Обсуждение: Odd problem with read_pg_options ...

Поиск
Список
Период
Сортировка

Odd problem with read_pg_options ...

От
"Marc G. Fournier"
Дата:
I reported earlier a SegFault when doing an initdb, and have narrowed it
down somewhat...still probing, but figured I'd see if I'm overlooking
something obvious...

From the command line, I'm running:

echo "vacuum" | postgres -o /dev/null -F -Q -D/home/centre/marc/pgsql/data

This is the stage that the first SegFault happens in initdb.

I added some debugging, and it turns out that the 'DataDir' variable isn't
being initialized at this point:

void
read_pg_options(SIGNAL_ARGS)
{
  int     fd;
  int     n;
  int     verbose;
  char    buffer[BUF_SIZE];
  char    c;
  char     *s,
         *p;

  printf("before sprintf()\n");
  printf("%s\n", DataDir);
  sprintf(buffer, "%s/%s", DataDir, "pg_options");
  printf("after sprintf()\n");
  if ((fd = open(buffer, O_RDONLY)) < 0)
    return;

=====================

Still looking, but uncovered a slight bug:

diff -cr postgres.c.orig postgres.c
*** postgres.c.orig     Tue Oct 13 16:47:00 1998
--- postgres.c  Tue Oct 13 16:47:33 1998
***************
*** 1052,1057 ****
--- 1052,1058 ----

                        case 'D':                       /* PGDATA directory */
                                DataDir = optarg;
+                               break;

                        case 'd':                       /* debug level */
                                flagQ = false;


Further into it...if I do a setenv PGDATA, it gets around the 'bug'...back
later...

Marc G. Fournier                               scrappy@hub.org
Systems Administrator @ hub.org
scrappy@{postgresql|isc}.org                       ICQ#7615664


Re: [HACKERS] Odd problem with read_pg_options ...

От
Massimo Dal Zotto
Дата:
>
>
> I reported earlier a SegFault when doing an initdb, and have narrowed it
> down somewhat...still probing, but figured I'd see if I'm overlooking
> something obvious...
>
> echo "vacuum" | postgres -o /dev/null -F -Q -D/home/centre/marc/pgsql/data
>
> This is the stage that the first SegFault happens in initdb.
>
> I added some debugging, and it turns out that the 'DataDir' variable isn't
> being initialized at this point:
>
> void
> read_pg_options(SIGNAL_ARGS)
> {
>   int     fd;
>   int     n;
>   int     verbose;
>   char    buffer[BUF_SIZE];
>   char    c;
>   char     *s,
>          *p;
>
>   printf("before sprintf()\n");
>   printf("%s\n", DataDir);
>   sprintf(buffer, "%s/%s", DataDir, "pg_options");
>   printf("after sprintf()\n");
>   if ((fd = open(buffer, O_RDONLY)) < 0)
>     return;
>
> =====================
>
> Still looking, but uncovered a slight bug:
>
> diff -cr postgres.c.orig postgres.c
> *** postgres.c.orig     Tue Oct 13 16:47:00 1998
> --- postgres.c  Tue Oct 13 16:47:33 1998
> ***************
> *** 1052,1057 ****
> --- 1052,1058 ----
>
>                         case 'D':                       /* PGDATA directory */
>                                 DataDir = optarg;
> +                               break;
>
>                         case 'd':                       /* debug level */
>                                 flagQ = false;
>
>
> Further into it...if I do a setenv PGDATA, it gets around the 'bug'...back
> later...
>
> Marc G. Fournier                               scrappy@hub.org
> Systems Administrator @ hub.org
> scrappy@{postgresql|isc}.org                       ICQ#7615664
>

The problem is that read_pg_options needs DataDir to read its file but
DataDir is set after read_pg_options if postgres is called interactively.
If postgres is forked by postgres DataDir is read from the PGDATA enviromnent
variable set by the postmaster and this explains while the bug disappears.
I have written this patch but I don't like it. Any better idea?

