)
+	: __clobber_all);
+}
+
+SEC("tc")
+__arch_x86_64 __arch_arm64
+__failure __msg("R2 !read_ok")
+__naked int aggregate_ret_kfunc_small_no_r2(void)
+{
+	asm volatile (
+	"r1 = 0;"
+	"r2 = 0;"
+	"call %[bpf_kfunc_call_test_ret_ii];"
+	"r0 = r2;"	/* R2 is not a return register for a <=8 byte struct */
+	"exit;"
+	:
+	: __imm(bpf_kfunc_call_test_ret_ii)
+	: __clobber_all);
+}
+
+/*
+ * A return value larger than 16 bytes does not fit in R0:R2 and is rejected by
+ * btf_distill_func_proto(), before the KF_FASTCALL and JIT-capability checks,
+ * so this behaves the same on every architecture.
+ */
+SEC("tc")
+__arch_x86_64 __arch_arm64
+__failure __msg("The function bpf_kfunc_call_test_ret_big return type STRUCT is unsupported")
+__naked int aggregate_ret_kfunc_too_big_fail(void)
+{
+	asm volatile (
+	"call %[bpf_kfunc_call_test_ret_big];"
+	"r0 = 0;"
+	"exit;"
+	:
+	: __imm(bpf_kfunc_call_test_ret_big)
+	: __clobber_all);
+}
+
+char _license[] SEC("license") = "GPL";
diff --git a/tools/testing/selftests/bpf/progs/aggregate_ret_target.c b/tools/testing/selftests/bpf/progs/aggregate_ret_target.c
new file mode 100644
index 0000000000000..cffd8d7d3241a
--- /dev/null
+++ b/tools/testing/selftests/bpf/progs/aggregate_ret_target.c
@@ -0,0 +1,29 @@
+// SPDX-License-Identifier: GPL-2.0
+/* Copyright (c) 2026 Meta Platforms, Inc. and affiliates. */
+#include <linux/bpf.h>
+#include <bpf/bpf_helpers.h>
+#include "bpf_misc.h"
+
+/* freplace target: a global subprogram returning 16 bytes in R0:R2. */
+__naked unsigned __int128 agg_ret_target_func(void)
+{
+	asm volatile (
+	"r0 = 0x1234;"
+	"r2 = 0x5678;"
+	"exit;"
+	);
+}
+
+SEC("tc")
+__naked int agg_ret_target(void)
+{
+	asm volatile (
+	"call %[agg_ret_target_func];"
+	"r0 = 0;"
+	"exit;"
+	:
+	: __imm(agg_ret_target_func)
+	: __clobber_all);
+}
+
+char _license[] SEC("license") = "GPL";
diff --git a/tools/testing/selftests/bpf/progs/compute_live_registers.c b/tools/testing/selftests/bpf/progs/compute_live_registers.c
index d055fc7b3b95d..0be9441ec273b 100644
--- a/tools/testing/selftests/bpf/progs/compute_live_registers.c
+++ b/tools/testing/selftests/bpf/progs/compute_live_registers.c
@@ -431,6 +431,36 @@ __naked void subprog1(void)
 		::: __clobber_all);
 }
 
+static __used __naked unsigned __int128 aux2(void)
+{
+	asm volatile (
+		"r0 = 1;"
+		"r2 = 2;"
+		"exit;"
+		::: __clobber_all);
+}
+
+SEC("socket")
+/* A program observing the pair needs the JIT; see bpf_compute_subprog_ret_regs(). */
+__load_if_JITed()
+__log_level(2)
+__msg("0: .12345.... (85) call pc+2")
+__msg("1: ..2....... (bf) r0 = r2")
+/* R2 is not read at the exit of this program, which returns an int, ... */
+__msg("2: 0......... (95) exit")
+__msg("3: .......... (b7) r0 = 1")
+__msg("4: 0......... (b7) r2 = 2")
+/* ... but it is at the exit of aux2(), which returns a register pair. */
+__msg("5: 0.2....... (95) exit")
+__naked void subprog_ret_reg_pair(void)
+{
+	asm volatile (
+		"call aux2;"
+		"r0 = r2;"
+		"exit;"
+		::: __clobber_all);
+}
+
 #if defined(__TARGET_ARCH_x86) || defined(__TARGET_ARCH_arm64)
 
 SEC("socket")
diff --git a/tools/testing/selftests/bpf/progs/exceptions_fail.c b/tools/testing/selftests/bpf/progs/exceptions_fail.c
index ac44d60e50666..9708efb93683b 100644
--- a/tools/testing/selftests/bpf/progs/exceptions_fail.c
+++ b/tools/testing/selftests/bpf/progs/exceptions_fail.c
@@ -60,7 +60,7 @@ __noinline int exception_cb_ok_arg_small(int a)
 
 SEC("?tc")
 __exception_cb(exception_cb_bad_ret_type1)
-__failure __msg("Global function exception_cb_bad_ret_type1() return value not void or scalar.")
