Skip to contentSkip to conversation

bpf: allow void* cast using bpf_rdonly_cast()

PATCH Superseded Eduard Zingerman 2 participants 9 messages Jun 25, 2025

Subsystems (4)

v1 of 3 Superseded Latest is v3 →

lore
Share & tools
lore

Eduard Zingerman

author maintainer
#0 lore raw

cover bpf: allow void* cast using bpf_rdonly_cast()

Currently, pointers returned by bpf_rdonly_cast() have a type of "pointer to btf id", and only casts to structure types are allowed. Access to memory pointed to by these pointers is done through BPF_PROBE_{MEM,MEMSX} instructions and does not produce errors on invalid memory access.

This patch set extends bpf_rdonly_cast() to allow casts to an equivalent of 'void *', effectively replacing bpf_probe_read_kernel() calls in situations where access to individual bytes or integers is necessary.

The mechanism was suggested and explored by Andrii Nakryiko in [1].

To help with detecting support for this feature, an enum bpf_features is added with intended usage as follows:

if (bpf_core_enum_value_exists(enum bpf_features,

                                 BPF_FEAT_RDONLY_CAST_TO_VOID))
    ...

[1] https://github.com/anakryiko/linux/tree/bpf-mem-cast

Changelog:
v1: https://lore.kernel.org/bpf/[email protected]/
v1 -> v2:
- renamed BPF_FEAT_TOTAL to __MAX_BPF_FEAT and moved patch introducing bpf_features enum to the start of the series (Alexei);
- dropped patch #3 allowing optout from CAP_SYS_ADMIN drop in prog_tests/verifier.c, use a separate runner in prog_tests/* instead.

Eduard Zingerman (3):
bpf: add bpf_features enum
bpf: allow void* cast using bpf_rdonly_cast() selftests/bpf: check operations on untrusted ro pointers to mem

 kernel/bpf/verifier.c                         |  79 ++++++++--
 .../bpf/prog_tests/mem_rdonly_untrusted.c     |   9 ++
 .../bpf/progs/mem_rdonly_untrusted.c          | 136 ++++++++++++++++++
 3 files changed, 212 insertions(+), 12 deletions(-)
 create mode 100644 tools/testing/selftests/bpf/prog_tests/mem_rdonly_untrusted.c
 create mode 100644 tools/testing/selftests/bpf/progs/mem_rdonly_untrusted.c
Signature · 1 line
2.47.1
author maintainer
#1 lore raw

PATCH 1/3 bpf: add bpf_features enum

This commit adds a kernel side enum for use in conjucntion with BTF CO-RE bpf_core_enum_value_exists. The goal of the enum is to assist with available BPF features detection. Intended usage looks as follows:

if (bpf_core_enum_value_exists(enum bpf_features, BPF_FEAT_<f>))

     ... use feature f ...
---
 kernel/bpf/verifier.c | 6 ++++++
 1 file changed, 6 insertions(+)
kernel/bpf/verifier.c +6 -0
index 279a64933262..71de4c9487d5 100644
@@ -44,6 +44,10 @@ static const struct bpf_verifier_ops * const bpf_verifier_ops[] = {
 #undef BPF_LINK_TYPE
 };
 
+enum bpf_features {
+	__MAX_BPF_FEAT = 0,
+};
+
 struct bpf_mem_alloc bpf_global_percpu_ma;
 static bool bpf_global_percpu_ma_set;
 
@@ -24388,6 +24392,8 @@ int bpf_check(struct bpf_prog **prog, union bpf_attr *attr, bpfptr_t uattr, __u3
 	u32 log_true_size;
 	bool is_priv;
 
+	BTF_TYPE_EMIT(enum bpf_features);
+
 	/* no program is valid */
 	if (ARRAY_SIZE(bpf_verifier_ops) == 0)
 		return -EINVAL;
Signature · 1 line
2.47.1
author maintainer
#2 lore raw

PATCH 2/3 bpf: allow void* cast using bpf_rdonly_cast()

In reply to Eduard Zingerman

Introduce support for bpf_rdonly_cast(v, 0), which casts the value v to an untyped, untrusted pointer, logically similar to a void *. The memory pointed to by such a pointer is treated as read-only. As with other untrusted pointers, memory access violations on loads return zero instead of causing a fault.

