Thread: ERROR: Memory exhausted in AllocSetAlloc(909324558)

I saw someone else also saw this error.  I am seeing it in 7.1.2 and I
think I may have tracked it down.  It's after a different operation but
perhaps it is related.

I have written a trigger.  It dies with that error when I make the
following call.

SPI_modifytuple (rel,rettuple,1,&targ_att,&newval,NULL)

I created targ_att as follows.

targ_att = SPI_fnumber (tupdesc, args[0]

This returns 23 in my case which is the correct field that I want to
modify.  The existing value for that field is NULL.  I looked at the
length of that fields with rel->rd_att->attrs[23]->attlen and it is -1.
I assume that that is OK for a NULL value.

I eventually followed this call to heaptuple.c.  In heap_formtuple()
there is a call to ComputeDataSize() with the existing tuple, the value
I am changing to and the nulls.  In that function the length (-1) is
added to the total length.  This causes the problem I am seeing.

Am I misunderstanding the call?  I thought that the nulls parameter was
to map out which values you were changing to were NULL.  It appears
to be the value of the existing value.  Why would I care about that value?
Is it up to me to find all the NULLS in an existing tuple before calling SPI
functions?

I called this function that way in an earlier version with no problem.

-- 
D'Arcy J.M. Cain <darcy@{druid|vex}.net>   |  Democracy is three wolves
http://www.druid.net/darcy/                |  and a sheep voting on
+1 416 425 1212     (DoD#0082)    (eNTP)   |  what's for dinner.

darcy@druid.net (D'Arcy J.M. Cain) writes:
> Am I misunderstanding the call?  I thought that the nulls parameter was
> to map out which values you were changing to were NULL.

heap_formtuple is for constructing a tuple from scratch.  It sounds like
you should be using heap_modifytuple instead.
        regards, tom lane

Thus spake Tom Lane
> darcy@druid.net (D'Arcy J.M. Cain) writes:
> > Am I misunderstanding the call?  I thought that the nulls parameter was
> > to map out which values you were changing to were NULL.
> 
> heap_formtuple is for constructing a tuple from scratch.  It sounds like
> you should be using heap_modifytuple instead.

But I am using SPI_modifytuple().  The rest came from tracing that
function.

-- 
D'Arcy J.M. Cain <darcy@{druid|vex}.net>   |  Democracy is three wolves
http://www.druid.net/darcy/                |  and a sheep voting on
+1 416 425 1212     (DoD#0082)    (eNTP)   |  what's for dinner.

darcy@druid.net (D'Arcy J.M. Cain) writes:
>> heap_formtuple is for constructing a tuple from scratch.  It sounds like
>> you should be using heap_modifytuple instead.

> But I am using SPI_modifytuple().  The rest came from tracing that
> function.

Hm.  Looks like the author of SPI_modifytuple didn't realize he could
use heap_modifytuple :-(.  But he is correctly extracting the old data.
I don't see anything wrong there, assuming that you are passing a
relation descriptor that matches the original tuple.
        regards, tom lane

Thus spake Tom Lane
> darcy@druid.net (D'Arcy J.M. Cain) writes:
> >> heap_formtuple is for constructing a tuple from scratch.  It sounds like
> >> you should be using heap_modifytuple instead.
> 
> > But I am using SPI_modifytuple().  The rest came from tracing that
> > function.
> 
> Hm.  Looks like the author of SPI_modifytuple didn't realize he could
> use heap_modifytuple :-(.  But he is correctly extracting the old data.
> I don't see anything wrong there, assuming that you are passing a
> relation descriptor that matches the original tuple.

Well, let's see.
   SPI_modifytuple (rel,rettuple,1,&targ_att,&newval,NULL)

Here are (I think) the relevant lines that get me the data for that call.

Datum
mk_cardnum(PG_FUNCTION_ARGS)
{   TriggerData *trigdata = (TriggerData *) fcinfo->context;   int         nargs;          /* # of arguments */   Datum
     newval;         /* new value of column */   char        **args;         /* arguments */   char        *relname;
  /* triggered relation name */   Relation    rel;            /* triggered relation */   HeapTuple   rettuple = NULL;
int        targ_att;   bool        isnull;   char        cardnum[48];
 
...   /* assume full error checking - leaving out for clarity */   rel = trigdata->tg_relation;   relname =
SPI_getrelname(rel);  targ_att = SPI_fnumber (tupdesc, args[0]);   newval = CStringGetDatum(cardnum);
 

-- 
D'Arcy J.M. Cain <darcy@{druid|vex}.net>   |  Democracy is three wolves
http://www.druid.net/darcy/                |  and a sheep voting on
+1 416 425 1212     (DoD#0082)    (eNTP)   |  what's for dinner.

darcy@druid.net (D'Arcy J.M. Cain) writes:
>     char        cardnum[48];
> ...
>     newval = CStringGetDatum(cardnum);

And you are trying to assign this datum to what kind of column?
(Hint: there isn't any kind for which a plain C string is valid data.)
        regards, tom lane

Thus spake Tom Lane
> darcy@druid.net (D'Arcy J.M. Cain) writes:
> >     char        cardnum[48];
> > ...
> >     newval = CStringGetDatum(cardnum);
> 
> And you are trying to assign this datum to what kind of column?
> (Hint: there isn't any kind for which a plain C string is valid data.)

Right.  That was just my latest try.  I did notice that CStringGetDatum was
just defined as PointerGetDatum.  I tried to find something specific for
text but nothing seemed to fit the bill.  I saw DatumGetTextP but that
detoasts a toasted object.  There were the various PG_RETURNxxx macros
but this is a user created strting, not an argument.

Do I have to manually create a varlena?  I'll try that next.

-- 
D'Arcy J.M. Cain <darcy@{druid|vex}.net>   |  Democracy is three wolves
http://www.druid.net/darcy/                |  and a sheep voting on
+1 416 425 1212     (DoD#0082)    (eNTP)   |  what's for dinner.

darcy@druid.net (D'Arcy J.M. Cain) writes:
> Do I have to manually create a varlena?  I'll try that next.

The best way to get from a C string to a valid datum is to invoke the
datatype's input conversion routine.  If you know you want text,
you could do

#define _textin(str) DirectFunctionCall1(textin, CStringGetDatum(str))

to produce a datum from a C string.
        regards, tom lane

Thus spake Tom Lane
> darcy@druid.net (D'Arcy J.M. Cain) writes:
> > Do I have to manually create a varlena?  I'll try that next.
> 
> The best way to get from a C string to a valid datum is to invoke the
> datatype's input conversion routine.  If you know you want text,
> you could do
> 
> #define _textin(str) DirectFunctionCall1(textin, CStringGetDatum(str))
> 
> to produce a datum from a C string.

Ah, that was it.  I had done a similar thing in another trigger with a
date type.  I should have thought of that.

Hmm.  Does this suggest more macros in fmgr.h or postgres.h?

-- 
D'Arcy J.M. Cain <darcy@{druid|vex}.net>   |  Democracy is three wolves
http://www.druid.net/darcy/                |  and a sheep voting on
+1 416 425 1212     (DoD#0082)    (eNTP)   |  what's for dinner.