+__failure __msg("Only void, scalar, or a scalar-only struct/union up to 16 bytes is supported.")
 int reject_exception_cb_type_1(struct __sk_buff *ctx)
 {
 	bpf_throw(0);
diff --git a/tools/testing/selftests/bpf/progs/freplace_ret_pair.c b/tools/testing/selftests/bpf/progs/freplace_ret_pair.c
new file mode 100644
index 0000000000000..12c15d293bd79
--- /dev/null
+++ b/tools/testing/selftests/bpf/progs/freplace_ret_pair.c
@@ -0,0 +1,12 @@
+// SPDX-License-Identifier: GPL-2.0
+/* Copyright (c) 2026 Meta Platforms, Inc. and affiliates. */
+#include <linux/bpf.h>
+#include <bpf/bpf_helpers.h>
+
+SEC("freplace/agg_ret_target_func")
+__u64 new_agg_ret_target_func(void)
+{
+	return 0;
+}
+
+char _license[] SEC("license") = "GPL";
diff --git a/tools/testing/selftests/bpf/progs/verifier_aggregate_ret.c b/tools/testing/selftests/bpf/progs/verifier_aggregate_ret.c
new file mode 100644
index 0000000000000..7851bade2b406
--- /dev/null
+++ b/tools/testing/selftests/bpf/progs/verifier_aggregate_ret.c
@@ -0,0 +1,178 @@
+// SPDX-License-Identifier: GPL-2.0
+/* Copyright (c) 2026 Meta Platforms, Inc. and affiliates. */
+#include <vmlinux.h>
+#include <bpf/bpf_helpers.h>
+#include "../test_kmods/bpf_testmod_kfunc.h"
+#include "bpf_misc.h"
+
+#if defined(__clang_major__) && __clang_major__ >= 23
+
+#define MIX_A	0xdeadbeefcafef00dULL
+#define MIX_B	0x0123456789abcdefULL
+
+typedef unsigned __int128 u128;
+
+struct pair {
+	__u64 lo;	/* R0 */
+	__u64 hi;	/* R2 */
+};
+
+union upair {
+	__u64 halves[2];
+	struct {
+		__u64 lo;	/* R0 */
+		__u64 hi;	/* R2 */
+	} parts;
+};
+
+static __noinline u128 make_i128(__u64 a, __u64 b)
+{
+	return ((u128)(a + b) << 64) | (a - b);
+}
+
+SEC("tc")
+__load_if_JITed()
+__success __retval(0)
+int aggregate_ret_int128_c_test(struct __sk_buff *skb)
+{
+	__u64 a = skb->len ^ MIX_A;
+	__u64 b = skb->len ^ MIX_B;
+	u128 v;
+
+	v = make_i128(a, b);
+	if ((__u64)(v >> 64) != a + b)
+		return 1;
+	if ((__u64)v != a - b)
+		return 2;
+
+	return 0;
+}
+
+static __noinline struct pair make_pair(__u64 a, __u64 b)
+{
+	struct pair p = { .lo = a + b, .hi = a - b };
+
+	return p;
+}
+
+SEC("tc")
+__load_if_JITed()
+__success __retval(0)
+int aggregate_ret_struct_c_test(struct __sk_buff *skb)
+{
+	__u64 a = skb->len ^ MIX_A;
+	__u64 b = skb->len ^ MIX_B;
+	struct pair p;
+
+	p = make_pair(a, b);
+	if (p.lo != a + b)
+		return 1;
+	if (p.hi != a - b)
+		return 2;
+
+	return 0;
+}
+
+__noinline struct pair make_pair_global(__u64 a, __u64 b)
+{
+	struct pair p = { .lo = a + b, .hi = a - b };
+
+	return p;
+}
+
+SEC("tc")
+__load_if_JITed()
+__success __retval(0)
+int aggregate_ret_global_struct_c_test(struct __sk_buff *skb)
+{
+	__u64 a = skb->len ^ MIX_A;
+	__u64 b = skb->len ^ MIX_B;
+	struct pair p;
+
+	p = make_pair_global(a, b);
+	if (p.lo != a + b)
+		return 1;
+	if (p.hi != a - b)
+		return 2;
+
+	return 0;
+}
