This is a very specific issue. To encounter the problem, the
instance needs to have done more then 1 in-place migration,
likely including fairly specific server versions.
The issue was reproduce using the following steps:
1) Initialize an instance using 7.31
2) Create the following simple table using the listed SQL
commands
create table t1 (c1 int, c2 int) lock mode row;
create unique index t1_idx1 on t1(c1);
insert into t1 values(1,1);
3) Do an in-place migration from 7.31 to 10 (this testing
specifically used 10.00.FC9)
4) After upgrade completes run the following SQL command
alter fragment on index t1_idx init in rootdbs (this should
detach the index from being in table as the server created all
indexes in 7.x but it also marks this index with a special old
style flag indicating that this detached index is fully
detached, meaning in the leaf pages the index key has the key
value and then fragid/rowid/delete flag entries)
5) Do another in-place migration up to 11.70.FC7W2
6) After upgrade completes run the following SQL
alter table t1 add constraint primary key(c1) constraint pk_t1;
At this point, the alter table command to also mark the old
unique index as a primary key, when it re-writes the partition
page, it loses the special old style flag, so the server no long
knows the index leaf pages should have fragid/rowid/delete flag
entries, so when you run oncheck -cI, it will report errors
trying to check the index.
If the system was in this state and then upgraded to 12.10, and
oncheck -cI was run, the server would then start throwing may
assertions that would looks like the following.
03/24/17 12:45:26 Assert Failed: Code assertion at rsbtcs.c:308
03/24/17 12:45:26 IBM Informix Dynamic Server Version
12.10.FC8W1
03/24/17 12:45:26 Who: Session(7040, informix@server, 19955,
113f64c68)
Thread(16085, oncheckm, 113f2b0b0, 9)
File: rsdebug.c Line: 3882
03/24/17 12:45:26 Results:
rowoff+BT_RIDL_FLAG_SIZE+sizeof(int4) bc_slotptr+cs->bc_slotlen
With a stack like this:
afhandler()
affail_interface()
rsam_assert()
ridl_ld_elem()
ckfnode()
ckfnode()
ckfnode()
ischeck_idx()
fm_indexcheck()
fmoncheck()
doidxcheck()
chkarg4()
sqmain()
startup()
At this point the index could be dropped and re-created which
would get around the issue. |
