[PATCH net v2] bpf: Fix build when CONFIG_BPF_SYSCALL is disabled

When CONFIG_DEBUG_INFO_BTF is enabled and CONFIG_BPF_SYSCALL is
disabled, the following compilation error can be seen:

  GEN     .version
  CHK     include/generated/compile.h
  UPD     include/generated/compile.h
  CC      init/version.o
  AR      init/built-in.a
  LD      vmlinux.o
  MODPOST vmlinux.symvers
  MODINFO modules.builtin.modinfo
  GEN     modules.builtin
  LD      .tmp_vmlinux.btf
ld: net/ipv4/tcp_cubic.o: in function `cubictcp_unregister':
net/ipv4/tcp_cubic.c:545: undefined reference to `bpf_tcp_ca_kfunc_list'
ld: net/ipv4/tcp_cubic.c:545: undefined reference to `unregister_kfunc_btf_id_set'
ld: net/ipv4/tcp_cubic.o: in function `cubictcp_register':
net/ipv4/tcp_cubic.c:539: undefined reference to `bpf_tcp_ca_kfunc_list'
ld: net/ipv4/tcp_cubic.c:539: undefined reference to `register_kfunc_btf_id_set'
  BTF     .btf.vmlinux.bin.o
pahole: .tmp_vmlinux.btf: No such file or directory
  LD      .tmp_vmlinux.kallsyms1
.btf.vmlinux.bin.o: file not recognized: file format not recognized
make: *** [Makefile:1187: vmlinux] Error 1

'bpf_tcp_ca_kfunc_list', 'register_kfunc_btf_id_set()' and
'unregister_kfunc_btf_id_set()' are only defined when
CONFIG_BPF_SYSCALL is enabled.

Fix that by moving those definitions somewhere that doesn't depend on
the bpf() syscall.

Fixes: 14f267d95fe4 ("bpf: btf: Introduce helpers for dynamic BTF set registration")
Signed-off-by: Vinicius Costa Gomes <vinicius.gomes@intel.com>
---
 kernel/bpf/btf.c  | 33 ---------------------------------
 kernel/bpf/core.c | 37 +++++++++++++++++++++++++++++++++++++
 2 files changed, 37 insertions(+), 33 deletions(-)

diff --git a/kernel/bpf/btf.c b/kernel/bpf/btf.c
index dbc3ad07e21b..cc9868376345 100644
--- a/kernel/bpf/btf.c
+++ b/kernel/bpf/btf.c

@@ -6344,33 +6344,8 @@ const struct bpf_func_proto bpf_btf_find_by_name_kind_proto = {
 
 BTF_ID_LIST_GLOBAL_SINGLE(btf_task_struct_ids, struct, task_struct)
 
-/* BTF ID set registration API for modules */
-
-struct kfunc_btf_id_list {
-	struct list_head list;
-	struct mutex mutex;
-};
-
 #ifdef CONFIG_DEBUG_INFO_BTF_MODULES
 
-void register_kfunc_btf_id_set(struct kfunc_btf_id_list *l,
-			       struct kfunc_btf_id_set *s)
-{
-	mutex_lock(&l->mutex);
-	list_add(&s->list, &l->list);
-	mutex_unlock(&l->mutex);
-}
-EXPORT_SYMBOL_GPL(register_kfunc_btf_id_set);
-
-void unregister_kfunc_btf_id_set(struct kfunc_btf_id_list *l,
-				 struct kfunc_btf_id_set *s)
-{
-	mutex_lock(&l->mutex);
-	list_del_init(&s->list);
-	mutex_unlock(&l->mutex);
-}
-EXPORT_SYMBOL_GPL(unregister_kfunc_btf_id_set);
-
 bool bpf_check_mod_kfunc_call(struct kfunc_btf_id_list *klist, u32 kfunc_id,
 			      struct module *owner)
 {

@@ -6390,11 +6365,3 @@ bool bpf_check_mod_kfunc_call(struct kfunc_btf_id_list *klist, u32 kfunc_id,
 }
 
 #endif
-
-#define DEFINE_KFUNC_BTF_ID_LIST(name)                                         \
-	struct kfunc_btf_id_list name = { LIST_HEAD_INIT(name.list),           \
-					  __MUTEX_INITIALIZER(name.mutex) };   \
-	EXPORT_SYMBOL_GPL(name)
-
-DEFINE_KFUNC_BTF_ID_LIST(bpf_tcp_ca_kfunc_list);
-DEFINE_KFUNC_BTF_ID_LIST(prog_test_kfunc_list);

diff --git a/kernel/bpf/core.c b/kernel/bpf/core.c
index 2405e39d800f..f2939a6d8199 100644
--- a/kernel/bpf/core.c
+++ b/kernel/bpf/core.c

