Hi, When I think about how to add a test case for v5 version patch, and I want to test if v5 version patch has memory leak. This thread [1] provided a way how to repeat the memory leak, so I used it to test v5 patch. I didn't found memory leak on v5 patch.
But I found other interesting issue. When changed whereClause in [1], the
query reported below error:
"ERROR could not find memoization table entry"
the query:
EXPLAIN analyze
select sum(q.id_table1)
from (
SELECT t2.*
FROM table1 t1
JOIN table2 t2
ON (t2.id_table1 + t2.id_table1) = t1.id) q;
But on v5 patch, it didn't report error.
I guess it is the same reason that data in probeslot was reset in Hash
function.
I debug the above query, and get this:
before
(gdb) p *(DatumGetNumeric(mstate->probeslot->tts_values[0]))
$1 = {vl_len_ = 48, choice = {n_header = 32770, n_long = {n_sign_dscale =
32770, n_weight = 60, n_data = 0x564632ebd708}, n_short = {n_header =
32770, n_data = 0x564632ebd706}}}
after
(gdb) p *(DatumGetNumeric(mstate->probeslot->tts_values[0]))
$2 = {vl_len_ = 264, choice = {n_header = 32639, n_long = {n_sign_dscale =
32639, n_weight = 32639, n_data = 0x564632ebd6a8}, n_short = {n_header =
32639, n_data = 0x564632ebd6a6}}}
So after call ResetExprContext() in Hash function, the data in probeslot is
corrupted. It is not sure what error will happen when executing on
corrupted data.
During debug, I learned that numeric_add doesn't have type check like
rangetype, so aboved query will not report "type with xxx does not exist".
And I realize that the test case added by Andrei Lepikhov in v3 is right.
So in v6 patch I add Andrei Lepikhov's test case. Thanks a lot.
Now I think the v6 version patch seems to be complete now.
[1]
https://www.postgresql.org/message-id/83281eed63c74e4f940317186372abfd%40cft.ru
--
Tender Wang
OpenPie: https://en.openpie.com/
v6-0001-Fix-wrong-used-ResetExprContext-in-ExecMemoize.patch
Description: Binary data
