Thread: Segmentation fault when using a set-returning C function from a view in 8.4.0

Hello,

I have created a set-returning C function and a view that selects all
the returned rows. When I use SELECT * FROM theview, the returned rows
look fine. But if I use, e.g., SELECT count(*) FROM theview or SELECT
sum(a) FROM theview, I get a segmentation fault.

LOG:  server process (PID 7099) was terminated by signal 11:
Segmentation fault

Is this a bug? SELECT count(*), sum(a) FROM thefunction() works fine.

I have created a small example that demonstrates the problem (see
below). If the C function only returns few rows, everything works. If
the function returns more rows (e.g., 5,000), I get the segmentation fault.

I have expericenced this on 8.4.0:
chr=# select version();                                                  version
-------------------------------------------------------------------------------------------------------------PostgreSQL
8.4.0on x86_64-unknown-linux-gnu, compiled by GCC gcc (GCC)
 
4.1.2 20070115 (SUSE Linux), 64-bit

chr=# show work_mem ;work_mem
----------256MB


EXAMPLE:

create table demotbl(a int, b int, c int);

create function demosrf() returns setof demotbl language 'c' as
'/tmp/demo.so';

create view demoview as
select * from demotbl
union all
select * from demosrf();


The C code is shown below.

#include "postgres.h"
#include "tupdesc.h"
#include "funcapi.h"
#include "fmgr.h"

#ifdef PG_MODULE_MAGIC
PG_MODULE_MAGIC;
#endif

PG_FUNCTION_INFO_V1(demosrf);
Datum
demosrf(PG_FUNCTION_ARGS)
{   ReturnSetInfo *rsinfo = (ReturnSetInfo *)fcinfo->resultinfo;   TupleDesc tupdesc;   Tuplestorestate *tupstore;
AttInMetadata*attinmeta;   int numberOfAttributes;   int i, j;   Datum *values;   bool *isnull;   extern int work_mem;
 
   if(rsinfo == NULL || !IsA(rsinfo, ReturnSetInfo))ereport(ERROR, (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
errmsg("Contextdoes not accept a set")));   if(!(rsinfo->allowedModes & SFRM_Materialize))ereport(ERROR,
(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),       errmsg("Materialize not allowed")));
 
   tupdesc = rsinfo->expectedDesc;   tupstore = tuplestore_begin_heap(false, false, work_mem);
   attinmeta = TupleDescGetAttInMetadata(tupdesc);   numberOfAttributes = attinmeta->tupdesc->natts;
   values = (Datum *)palloc(numberOfAttributes * sizeof(Datum));   isnull = (bool *)palloc(numberOfAttributes *
sizeof(bool));
   // Create rows   for(i = 0; i < 10000; i++) {for(j = 0; j < numberOfAttributes; j++) {    isnull[j] = false;
values[j]= Int32GetDatum(i);}tuplestore_putvalues(tupstore, tupdesc, values, isnull);   }
 
   rsinfo->returnMode = SFRM_Materialize;   rsinfo->setResult = tupstore;   return (Datum)0;
}

Best regards,
Christian Thomsen

Christian Thomsen <chr@cs.aau.dk> writes:
> I have created a set-returning C function and a view that selects all
> the returned rows. When I use SELECT * FROM theview, the returned rows
> look fine. But if I use, e.g., SELECT count(*) FROM theview or SELECT
> sum(a) FROM theview, I get a segmentation fault.

> LOG:  server process (PID 7099) was terminated by signal 11:
> Segmentation fault

> Is this a bug?

Yeah, in your code: you've violated multiple rules about set-returning
functions.  Offhand:* using expectedDesc without checking for NULL* creating the tuplestore in the wrong context*
failingto set setDesc
 

Also, blindly using expectedDesc instead of constructing your own tuple
descriptor is rather badly missing the point.  This code will *only*
work for a tupdesc consisting of some number of integer columns; if
you're passed something else it will fail in more or less horrible ways.
What you're supposed to do is construct a tupdesc that accurately
describes what you're returning, and use that for the tupstore and pass
it back as setDesc.  Then the core code can verify it matches
expectedDesc.  Passing in expectedDesc is only useful for functions that
can work with a variety of actual output tupledescriptors.
        regards, tom lane