Skip to content
This repository
Fetching contributors…

Octocat-spinner-32-eaf2f5

Cannot retrieve contributors at this time

file 504 lines (377 sloc) 18.17 kb

MagLev implementation of Ruby C extensions API.

This document explains the state of C extensions on Maglev. This is meant to become an exhaustive listing of the restrictions in the C API layer, their reasons, and how to work around them.

We also provide code examples for some workarounds at the end of this document.

If more restrictions are found, or you have code examples, please feel free to add them.

Objectives

Provide a high performance implementation of the Ruby C extension API without compromising garbage collector performance, and without introducing expensive object identifer translations.

The Ruby C VALUE is the same as the OopType from the Gemstone C API( GCI) The Maglev VM already knows how to associate an object identifier with an object. This identifier is used both for object identity of persistent objects on disk , and as the identifier exported through GCI function calls. We use this same object identifier for VALUE's in the Ruby C API.

Restrictions

Special objects

  • Qnil, Qfalse, Qtrue

    Problem: Maglev does not support C comparisons expecting Qfalse to have C integer value zero. Qnil, Qtrue, Qfalse are all non-zero values in Maglev.

    Workaround: You must use the RTEST macro before doing a C comparision of a VALUE to zero. Based on looking at a few C extensions, it would appear that all uses of C expressions of the form x ? a : b need to be examined for missing RTEST(x). If you want to test for Qnil, use the NIL_P macro.

  • NULL

    Problem: Casting the C value NULL to the type VALUE is not supported. In Maglev (VALUE)0 is not a legal object identifier. Attempts to use it will raise a StandardError, 'object does not exist'.

    Workaround: In most cases Qnil should be used instead of (VALUE)NULL. Corresponding changes to use RTEST instead of a C compare to zero or other code changes may be required.

Strings

  • RSTRING_PTR

    Problem: Has a result type of const char* in Maglev. Maglev will replicate the body of a String object to C memory on request. That body is then read-only from C.

    Workaround: To change bytes in a String object, you must use one of

    void rb_store_string_l(VALUE str, long offset, const char* value, size_t len);       rb_store_string(VALUE str, long offset, const char* value);    
    VALUE rb_str_buf_append(VALUE str, VALUE val);
    VALUE rb_str_buf_cat(VALUE str, const char* bytes, long len);
    VALUE rb_str_buf_cat2(VALUE str, const char* cstring);
    VALUE rb_str_append(VALUE str, VALUE arg);
    

    After a String has been modified by one of the above store functions, or by any execution of Ruby code, a subsequent RSTRING_PTR, rb_str_ptr, or rb_str_ptr_readonly (all of which call rb_str2cstr) will return a pointer to updated C memory. If the String has grown since the previous RSTRING_PTR call, the returned memory may be a different C memory pointer, otherwise contents of the previously allocated C memory will be updated by the rb_str2cstr() implementation to reflect the current contents of the String .

    The lifetime of the C memory referenced by the result of RSTRING_PTR is the same as the garbage collection lifetime of the corresponding String object.

    C code must copy string data to it's own C memory if it expects indefinite lifetime of the C memory, or to control lifetime by free(). Based on the "Lifetimes of temporary objects" below, C code needs to copy the result of RSTRING_PTR to other C memory if the C memory is needed beyond the lifetime of the current Ruby to C extension invocation, unless it is known that the String will be kept alive by longer lived Ruby objects, and will not be changed.

    The result of RSTRING_PTR must not be passed to free()

  • rb_str_set_len

    Problem: This will change the logical size of the string, and if it is used to shrink the string and then grow it all bytes beyond the size produced by the shrink will be zero after the string is grown again, because Maglev doesn't keep an underlying char* buffer that is still larger than the logical size

    Workaround: Do not rely on this MRI behavior

  • rb_str_buf_new(x)

    Problem: The argument to rb_str_buf_new is ignored, which makes this equivalent to rb_str_new("", 0). The physical size of a String object is not accessible from Ruby, Smalltalk, or Ruby C extensions, so the logical and physical sizes are always the same from the point of view of C extensions.

    Workaround: Since the result of RSTRING_PTR is read-only, one of the rb_str_buf_append or rb_str_buf_cat* functions must be used to append to the result of rb_str_buf_new .

