BTF support for tracepoints defined in modules

For tracepoints defined in vmlinux, we support usage of BTF for type
parsing which can automatically resolve types from tracepoint defs.
Thanks to this feature, users do not have to include headers with
necessary types manually.

This commit enables the same for tracepoints defined in kernel modules.
Since parsing BTF for modules is already supported for kfuncs, it is
sufficient to add the tracepoint subsystem name to the list of modules.
If the name of the subsystem does not correspond to a loaded module, it
will be ignored.

Note that this is only supported for single-subsystem tracepoints at the
moment. The reason is that tracepoints require dumping of C definitions
from BTF which is not possible for multiple modules at once (as it could
introduce type redefinitions).

.github/workflows/ci.yml

@@ -153,8 +153,10 @@ jobs:
       run: >
         docker load --input /tmp/docker-save/i.tar
     - name: Load kernel modules
-      # nf_tables is necessary for testing kernel modules BTF support
-      run: sudo modprobe nf_tables
+      # nf_tables and xfs are necessary for testing kernel modules BTF support
+      run: |
+        sudo modprobe nf_tables
+        sudo modprobe xfs
     - name: Build and test
       env: ${{matrix.env}}
       run: >

src/driver.cpp

@@ -78,8 +78,9 @@ void Driver::error(const std::string &m)
   failed_ = true;
 }
 
-// Retrieves the list of kernel modules for all attached-to functions.
-// Currently modules are only important for k(ret)func probes.
+// Retrieves the list of kernel modules for all attachpoints. Will be used to
+// identify modules whose BTF we need to parse.
+// Currently, this is useful for k(ret)func and tracepoint probes.
 std::set<std::string> Driver::list_modules() const
 {
   std::set<std::string> modules;
@@ -102,6 +103,13 @@ std::set<std::string> Driver::list_modules() const
         else
           modules.insert(ap->target);
       }
+      else if (probe_type == ProbeType::tracepoint)
+      {
+        // For now, we support this for a single target only since tracepoints
+        // need dumping of C definitions BTF and that is not available for
+        // multiple modules at once.
+        modules.insert(ap->target);
+      }
     }
   }
   return modules;

tests/runtime/btf

@@ -62,6 +62,13 @@ REQUIRES lsmod | grep '^nf_tables'
 REQUIRES /usr/sbin/nft --help
 AFTER nft delete table bpftrace
 
+NAME kernel_module_tracepoint
+PROG tracepoint:xfs:xfs_setfilesize { print(args->offset) } i:ms:1 { exit(); }
+EXPECT Attaching 2 probes...
+TIMEOUT 5
+REQUIRES_FEATURE btf
+REQUIRES lsmod | grep "^xfs"
+
 # args->ctxt has type 'struct x86_emulate_ctxt' which is forward-defined in
 # 'vmlinux' BTF and fully defined in 'kvm' BTF.
 # This tests checks that the correct BTF definition is pulled from 'kvm'.