+
+static __noinline union upair make_upair(__u64 a, __u64 b)
+{
+	union upair p;
+
+	p.halves[0] = a + b;
+	p.halves[1] = a - b;
+	return p;
+}
+
+SEC("tc")
+__load_if_JITed()
+__success __retval(0)
+int aggregate_ret_union_c_test(struct __sk_buff *skb)
+{
+	__u64 a = skb->len ^ MIX_A;
+	__u64 b = skb->len ^ MIX_B;
+	union upair p;
+
+	p = make_upair(a, b);
+	if (p.parts.lo != a + b)
+		return 1;
+	if (p.parts.hi != a - b)
+		return 2;
+
+	return 0;
+}
+
+SEC("tc")
+__arch_x86_64 __arch_arm64
+__load_if_JITed()
+__success __retval(0)
+int aggregate_ret_kfunc_int128_c_test(struct __sk_buff *skb)
+{
+	__u64 a = skb->len ^ MIX_A;
+	__u64 b = skb->len ^ MIX_B;
+	u128 v;
+
+	v = bpf_kfunc_call_test_i128(a, b);
+	if ((__u64)(v >> 64) != a + b)
+		return 1;
+	if ((__u64)v != a - b)
+		return 2;
+
+	return 0;
+}
+
+SEC("tc")
+__arch_x86_64 __arch_arm64
+__load_if_JITed()
+__success __retval(0)
+int aggregate_ret_kfunc_struct_c_test(struct __sk_buff *skb)
+{
+	__u64 a = skb->len ^ MIX_A;
+	__u64 b = skb->len ^ MIX_B;
+	struct prog_test_ret_pair p;
+
+	p = bpf_kfunc_call_test_ret_pair(a, b);
+	if (p.lo != a + b)
+		return 1;
+	if (p.hi != a - b)
+		return 2;
+
+	return 0;
+}
+
+#else
+
+SEC("socket")
+__description("verifier_aggregate_ret: needs LLVM 23, dummy test")
+__success
+int dummy_test(void)
+{
+	return 0;
+}
+
+#endif
+
+char _license[] SEC("license") = "GPL";
diff --git a/tools/testing/selftests/bpf/test_kmods/bpf_testmod.c b/tools/testing/selftests/bpf/test_kmods/bpf_testmod.c
index 9366a3c578f13..20a9b9f20e96f 100644
--- a/tools/testing/selftests/bpf/test_kmods/bpf_testmod.c
+++ b/tools/testing/selftests/bpf/test_kmods/bpf_testmod.c
@@ -954,6 +954,53 @@ __bpf_kfunc int bpf_kfunc_call_test5(u8 a, u16 b, u32 c)
 	return 0;
 }
 
+#if defined(__x86_64__) || defined(__aarch64__)
+__bpf_kfunc __int128 bpf_kfunc_call_test_i128(u64 a, u64 b)
+{
+	return (__int128)(((unsigned __int128)(a + b) << 64) | (a - b));
+}
+
+__bpf_kfunc struct prog_test_ret_pair bpf_kfunc_call_test_ret_pair(u64 a, u64 b)
+{
+	struct prog_test_ret_pair r = { .lo = a + b, .hi = a - b };
+
+	return r;
+}
+
+__bpf_kfunc struct prog_test_ret_pair bpf_kfunc_call_test_ret_fastcall(u64 a, u64 b)
+{
+	struct prog_test_ret_pair r = { .lo = a + b, .hi = a - b };
+
+	return r;
+}
+
+__bpf_kfunc struct prog_test_ret_ptr bpf_kfunc_call_test_ret_ptr(u64 tag)
+{
+	struct prog_test_ret_ptr r = { .p = NULL, .tag = tag };
+
+	return r;
+}
+
+__bpf_kfunc struct prog_test_ret_ii bpf_kfunc_call_test_ret_ii(int a, int b)
+{
+	struct prog_test_ret_ii r = { .a = a, .b = b };
+
+	return r;
+}
+#endif /* __x86_64__ || __aarch64__ */
+
+/*
+ * Takes no argument on purpose: with no arguments there is nothing for the sret
+ * pointer to displace, so this needs no architecture guard even though it
+ * returns 24 bytes. See the comment on bpf_kfunc_call_test_i128() above.
