Skip to contentSkip to conversation

iptables: open eBPF programs in read only mode

PATCH Superseded Maciej Żenczykowski 4 participants 12 messages Mar 20, 2020 – Mar 31, 2020

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 iptables: open eBPF programs in read only mode

From: Maciej Żenczykowski <[email protected]>

Adjust the mode eBPF programs are opened in so 0400 pinned bpf programs work without requiring CAP_DAC_OVERRIDE.

This matches Linux 5.2's:
commit e547ff3f803e779a3898f1f48447b29f43c54085
Author: Chenbo Feng <[email protected]>
Date: Tue May 14 19:42:57 2019 -0700

bpf: relax inode permission check for retrieving bpf program

For iptable module to load a bpf program from a pinned location, it only retrieve a loaded program and cannot change the program content so requiring a write permission for it might not be necessary. Also when adding or removing an unrelated iptable rule, it might need to flush and reload the xt_bpf related rules as well and triggers the inode permission check. It might be better to remove the write premission check for the inode so we won't need to grant write access to all the processes that flush and restore iptables rules.

kernel/bpf/inode.c:
- int ret = inode_permission(inode, MAY_READ | MAY_WRITE);
+ int ret = inode_permission(inode, MAY_READ);

In practice, AFAICT, the xt_bpf match .fd field isn't even used by new kernels, but I believe it might be needed for compatibility with old ones (though I'm pretty sure table modifications on them will outright fail).

Test: builds, passes Android test suite (albeit on an older iptables base),
git grep bpf_obj_get - finds no other users

Cc Chenbo Feng <[email protected]>
Cc Alexei Starovoitov <[email protected]>
Cc Willem de Bruijn <[email protected]>
---
 extensions/libxt_bpf.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)
extensions/libxt_bpf.c +3 -2
index 92958247..bf46505c 100644
@@ -61,12 +61,13 @@ static const struct xt_option_entry bpf_opts_v1[] = {
 	XTOPT_TABLEEND,
 };
 
-static int bpf_obj_get(const char *filepath)
+static int bpf_obj_get_readonly(const char *filepath)
 {
 #if defined HAVE_LINUX_BPF_H && defined __NR_bpf && defined BPF_FS_MAGIC
 	union bpf_attr attr;
 
 	memset(&attr, 0, sizeof(attr));
+	attr.file_flags = BPF_F_RDONLY;
 	attr.pathname = (__u64) filepath;
 
 	return syscall(__NR_bpf, BPF_OBJ_GET, &attr, sizeof(attr));
@@ -125,7 +126,7 @@ static void bpf_parse_string(struct sock_filter *pc, __u16 *lenp, __u16 len_max,
 static void bpf_parse_obj_pinned(struct xt_bpf_info_v1 *bi,
 				 const char *filepath)
 {
-	bi->fd = bpf_obj_get(filepath);
+	bi->fd = bpf_obj_get_readonly(filepath);
 	if (bi->fd < 0)
 		xtables_error(PARAMETER_PROBLEM,
 			      "bpf: failed to get bpf object");
Signature · 1 line
2.25.1.696.g5e7596f4ac-goog
#1 lore raw
Quoted 4 lines · Maciej Żenczykowski

From: Maciej Żenczykowski <[email protected]>

Adjust the mode eBPF programs are opened in so 0400 pinned bpf programs work without requiring CAP_DAC_OVERRIDE.

Unfortunately this is breaking stuff:

libxt_bpf.c: In function ‘bpf_obj_get_readonly’:
libxt_bpf.c:70:6: error: ‘union bpf_attr’ has no member named ‘file_flags’
70 | attr.file_flags = BPF_F_RDONLY;
| ^
libxt_bpf.c:70:20: error: ‘BPF_F_RDONLY’ undeclared (first use in this function)

   70 |  attr.file_flags = BPF_F_RDONLY;
      |                    ^~~~~~~~~~~~
#2 lore raw

I don't get it. It builds for me.

And it doesn't if I insert an intentional syntax error in the same line, so I'm definitely compiling exactly that code.

