PR33428: imply kernel<vmlinux.h> in kernel-side @cast() for type info

Linux 5.7+ includes a vmlinux.h file generated during build that includes most or all kernel-side data structure declarations, mainly for use by BPF compilers. By processing that into DWARF via the typequery mechanism, we can generate kernel data structure debuginfo for use in many @cast operations, without having to painstakingly enumerate and track the exact kernel header files containing those declarations. This search is active and preferred by default for all kernel-space @cast operations. It can be disabled with "--compatible=5.3" . 

5 files changed, 76 insertions, 13 deletions

diff --git a/NEWS b/NEWS
index f1d78e8fa..76c9676a6 100644
--- a/NEWS
+++ b/NEWS
@@ -1,5 +1,13 @@
 * What's new in version 5.4
 
+- The kernel-context @cast operator now implicitly searches a kernel's
+ <vmlinux.h> generated header file first, if available (kernel 5.7+),
+ for type declarations. This can make some debuginfo access
+ unnecessary, and thus processing faster. It can work around kernel
+ API changes where type declarations move between headers.
+ Use --compatible=5.3 to disable this behaviour.
+ Use @cast(..., "kernel<vmlinux.h>") manually if desired.
+
 - Type checking and autocast processing have been made more thorough,
  so elided variables are checked more and @defined() tests may be
  more complicated. Preexisting scripts that rely on elision for
diff --git a/buildrun.cxx b/buildrun.cxx
index ad299d0fc..1f56a976b 100644
--- a/buildrun.cxx
+++ b/buildrun.cxx
@@ -1201,9 +1201,19 @@ make_typequery_kmod(systemtap_session& s, const vector<string>& headers, string&
  // full kernel build tree, it's possible to get at types that aren't in the
  // normal include path, e.g.:
  // @cast(foo, "bsd_acct_struct", "kernel<kernel/acct.c>")->...
- omf << "CFLAGS_" << basename << ".o :=";
+ omf << "CFLAGS_" << basename << ".o := -w "; // suppress warnings
+ bool no_vmlinux_h = true;
  for (size_t i = 0; i < headers.size(); ++i)
- omf << " -include " << lex_cast_qstring(headers[i]); // XXX right quoting?
+ {
+ const string& h = headers[i];
+ if (h == string("vmlinux.h")) // PR33428: vmlinux.h special case
+ {
+ omf << " -include " << lex_cast_qstring(s.kernel_build_tree) << "/" << lex_cast_qstring("vmlinux.h");
+ no_vmlinux_h = false;
+ }
+ else
+ omf << " -include " << lex_cast_qstring(h); // XXX right quoting?
+ }
  omf << endl;
 
  omf << "obj-m := " + basename + ".o" << endl;
@@ -1213,10 +1223,17 @@ make_typequery_kmod(systemtap_session& s, const vector<string>& headers, string&
  string source(dir + "/" + basename + ".c");
  ofstream osrc(source.c_str());
 
- // this is mandated by linux kbuild as of 5.11+
- osrc << "#include <linux/module.h>" << endl;
- osrc << "MODULE_LICENSE(\"GPL\");" << endl;
- osrc << "MODULE_DESCRIPTION(\"" << basename << "\");" << endl;
+ if (no_vmlinux_h) {
+ // this is mandated by linux kbuild as of 5.11+
+ osrc << "#include <linux/module.h>" << endl;
+ osrc << "MODULE_LICENSE(\"GPL\");" << endl;
+ osrc << "MODULE_DESCRIPTION(\"" << basename << "\");" << endl;
+ } else {
+ // PR33428: do the same as the above, but without using linux/module.h macro infrastructure, because
+ // vmlinux.h conflicts with kernel headers.
+ osrc << "static const char modinfo[] __attribute__((__used__)) __attribute__((__section__(\".modinfo\"))) = \"license=GPL\\0description=typequery\";" << endl;
+ }
+ 
  
  osrc.close();
 
