On 3/15/21 4:33 PM, Daniel Borkmann wrote:
On 3/14/21 4:58 AM, Yonghong Song wrote:
[...]
This patch explicited add an expression like
(void)BPF_TCP_ESTABLISHED
to enable generation of debuginfo for the anonymous
enum which also includes BPF_TCP_CLOSE. I put
this explicit type generation in kernel/bpf/core.c
to (1) avoid polute net/ipv4/tcp.c and more importantly
(2) provide a central place to add other types (e.g. in
bpf/btf uapi header) if they are not referenced in the kernel
or generated in vmlinux dwarf.
Signed-off-by: Yonghong Song <yhs@xxxxxx>
---
include/linux/btf.h | 1 +
kernel/bpf/core.c | 19 +++++++++++++++++++
2 files changed, 20 insertions(+)
diff --git a/include/linux/btf.h b/include/linux/btf.h
index 7fabf1428093..9c1b52738bbe 100644
--- a/include/linux/btf.h
+++ b/include/linux/btf.h
@@ -9,6 +9,7 @@
#include <uapi/linux/bpf.h>
#define BTF_TYPE_EMIT(type) ((void)(type *)0)
+#define BTF_TYPE_EMIT_ENUM(enum_val) ((void)enum_val)
struct btf;
struct btf_member;
diff --git a/kernel/bpf/core.c b/kernel/bpf/core.c
index 3a283bf97f2f..60551bf68ece 100644
--- a/kernel/bpf/core.c
+++ b/kernel/bpf/core.c
@@ -2378,3 +2378,22 @@ EXPORT_SYMBOL(bpf_stats_enabled_key);
EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_exception);
EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_bulk_tx);
+
+static int __init bpf_emit_btf_type(void)
+{
+ /* bpf uapi header bpf.h defines an anonymous enum with values
+ * BPF_TCP_* used by bpf programs. Currently gcc built vmlinux
+ * is able to emit this enum in dwarf due to the following
+ * BUILD_BUG_ON test in net/ipv4/tcp.c:
+ * BUILD_BUG_ON((int)BPF_TCP_ESTABLISHED != (int)TCP_ESTABLISHED);
+ * clang built vmlinux does not have this enum in dwarf
+ * since clang removes the above code before generating
IR/debuginfo.
+ * Let us explicitly emit the type debuginfo to ensure the
+ * above-mentioned anonymous enum in the vmlinux dwarf and hence BTF
+ * regardless of which compiler is used.
+ */
+ BTF_TYPE_EMIT_ENUM(BPF_TCP_ESTABLISHED);
+
+ return 0;
+}
+late_initcall(bpf_emit_btf_type);
Does this have to be late_initcall() given this adds minor init call
overhead, what if this would be exported as symbol for modules instead?
If issuing types in module, if I understand correctly, it will not be in
main vmlinux btf, so programs will not be able to use unless module
is loaded. I would prefer such types always available in vmlinux btf.
I am using a separate late_initcall just to cleaner codes. But
this BTF_TYPE_EMIT_ENUM can be in any init call.
$ grep _initcall *.c
btf.c:fs_initcall(btf_module_init);
core.c:pure_initcall(bpf_jit_charge_init);
cpumap.c:subsys_initcall(cpu_map_init);
devmap.c:subsys_initcall(dev_map_init);
inode.c:fs_initcall(bpf_init);
map_iter.c:late_initcall(bpf_map_iter_init);
net_namespace.c:subsys_initcall(netns_bpf_init);
prog_iter.c:late_initcall(bpf_prog_iter_init);
stackmap.c:subsys_initcall(stack_map_init);
sysfs_btf.c:subsys_initcall(btf_vmlinux_init);
task_iter.c:late_initcall(task_iter_init);
trampoline.c:late_initcall(init_trampolines);
$
I think we can use any above in kernel/bpf directory. This way,
we will have 0 runtime overhead as the code will be optimized away.
Any preference?
Thanks,
Daniel