#3 lore raw

I think your build system's kernel headers are old.

Linux 4.15-rc1 commit 6e71b04a82248ccf13a94b85cbc674a9fefe53f5

Author: Chenbo Feng <[email protected]>
Date:   Wed Oct 18 13:00:22 2017 -0700

bpf: Add file mode configuration into bpf maps

include/uapi/linux/bpf.h +5 -0
@@ -218,6 +218,10 @@ enum bpf_attach_type {
 
 #define BPF_OBJ_NAME_LEN 16U
 
+/* Flags for accessing BPF object */
+#define BPF_F_RDONLY           (1U << 3)
+#define BPF_F_WRONLY           (1U << 4)
+
 union bpf_attr {
        struct { /* anonymous struct used by BPF_MAP_CREATE command */
                __u32   map_type;       /* one of enum bpf_map_type */
@@ -260,6 +264,7 @@ union bpf_attr {
        struct { /* anonymous struct used by BPF_OBJ_* commands */
                __aligned_u64   pathname;
                __u32           bpf_fd;
+               __u32           file_flags;
        };
 
        struct { /* anonymous struct used by BPF_PROG_ATTACH/DETACH commands */
Quoted 5 lines · Maciej Żenczykowski

I don't get it. It builds for me.

And it doesn't if I insert an intentional syntax error in the same line, so I'm definitely compiling exactly that code.

#4 lore raw

I guess maybe we could wrap it in a

#ifdef BPF_F_RDONLY
attr.file_flags = BPF_F_RDONLY;
#endif

if we want to continue supporting building against pre-4.15 kernel headers...

#5 lore raw

Ugh, and I guess that on a pre-4.15 kernel it would cause failures due to unknown flag...

Maybe we need to try with flag and fallback to without...

#6 lore raw

PATCH v2 1/1 iptables: open eBPF programs in read only mode

From: Maciej Żenczykowski <[email protected]>

Adjust the mode eBPF programs are opened in so 0400 pinned bpf programs work without requiring CAP_DAC_OVERRIDE.

This matches Linux 5.2's:
commit e547ff3f803e779a3898f1f48447b29f43c54085
Author: Chenbo Feng <[email protected]>
Date: Tue May 14 19:42:57 2019 -0700

bpf: relax inode permission check for retrieving bpf program

For iptable module to load a bpf program from a pinned location, it only retrieve a loaded program and cannot change the program content so requiring a write permission for it might not be necessary. Also when adding or removing an unrelated iptable rule, it might need to flush and reload the xt_bpf related rules as well and triggers the inode permission check. It might be better to remove the write premission check for the inode so we won't need to grant write access to all the processes that flush and restore iptables rules.

kernel/bpf/inode.c:
- int ret = inode_permission(inode, MAY_READ | MAY_WRITE);
+ int ret = inode_permission(inode, MAY_READ);

In practice, AFAICT, the xt_bpf match .fd field isn't even used by new kernels, but I believe it might be needed for compatibility with old ones (though I'm pretty sure table modifications on them will outright fail).

Test: builds, passes Android test suite (albeit on an older iptables base),
git grep bpf_obj_get - finds no other users

Cc Chenbo Feng <[email protected]>
Cc Alexei Starovoitov <[email protected]>
Cc Willem de Bruijn <[email protected]>
---
 extensions/libxt_bpf.c | 15 +++++++++++++--
 1 file changed, 13 insertions(+), 2 deletions(-)
extensions/libxt_bpf.c +13 -2
index 92958247..44cdd5cb 100644
@@ -61,11 +61,22 @@ static const struct xt_option_entry bpf_opts_v1[] = {
 	XTOPT_TABLEEND,
 };
 
-static int bpf_obj_get(const char *filepath)
+static int bpf_obj_get_readonly(const char *filepath)
 {
 #if defined HAVE_LINUX_BPF_H && defined __NR_bpf && defined BPF_FS_MAGIC
 	union bpf_attr attr;
+	// file_flags && BPF_F_RDONLY requires Linux 4.15+ uapi kernel headers
+#ifdef BPF_F_RDONLY
+	int fd;
 
+	memset(&attr, 0, sizeof(attr));
+	attr.pathname = (__u64) filepath;
+	attr.file_flags = BPF_F_RDONLY;
+	fd = syscall(__NR_bpf, BPF_OBJ_GET, &attr, sizeof(attr));
+	if (fd >= 0) return fd;
+
+	// on any error fallback to default R/W access for pre-4.15-rc1 kernels
+#endif
 	memset(&attr, 0, sizeof(attr));
 	attr.pathname = (__u64) filepath;
 
@@ -125,7 +136,7 @@ static void bpf_parse_string(struct sock_filter *pc, __u16 *lenp, __u16 len_max,
 static void bpf_parse_obj_pinned(struct xt_bpf_info_v1 *bi,
 				 const char *filepath)
 {
-	bi->fd = bpf_obj_get(filepath);
+	bi->fd = bpf_obj_get_readonly(filepath);
 	if (bi->fd < 0)
 		xtables_error(PARAMETER_PROBLEM,
 			      "bpf: failed to get bpf object");
Signature · 1 line
2.25.1.696.g5e7596f4ac-goog
maintainer
#7 lore raw
Quoted 5 lines · On Thu, 26 Mar 2020 07:28:03 -0700 Maciej Żenczykowski wrote:
extensions/libxt_bpf.c +0 -0
index 92958247..44cdd5cb 100644

extensions/libxt_bpf.c static const struct xt_option_entry bpf_opts_v1[] = { · line 69

Quoted 10 lines
extensions/libxt_bpf.c +3 -1
@@ -61,11 +61,22 @@ static const struct xt_option_entry bpf_opts_v1[] = {
 	XTOPT_TABLEEND,
 };
 
-static int bpf_obj_get(const char *filepath)
+static int bpf_obj_get_readonly(const char *filepath)
 {
 #if defined HAVE_LINUX_BPF_H && defined __NR_bpf && defined BPF_FS_MAGIC
 	union bpf_attr attr;
+	// file_flags && BPF_F_RDONLY requires Linux 4.15+ uapi kernel headers
+#ifdef BPF_F_RDONLY

FWIW the BPF subsystem is about to break uAPI backward-compat and replace the defines with enums. See commit 1aae4bdd7879 ("bpf: Switch BPF UAPI #define constants used from BPF program side to enums").

Quoted 10 lines
extensions/libxt_bpf.c +9 -0
+	int fd;
 
+	memset(&attr, 0, sizeof(attr));
+	attr.pathname = (__u64) filepath;
+	attr.file_flags = BPF_F_RDONLY;
+	fd = syscall(__NR_bpf, BPF_OBJ_GET, &attr, sizeof(attr));
+	if (fd >= 0) return fd;
+
+	// on any error fallback to default R/W access for pre-4.15-rc1 kernels
+#endif
#8 lore raw
Quoted 3 lines

FWIW the BPF subsystem is about to break uAPI backward-compat and replace the defines with enums. See commit 1aae4bdd7879 ("bpf: Switch BPF UAPI #define constants used from BPF program side to enums").

Shouldn't it do what is normally done in such a case? #define BPF_F_RDONLY BPF_F_RDONLY

#9 lore raw
Quoted 7 lines · Maciej Żenczykowski
Quoted 4 lines

FWIW the BPF subsystem is about to break uAPI backward-compat and replace the defines with enums. See commit 1aae4bdd7879 ("bpf: Switch BPF UAPI #define constants used from BPF program side to enums").

Shouldn't it do what is normally done in such a case? #define BPF_F_RDONLY BPF_F_RDONLY

No. just update the headers.

#10 lore raw
In reply to Maciej Żenczykowski
Quoted 7 lines · Maciej Żenczykowski

I guess maybe we could wrap it in a

#ifdef BPF_F_RDONLY
attr.file_flags = BPF_F_RDONLY;
#endif

if we want to continue supporting building against pre-4.15 kernel headers...

You can probably add a cached copy of this header file to the iptables tree via your patch like. This is done in other existing extensions to not rely on the available kernel headers.

