2007-03-20 Frank Ch. Eigler <fche@elastic.org>

	PR 4224.
	* tapsets.cxx (add_probe_point): Make kernel implicitly relocated
	relative to the _stext symbol.
	(dwarf_derived_probe ctor, emit_module_decls): Cooperate.
	(lookup_symbol_address): New function.
	(dwarf_builder::build): Call it thrice.
	(in_kprobes_function): Simplify.
	* session.h (systemtap_session): Rename cached symbol addresses.
	* translate.cxx, elaborate.cxx: Corresponding tweaks.

2007-03-20  Frank Ch. Eigler  <fche@elastic.org>

	PR 4224.
	* sym.c (_stp_module_relocate): Support kernel relocations.

2007-03-20  Frank Ch. Eigler  <fche@elastic.org>

	* symbols.c (_stp_do_symbols): Add cautionary blurb for important
	setup of _stp_modules[0]->text.

9 files changed, 124 insertions, 80 deletions

diff --git a/ChangeLog b/ChangeLog
index 19318d35..1e277306 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,3 +1,15 @@
+2007-03-20  Frank Ch. Eigler  <fche@elastic.org>
+
+	PR 4224.
+	* tapsets.cxx (add_probe_point): Make kernel implicitly relocated
+	relative to the _stext symbol.
+	(dwarf_derived_probe ctor, emit_module_decls): Cooperate.
+	(lookup_symbol_address): New function.
+	(dwarf_builder::build): Call it thrice.
+	(in_kprobes_function): Simplify.
+	* session.h (systemtap_session): Rename cached symbol addresses.
+	* translate.cxx, elaborate.cxx: Corresponding tweaks.
+
 2007-03-19  David Smith  <dsmith@redhat.com>
 
 	PR 4146 (partial fix)
diff --git a/elaborate.cxx b/elaborate.cxx
index ef019016..c2e57856 100644
--- a/elaborate.cxx
+++ b/elaborate.cxx
@@ -973,9 +973,7 @@ systemtap_session::systemtap_session ():
   mark_derived_probes(0), 
   hrtimer_derived_probes(0), 
   perfmon_derived_probes(0), 
-  op (0), up (0),
-  kprobes_text_initialized (false),
-  kprobes_text_start (0), kprobes_text_end (0)
+  op (0), up (0)
 {
 }
 
diff --git a/runtime/ChangeLog b/runtime/ChangeLog
index 2b068b5c..4c714b01 100644
--- a/runtime/ChangeLog
+++ b/runtime/ChangeLog
@@ -1,3 +1,8 @@
+2007-03-20  Frank Ch. Eigler  <fche@elastic.org>
+
+	PR 4224.
+	* sym.c (_stp_module_relocate): Support kernel relocations.
+
 2007-03-19  Frank Ch. Eigler  <fche@elastic.org>
 
 	* autoconf-hrtimer-rel.c: New file.
diff --git a/runtime/sym.c b/runtime/sym.c
index a5bc0195..70efd82d 100644
--- a/runtime/sym.c
+++ b/runtime/sym.c
@@ -34,24 +34,44 @@ static unsigned long _stp_module_relocate (const char *module, const char *secti
 
 	if (last) {
 		if (!strcmp (module, last->name) && !strcmp (section, last_sec->symbol)) {
-			STP_UNLOCK_MODULES;
-			return offset + last_sec->addr;
+                  /* XXX: But is this enough protection?  What if the module `last' is
+                     unloaded sometime between the last relocate call and this one?  Do
+                     the last/last_sec pointers become invalid to traverse like that? */
+                  STP_UNLOCK_MODULES;
+                  return offset + last_sec->addr;
 		}
 	}
 