@@ -2456,6 +2456,43 @@ int __weak bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
 DEFINE_STATIC_KEY_FALSE(bpf_stats_enabled_key);
 EXPORT_SYMBOL(bpf_stats_enabled_key);
 
+/* BTF ID set registration API for modules */
+
+struct kfunc_btf_id_list {
+	struct list_head list;
+	struct mutex mutex;
+};
+
+#ifdef CONFIG_DEBUG_INFO_BTF_MODULES
+
+void register_kfunc_btf_id_set(struct kfunc_btf_id_list *l,
+			       struct kfunc_btf_id_set *s)
+{
+	mutex_lock(&l->mutex);
+	list_add(&s->list, &l->list);
+	mutex_unlock(&l->mutex);
+}
+EXPORT_SYMBOL_GPL(register_kfunc_btf_id_set);
+
+void unregister_kfunc_btf_id_set(struct kfunc_btf_id_list *l,
+				 struct kfunc_btf_id_set *s)
+{
+	mutex_lock(&l->mutex);
+	list_del_init(&s->list);
+	mutex_unlock(&l->mutex);
+}
+EXPORT_SYMBOL_GPL(unregister_kfunc_btf_id_set);
+
+#endif
+
+#define DEFINE_KFUNC_BTF_ID_LIST(name)                                         \
+	struct kfunc_btf_id_list name = { LIST_HEAD_INIT(name.list),           \
+					  __MUTEX_INITIALIZER(name.mutex) };   \
+	EXPORT_SYMBOL_GPL(name)
+
+DEFINE_KFUNC_BTF_ID_LIST(bpf_tcp_ca_kfunc_list);
+DEFINE_KFUNC_BTF_ID_LIST(prog_test_kfunc_list);
+
 /* All definitions of tracepoints related to BPF. */
 #define CREATE_TRACE_POINTS
 #include <linux/bpf_trace.h>

-- 
2.32.0

On Thu, Nov 11, 2021 at 02:24:18AM IST, Vinicius Costa Gomes wrote:

When CONFIG_DEBUG_INFO_BTF is enabled and CONFIG_BPF_SYSCALL is
disabled, the following compilation error can be seen:

  GEN     .version
  CHK     include/generated/compile.h
  UPD     include/generated/compile.h
  CC      init/version.o
  AR      init/built-in.a
  LD      vmlinux.o
  MODPOST vmlinux.symvers
  MODINFO modules.builtin.modinfo
  GEN     modules.builtin
  LD      .tmp_vmlinux.btf
ld: net/ipv4/tcp_cubic.o: in function `cubictcp_unregister':
net/ipv4/tcp_cubic.c:545: undefined reference to `bpf_tcp_ca_kfunc_list'
ld: net/ipv4/tcp_cubic.c:545: undefined reference to `unregister_kfunc_btf_id_set'
ld: net/ipv4/tcp_cubic.o: in function `cubictcp_register':
net/ipv4/tcp_cubic.c:539: undefined reference to `bpf_tcp_ca_kfunc_list'
ld: net/ipv4/tcp_cubic.c:539: undefined reference to `register_kfunc_btf_id_set'
  BTF     .btf.vmlinux.bin.o
pahole: .tmp_vmlinux.btf: No such file or directory
  LD      .tmp_vmlinux.kallsyms1
.btf.vmlinux.bin.o: file not recognized: file format not recognized
make: *** [Makefile:1187: vmlinux] Error 1

'bpf_tcp_ca_kfunc_list', 'register_kfunc_btf_id_set()' and
'unregister_kfunc_btf_id_set()' are only defined when
CONFIG_BPF_SYSCALL is enabled.

Fix that by moving those definitions somewhere that doesn't depend on
the bpf() syscall.

Fixes: 14f267d95fe4 ("bpf: btf: Introduce helpers for dynamic BTF set registration")
Signed-off-by: Vinicius Costa Gomes <vinicius.gomes@intel.com>

Thanks for the fix.

But instead of moving this to core.c, you can probably make the btf.h
declaration conditional on CONFIG_BPF_SYSCALL, since this is not useful in
isolation (only used by verifier for module kfunc support). For the case of
kfunc_btf_id_list variables, just define it as an empty struct and static
variables, since the definition is still inside btf.c. So it becomes a noop for
!CONFIG_BPF_SYSCALL.