Technically:
- The resulting pointer is represented as a register of type PTR_TO_MEM | MEM_RDONLY | PTR_UNTRUSTED with size zero.
- Offsets within such pointers are not tracked.
- Same load instructions are allowed to have both PTR_TO_MEM | MEM_RDONLY | PTR_UNTRUSTED and PTR_TO_BTF_ID as the base pointer types.
In such cases, bpf_insn_aux_data->ptr_type is considered the weaker of the two: PTR_TO_MEM | MEM_RDONLY | PTR_UNTRUSTED.

The following constraints apply to the new pointer type:
- can be used as a base for LDX instructions;
- can't be used as a base for ST/STX or atomic instructions;
- can't be used as parameter for kfuncs or helpers.

These constraints are enforced by existing handling of MEM_RDONLY flag and PTR_TO_MEM of size zero.

---
 kernel/bpf/verifier.c | 75 +++++++++++++++++++++++++++++++++++--------
 1 file changed, 62 insertions(+), 13 deletions(-)
kernel/bpf/verifier.c +62 -13
index 71de4c9487d5..6b2c38b7a7b6 100644
@@ -45,7 +45,8 @@ static const struct bpf_verifier_ops * const bpf_verifier_ops[] = {
 };
 
 enum bpf_features {
-	__MAX_BPF_FEAT = 0,
+	BPF_FEAT_RDONLY_CAST_TO_VOID = 0,
+	__MAX_BPF_FEAT = 1,
 };
 
 struct bpf_mem_alloc bpf_global_percpu_ma;
@@ -7539,6 +7540,7 @@ static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regn
 		}
 	} else if (base_type(reg->type) == PTR_TO_MEM) {
 		bool rdonly_mem = type_is_rdonly_mem(reg->type);
+		bool rdonly_untrusted = rdonly_mem && (reg->type & PTR_UNTRUSTED);
 
 		if (type_may_be_null(reg->type)) {
 			verbose(env, "R%d invalid mem access '%s'\n", regno,
@@ -7558,8 +7560,13 @@ static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regn
 			return -EACCES;
 		}
 
-		err = check_mem_region_access(env, regno, off, size,
-					      reg->mem_size, false);
+		/*
+		 * Accesses to untrusted PTR_TO_MEM are done through probe
+		 * instructions, hence no need to check bounds in that case.
+		 */
+		if (!rdonly_untrusted)
+			err = check_mem_region_access(env, regno, off, size,
+						      reg->mem_size, false);
 		if (!err && value_regno >= 0 && (t == BPF_READ || rdonly_mem))
 			mark_reg_unknown(env, regs, value_regno);
 	} else if (reg->type == PTR_TO_CTX) {
@@ -13606,16 +13613,24 @@ static int check_special_kfunc(struct bpf_verifier_env *env, struct bpf_kfunc_ca
 		regs[BPF_REG_0].btf_id = meta->ret_btf_id;
 	} else if (meta->func_id == special_kfunc_list[KF_bpf_rdonly_cast]) {
 		ret_t = btf_type_by_id(desc_btf, meta->arg_constant.value);
-		if (!ret_t || !btf_type_is_struct(ret_t)) {
+		if (!ret_t) {
+			verbose(env, "Unknown type ID %lld passed to kfunc bpf_rdonly_cast\n",
+				meta->arg_constant.value);
+			return -EINVAL;
+		} else if (btf_type_is_struct(ret_t)) {
+			mark_reg_known_zero(env, regs, BPF_REG_0);
+			regs[BPF_REG_0].type = PTR_TO_BTF_ID | PTR_UNTRUSTED;
+			regs[BPF_REG_0].btf = desc_btf;
+			regs[BPF_REG_0].btf_id = meta->arg_constant.value;
+		} else if (btf_type_is_void(ret_t)) {
+			mark_reg_known_zero(env, regs, BPF_REG_0);
+			regs[BPF_REG_0].type = PTR_TO_MEM | MEM_RDONLY | PTR_UNTRUSTED;
+			regs[BPF_REG_0].mem_size = 0;
+		} else {
 			verbose(env,
-				"kfunc bpf_rdonly_cast type ID argument must be of a struct\n");
+				"kfunc bpf_rdonly_cast type ID argument must be of a struct or void\n");
 			return -EINVAL;
 		}
-
-		mark_reg_known_zero(env, regs, BPF_REG_0);
-		regs[BPF_REG_0].type = PTR_TO_BTF_ID | PTR_UNTRUSTED;
-		regs[BPF_REG_0].btf = desc_btf;
-		regs[BPF_REG_0].btf_id = meta->arg_constant.value;
 	} else if (meta->func_id == special_kfunc_list[KF_bpf_dynptr_slice] ||
 		   meta->func_id == special_kfunc_list[KF_bpf_dynptr_slice_rdwr]) {
 		enum bpf_type_flag type_flag = get_dynptr_type_flag(meta->initialized_dynptr.type);
@@ -14414,6 +14429,13 @@ static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
 		return -EACCES;
 	}
 
+	/*
+	 * Accesses to untrusted PTR_TO_MEM are done through probe
+	 * instructions, hence no need to track offsets.
+	 */
+	if (base_type(ptr_reg->type) == PTR_TO_MEM && (ptr_reg->type & PTR_UNTRUSTED))
+		return 0;
+
 	switch (base_type(ptr_reg->type)) {
 	case PTR_TO_CTX:
 	case PTR_TO_MAP_VALUE:
@@ -19571,10 +19593,27 @@ static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
 			       !reg_type_mismatch_ok(prev));
 }
 
+static bool is_ptr_to_mem_or_btf_id(enum bpf_reg_type type)
+{
+	switch (base_type(type)) {
+	case PTR_TO_MEM:
+	case PTR_TO_BTF_ID:
+		return true;
+	default:
+		return false;
+	}
+}
+
+static bool is_ptr_to_mem(enum bpf_reg_type type)
+{
+	return base_type(type) == PTR_TO_MEM;
+}
+
 static int save_aux_ptr_type(struct bpf_verifier_env *env, enum bpf_reg_type type,
 			     bool allow_trust_mismatch)
 {
 	enum bpf_reg_type *prev_type = &env->insn_aux_data[env->insn_idx].ptr_type;
+	enum bpf_reg_type merged_type;
 
 	if (*prev_type == NOT_INIT) {
 		/* Saw a valid insn
@@ -19591,15 +19630,24 @@ static int save_aux_ptr_type(struct bpf_verifier_env *env, enum bpf_reg_type typ
 		 * Reject it.
 		 */
 		if (allow_trust_mismatch &&
-		    base_type(type) == PTR_TO_BTF_ID &&
-		    base_type(*prev_type) == PTR_TO_BTF_ID) {
+		    is_ptr_to_mem_or_btf_id(type) &&
+		    is_ptr_to_mem_or_btf_id(*prev_type)) {
 			/*
 			 * Have to support a use case when one path through
 			 * the program yields TRUSTED pointer while another
 			 * is UNTRUSTED. Fallback to UNTRUSTED to generate
 			 * BPF_PROBE_MEM/BPF_PROBE_MEMSX.
+			 * Same behavior of MEM_RDONLY flag.
 			 */
-			*prev_type = PTR_TO_BTF_ID | PTR_UNTRUSTED;
+			if (is_ptr_to_mem(type) || is_ptr_to_mem(*prev_type))
+				merged_type = PTR_TO_MEM;
+			else
+				merged_type = PTR_TO_BTF_ID;
+			if ((type & PTR_UNTRUSTED) || (*prev_type & PTR_UNTRUSTED))
+				merged_type |= PTR_UNTRUSTED;
+			if ((type & MEM_RDONLY) || (*prev_type & MEM_RDONLY))
+				merged_type |= MEM_RDONLY;
+			*prev_type = merged_type;
 		} else {
 			verbose(env, "same insn cannot be used with different pointers\n");
 			return -EINVAL;
@@ -21207,6 +21255,7 @@ static int convert_ctx_accesses(struct bpf_verifier_env *env)
 		 * for this case.
 		 */
 		case PTR_TO_BTF_ID | MEM_ALLOC | PTR_UNTRUSTED:
+		case PTR_TO_MEM | MEM_RDONLY | PTR_UNTRUSTED:
 			if (type == BPF_READ) {
 				if (BPF_MODE(insn->code) == BPF_MEM)
 					insn->code = BPF_LDX | BPF_PROBE_MEM |
Signature · 1 line
2.47.1
#3 lore raw
Quoted 37 lines · Eduard Zingerman

Introduce support for bpf_rdonly_cast(v, 0), which casts the value v to an untyped, untrusted pointer, logically similar to a void *. The memory pointed to by such a pointer is treated as read-only. As with other untrusted pointers, memory access violations on loads return zero instead of causing a fault.