There is no parity between userspace iptables and kernel version, it is good if you make sure this compiles for older kernels are still supported.

Thank you.

#11 lore raw

PATCH v3 1/1 iptables: open eBPF programs in read only mode

In reply to Maciej Żenczykowski

From: Maciej Żenczykowski <[email protected]>

Adjust the mode eBPF programs are opened in so 0400 pinned bpf programs work without requiring CAP_DAC_OVERRIDE.

This matches Linux 5.2's:
commit e547ff3f803e779a3898f1f48447b29f43c54085
Author: Chenbo Feng <[email protected]>
Date: Tue May 14 19:42:57 2019 -0700

bpf: relax inode permission check for retrieving bpf program

For iptable module to load a bpf program from a pinned location, it only retrieve a loaded program and cannot change the program content so requiring a write permission for it might not be necessary. Also when adding or removing an unrelated iptable rule, it might need to flush and reload the xt_bpf related rules as well and triggers the inode permission check. It might be better to remove the write premission check for the inode so we won't need to grant write access to all the processes that flush and restore iptables rules.

kernel/bpf/inode.c:
- int ret = inode_permission(inode, MAY_READ | MAY_WRITE);
+ int ret = inode_permission(inode, MAY_READ);

In practice, AFAICT, the xt_bpf match .fd field isn't even used by new kernels, but I believe it might be needed for compatibility with old ones (though I'm pretty sure table modifications on them will outright fail).

Test: builds, passes Android test suite (albeit on an older iptables base),
git grep bpf_obj_get - finds no other users

Cc Chenbo Feng <[email protected]>
Cc Alexei Starovoitov <[email protected]>
Cc Willem de Bruijn <[email protected]>
---
 extensions/libxt_bpf.c | 25 ++++++++++++++++++-------
 1 file changed, 18 insertions(+), 7 deletions(-)
extensions/libxt_bpf.c +18 -7
index 92958247..4aea477a 100644
@@ -61,14 +61,25 @@ static const struct xt_option_entry bpf_opts_v1[] = {
 	XTOPT_TABLEEND,
 };
 
-static int bpf_obj_get(const char *filepath)
+static int bpf_obj_get_readonly(const char *filepath)
 {
 #if defined HAVE_LINUX_BPF_H && defined __NR_bpf && defined BPF_FS_MAGIC
-	union bpf_attr attr;
-
-	memset(&attr, 0, sizeof(attr));
-	attr.pathname = (__u64) filepath;
-
+	// union bpf_attr includes this in an anonymous struct, but the
+	// file_flags field and the BPF_F_RDONLY constant are only present
+	// in Linux 4.15+ kernel headers (include/uapi/linux/bpf.h)
+	struct {   // this part of union bpf_attr is for BPF_OBJ_* commands
+		__aligned_u64	pathname;
+		__u32		bpf_fd;
+		__u32		file_flags;
+	} attr = {
+		.pathname = (__u64)filepath,
+		.file_flags = (1U << 3),   // BPF_F_RDONLY
+	};
+	int fd = syscall(__NR_bpf, BPF_OBJ_GET, &attr, sizeof(attr));
+	if (fd >= 0) return fd;
+
+	// on any error fallback to default R/W access for pre-4.15-rc1 kernels
+	attr.file_flags = 0;
 	return syscall(__NR_bpf, BPF_OBJ_GET, &attr, sizeof(attr));
 #else
 	xtables_error(OTHER_PROBLEM,
@@ -125,7 +136,7 @@ static void bpf_parse_string(struct sock_filter *pc, __u16 *lenp, __u16 len_max,
 static void bpf_parse_obj_pinned(struct xt_bpf_info_v1 *bi,
 				 const char *filepath)
 {
-	bi->fd = bpf_obj_get(filepath);
+	bi->fd = bpf_obj_get_readonly(filepath);
 	if (bi->fd < 0)
 		xtables_error(PARAMETER_PROBLEM,
 			      "bpf: failed to get bpf object");
Signature · 1 line
2.26.0.rc2.310.g2932bb562d-goog