*** src/backend/utils/init/globals.c.orig    Thu Oct 15 00:13:03 1998
--- src/backend/utils/init/globals.c    Thu Oct 15 00:13:07 1998
***************
*** 46,52 ****
  struct Port *MyProcPort;
  long        MyCancelKey;

! char       *DataDir;

   /*
    * The PGDATA directory user says to use, or defaults to via environment
--- 46,52 ----
  struct Port *MyProcPort;
  long        MyCancelKey;

! char       *DataDir = NULL;

   /*
    * The PGDATA directory user says to use, or defaults to via environment
*** src/backend/utils/misc/trace.c.orig    Thu Sep  3 09:00:39 1998
--- src/backend/utils/misc/trace.c    Thu Oct 15 00:18:49 1998
***************
*** 343,348 ****
--- 343,353 ----
      char       *s,
                 *p;

+     if (!DataDir) {
+         fprintf(stderr, "read_pg_options: DataDir not defined\n");
+         return;
+     }
+
      sprintf(buffer, "%s/%s", DataDir, "pg_options");
      if ((fd = open(buffer, O_RDONLY)) < 0)
          return;
*** src/backend/tcop/postgres.c.orig    Tue Sep  1 09:01:27 1998
--- src/backend/tcop/postgres.c    Thu Oct 15 00:23:24 1998
***************
*** 1049,1055 ****
--- 1049,1061 ----
                  break;

              case 'D':            /* PGDATA directory */
+                     if (!DataDir) {
+                     DataDir = optarg;
+                     /* must be done after DataDir is defined */
+                     read_pg_options(0);
+                 }
                  DataDir = optarg;
+                 break;

              case 'd':            /* debug level */
                  flagQ = false;

--
Massimo Dal Zotto

+----------------------------------------------------------------------+
|  Massimo Dal Zotto                email:  dz@cs.unitn.it             |
|  Via Marconi, 141                 phone:  ++39-461-534251            |
|  38057 Pergine Valsugana (TN)     www:  http://www.cs.unitn.it/~dz/  |
|  Italy                            pgp:  finger dz@tango.cs.unitn.it  |
+----------------------------------------------------------------------+

Re: [HACKERS] Odd problem with read_pg_options ...

От
Bruce Momjian
Дата:
> The problem is that read_pg_options needs DataDir to read its file but
> DataDir is set after read_pg_options if postgres is called interactively.
> If postgres is forked by postgres DataDir is read from the PGDATA enviromnent
> variable set by the postmaster and this explains while the bug disappears.
> I have written this patch but I don't like it. Any better idea?

I have applied it.  You can improve it later, if you wish.

>
> *** src/backend/utils/init/globals.c.orig    Thu Oct 15 00:13:03 1998
> --- src/backend/utils/init/globals.c    Thu Oct 15 00:13:07 1998
> ***************
> *** 46,52 ****
>   struct Port *MyProcPort;
>   long        MyCancelKey;
>
> ! char       *DataDir;
>
>    /*
>     * The PGDATA directory user says to use, or defaults to via environment
> --- 46,52 ----
>   struct Port *MyProcPort;
>   long        MyCancelKey;
>
> ! char       *DataDir = NULL;
>
>    /*
>     * The PGDATA directory user says to use, or defaults to via environment
> *** src/backend/utils/misc/trace.c.orig    Thu Sep  3 09:00:39 1998
> --- src/backend/utils/misc/trace.c    Thu Oct 15 00:18:49 1998
> ***************
> *** 343,348 ****
> --- 343,353 ----
>       char       *s,
>                  *p;
>
> +     if (!DataDir) {
> +         fprintf(stderr, "read_pg_options: DataDir not defined\n");
> +         return;
> +     }
> +
>       sprintf(buffer, "%s/%s", DataDir, "pg_options");
>       if ((fd = open(buffer, O_RDONLY)) < 0)
>           return;
> *** src/backend/tcop/postgres.c.orig    Tue Sep  1 09:01:27 1998
> --- src/backend/tcop/postgres.c    Thu Oct 15 00:23:24 1998
> ***************
> *** 1049,1055 ****
> --- 1049,1061 ----
>                   break;
>
>               case 'D':            /* PGDATA directory */
> +                     if (!DataDir) {
> +                     DataDir = optarg;
> +                     /* must be done after DataDir is defined */
> +                     read_pg_options(0);
> +                 }
>                   DataDir = optarg;
> +                 break;
>
>               case 'd':            /* debug level */
>                   flagQ = false;
>
> --
> Massimo Dal Zotto
>
> +----------------------------------------------------------------------+
> |  Massimo Dal Zotto                email:  dz@cs.unitn.it             |
> |  Via Marconi, 141                 phone:  ++39-461-534251            |
> |  38057 Pergine Valsugana (TN)     www:  http://www.cs.unitn.it/~dz/  |
> |  Italy                            pgp:  finger dz@tango.cs.unitn.it  |
> +----------------------------------------------------------------------+
>
>


--
  Bruce Momjian                        |  http://www.op.net/~candle
  maillist@candle.pha.pa.us            |  (610) 853-3000
  +  If your life is a hard drive,     |  830 Blythe Avenue
  +  Christ can be your backup.        |  Drexel Hill, Pennsylvania 19026