 	/* need to scan all modules */
-	for (i = 1; i < _stp_num_modules; i++) { /* XXX: why start at i=1? */
-		last = _stp_modules[i];
-		if (strcmp(module, last->name))
-			continue;
-		for (j = 0; j < (int)last->num_sections; j++) {
-			last_sec = &last->sections[j];
-			if (!strcmp (section, last_sec->symbol)) {
-				STP_UNLOCK_MODULES;
-				return offset + last_sec->addr;
-			}
-		}
-	}
+        if (! strcmp (module, "kernel"))
+          {
+            STP_UNLOCK_MODULES;
+
+            /* See also transport/symbols.c (_stp_do_symbols). */
+            if (strcmp (section, "_stext"))
+              return 0;
+            else
+              return offset + _stp_modules[0]->text;
+
+            /* NB: we could also use _stp_kallsyms_lookup_name (section); */
+            /* If _stp_kallsyms_lookup_name also returned the symbol,
+               we could set last & last_sym and take some advantage of
+               caching.  But OTOH the advantage would be tiny in comparison
+               to the hard-coded calculation above. */
+          }
+        else /* relocatable module */
+          for (i = 1; i < _stp_num_modules; i++) { /* skip over [0]=kernel */
+            last = _stp_modules[i];
+            if (strcmp(module, last->name))
+              continue;
+            for (j = 0; j < (int)last->num_sections; j++) {
+              last_sec = &last->sections[j];
+              if (!strcmp (section, last_sec->symbol)) {
+                STP_UNLOCK_MODULES;
+                return offset + last_sec->addr;
+              }
+            }
+          }
 	STP_UNLOCK_MODULES;
 	last = NULL;
 	return 0;
@@ -65,7 +85,7 @@ static unsigned long _stp_kallsyms_lookup_name(const char *name)
 
 	while (num--) {
 		if (strcmp(name, s->symbol) == 0)
-			return s->addr;
+                    	return s->addr;
 		s++;
 	}
 	return 0;
diff --git a/runtime/transport/ChangeLog b/runtime/transport/ChangeLog
index dcee1d79..90661baf 100644
--- a/runtime/transport/ChangeLog
+++ b/runtime/transport/ChangeLog
@@ -1,3 +1,8 @@
+2007-03-20  Frank Ch. Eigler  <fche@elastic.org>
+
+	* symbols.c (_stp_do_symbols): Add cautionary blurb for important
+	setup of _stp_modules[0]->text.
+
 2007-03-20  Martin Hunt  <hunt@redhat.com>
 
 	* symbols.c (_stp_alloc_module): Fix up error
diff --git a/runtime/transport/symbols.c b/runtime/transport/symbols.c
index 9d5163d0..ee5e967e 100644
--- a/runtime/transport/symbols.c
+++ b/runtime/transport/symbols.c
@@ -203,6 +203,7 @@ static int _stp_do_symbols(const char __user *buf, int count)
 		for (i = 0; i < _stp_modules[0]->num_symbols; i++) 
 			s[i].symbol += (long)_stp_modules[0]->symbol_data;
 		_stp_symbol_state = 3;
+                /* NB: this mapping is used by kernel/_stext pseudo-relocations. */
 		_stp_modules[0]->text = _stp_kallsyms_lookup_name("_stext");
 		_stp_modules_by_addr[0] = _stp_modules[0];
 		break;
diff --git a/session.h b/session.h
index 956d4394..7798b32f 100644
--- a/session.h
+++ b/session.h
@@ -93,7 +93,7 @@ struct systemtap_session
   bool merge;
   int buffer_size;
   unsigned perfmon;
-  bool symtab;
+  bool symtab; /* true: emit symbol table at translation time; false: let staprun do it. */
   bool prologue_searching;
 
   // Cache data
@@ -136,10 +136,11 @@ struct systemtap_session
   translator_output* op;
   unparser* up;
 