I am also not sure whether BTF is useful without BPF support, but maybe I'm
missing some usecase.

That's just my opinion however, I'll defer to BPF maintainers.

--
Kartikeya

On Wed, Nov 10, 2021 at 1:25 PM Kumar Kartikeya Dwivedi
[off-list ref] wrote:

On Thu, Nov 11, 2021 at 02:24:18AM IST, Vinicius Costa Gomes wrote:

quoted

When CONFIG_DEBUG_INFO_BTF is enabled and CONFIG_BPF_SYSCALL is
disabled, the following compilation error can be seen:

  GEN     .version
  CHK     include/generated/compile.h
  UPD     include/generated/compile.h
  CC      init/version.o
  AR      init/built-in.a
  LD      vmlinux.o
  MODPOST vmlinux.symvers
  MODINFO modules.builtin.modinfo
  GEN     modules.builtin
  LD      .tmp_vmlinux.btf
ld: net/ipv4/tcp_cubic.o: in function `cubictcp_unregister':
net/ipv4/tcp_cubic.c:545: undefined reference to `bpf_tcp_ca_kfunc_list'
ld: net/ipv4/tcp_cubic.c:545: undefined reference to `unregister_kfunc_btf_id_set'
ld: net/ipv4/tcp_cubic.o: in function `cubictcp_register':
net/ipv4/tcp_cubic.c:539: undefined reference to `bpf_tcp_ca_kfunc_list'
ld: net/ipv4/tcp_cubic.c:539: undefined reference to `register_kfunc_btf_id_set'
  BTF     .btf.vmlinux.bin.o
pahole: .tmp_vmlinux.btf: No such file or directory
  LD      .tmp_vmlinux.kallsyms1
.btf.vmlinux.bin.o: file not recognized: file format not recognized
make: *** [Makefile:1187: vmlinux] Error 1

'bpf_tcp_ca_kfunc_list', 'register_kfunc_btf_id_set()' and
'unregister_kfunc_btf_id_set()' are only defined when
CONFIG_BPF_SYSCALL is enabled.

Fix that by moving those definitions somewhere that doesn't depend on
the bpf() syscall.

Fixes: 14f267d95fe4 ("bpf: btf: Introduce helpers for dynamic BTF set registration")
Signed-off-by: Vinicius Costa Gomes <vinicius.gomes@intel.com>

Thanks for the fix.

But instead of moving this to core.c, you can probably make the btf.h
declaration conditional on CONFIG_BPF_SYSCALL, since this is not useful in
isolation (only used by verifier for module kfunc support). For the case of
kfunc_btf_id_list variables, just define it as an empty struct and static
variables, since the definition is still inside btf.c. So it becomes a noop for
!CONFIG_BPF_SYSCALL.

I am also not sure whether BTF is useful without BPF support, but maybe I'm
missing some usecase.

Unlikely. I would just disallow such config instead of sprinkling
the code with ifdefs.

Hi,

Kumar Kartikeya Dwivedi [off-list ref] writes:

On Thu, Nov 11, 2021 at 02:24:18AM IST, Vinicius Costa Gomes wrote:

quoted

When CONFIG_DEBUG_INFO_BTF is enabled and CONFIG_BPF_SYSCALL is
disabled, the following compilation error can be seen:

  GEN     .version
  CHK     include/generated/compile.h
  UPD     include/generated/compile.h
  CC      init/version.o
  AR      init/built-in.a
  LD      vmlinux.o
  MODPOST vmlinux.symvers
  MODINFO modules.builtin.modinfo
  GEN     modules.builtin
  LD      .tmp_vmlinux.btf
ld: net/ipv4/tcp_cubic.o: in function `cubictcp_unregister':
net/ipv4/tcp_cubic.c:545: undefined reference to `bpf_tcp_ca_kfunc_list'
ld: net/ipv4/tcp_cubic.c:545: undefined reference to `unregister_kfunc_btf_id_set'
ld: net/ipv4/tcp_cubic.o: in function `cubictcp_register':
net/ipv4/tcp_cubic.c:539: undefined reference to `bpf_tcp_ca_kfunc_list'
ld: net/ipv4/tcp_cubic.c:539: undefined reference to `register_kfunc_btf_id_set'
  BTF     .btf.vmlinux.bin.o
pahole: .tmp_vmlinux.btf: No such file or directory
  LD      .tmp_vmlinux.kallsyms1
.btf.vmlinux.bin.o: file not recognized: file format not recognized
make: *** [Makefile:1187: vmlinux] Error 1

'bpf_tcp_ca_kfunc_list', 'register_kfunc_btf_id_set()' and
'unregister_kfunc_btf_id_set()' are only defined when
CONFIG_BPF_SYSCALL is enabled.