Arrays

  • RARRAY_PTR

    Problem: Unsupported. In MRI this macro returns a pointer to the body of an Array. Arrays are Smalltalk objects on Maglev, so we cannot simply expose them as a C array. RARRAY_PTR is not supported in Maglev because the number of uses in a typical C extension is small, and if it were supported it could only be via const VALUE* RARRAY_PTR(VALUE ary); semantics - updates would have to go through a function call to rb_ary_store. It is easier to change a C extension to use rb_ary_entry instead of RARRAY_PTR than it would to change C code to substitute a function calls for a RSTRING_PTR calls, because RARRAY_PTR is usually used to access individual elements of an Array. Thus we have chosen not to implement object memory to C memory replication of Array bodies.

    Workaround: To access elements of an Array in Maglev, use rb_ary_entry or rb_ary_store which are also available in JRuby and Rubinius.

Hash

  • rb_iterate_each_pair

    Problem: Only supports iteration on Array arguments for now

    Workaround: None

  • RHASH, RHASH_TBL

    Problem: Like on JRuby and Rubinius, these macros aren't supported.

  • st.h

    Problem: The Maglev VM does not implement the MRI st.h API.

    WorkaroundL A C extension which needs to use the C hash implementation of st.h must ship and build it's own copy of st.c. (observed in nokogiri)

Structs

  • DATA_PTR

    Problem: Direct access is not supported.

    Workaround: To change the C data wrapped by an object, use

    void rb_rdata_store(VALUE obj, void *p);
    

    To fetch the C data use

    void* rb_rdata_fetch(VALUE obj);
    

    rb_rdata_store will run the free function previously installed by Data_Wrap_Struct if appropriate, and will decide how to allocate the meta information for the object to hold the C data pointer if no C pointer was previously associated with the object.

  • Data_Wrap_Struct

    Problem: Calling Data_Wrap_Struct with a non-NULL mark function is not supported yet.

    Workaround: If the referenced object(s) to be kept alive exist at the point of the Data_Wrap_Struct call, use rb_ivar_set to create an object reference to them.

Test-functions

  • rb_obj_is_kind_of, rb_obj_is_instance_of, rb_ivar_defined, ...

    Problem: MRI, JRuby and Rubinius all define

    VALUE rb_obj_is_kind_of(VALUE, VALUE);
    VALUE rb_obj_is_instance_of(VALUE, VALUE);
    VALUE rb_ivar_defined(VALUE, ID);
    

    and they are commented as /** Returns true-ish ... */. The common usage pattern, however, appears to be like

    if (! (rb_obj_is_kind_of(rb_node, cNokogiriXmlNode)))  
    

    which fails to apply RTEST to the result of rb_obj_is_* functions before doing the C comparision to zero (see also restrictions on special values Qfalse and Qtrue, above).

    Workaround: Add RTEST macro calls as needed around the results of these functions. Use MAGLEV_LINT (see below) to help find possibly bad uses.

Globals and GC

  • rb_global_variable, rb_gc_register_address, rb_gc_unregister_address

    Problem: To implement these requires supporting an arbitrary number of garbage collection roots, which reduces garbage collector performance. A fast garbage collector needs to have a small root set, so that a scavenge only needs to examine a minimum number of young objects. In the Maglev VM, the root set includes the stacks of recently active Ruby Threads, and only a small number of well known objects needed by the VM. All other objects are reachable from those well known objects, and thus only need to be examined by a scavenge if they have been recently created or dirtied since a previous scavenge.

    Workarounds:

    • To define global variables, use rb_gv_set instead.
    • To protect a temporary C array of VALUE from garbage collection, create an Array object and store values into it with rb_ary_store
    • Maglev defines the additional function rb_funcall2_ to take a Ruby array argument

      VALUE rb_funcall2(VALUE obj, ID meth, int cnt, VALUE* args);
      

      VALUE rb_funcall2_(VALUE obj, ID meth, int cnt, VALUE args_array);

  • rb_gv_get

    Problem: Retrieving $~ is not supported from C

    Workaround: None.

  • Object lifetime

    Problem: C-references aren't kept alive/valid after returning from C extension code.

    Each call from Ruby into a C extension initializes an IdentitySet known to the garbage collector that is kept alive for the duration of that entry into a C extension. This IdentitySet is the 'export set' for that entry into the C extension. The IdentitySet is dereferenced upon return from C back to Ruby.

    The following are automatically added to the export set, to ensure objects stay alive at least for the duration of the Ruby to C call .

    • Newly created objects, such as results of rb_ary_new* , rb_str_new*, rb_str_buf_new* .
    • Objects returned from Ruby execution, such as results from rb_funcall*
    • Objects for which C memory pointers were returned by RSTRING_PTR or rb_str2cstr.

