Improvements of BTF sanitizing for old kernels

Improvements of BTF sanitizing for old kernels

[Timur Chernykh]

Hello everyone. Recently I've encountered some issues with loading on debian 10 with kernel version `4.19.0-[17-22]`.

This first one consists of when the kernel loads BTF with specified min-CORE and libbpf does some sanitizing on those, then it "translates" func_proto to enum. But if `func_proto` has no names for it's parameters then kernel verifier fails with "Invalid name" error. This error caused by enum members must has a valid C identifier, but there's might be no names generated.

The second issue this PR supposed to fix is missed check whether kernel supports the `kind flag` or not. Let's say we're compiling the BPF object on new kernel with a new compiler. There's 99% percent chance that compiler will use this `kind flag`. But of we're loading on old kernels that have no support of this, so the loading will fail in multiple places.

- The type metadata check. Old BTF_INFO_MASK prohibits 31d bit
- Integer bitfield verification. Because old kernel waits for an offset-only encoding for the btf_member->offset field, but it's encoded in different way in case when the `kind flag` is set.

This PR contains:
- Enum names generation during sanitizing process for `func_proto` kind, when it being translate to `enum` kind.
- The feature check whether kernel supports the kind flag or not
- Kind flag sanitizing if kernel doesn't support one
- Struct/enum bitfield members sanitizing by generation a proper replacement the type of bitfield with corresponding integer type with same bit size

From: Timur Chernykh <tim.cherry.co@gmail.com>
To: bpf@vger.kernel.org
Cc: 
Bcc: 
Reply-To: 
Subject: Improvements of BTF sanitizing for old kernels
In-Reply-To: 

Improvements of BTF sanitizing for old kernels

[Timur Chernykh]

I'm very sorry, but I found a small typo in 2nd patch, here's a
quick fix for one.

From: Timur Chernykh <tim.cherry.co@gmail.com>
To: bpf@vger.kernel.org
Subject: Improvements of BTF sanitizing for old kernels

[PATCH 2/2] libbpf: add check if kernel supports kind flag and fix the bitfield members in union and structs if not

[Timur Chernykh]

Signed-off-by: Timur Chernykh <tim.cherry.co@gmail.com>
---
 tools/lib/bpf/features.c        | 30 ++++++++++++++++
 tools/lib/bpf/libbpf.c          | 62 ++++++++++++++++++++++++++++++++-
 tools/lib/bpf/libbpf_internal.h |  2 ++
 3 files changed, 93 insertions(+), 1 deletion(-)

diff --git a/tools/lib/bpf/features.c b/tools/lib/bpf/features.c
index 760657f5224c..dfab65f30f0c 100644
--- a/tools/lib/bpf/features.c
+++ b/tools/lib/bpf/features.c
@@ -507,6 +507,33 @@ static int probe_kern_arg_ctx_tag(int token_fd)
 	return probe_fd(prog_fd);
 }
 
+static int probe_kern_btf_type_kind_flag(int token_fd)
+{
+	const char strs[] = "\0bpf_spin_lock\0val\0cnt\0l";
+	/* struct bpf_spin_lock {
+	 *   int val;
+	 * };
+	 * struct val {
+	 *   int cnt;
+	 *   struct bpf_spin_lock l;
+	 * };
+	 */
+	__u32 types[] = {
+		/* int */
+		BTF_TYPE_INT_ENC(0, BTF_INT_SIGNED, 0, 32, 4),  /* [1] */
+		/* struct bpf_spin_lock */                      /* [2] */
+		BTF_TYPE_ENC(1, BTF_INFO_ENC(BTF_KIND_STRUCT, 1 /* kind bit */, 1), 4),
+		BTF_MEMBER_ENC(15, 1, 0), /* int val; */
+		/* struct val */                                /* [3] */
+		BTF_TYPE_ENC(15, BTF_INFO_ENC(BTF_KIND_STRUCT, 1 /* kind bit */, 2), 8),
+		BTF_MEMBER_ENC(19, 1, 0), /* int cnt; */
+		BTF_MEMBER_ENC(23, 2, 32),/* struct bpf_spin_lock l; */
+	    };
+
+	return probe_fd(libbpf__load_raw_btf((char *)types, sizeof(types),
+			     strs, sizeof(strs), token_fd));
+}
+
 typedef int (*feature_probe_fn)(int /* token_fd */);
 
 static struct kern_feature_cache feature_cache;
@@ -582,6 +609,9 @@ static struct kern_feature_desc {
 	[FEAT_BTF_QMARK_DATASEC] = {
 		"BTF DATASEC names starting from '?'", probe_kern_btf_qmark_datasec,
 	},
+	[FEAT_BTF_TYPE_KIND_FLAG] = {
+		"BTF btf_type can have the kind flags set", probe_kern_btf_type_kind_flag,
+	},
 };
 
 bool feat_supported(struct kern_feature_cache *cache, enum kern_feature_id feat_id)
diff --git a/tools/lib/bpf/libbpf.c b/tools/lib/bpf/libbpf.c
index 8e1edba443dd..13ed8f03cc65 100644
--- a/tools/lib/bpf/libbpf.c
+++ b/tools/lib/bpf/libbpf.c
@@ -3113,9 +3113,10 @@ static bool btf_needs_sanitization(struct bpf_object *obj)
 	bool has_type_tag = kernel_supports(obj, FEAT_BTF_TYPE_TAG);
 	bool has_enum64 = kernel_supports(obj, FEAT_BTF_ENUM64);
 	bool has_qmark_datasec = kernel_supports(obj, FEAT_BTF_QMARK_DATASEC);
+    bool has_kind_bit_support = kernel_supports(obj, FEAT_BTF_TYPE_KIND_FLAG);
 