Technically:
- The resulting pointer is represented as a register of type PTR_TO_MEM | MEM_RDONLY | PTR_UNTRUSTED with size zero.
- Offsets within such pointers are not tracked.
- Same load instructions are allowed to have both PTR_TO_MEM | MEM_RDONLY | PTR_UNTRUSTED and PTR_TO_BTF_ID as the base pointer types.
In such cases, bpf_insn_aux_data->ptr_type is considered the weaker of the two: PTR_TO_MEM | MEM_RDONLY | PTR_UNTRUSTED.

The following constraints apply to the new pointer type:
- can be used as a base for LDX instructions;
- can't be used as a base for ST/STX or atomic instructions;
- can't be used as parameter for kfuncs or helpers.

These constraints are enforced by existing handling of MEM_RDONLY flag and PTR_TO_MEM of size zero.

---
 kernel/bpf/verifier.c | 75 +++++++++++++++++++++++++++++++++++--------
 1 file changed, 62 insertions(+), 13 deletions(-)
kernel/bpf/verifier.c +0 -0
index 71de4c9487d5..6b2c38b7a7b6 100644

kernel/bpf/verifier.c static const struct bpf_verifier_ops * const bpf_verifier_ops[] = { · line 49

Quoted 6 lines
kernel/bpf/verifier.c +2 -1
@@ -45,7 +45,8 @@ static const struct bpf_verifier_ops * const bpf_verifier_ops[] = {
 };
 
 enum bpf_features {
-       __MAX_BPF_FEAT = 0,
+       BPF_FEAT_RDONLY_CAST_TO_VOID = 0,
+       __MAX_BPF_FEAT = 1,

and the idea is to manually adjust it every time?! That's way too much churn.
Either remove it or keep it without assignment.
Just as __MAX_BPF_FEAT. Like similar thing in enum bpf_link_type.

--
pw-bot: cr

author maintainer
#4 lore raw

On Tue, 2025-06-24 at 17:11 -0700, Alexei Starovoitov wrote:

[...]

Quoted 9 lines

kernel/bpf/verifier.c static const struct bpf_verifier_ops * const bpf_verifier_ops[] = { · line 49

Quoted 4 lines

enum bpf_features {

kernel/bpf/verifier.c +2 -1
-       __MAX_BPF_FEAT = 0,
+       BPF_FEAT_RDONLY_CAST_TO_VOID = 0,
+       __MAX_BPF_FEAT = 1,

and the idea is to manually adjust it every time?! That's way too much churn.
Either remove it or keep it without assignment.
Just as __MAX_BPF_FEAT. Like similar thing in enum bpf_link_type.

I probably did not understand your previous message:

Quoted 9 lines

Review

Quoted 3 lines
(quoted diff) +3 -0
+enum bpf_features {
+       BPF_FEAT_RDONLY_CAST_TO_VOID = 0,
+       BPF_FEAT_TOTAL,

I don't see the value of 'total', but not strongly against it. But pls be consistent with __MAX_BPF_CMD, __MAX_BPF_MAP_TYPE, ... Say, __MAX_BPF_FEAT ?

Quoted 5 lines

Also it's better to introduce this enum in some earlier patch, and then always add BTF_FEAT_... to this enum
in the same patch that adds the feature to make
sure backports won't screw it up.
Another rule should be to always assign a number to it.

Specifically: "Another rule should be to always assign a number to it." The BPF_FEAT_RDONLY_CAST_TO_VOID already had a number, so I assumed you were talking about __MAX_BPF_FEAT.
What did you mean?

#5 lore raw
Quoted 33 lines · Eduard Zingerman

On Tue, 2025-06-24 at 17:11 -0700, Alexei Starovoitov wrote:

[...]

Quoted 9 lines

kernel/bpf/verifier.c static const struct bpf_verifier_ops * const bpf_verifier_ops[] = { · line 49

Quoted 4 lines

enum bpf_features {

kernel/bpf/verifier.c +2 -1
-       __MAX_BPF_FEAT = 0,
+       BPF_FEAT_RDONLY_CAST_TO_VOID = 0,
+       __MAX_BPF_FEAT = 1,

and the idea is to manually adjust it every time?! That's way too much churn.
Either remove it or keep it without assignment. Just as __MAX_BPF_FEAT. Like similar thing in enum /.

