Michael Paquier <michael@paquier.xyz> writes:
> On Fri, Sep 15, 2023 at 08:00:00PM +0300, Alexander Lakhin wrote:
>> + /*
>> + * When we have no detail from the error data, an exception message,
>> + * if not empty, goes in the detail.
>> + */
>> + if (detail == NULL && xmsg != NULL)
>> detail = xmsg;
> Yeah, I think that you're right here to enforce the use of a DETAIL
> message if it exists instead of the xmsg.
I dunno about that. We're trying to shove three pounds of stuff into
a two-pound sack, which isn't going to work perfectly in any case,
but I'm not sure this is the right priority order for the available
strings.
In the original example with an infinite-recursion error, what we
end up with at the point of the Assert is
(gdb) i locals
save_errno = 2
xmsg = 0x666a8d8 "spiexceptions.StatementTooComplex: error fetching next item from iterator"
tbmsg = 0x666ace8 "Traceback (most recent call last):"
tb_depth = 1
emsg = {data = 0x666b0f8 "error fetching next item from iterator", len = 38,
maxlen = 1024, cursor = 0}
exc = 0x18fd448
val = 0x7f1fefe39e88
tb = 0x7f1fef858608
primary = 0x666b0f8 "error fetching next item from iterator"
sqlerrcode = 16777477
detail = 0x7f1fefe21920 "spiexceptions.StatementTooComplex: stack depth limit exceeded"
hint = 0x7f1ffd5c6f78 "Increase the configuration parameter \"max_stack_depth\" (currently 2048kB), after ensuring the
platform'sstack depth limit is adequate."
query = 0x0
position = 0
schema_name = 0x0
table_name = 0x0
column_name = 0x0
datatype_name = 0x0
Clearly, "detail" offers the key insight here and we don't want to
lose it; but xmsg seems rather duplicative of the fmt string
(which is what is reflected into emsg and primary here).
On the other hand, in Alexander's newer test case we find
(gdb) i locals
save_errno = 2
xmsg = 0x191ef50 "spiexceptions.WrongObjectType: cannot open relation \"tst\""
tbmsg = 0x191f360 "Traceback (most recent call last):"
tb_depth = 1
emsg = {data = 0x191f770 "error fetching next item from iterator", len = 38,
maxlen = 1024, cursor = 0}
exc = 0x18f02f8
val = 0x7f1fefda7e88
tb = 0x7f1fefcfd3c8
primary = 0x191f770 "error fetching next item from iterator"
sqlerrcode = 151027844
detail = 0x7f1fefe1fa20 "This operation is not supported for composite types."
hint = 0x0
query = 0x7f1fefcfc3c0 "SELECT * FROM tst"
position = 15
schema_name = 0x0
table_name = 0x0
column_name = 0x0
datatype_name = 0x0
Here, xmsg seems strictly more useful than fmt, but the patch as
proposed would discard xmsg.
I wonder if we shouldn't simply say the rule is that xmsg, if set,
overrides fmt. This'd likely have some impact on existing test
cases, but perhaps the net result would be no worse?
regards, tom lane