Fix that by moving those definitions somewhere that doesn't depend on
the bpf() syscall.

Fixes: 14f267d95fe4 ("bpf: btf: Introduce helpers for dynamic BTF set registration")
Signed-off-by: Vinicius Costa Gomes <vinicius.gomes@intel.com>

Thanks for the fix.

But instead of moving this to core.c, you can probably make the btf.h
declaration conditional on CONFIG_BPF_SYSCALL, since this is not useful in
isolation (only used by verifier for module kfunc support). For the case of
kfunc_btf_id_list variables, just define it as an empty struct and static
variables, since the definition is still inside btf.c. So it becomes a noop for
!CONFIG_BPF_SYSCALL.

I am also not sure whether BTF is useful without BPF support, but maybe I'm
missing some usecase.

From my side, you are not missing anything, it was just random chance
that I had a 'x86_64_defconfig + debug + BTF' .config laying around and
the build broke with it. I don't have any real usecases for this
combination.


Cheers,
-- 
Vinicius

Alexei Starovoitov [off-list ref] writes:

quoted

Thanks for the fix.

But instead of moving this to core.c, you can probably make the btf.h
declaration conditional on CONFIG_BPF_SYSCALL, since this is not useful in
isolation (only used by verifier for module kfunc support). For the case of
kfunc_btf_id_list variables, just define it as an empty struct and static
variables, since the definition is still inside btf.c. So it becomes a noop for
!CONFIG_BPF_SYSCALL.

I am also not sure whether BTF is useful without BPF support, but maybe I'm
missing some usecase.

Unlikely. I would just disallow such config instead of sprinkling
the code with ifdefs.

Is something like this what you have in mind?

diff --git a/lib/Kconfig.debug b/lib/Kconfig.debug
index 6fdbf9613aec..eae860c86e26 100644
--- a/lib/Kconfig.debug
+++ b/lib/Kconfig.debug

@@ -316,6 +316,7 @@ config DEBUG_INFO_BTF
 	bool "Generate BTF typeinfo"
 	depends on !DEBUG_INFO_SPLIT && !DEBUG_INFO_REDUCED
 	depends on !GCC_PLUGIN_RANDSTRUCT || COMPILE_TEST
+	depends on BPF_SYSCALL
 	help
 	  Generate deduplicated BTF type information from DWARF debug info.
 	  Turning this on expects presence of pahole tool, which will convert

Cheers,
-- 
Vinicius

On Thu, Nov 11, 2021 at 05:21:53AM IST, Vinicius Costa Gomes wrote:

BTW, you will need a little more than that, I suspect the compiler optimizes out
the register/unregister call so we don't see a build failure, but adding a side
effect gives me errors, so something like this should resolve the problem (since
kfunc_btf_id_list variable definition is behind CONFIG_BPF_SYSCALL).

diff --git a/include/linux/btf.h b/include/linux/btf.h
index 203eef993d76..e9881ef9e9aa 100644
--- a/include/linux/btf.h
+++ b/include/linux/btf.h

@@ -254,6 +254,8 @@ void unregister_kfunc_btf_id_set(struct kfunc_btf_id_list *l,
                                 struct kfunc_btf_id_set *s);
 bool bpf_check_mod_kfunc_call(struct kfunc_btf_id_list *klist, u32 kfunc_id,
                              struct module *owner);
+extern struct kfunc_btf_id_list bpf_tcp_ca_kfunc_list;
+extern struct kfunc_btf_id_list prog_test_kfunc_list;
 #else
 static inline void register_kfunc_btf_id_set(struct kfunc_btf_id_list *l,
                                             struct kfunc_btf_id_set *s)

@@ -268,13 +270,13 @@ static inline bool bpf_check_mod_kfunc_call(struct kfunc_btf_id_list *klist,
 {
        return false;
 }
+struct kfunc_btf_id_list {};
+static struct kfunc_btf_id_list bpf_tcp_ca_kfunc_list __maybe_unused;
+static struct kfunc_btf_id_list prog_test_kfunc_list __maybe_unused;
+
 #endif

 #define DEFINE_KFUNC_BTF_ID_SET(set, name)                                     \
        struct kfunc_btf_id_set name = { LIST_HEAD_INIT(name.list), (set),     \
                                         THIS_MODULE }
-
-extern struct kfunc_btf_id_list bpf_tcp_ca_kfunc_list;
-extern struct kfunc_btf_id_list prog_test_kfunc_list;
-
 #endif

--

Kartikeya

Hi Kartikeya,

Kumar Kartikeya Dwivedi [off-list ref] writes:

I could not reproduce the build failure here even when adding some side
effects, but I didn't try very hard.