+ */
+__bpf_kfunc struct prog_test_ret_big bpf_kfunc_call_test_ret_big(void)
+{
+	struct prog_test_ret_big r = { .a = 1, .b = 2, .c = 3 };
+
+	return r;
+}
+
 __bpf_kfunc u64 bpf_kfunc_call_stack_arg(u64 a, u64 b, u64 c, u64 d,
 					 u64 e, u64 f, u64 g, u64 h,
 					 u64 i, u64 j)
@@ -1487,6 +1534,14 @@ BTF_ID_FLAGS(func, bpf_kfunc_call_test2)
 BTF_ID_FLAGS(func, bpf_kfunc_call_test3)
 BTF_ID_FLAGS(func, bpf_kfunc_call_test4)
 BTF_ID_FLAGS(func, bpf_kfunc_call_test5)
+#if defined(__x86_64__) || defined(__aarch64__)
+BTF_ID_FLAGS(func, bpf_kfunc_call_test_i128)
+BTF_ID_FLAGS(func, bpf_kfunc_call_test_ret_pair)
+BTF_ID_FLAGS(func, bpf_kfunc_call_test_ret_fastcall, KF_FASTCALL)
+BTF_ID_FLAGS(func, bpf_kfunc_call_test_ret_ptr)
+BTF_ID_FLAGS(func, bpf_kfunc_call_test_ret_ii)
+#endif
+BTF_ID_FLAGS(func, bpf_kfunc_call_test_ret_big)
 BTF_ID_FLAGS(func, bpf_kfunc_call_stack_arg)
 BTF_ID_FLAGS(func, bpf_kfunc_call_stack_arg_ptr)
 BTF_ID_FLAGS(func, bpf_kfunc_call_stack_arg_mix)
diff --git a/tools/testing/selftests/bpf/test_kmods/bpf_testmod_kfunc.h b/tools/testing/selftests/bpf/test_kmods/bpf_testmod_kfunc.h
index 7d81070eefe7f..65e693ada7364 100644
--- a/tools/testing/selftests/bpf/test_kmods/bpf_testmod_kfunc.h
+++ b/tools/testing/selftests/bpf/test_kmods/bpf_testmod_kfunc.h
@@ -55,6 +55,27 @@ struct prog_test_big_arg {
 	__u64 b;
 };
 
+struct prog_test_ret_pair {	/* 16 bytes: R0:R2 */
+	__u64 lo;
+	__u64 hi;
+};
+
+struct prog_test_ret_ii {	/* 8 bytes: R0 only */
+	int a;
+	int b;
+};
+
+struct prog_test_ret_ptr {	/* 16 bytes: contains a pointer */
+	void *p;
+	__u64 tag;
+};
+
+struct prog_test_ret_big {	/* 24 bytes: too large for R0:R2 */
+	__u64 a;
+	__u64 b;
+	__u64 c;
+};
+
 struct prog_test_fail1 {
 	void *p;
 	int x;
@@ -131,6 +152,14 @@ int bpf_kfunc_call_test2(struct sock *sk, __u32 a, __u32 b) __ksym;
 struct sock *bpf_kfunc_call_test3(struct sock *sk) __ksym;
 long bpf_kfunc_call_test4(signed char a, short b, int c, long d) __ksym;
 int bpf_kfunc_call_test5(__u8 a, __u16 b, __u32 c) __ksym;
+#ifdef __SIZEOF_INT128__
+__int128 bpf_kfunc_call_test_i128(__u64 a, __u64 b) __ksym;
+#endif
+struct prog_test_ret_pair bpf_kfunc_call_test_ret_pair(__u64 a, __u64 b) __ksym;
+struct prog_test_ret_pair bpf_kfunc_call_test_ret_fastcall(__u64 a, __u64 b) __ksym;
+struct prog_test_ret_ii bpf_kfunc_call_test_ret_ii(int a, int b) __ksym;
+struct prog_test_ret_ptr bpf_kfunc_call_test_ret_ptr(__u64 tag) __ksym;
+struct prog_test_ret_big bpf_kfunc_call_test_ret_big(void) __ksym;
 __u64 bpf_kfunc_call_stack_arg(__u64 a, __u64 b, __u64 c, __u64 d,
 			       __u64 e, __u64 f, __u64 g, __u64 h,
 			       __u64 i, __u64 j) __ksym;