-  // kprobes_text data
-  bool kprobes_text_initialized;
-  Dwarf_Addr kprobes_text_start;
-  Dwarf_Addr kprobes_text_end;
+  // some symbol addresses
+  // XXX: these belong elsewhere; perhaps the dwflpp instance
+  Dwarf_Addr sym_kprobes_text_start;
+  Dwarf_Addr sym_kprobes_text_end;
+  Dwarf_Addr sym_stext;
 
   std::set<std::string> seen_errors;
   unsigned num_errors () { return seen_errors.size(); }
diff --git a/tapsets.cxx b/tapsets.cxx
index eb66f2bc..78d47dc7 100644
--- a/tapsets.cxx
+++ b/tapsets.cxx
@@ -2278,55 +2278,11 @@ static int query_kernel_module (Dwfl_Module *, void **, const char *,
 static bool
 in_kprobes_function(systemtap_session& sess, Dwarf_Addr addr)
 {
-  if (! sess.kprobes_text_initialized)
-    {
-      // Only attempt kprobes_text_start/kprobes_text_end
-      // initialization once
-      sess.kprobes_text_initialized = true;
-
-      dwflpp *kernel_dw = new dwflpp(sess);
-      assert(kernel_dw);
-      kernel_dw->setup(true);
-
-      Dwfl_Module *m = NULL;
-      kernel_dw->iterate_over_modules(&query_kernel_module, &m);
-      assert(m);
-      kernel_dw->focus_on_module(m);
-
-      // Look through the symbol table for "__kprobes_text_{start,end}"
-      int syments = dwfl_module_getsymtab(kernel_dw->module);
-      assert(syments);
-      for (int i = 1; i < syments; ++i)
-        {
-	  GElf_Sym sym;
-	  const char *name = dwfl_module_getsym(kernel_dw->module, i,
-						&sym, NULL);
-
-	  // Look for a symbol that starts with "__kprobes_text_"
-	  if (name != NULL
-	      && strncmp(name, "__kprobes_text_", 15) == 0)
-	    {
-	      // Match either "__kprobes_text_start" or "__kprobes_text_end"
-	      if (strcmp(name, "__kprobes_text_start") == 0)
-		sess.kprobes_text_start = sym.st_value;
-	      else if (strcmp(name, "__kprobes_text_end") == 0)
-		sess.kprobes_text_end = sym.st_value;
-
-	      // If we've got both values, quit processing symbols.
-	      if (sess.kprobes_text_start != 0 && sess.kprobes_text_end != 0)
-		i = syments;
-	    }
-	}
-
-      if (kernel_dw)
-	delete kernel_dw;
-    }
-
-  if (sess.kprobes_text_start != 0 && sess.kprobes_text_end != 0)
+  if (sess.sym_kprobes_text_start != 0 && sess.sym_kprobes_text_end != 0)
     {
       // If the probe point address is anywhere in the __kprobes
       // address range, we can't use this probe point.
-      if (addr >= sess.kprobes_text_start && addr < sess.kprobes_text_end)
+      if (addr >= sess.sym_kprobes_text_start && addr < sess.sym_kprobes_text_end)
 	return true;
     }
   return false;
@@ -2455,6 +2411,13 @@ dwarf_query::add_probe_point(const string& funcname,
   if (sess.verbose > 1)
     clog << endl;
 
+  if (module == TOK_KERNEL)
+    {
+      // PR 4224: adapt to relocatable kernel by subtracting the _stext address here.
+      reloc_addr = addr - sess.sym_stext;
+      reloc_section = "_stext"; // a deliberate eyesore
+    }
+
   if (! bad)
     {
       probe = new dwarf_derived_probe(funcname, filename, line, 
@@ -2938,9 +2901,9 @@ query_module (Dwfl_Module *mod __attribute__ ((unused)),
 