As you are more familiar with the code, I would be glad if you could
take it from here and propose a patch.


Cheers,
-- 
Vinicius

On Thu, Nov 11, 2021 at 07:34:38AM IST, Vinicius Costa Gomes wrote:

Hi Kartikeya,

Kumar Kartikeya Dwivedi [off-list ref] writes:

quoted

On Thu, Nov 11, 2021 at 05:21:53AM IST, Vinicius Costa Gomes wrote:

quoted

Alexei Starovoitov [off-list ref] writes:

quoted

quoted

Thanks for the fix.

But instead of moving this to core.c, you can probably make the btf.h
declaration conditional on CONFIG_BPF_SYSCALL, since this is not useful in
isolation (only used by verifier for module kfunc support). For the case of
kfunc_btf_id_list variables, just define it as an empty struct and static
variables, since the definition is still inside btf.c. So it becomes a noop for
!CONFIG_BPF_SYSCALL.

I am also not sure whether BTF is useful without BPF support, but maybe I'm
missing some usecase.

Unlikely. I would just disallow such config instead of sprinkling
the code with ifdefs.

Is something like this what you have in mind?

diff --git a/lib/Kconfig.debug b/lib/Kconfig.debug
index 6fdbf9613aec..eae860c86e26 100644
--- a/lib/Kconfig.debug
+++ b/lib/Kconfig.debug

@@ -316,6 +316,7 @@ config DEBUG_INFO_BTF
 	bool "Generate BTF typeinfo"
 	depends on !DEBUG_INFO_SPLIT && !DEBUG_INFO_REDUCED
 	depends on !GCC_PLUGIN_RANDSTRUCT || COMPILE_TEST
+	depends on BPF_SYSCALL
 	help
 	  Generate deduplicated BTF type information from DWARF debug info.
 	  Turning this on expects presence of pahole tool, which will convert

BTW, you will need a little more than that, I suspect the compiler optimizes out
the register/unregister call so we don't see a build failure, but adding a side
effect gives me errors, so something like this should resolve the problem (since
kfunc_btf_id_list variable definition is behind CONFIG_BPF_SYSCALL).

diff --git a/include/linux/btf.h b/include/linux/btf.h
index 203eef993d76..e9881ef9e9aa 100644
--- a/include/linux/btf.h
+++ b/include/linux/btf.h

@@ -254,6 +254,8 @@ void unregister_kfunc_btf_id_set(struct kfunc_btf_id_list *l,
                                 struct kfunc_btf_id_set *s);
 bool bpf_check_mod_kfunc_call(struct kfunc_btf_id_list *klist, u32 kfunc_id,
                              struct module *owner);
+extern struct kfunc_btf_id_list bpf_tcp_ca_kfunc_list;
+extern struct kfunc_btf_id_list prog_test_kfunc_list;
 #else
 static inline void register_kfunc_btf_id_set(struct kfunc_btf_id_list *l,
                                             struct kfunc_btf_id_set *s)

@@ -268,13 +270,13 @@ static inline bool bpf_check_mod_kfunc_call(struct kfunc_btf_id_list *klist,
 {
        return false;
 }
+struct kfunc_btf_id_list {};
+static struct kfunc_btf_id_list bpf_tcp_ca_kfunc_list __maybe_unused;
+static struct kfunc_btf_id_list prog_test_kfunc_list __maybe_unused;
+
 #endif

 #define DEFINE_KFUNC_BTF_ID_SET(set, name)                                     \
        struct kfunc_btf_id_set name = { LIST_HEAD_INIT(name.list), (set),     \
                                         THIS_MODULE }
-
-extern struct kfunc_btf_id_list bpf_tcp_ca_kfunc_list;
-extern struct kfunc_btf_id_list prog_test_kfunc_list;
-
 #endif

I could not reproduce the build failure here even when adding some side
effects, but I didn't try very hard.

As you are more familiar with the code, I would be glad if you could
take it from here and propose a patch.

Sure, no worries!

Cheers,
--
Vinicius

--
Kartikeya