    Workaround: For a newly created object to stay alive after returning from C to Ruby it must have been stored as the value of a Ruby global variable, Ruby constant, or stored into an instance variable of some other object reachable from top level Ruby state, or be reachable from the VALUE returned from C back to Ruby.

Exception Handling

  • rb_rescue

    Problem: In Ruby code invoked from the C rescue function (second procedure arg to rb_rescue or rb_rescue2, $! does not reflect the exception being rescued by the application's C code. $! only reflects exceptions being rescued by Ruby rescue blocks.

    Workaround: None

  • rb_eException

    Problem: The maglev class Exception does not define the instance variables 'mesg' or 'backtrace', so they cannot be accessed directly from C (observed in Psych)

    Workaround: Initialization of instances of Exception should use rb_obj_call_init, instead of rb_iv_set. Accessing the backtrace of an Exception should use an appropriate variant of rb_funcall to call a public method in Exception.

Numerics

  • rb_big2ulong

    Problem: Raises a RangeError when passed a negative argument.

    Workaround: Argument checking.

Regexps

  • rb_reg_nth_match

    Problem: Regexp matches invoked from C do not set $1, $2 in the Ruby code which called the C code.

    Workaround: Retrieve and store or pass the matches explicitely

Lint

Many of the aforementioned problems aren't immediately obvious, and might not lead compiler errors or SEGVs straight away, but rather create subtle and hard-to-track bugs in C extension execution. To help find possible causes for C extension bugs on Maglev, the MAGLEV_LINT macro can be used.

Maglev defines some functions for which we have seen problematic usage patterns (like rb_obj_is_*, rb_*_defined) only if MAGLEV_LINT is undefined. Thus, to check your extension for those functions and the probable bad usage of them, you can use use -DMAGLEV_LINT=1 on a C compile to trigger compiler errors. Then use this document to apply the appropriate workarounds as needed.

Bugs and status of rubyspecs (as of 17 Feb 2011)

For the rubyspecs in optional/capi/ , the following specs are skipped with 'not_compliant_on :maglev do' because they are not implemented at all (and some of them never might be)

  • rb_ary_new2_assign
  • rb_ary_to_s
  • rb_block_proc
  • rb_call_super
  • rb_define_hooked_variable
  • rb_define_readonly_variable
  • rb_define_variable
  • rb_exec_recursive
  • rb_hash
  • rb_hash_delete_if
  • rb_hash_foreach
  • rb_mem_clear
  • rb_num_zerodiv
  • rb_reg_match
  • rb_set_kcode
  • rb_str_buf_new
  • rb_str2cstr_replace
  • rb_struct_aref
  • rb_struct_aset
  • rb_struct_new
  • rb_str_hash
  • rb_throw
  • rb_big2ll
  • rb_ll2inum
  • RARRAY
  • RARRAY_PTR_assign
  • RBIGNUM_SIGN
  • RFLOAT
  • RSTRING
  • uses of RSTRING_PTR which attempt to change the string
  • uses of STR2CSTR which attempt to change the string

The following functions have known bugs that we will fix

  • rb_cvar_defined has some failures
  • rb_const_get_from , rb_const_get_at fail to call const_missing
  • The protections of methods installed by rb_define_private_method,
  • rb_define_protected_method is not correct.
  • rb_is_type_data(Time.now) returns false ; it should be true

Code samples

Following are some code examples of changes required to use C extensions on Maglev.

examples of (VALUE)NULL problems
--- a/ext/nokogiri/xml_syntax_error.c
+++ b/ext/nokogiri/xml_syntax_error.c
@@ -3,19 +3,19 @@
 void Nokogiri_error_array_pusher(void * ctx, xmlErrorPtr error)
 {
   VALUE list = (VALUE)ctx;
-  rb_ary_push(list,  Nokogiri_wrap_xml_syntax_error((VALUE)NULL, error));
+  rb_ary_push(list,  Nokogiri_wrap_xml_syntax_error( Qnil , error));
 }

 void Nokogiri_error_raise(void * ctx, xmlErrorPtr error)
 {
-  rb_exc_raise(Nokogiri_wrap_xml_syntax_error((VALUE)NULL, error));
+  rb_exc_raise(Nokogiri_wrap_xml_syntax_error( Qnil , error));
 }

 VALUE Nokogiri_wrap_xml_syntax_error(VALUE klass, xmlErrorPtr error)
 {
   VALUE msg, e;

-  if(!klass) klass = cNokogiriXmlSyntaxError;
+  if( klass == Qnil ) klass = cNokogiriXmlSyntaxError;

