Re: [RFC][PATCH 2/4] kprobe-based symbol resolution for stap-translator

[Prerna Saxena <prerna at linux dot vnet dot ibm dot com>]

Prerna Saxena wrote:
> Hi,
> Here's a set of patches which enable the stap-translator to utilize 
> kprobes for resolving function addresses.
> ( Similar to James Bottomley's patch sent out last july )
> In place of resolving probe points in semantic pass (Pass 2 ) by 
> consulting vmlinux/debuginfo, this approach defers symbol resolution 
> to the generated kprobes module. The kprobes module is passed the name 
> of the function to be probed, which gets resolved against the kernel 
> symbol tables to insert probes.
>
> This construct can be invoked using a new switch "--ksym" .
>
> In its present form, it is capable of probing function entry & returns 
> (non-inlines). It does /*not*/ support :
>    *  a wildcard - based search for module/function names
>    *  probing  select/all functions in a given source file
>    *  probing inline functions.
>    *  statement probes
>
> Known issues :
> 1. Apparently systemtap modules pick up build-id from the debuginfo 
> files. Since debuginfo lookup is completely bypassed here, the 
> generated stap modules fail a consistency check later owing to 
> incorrect build id. For now, patch 4 comments out this check and the 
> stap modules run fine, but I'd appreciate some pointers on how to set 
> it right. :-)
>
> 2. An incorrect indentation parameter passed to the translated output 
> in pass 3 causes stap to abort due to assert failure. Patch 4 corrects 
> this as well.
>
> Patches :
> 1. kallsym_patch_1 , kallsym_patch_2 : introduce changes to session.h 
> / main.cxx for the new switch "--ksym"
> 2. kallsym_patch_3 : changes to tapsets.cxx.
> 3. kallsym_patch_4 : workarounds for known problems.
>
> I'm working on fine-tuning its capabilities.....looking fwd to 
> suggestions... :-)

--
Prerna Saxena

Linux Technology Centre,
IBM Systems and Technology Lab,
Bangalore, India 

Index: git-02-mar/main.cxx
===================================================================
--- git-02-mar.orig/main.cxx
+++ git-02-mar/main.cxx
@@ -131,6 +131,7 @@ usage (systemtap_session& s, int exitcod
     << "   --kelf     make do with symbol table from vmlinux" << endl
     << "   --kmap[=FILE]" << endl
     << "              make do with symbol table from nm listing" << endl
+    << "   --ksym     defer symbol resolution to kprobes " <<endl
 #endif
   // Formerly present --ignore-{vmlinux,dwarf} options are for testsuite use
   // only, and don't belong in the eyesight of a plain user.
@@ -428,6 +429,7 @@ main (int argc, char * const argv [])
 #define LONG_OPT_IGNORE_VMLINUX 3
 #define LONG_OPT_IGNORE_DWARF 4
 #define LONG_OPT_VERBOSE_PASS 5
+#define LONG_OPT_KALLSYMS 6
       // NB: also see find_hash(), usage(), switch stmt below, stap.1 man page
       static struct option long_options[] = {
         { "kelf", 0, &long_opt, LONG_OPT_KELF },
@@ -435,6 +437,7 @@ main (int argc, char * const argv [])
         { "ignore-vmlinux", 0, &long_opt, LONG_OPT_IGNORE_VMLINUX },
         { "ignore-dwarf", 0, &long_opt, LONG_OPT_IGNORE_DWARF },
         { "vp", 1, &long_opt, LONG_OPT_VERBOSE_PASS },
+        { "ksym", 0, &long_opt, LONG_OPT_KALLSYMS },
         { NULL, 0, NULL, 0 }
       };
       int grc = getopt_long (argc, argv, "hVMvtp:I:e:o:R:r:m:kgPc:x:D:bs:uqwl:d:L:F",
@@ -675,6 +678,9 @@ main (int argc, char * const argv [])
 	    case LONG_OPT_IGNORE_DWARF:
 	      s.ignore_dwarf = true;
 	      break;
+	    case LONG_OPT_KALLSYMS:
+	      s.consult_kallsym = true;
+              break;
 	    case LONG_OPT_VERBOSE_PASS:
               {
                 bool ok = true;