I probably did not understand your previous message:

Quoted 9 lines

Review

Quoted 3 lines
(quoted diff) +3 -0
+enum bpf_features {
+       BPF_FEAT_RDONLY_CAST_TO_VOID = 0,
+       BPF_FEAT_TOTAL,

I don't see the value of 'total', but not strongly against it. But pls be consistent with __MAX_BPF_CMD, __MAX_BPF_MAP_TYPE, ... Say, __MAX_BPF_FEAT ?

Quoted 5 lines

Also it's better to introduce this enum in some earlier patch, and then always add BTF_FEAT_... to this enum
in the same patch that adds the feature to make
sure backports won't screw it up.
Another rule should be to always assign a number to it.

Quoted 4 lines

Specifically: "Another rule should be to always assign a number to it." The BPF_FEAT_RDONLY_CAST_TO_VOID already had a number, so I assumed you were talking about __MAX_BPF_FEAT.
What did you mean?

I mean to add " = 123," to actual features, so when they're backported the number stays the same.
Not to __MAX_BPF_FEAT.

I doubt it matters though,
since bpf progs suppose to use bpf_core_enum_value_exists(enum bpf_features, name) that doesn't care about the actual id.

In bpf helpers we got burned by broken backports and added constants to ___BPF_FUNC_MAPPER macro.
Here I don't see it ever matter.
Just like I don't think __MAX_BPF_FEAT is needed, but if we follow old steps, then let's do both __MAX_BPF_FEAT without number and every feature with the number. The end result will look like bpf_link_type.

author maintainer
#6 lore raw

On Tue, 2025-06-24 at 17:21 -0700, Alexei Starovoitov wrote:

[...]

Quoted 9 lines

I mean to add " = 123," to actual features, so when they're backported the number stays the same.
Not to __MAX_BPF_FEAT.

I doubt it matters though,
since bpf progs suppose to use bpf_core_enum_value_exists(enum bpf_features, name) that doesn't care about the actual id.

Quoted 7 lines

In bpf helpers we got burned by broken backports and added constants to ___BPF_FUNC_MAPPER macro.
Here I don't see it ever matter.
Just like I don't think __MAX_BPF_FEAT is needed, but if we follow old steps, then let's do both __MAX_BPF_FEAT without number and every feature with the number. The end result will look like bpf_link_type.

Understood, thank you, will respin.

author maintainer
#7 lore raw

PATCH 3/3 selftests/bpf: check operations on untrusted ro pointers to mem

In reply to Eduard Zingerman

The following cases are tested:
- it is ok to load memory at any offset from rdonly_untrusted_mem;
- rdonly_untrusted_mem offset/bounds are not tracked;
- writes into rdonly_untrusted_mem are forbidden;
- atomic operations on rdonly_untrusted_mem are forbidden;
- rdonly_untrusted_mem can't be passed as a memory argument of a helper of kfunc;
- it is ok to use PTR_TO_MEM and PTR_TO_BTF_ID in a same load instruction.

---
 .../bpf/prog_tests/mem_rdonly_untrusted.c     |   9 ++
 .../bpf/progs/mem_rdonly_untrusted.c          | 136 ++++++++++++++++++
 2 files changed, 145 insertions(+)
 create mode 100644 tools/testing/selftests/bpf/prog_tests/mem_rdonly_untrusted.c
 create mode 100644 tools/testing/selftests/bpf/progs/mem_rdonly_untrusted.c
tools/testing/selftests/bpf/prog_tests/mem_rdonly_untrusted.c new file +9 -0
new file mode 100644
index 000000000000..40d4f687bd9c
@@ -0,0 +1,9 @@
+// SPDX-License-Identifier: GPL-2.0-only
+
+#include <test_progs.h>
+#include "mem_rdonly_untrusted.skel.h"
+
+void test_mem_rdonly_untrusted(void)
+{
+	RUN_TESTS(mem_rdonly_untrusted);
+}
tools/testing/selftests/bpf/progs/mem_rdonly_untrusted.c new file +136 -0

136 lines not on this page · raw message

Signature · 1 line
2.47.1