----- ext/nokogiri/xml_node.c

@@ -1113,7 +1115,7 @@ static VALUE new(int argc, VALUE *argv, VALUE klass)
   NOKOGIRI_ROOT_NODE(node);

   rb_node = Nokogiri_wrap_xml_node(
-      klass == cNokogiriXmlNode ? (VALUE)NULL : klass,
+      klass == cNokogiriXmlNode ? Qnil : klass,
examples of missing RTEST
-----  ext/psych/emitter.c

@@ -160,7 +163,7 @@ static VALUE start_document(VALUE self, VALUE version, VALUE tags, VALUE imp
            (RARRAY_LEN(version) > 0) ? &version_directive : NULL,
            head,
            tail,
-           imp ? 1 : 0
+           RTEST(imp) ? 1 : 0


@@ -182,7 +185,7 @@ static VALUE end_document(VALUE self, VALUE imp)
-    yaml_document_end_event_initialize(&event, imp ? 1 : 0);
+    yaml_document_end_event_initialize(&event, RTEST(imp) ? 1 : 0);
tatic VALUE scalar(
        Check_Type(tag, T_STRING);
        tag = rb_str_export_to_enc(tag, encoding);
     }
+#else
+    if(!NIL_P(anchor)) {
+        Check_Type(anchor, T_STRING);
+    }
+    if(!NIL_P(tag)) {
+        Check_Type(tag, T_STRING);
+    }
 #endif
psych contains code dependent on the name of an MRI instVar
--- a/ext/psych/to_ruby.c
+++ b/ext/psych/to_ruby.c
@@ -10,7 +10,8 @@ static VALUE build_exception(VALUE self, VALUE klass, VALUE mesg)
 {
     VALUE e = rb_obj_alloc(klass);

-    rb_iv_set(e, "mesg", mesg);
+    // rb_iv_set(e, "mesg", mesg);  // MRI implementation dependent
+    rb_obj_call_init(e, 1, &mesg);
avoidance of rb_gc_register_address , from nokogiri
+++ b/ext/nokogiri/xml_xpath_context.c
@@ -53,7 +53,7 @@ static void ruby_funcall(xmlXPathParserContextPtr ctx, int nargs)
 {
   VALUE xpath_handler = Qnil;
   VALUE result;
-  VALUE *argv;
+  // VALUE *argv;
   VALUE doc;
   VALUE node_set = Qnil;
   xmlNodeSetPtr xml_node_set = NULL;
@@ -68,10 +68,11 @@ static void ruby_funcall(xmlXPathParserContextPtr ctx, int nargs)

   xpath_handler = (VALUE)(ctx->context->userData);

-  argv = (VALUE *)calloc((size_t)nargs, sizeof(VALUE));
-  for (i = 0 ; i < nargs ; ++i) {
-    rb_gc_register_address(&argv[i]);
-  }
+  //argv = (VALUE *)calloc((size_t)nargs, sizeof(VALUE));
+  //for (i = 0 ; i < nargs ; ++i) {
+  //  rb_gc_register_address(&argv[i]);
+  //}
+  VALUE argv = rb_ary_new2(nargs);

   doc = DOC_RUBY_OBJECT(ctx->context->doc);

@@ -80,46 +81,51 @@ static void ruby_funcall(xmlXPathParserContextPtr ctx, int nargs)
     obj = valuePop(ctx);
     switch(obj->type) {
       case XPATH_STRING:
-        argv[i] = NOKOGIRI_STR_NEW2(obj->stringval);
+        rb_ary_store( argv, i, NOKOGIRI_STR_NEW2(obj->stringval));
         break;
......
example for missing Data_Wrap_Struct mark-function support from nokogiri
static VALUE sym_iv_doc = Qnil;

void init_xml_node() {
  ...
  sym_iv_doc = rb_intern("@doc");
  ...
}

VALUE Nokogiri_wrap_xml_node(VALUE klass, xmlNodePtr node) {
  ...
  if (DOC_RUBY_OBJECT_TEST(node->doc)) {  // maglev workaround , no gc mark yet
    VALUE ref = DOC_RUBY_OBJECT(node->doc);
    if (ref != 0 && ref != Qnil && sym_iv_doc != Qnil) {
      rb_node = Data_Wrap_Struct(klass, NULL, debug_node_dealloc, node) ;
      rb_ivar_set(rb_node, sym_iv_doc, ref);
    }
  }
  ...
}
Something went wrong with that request. Please try again.