 	return !has_func || !has_datasec || !has_func_global || !has_float ||
-	       !has_decl_tag || !has_type_tag || !has_enum64 || !has_qmark_datasec;
+	       !has_decl_tag || !has_type_tag || !has_enum64 || !has_qmark_datasec || !has_kind_bit_support;
 }
 
 static int bpf_object__sanitize_btf(struct bpf_object *obj, struct btf *btf)
@@ -3128,6 +3129,7 @@ static int bpf_object__sanitize_btf(struct bpf_object *obj, struct btf *btf)
 	bool has_type_tag = kernel_supports(obj, FEAT_BTF_TYPE_TAG);
 	bool has_enum64 = kernel_supports(obj, FEAT_BTF_ENUM64);
 	bool has_qmark_datasec = kernel_supports(obj, FEAT_BTF_QMARK_DATASEC);
+	bool has_kind_bit_support = kernel_supports(obj, FEAT_BTF_TYPE_KIND_FLAG);
 
 	char name_gen_buff[32] = {0};
 	int enum64_placeholder_id = 0;
@@ -3263,6 +3265,64 @@ static int bpf_object__sanitize_btf(struct bpf_object *obj, struct btf *btf)
 				m->type = enum64_placeholder_id;
 				m->offset = 0;
 			}
+		} else if (!has_kind_bit_support && (btf_is_struct(t) || btf_is_union(t) || btf_is_fwd(t) || btf_is_enum(t) || btf_is_enum64(t))) {
+			const uint16_t members_cnt = btf_vlen(t);
+
+			/* type encoded with a kind flag */
+		    if (t->info != BTF_INFO_ENC(btf_kind(t), 1, members_cnt)) {
+		        continue;
+		    }
+
+		    /* unset kind flag anyway */
+		    t->info = BTF_INFO_ENC(btf_kind(t), 0, btf_vlen(t));
+
+		    /* structs an unions has a different bitfield processing behaviour is kind flag is set */
+		    if (btf_is_struct(t) || btf_is_union(t)) {
+		        struct btf_member* members = btf_members(t);
+				struct btf_type* new_int_type = NULL;
+				int new_int_type_id;
+				__u32* new_int_type_data;
+				int encoding = 0;
+		        int nmember;
+
+		        for (nmember = 0; nmember < members_cnt; nmember++) {
+		            struct btf_member* member = &members[nmember];
+		            const struct btf_type* member_type = btf_type_by_id(btf, member->type);
+
+		            while (btf_is_typedef(member_type)) { /* unwrap typedefs */
+		                member_type = btf_type_by_id(btf, member_type->type);
+		            }
+
+		            /* bitfields can be only int or enum values */
+		            if (!(btf_is_int(member_type) || btf_is_enum(member_type))) {
+		                continue;
+		            }
+
+		            encoding = btf_int_encoding(member_type);
+		            if (btf_is_enum(member_type) && member_type->info & 0x80000000 /* kind flag */) {
+		                /* enum value encodes integer signed/unsigned info in the kind flag */
+		                encoding = BTF_INT_SIGNED;
+		            }
+
+		            /* create new integral type with the same info */
+		            snprintf(name_gen_buff, sizeof(name_gen_buff), "__int_%d_%d", i, nmember);
+		            new_int_type_id = btf__add_int(btf, name_gen_buff, member_type->size, encoding);
+
+		            if (new_int_type_id < 0) {
+		                pr_warn("Error adding integer type for a bitfield %d of [%d]", nmember, i);
+		                return new_int_type_id;
+		            }
+
+		            new_int_type = btf_type_by_id(btf, new_int_type_id);
+
+		            /* encode int in legacy way, keep offset 0 and specify bit size as set in the member */
+		            new_int_type_data = (__u32*)(new_int_type + 1);
+		            *new_int_type_data = BTF_INT_ENC(encoding, 0, BTF_MEMBER_BITFIELD_SIZE(member->offset));
+
+		            member->type = new_int_type_id;
+		            member->offset = BTF_MEMBER_BIT_OFFSET(member->offset) /* old kernels looks only on offset */;
+		        }
+		    }
 		}
 	}
 
diff --git a/tools/lib/bpf/libbpf_internal.h b/tools/lib/bpf/libbpf_internal.h
index 76669c73dcd1..6369c5520fce 100644
--- a/tools/lib/bpf/libbpf_internal.h
+++ b/tools/lib/bpf/libbpf_internal.h
@@ -380,6 +380,8 @@ enum kern_feature_id {
 	FEAT_ARG_CTX_TAG,
 	/* Kernel supports '?' at the front of datasec names */
 	FEAT_BTF_QMARK_DATASEC,
+	/* Kernel supports kind flag */
+	FEAT_BTF_TYPE_KIND_FLAG,
 	__FEAT_CNT,
 };
 
-- 
2.49.0

Improvements of BTF sanitizing for old kernels

[Timur Chernykh]

I've done fixing comments for first version of patch set (thanks for
Mykyta).

Here's a second version. 

From: Timur Chernykh <tim.cherry.co@gmail.com>
To: bpf@vger.kernel.org
Cc: mykyta.yatsenko5@gmail.com
Bcc: 
Reply-To: 20250331201016.345704-1-tim.cherry.co@gmail.com
Subject: Improvements of BTF sanitizing for old kernels
In-Reply-To: