This patch adds Python dtrace support. Note it is necessary to modify test_sys.py to add an integer to the frameobject structure size since this patch adds "int f_calllineno" to the structure, so this test does not fail. --- Python-2.7.6/Makefile.pre.in.~1~ 2013-11-09 23:36:41.000000000 -0800 +++ Python-2.7.6/Makefile.pre.in 2014-05-14 12:54:43.824219677 -0700 @@ -218,6 +218,7 @@ # Used of signalmodule.o is not available SIGNAL_OBJS= @SIGNAL_OBJS@ +DTRACE_OBJS=Python/dtrace.o Python/phelper.o ########################################################################## # Grammar @@ -341,6 +342,7 @@ Python/formatter_unicode.o \ Python/formatter_string.o \ Python/$(DYNLOADFILE) \ + $(DTRACE_OBJS) \ $(LIBOBJS) \ $(MACHDEP_OBJS) \ $(THREADOBJ) @@ -664,6 +666,18 @@ Python/formatter_string.o: $(srcdir)/Python/formatter_string.c \ $(STRINGLIB_HEADERS) +Python/phelper.o: $(srcdir)/Python/phelper.d + dtrace -o $@ -DPHELPER $(DFLAGS) $(CPPFLAGS) -C -G -s $(srcdir)/Python/phelper.d + +Python/python.h: $(srcdir)/Python/python.d + dtrace -o $@ $(DFLAGS) -C -h -s $(srcdir)/Python/python.d + +Python/ceval.o: Python/ceval.c Python/python.h + $(CC) -c -I. -IPython $(BASECFLAGS) $(EXTRA_CFLAGS) $(CPPFLAGS) $(CFLAGSFORSHARED) $(CFLAGS) -DPy_BUILD_CORE -o $@ $< + +Python/dtrace.o: $(srcdir)/Python/python.d Python/ceval.o + dtrace -o $@ $(DFLAGS) -C -G -s $(srcdir)/Python/python.d Python/ceval.o + ############################################################################ # Header files --- Python-2.7.6/Include/frameobject.h.~1~ 2013-11-09 23:36:39.000000000 -0800 +++ Python-2.7.6/Include/frameobject.h 2014-05-14 13:03:19.938777249 -0700 @@ -44,6 +44,7 @@ PyCode_Addr2Line to calculate the line from the current bytecode index. */ int f_lineno; /* Current line number */ + int f_calllineno; /* line number of call site */ int f_iblock; /* index in f_blockstack */ PyTryBlock f_blockstack[CO_MAXBLOCKS]; /* for try and loop blocks */ PyObject *f_localsplus[1]; /* locals+stack, dynamically sized */ --- Python-2.7.6/Objects/frameobject.c.~1~ 2013-11-09 23:36:41.000000000 -0800 +++ Python-2.7.6/Objects/frameobject.c 2014-05-14 12:56:06.970076859 -0700 @@ -738,6 +738,7 @@ f->f_tstate = tstate; f->f_lasti = -1; + f->f_calllineno = code->co_firstlineno; f->f_lineno = code->co_firstlineno; f->f_iblock = 0; --- Python-2.7.6/Python/ceval.c.~1~ 2013-11-09 23:36:41.000000000 -0800 +++ Python-2.7.6/Python/ceval.c 2014-05-14 13:04:01.334853206 -0700 @@ -19,6 +19,11 @@ #include +#define HAVE_DTRACE +#ifdef HAVE_DTRACE +#include "python.h" +#endif + #ifndef WITH_TSC #define READ_TIMESTAMP(var) @@ -672,6 +677,55 @@ NULL); } +#ifdef HAVE_DTRACE +static void +dtrace_entry(PyFrameObject *f) +{ + const char *filename; + const char *fname; + int lineno; + + filename = PyString_AsString(f->f_code->co_filename); + fname = PyString_AsString(f->f_code->co_name); + lineno = PyCode_Addr2Line(f->f_code, f->f_lasti); + + PYTHON_FUNCTION_ENTRY((char *)filename, (char *)fname, lineno); + + /* + * Currently a USDT tail-call will not receive the correct arguments. + * Disable the tail call here. + */ +#if defined(__sparc) + __asm__("nop"); +#endif +} + +static void +dtrace_return(PyFrameObject *f) +{ + const char *filename; + const char *fname; + int lineno; + + filename = PyString_AsString(f->f_code->co_filename); + fname = PyString_AsString(f->f_code->co_name); + lineno = PyCode_Addr2Line(f->f_code, f->f_lasti); + PYTHON_FUNCTION_RETURN((char *)filename, (char *)fname, lineno); + + /* + * Currently a USDT tail-call will not receive the correct arguments. + * Disable the tail call here. + */ +#if defined(__sparc) + __asm__("nop"); +#endif +} +#else +#define PYTHON_FUNCTION_ENTRY_ENABLED 0 +#define PYTHON_FUNCTION_RETURN_ENABLED 0 +#define dtrace_entry() +#define dtrace_return() +#endif /* Interpreter main loop */ @@ -683,9 +737,84 @@ return PyEval_EvalFrameEx(f, 0); } +/* + * These shenanigans look like utter madness, but what we're actually doing is + * making sure that the ustack helper will see the PyFrameObject pointer on the + * stack. We have two tricky cases: + * + * amd64 + * + * We use up the six registers for passing arguments, meaning the call can't + * use a register for passing 'f', and has to push it onto the stack in a known + * location. + * + * And how does "throwflag" figure in to this? -PN + * + * SPARC + * + * Here the problem is that (on 32-bit) the compiler is re-using %i0 before + * some calls inside PyEval_EvalFrameReal(), which means that when it's saved, + * it's just some junk value rather than the real first argument. So, instead, + * we trace our proxy PyEval_EvalFrame(), where we 'know' the compiler won't + * decide to re-use %i0. We also need to defeat optimization of our proxy. + */ + +#if defined(HAVE_DTRACE) + +#if defined(__amd64) +PyObject *PyEval_EvalFrameExReal(long, long, long, long, long, long, + PyFrameObject *, int throwflag); + + + +PyObject * +PyEval_EvalFrameEx(PyFrameObject *f, int throwflag) +{ + volatile PyObject *f2; + f2 = PyEval_EvalFrameExReal(0, 0, 0, 0, 0, 0, f, throwflag); + return (PyObject *)f2; +} + +PyObject * +PyEval_EvalFrameExReal(long a1, long a2, long a3, long a4, long a5, long a6, + PyFrameObject *f, int throwflag) +{ + +#elif defined(__sparc) + +PyObject *PyEval_EvalFrameExReal(PyFrameObject *f, int throwflag); + +volatile int dummy; + +PyObject * +PyEval_EvalFrameEx(PyFrameObject *f, int throwflag) +{ + volatile PyObject *f2; + f2 = PyEval_EvalFrameExReal(f, throwflag); + dummy = f->ob_refcnt; + return (PyObject *)f2; +} + +PyObject * +PyEval_EvalFrameExReal(PyFrameObject *f, int throwflag) +{ + +#else /* __amd64 || __sparc */ + PyObject * PyEval_EvalFrameEx(PyFrameObject *f, int throwflag) { + +#endif /* __amd64 || __sparc */ + +#else /* don't HAVE_DTRACE */ + +PyObject * +PyEval_EvalFrameexEx(PyFrameObject *f, int throwflag)) +{ + +#endif /* HAVE_DTRACE */ + #ifdef DYNAMIC_EXECUTION_PROFILE #undef USE_COMPUTED_GOTOS #endif @@ -910,6 +1039,11 @@ } } +#ifdef HAVE_DTRACE + if (PYTHON_FUNCTION_ENTRY_ENABLED()) + dtrace_entry(f); +#endif + co = f->f_code; names = co->co_names; consts = co->co_consts; @@ -2660,6 +2794,9 @@ PyObject **sp; PCALL(PCALL_ALL); sp = stack_pointer; +#ifdef HAVE_DTRACE + f->f_calllineno = PyCode_Addr2Line(f->f_code, f->f_lasti); +#endif #ifdef WITH_TSC x = call_function(&sp, oparg, &intr0, &intr1); #else @@ -2701,6 +2838,9 @@ } else Py_INCREF(func); sp = stack_pointer; +#ifdef HAVE_DTRACE + f->f_calllineno = PyCode_Addr2Line(f->f_code, f->f_lasti); +#endif READ_TIMESTAMP(intr0); x = ext_do_call(func, &sp, flags, na, nk); READ_TIMESTAMP(intr1); @@ -3001,6 +3141,10 @@ /* pop frame */ exit_eval_frame: +#ifdef HAVE_DTRACE + if (PYTHON_FUNCTION_RETURN_ENABLED()) + dtrace_return(f); +#endif Py_LeaveRecursiveCall(); tstate->frame = f->f_back; --- /dev/null +++ Python-2.6.4/Python/phelper.d @@ -0,0 +1,139 @@ + +/* + * Python ustack helper. This relies on the first argument (PyFrame *) being + * on the stack; see Python/ceval.c for the contortions we go through to ensure + * this is the case. + * + * On x86, the PyFrame * is two slots up from the frame pointer; on SPARC, it's + * eight. + */ + +/* + * Yes, this is as gross as it looks. DTrace cannot handle static functions, + * and our stat_impl.h has them in ILP32. + */ +#define _SYS_STAT_H + +#include +#include + +#include "pyport.h" +#include "object.h" +#include "pystate.h" +#include "pyarena.h" +#include "pythonrun.h" +#include "compile.h" +#include "frameobject.h" +#include "stringobject.h" + +#if defined(__i386) +#define startframe PyEval_EvalFrameEx +#define endframe PyEval_EvalCodeEx +#elif defined(__amd64) +#define PyEval_EvalFrameEx PyEval_EvalFrameExReal +#define startframe PyEval_EvalFrameExReal +#define endframe PyEval_EvalCodeEx +#elif defined(__sparc) +#define PyEval_EvalFrameEx PyEval_EvalFrameExReal +#define startframe PyEval_EvalFrameEx +#define endframe PyEval_EvalFrameExReal +#endif + +#ifdef __sparcv9 +#define STACK_BIAS (2048-1) +#else +#define STACK_BIAS 0 +#endif + +/* + * Not defining PHELPER lets us test this code as a normal D script. + */ +#ifdef PHELPER + +#define at_evalframe(addr) \ + ((uintptr_t)addr >= ((uintptr_t)&``startframe) && \ + (uintptr_t)addr < ((uintptr_t)&``endframe)) +#define probe dtrace:helper:ustack: +#define print_result(r) (r) + +#if defined(__i386) || defined(__amd64) +#define frame_ptr_addr ((uintptr_t)arg1 + sizeof(uintptr_t) * 2) +#elif defined(__sparc) +#define frame_ptr_addr ((uintptr_t)arg1 + STACK_BIAS + sizeof(uintptr_t) * 8) +#else +#error unknown architecture +#endif + +#else /* PHELPER */ + +#define at_evalframe(addr) (1) +#define probe pid$target::PyEval_EvalFrame:entry +#define print_result(r) (trace(r)) + +#if defined(__i386) || defined(__amd64) +#define frame_ptr_addr ((uintptr_t)uregs[R_SP] + sizeof(uintptr_t)) +#elif defined(__sparc) +/* + * Not implemented: we could just use R_I0, but what's the point? + */ +#else +#error unknown architecture +#endif + +#endif /* PHELPER */ + +extern uintptr_t PyEval_EvalFrameEx; +extern uintptr_t PyEval_EvalCodeEx; + +#define copyin_obj(addr, obj) ((obj *)copyin((uintptr_t)addr, sizeof(obj))) +#define pystr_addr(addr) ((char *)addr + offsetof(PyStringObject, ob_sval)) +#define copyin_str(dest, addr, obj) \ + (copyinto((uintptr_t)pystr_addr(addr), obj->ob_size, (dest))) +#define add_str(addr, obj) \ + copyin_str(this->result + this->pos, addr, obj); \ + this->pos += obj->ob_size; \ + this->result[this->pos] = '\0'; +#define add_digit(nr, div) ((nr / div) ? \ + (this->result[this->pos++] = '0' + ((nr / div) % 10)) : \ + (this->result[this->pos] = '\0')) +#define add_char(c) (this->result[this->pos++] = c) + +probe /at_evalframe(arg0)/ +{ + this->framep = *(uintptr_t *)copyin(frame_ptr_addr, sizeof(uintptr_t)); + this->frameo = copyin_obj(this->framep, PyFrameObject); + this->codep = this->frameo->f_code; + this->lineno = this->frameo->f_calllineno; + this->codeo = copyin_obj(this->codep, PyCodeObject); + this->filenamep = this->codeo->co_filename; + this->fnamep = this->codeo->co_name; + this->filenameo = copyin_obj(this->filenamep, PyStringObject); + this->fnameo = copyin_obj(this->fnamep, PyStringObject); + + this->len = 1 + this->filenameo->ob_size + 1 + 5 + 2 + + this->fnameo->ob_size + 1 + 1; + + this->result = (char *)alloca(this->len); + this->pos = 0; + + add_char('@'); + add_str(this->filenamep, this->filenameo); + add_char(':'); + add_digit(this->lineno, 10000); + add_digit(this->lineno, 1000); + add_digit(this->lineno, 100); + add_digit(this->lineno, 10); + add_digit(this->lineno, 1); + add_char(' '); + add_char('('); + add_str(this->fnamep, this->fnameo); + add_char(')'); + this->result[this->pos] = '\0'; + + print_result(stringof(this->result)); +} + +probe /!at_evalframe(arg0)/ +{ + NULL; +} diff --git Python-2.6.4/Python/python.d Python-2.6.4/Python/python.d new file mode 100644 --- /dev/null +++ Python-2.6.4/Python/python.d @@ -0,0 +1,10 @@ +provider python { + probe function__entry(const char *, const char *, int); + probe function__return(const char *, const char *, int); +}; + +#pragma D attributes Evolving/Evolving/Common provider python provider +#pragma D attributes Private/Private/Common provider python module +#pragma D attributes Private/Private/Common provider python function +#pragma D attributes Evolving/Evolving/Common provider python name +#pragma D attributes Evolving/Evolving/Common provider python args --- Python-2.7.6/Lib/test/test_sys.py.~1~ 2013-11-09 23:36:40.000000000 -0800 +++ Python-2.7.6/Lib/test/test_sys.py 2014-05-14 13:07:05.332748121 -0700 @@ -612,7 +612,7 @@ nfrees = len(x.f_code.co_freevars) extras = x.f_code.co_stacksize + x.f_code.co_nlocals +\ ncells + nfrees - 1 - check(x, vsize('12P3i' + CO_MAXBLOCKS*'3i' + 'P' + extras*'P')) + check(x, vsize('12P4i' + CO_MAXBLOCKS*'3i' + 'P' + extras*'P')) # function def func(): pass check(func, size('9P'))