Re: fmgr_oldstyle in extensions

Hi Paul,

On 2014-09-23 09:55:32 -0700, Paul Ramsey wrote:
> I’m trying to implement an spgist index in the PostGIS extension, which seems like it should work, but is thus far not working for what appear (to me) to be issues in the way spgist expects to pass pointers to user-defined functions.
>
> Right before anything happens, spgist calls a ‘config’ routine in the user code. This is the code from the example implementation (which is internal to pgsql, not a run-time add-on like postgis)
>
>  Datum
>  spg_quad_config(PG_FUNCTION_ARGS)
>  {
>  /* spgConfigIn *cfgin = (spgConfigIn *) PG_GETARG_POINTER(0); */
>  spgConfigOut *cfg = (spgConfigOut *) PG_GETARG_POINTER(1); 
>
>  cfg->prefixType = POINTOID;
>  cfg->labelType = VOIDOID; /* we don't need node labels */
>  cfg->canReturnData = true;
>  cfg->longValuesOK = false;
>  PG_RETURN_VOID();
>  }
>
> It is called from the spgist core in the spgGetCache() function, via 
>
>  FunctionCall2Coll(procinfo,
>    index->rd_indcollation[0],
>    PointerGetDatum(&in),
>    PointerGetDatum(&cache->config));
>
> The part the user function cares about is the pointer to cache->config. 
>
> In the core code, the call stack to the user function goes like this:
>
> postgres`spg_quad_config(fcinfo=0x00007fff5f0faf90) + 18 at spgquadtreeproc.c:29
> postgres`FunctionCall2Coll(flinfo=0x00007ff59a804cc8, collation=0, arg1=140734788252568, arg2=140692835814944) + 150 at fmgr.c:1327
>  postgres`spgGetCache(index=0x000000010254be68) + 220 at spgutils.c:71
>
> So, spgGetCache to FunctionCall2Coll to spg_quad_config.
>
> In my user-defined version of the same thing (I just copied the C code and wrote a CREATE OPERATOR CLASS for it) the call stack is this
>
> postgis-2.2.so`gserialized_spgist_quadtree_2d_config(fcinfo=0x00007fff5f0fb398) + 30 at gserialized_spgist_2d.c:60
> postgres`fmgr_oldstyle(fcinfo=0x00007fff5f0faf90) + 424 at fmgr.c:678
> postgres`FunctionCall2Coll(flinfo=0x00007ff59a039cc8, collation=0, arg1=140734788252568, arg2=140692827643424) + 150 at fmgr.c:1327
> postgres`spgGetCache(index=0x000000010254be68) + 220 at spgutils.c:71
>
> So, spgGetCache to FunctionCall2Coll to fmgr_oldstyle to 
> gserialized_spgist_quadtree_2d_config! 
>
> On the way through fmgr_oldstyle things go very much awry and the gserialized_spgist_quadtree_2d_config doesn’t get called with a pointer to an fcinfo, but with a pointer to the block of memory occupied by the first argument, and from there things naturally stop working.
>
> So, I’m wondering what I could have done that is causing my calls to route through fmgr_oldstyle instead of the usual path?

You forgot to add a PG_FUNCTION_INFO_V1(yourfunc); for the function.

I really, really think we should get rid of v0 functions. I've seen
errors like this far too many, and they really don't buy us much.

Greetings,

Andres Freund

--
Andres Freund http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services