diff --git a/man/stap.1.in b/man/stap.1.in
index 80c9d50d0..0a069eba7 100644
--- a/man/stap.1.in
+++ b/man/stap.1.in
@@ -1899,6 +1899,10 @@ surrounded by angle brackets, in case normal debuginfo is not available. For
 kernel headers, prefix it with "kernel" to use the appropriate build system.
 All other headers are built with default GCC parameters into a user module.
 Multiple headers may be specified in sequence to resolve a codependency.
+.PP
+As of systemtap version 5.4, the kernel "<vmlinux.h>" header is tried before
+any specified kernel headers or modules, which may make it unnecessary to list
+any explicit module or header file names for kernel data structures.
 .SAMPLE
 @cast(tv, "timeval", "<sys/time.h>")\->tv_sec
 @cast(task, "task_struct", "kernel<linux/sched.h>")\->tgid
diff --git a/tapsets.cxx b/tapsets.cxx
index 38c8a49b2..d267b6506 100644
--- a/tapsets.cxx
+++ b/tapsets.cxx
@@ -500,10 +500,10 @@ static const string TOK_LIBRARY("library");
 static const string TOK_PLT("plt");
 static const string TOK_METHOD("method");
 static const string TOK_CLASS("class");;
-static const string TOK_CALLEE("callee");;
-static const string TOK_CALLEES("callees");;
-static const string TOK_NEAREST("nearest");;
-
+static const string TOK_CALLEE("callee");
+static const string TOK_CALLEES("callees");
+static const string TOK_NEAREST("nearest");
+static const string TOK_KERNEL_VMLINUX_H("kernel<vmlinux.h>");
 
 
 struct dwarf_query; // forward decl
@@ -4853,7 +4853,7 @@ dwarf_var_expanding_visitor::visit_target_symbol (target_symbol *e)
 void
 dwarf_var_expanding_visitor::visit_cast_op (cast_op *e)
 {
- // Fill in our current module context if needed
+ // Fill in our current module context if needed, i.e., absent third field in @cast() 
  if (e->module.empty())
  {
  // Backward compatibility for @cast() ops, sans module string,
@@ -4862,7 +4862,16 @@ dwarf_var_expanding_visitor::visit_cast_op (cast_op *e)
  if (strverscmp(sess.compatible.c_str(), "4.3") < 0)
  e->module = "kernel";
  else
- e->module = q.dw.module_name;
+ {
+ // absolute /user/space/path/name xor kernel xor kernel-module name
+ if (is_user_module (q.dw.module_name))
+ e->module = q.dw.module_name;
+ else if ((strverscmp(sess.compatible.c_str(), "5.4") >= 0) && // default on new enough systemtap
+ (strverscmp(sess.kernel_base_release.c_str(), "6.7") >= 0)) // for new enough kernel to have a vmlinux.h
+ e->module = string(TOK_KERNEL_VMLINUX_H) + string(":") + q.dw.module_name; // PR33428: prefix
+ else
+ e->module = q.dw.module_name; 
+ }
  }
  
  var_expanding_visitor::visit_cast_op(e);
@@ -5153,10 +5162,27 @@ void dwarf_cast_expanding_visitor::visit_cast_op (cast_op* e)
 
  // split the module string by ':' for alternatives
  vector<string> modules;
+
+ // PR33428: prepend "kernel<vmlinux.h>" to the list if there is any
+ // "kernel" or "kernel<FILE>" component.
+ if ((strverscmp(sess.compatible.c_str(), "5.4") >= 0) && // default on new enough systemtap
+ (strverscmp(sess.kernel_base_release.c_str(), "6.7") >= 0) && // for new enough kernel to have a vmlinux.h
+ (e->module.find(TOK_KERNEL_VMLINUX_H) == string::npos)) // don't prepend again; might already be here from implicit "" expansion
+ {
+ if (e->module.starts_with("kernel")) // right at the front?
+ e->module = string(TOK_KERNEL_VMLINUX_H) + string(":") + e->module;
+ else {
+ string::size_type p = e->module.find(":kernel"); // in the middle?
+ if (p != string::npos)
+ e->module.insert(p, TOK_KERNEL_VMLINUX_H + string (":"));
+ }
+ }
+ 
  tokenize(e->module, modules, ":");
- bool userspace_p=false; // PR10601
+ 
  for (unsigned i = 0; !result && i < modules.size(); ++i)
  {
+ bool userspace_p=false; // PR10601
  string& module = modules[i];
  filter_special_modules(module);
 
diff --git a/testsuite/semok/pr33428.stp b/testsuite/semok/pr33428.stp
new file mode 100644
index 000000000..2557b00a3
--- /dev/null
+++ b/testsuite/semok/pr33428.stp
@@ -0,0 +1,8 @@
+#! stap -p2
+
+probe module("nfsv3").function("nfs3_proc_access"), kernel.function("do_sys_open")
+{
+ print(@cast(0xf00d, "struct mount")$, "kernel<vmlinux.h") // explicit
+ print(@cast(0xbeef, "struct mount")$) // implicit "nfsv3"->"kernel<vmlinux.h>:nfsv3"
+ print(@cast(0xdead, "struct mount")$,"kernel<linux/sched.h>") // explicit ->"kernel<vmlinux.h>:kernel<linux/sched.h>"
+}