       if (q->sess.verbose>2)
 	clog << "focused on module '" << q->dw.module_name
-	     << " = [" << hex << q->dw.module_start
-	     << "-" << q->dw.module_end
-	     << ", bias " << q->dw.module_bias << "]" << dec 
+	     << " = [0x" << hex << q->dw.module_start
+	     << "-0x" << q->dw.module_end
+	     << ", bias 0x" << q->dw.module_bias << "]" << dec 
              << " file " << debug_filename
              << " ELF machine " << expect_machine
              << " (code " << elf_machine << ")"
@@ -3495,11 +3458,9 @@ dwarf_derived_probe::dwarf_derived_probe(const string& funcname,
     maxactive_val (q.maxactive_val)
 {
   // Assert relocation invariants
-  if (module == TOK_KERNEL && section != "") 
-    throw semantic_error ("relocation requested against kernel", q.base_loc->tok);
-  if (module != TOK_KERNEL && section == "") 
-    throw semantic_error ("missing relocation base against module", q.base_loc->tok);
-  if (module != TOK_KERNEL && section != "" && dwfl_addr == addr) // addr should be an offset
+  if (section == "") 
+    throw semantic_error ("missing relocation base against", q.base_loc->tok);
+  if (section != "" && dwfl_addr == addr) // addr should be an offset
     throw semantic_error ("inconsistent relocation address", q.base_loc->tok);
 
   // Make a target-variable-expanded copy of the probe body
@@ -3734,9 +3695,9 @@ dwarf_derived_probe_group::emit_module_decls (systemtap_session& s)
         s.op->line() << " .return_p=1,";
       if (p->has_maxactive)
         s.op->line() << " .maxactive_p=1,";
-      if (p->module != TOK_KERNEL) // relocation info
-        s.op->line() << " .module=\"" << p->module << "\", .section=\"" << p->section << "\",";
       s.op->line() << " .address=0x" << hex << p->addr << dec << "UL,";
+      s.op->line() << " .module=\"" << p->module << "\",";
+      s.op->line() << " .section=\"" << p->section << "\",";
       if (p->has_maxactive)
 	s.op->line() << " .maxactive_val=" << p->maxactive_val << "UL,";
       s.op->line() << " .pp=" << lex_cast_qstring (*p->sole_location()) << ",";
@@ -3827,6 +3788,26 @@ dwarf_derived_probe_group::emit_module_exit (systemtap_session& s)
 }
 
 
+
+static Dwarf_Addr
+lookup_symbol_address (Dwfl_Module *m, const char* wanted)
+{
+  int syments = dwfl_module_getsymtab(m);
+  assert(syments);
+  for (int i = 1; i < syments; ++i)
+    {
+      GElf_Sym sym;
+      const char *name = dwfl_module_getsym(m, i, &sym, NULL);
+      if (name != NULL && strcmp(name, wanted) == 0)
+        return sym.st_value;
+    }
+
+  return 0;
+}
+
+
+
+
 void
 dwarf_builder::build(systemtap_session & sess,
 		     probe * base,
@@ -3852,6 +3833,26 @@ dwarf_builder::build(systemtap_session & sess,
 	  kern_dw->setup(true);
 	}
       dw = kern_dw;
+
+
+      Dwfl_Module* km = 0;
+      dw->iterate_over_modules(&query_kernel_module, &km);
+      if (km)
+        {
+          sess.sym_kprobes_text_start = lookup_symbol_address (km, "__kprobes_text_start");
+          sess.sym_kprobes_text_end = lookup_symbol_address (km, "__kprobes_text_end");
+          sess.sym_stext = lookup_symbol_address (km, "_stext");
+
+          if (sess.verbose > 2)
+            {
+              clog << "control symbols:"
+                // abbreviate the names - they're for our debugging only anyway
+                   << " kts: 0x" << hex << sess.sym_kprobes_text_start
+                   << " kte: 0x" << sess.sym_kprobes_text_end
+                   << " stext: 0x" << sess.sym_stext
+                   << dec << endl;
+            }
+        }
     }
   else
     {
diff --git a/translate.cxx b/translate.cxx
index 89a961e3..d3f9467a 100644
--- a/translate.cxx
+++ b/translate.cxx
@@ -3986,8 +3986,9 @@ emit_symbol_data (systemtap_session& s)
  
   if (s.symtab == false)
     {
+      s.op->newline() << "/* filled in by runtime */";
       s.op->newline() << "struct stap_symbol *stap_symbols;";
-      s.op->newline() << "unsigned stap_num_symbols = 0;\n";
+      s.op->newline() << "unsigned stap_num_symbols;\n";
       return 0;
     }