Обсуждение: cannot freeze committed xmax
Hi hackers,
The following error was encountered by our customers:
Them have very huge catalog (size of pg_class relation is more than
30Gb) blowned by temporary relations.
When them try to vacuum it, the following error is reported:
vacuum full analyze pg_catalog.pg_class;
ERROR: cannot freeze committed xmax 596099954
The following records are present in pg_class:
(standard input)-10436009-<Data> ------
(standard input)-10436010- Item 1 -- Length: 229 Offset: 7936
(0x1f00) Flags: NORMAL
(standard input):10436011: XMIN: 596098791 XMAX: 596099954 CID|XVAC:
1 OID: 930322390
(standard input)-10436012- Block Id: 108700 linp Index: 17 Attributes:
33 Size: 32
(standard input)-10436013- infomask: 0x290b
(HASNULL|HASVARWIDTH|HASOID|XMIN_COMMITTED|XMAX_INVALID|UPDATED)
(standard input)-10436014- t_bits: [0]: 0xff [1]: 0xff [2]: 0xff [3]: 0x7f
(standard input)-10436015- [4]: 0x00
(standard input)-10436016-
(standard input)-10436017- Item 2 -- Length: 184 Offset: 7752
(0x1e48) Flags: NORMAL
(standard input):10436018: XMIN: 596098791 XMAX: 596099954 CID|XVAC:
2 OID: 930322393
(standard input)-10436019- Block Id: 108700 linp Index: 18 Attributes:
33 Size: 32
(standard input)-10436020- infomask: 0x2909
(HASNULL|HASOID|XMIN_COMMITTED|XMAX_INVALID|UPDATED)
(standard input)-10436021- t_bits: [0]: 0xff [1]: 0xff [2]: 0xff [3]: 0x3f
(standard input)-10436022- [4]: 0x00
(standard input)-10436023-
(standard input)-10436024- Item 3 -- Length: 184 Offset: 7568
(0x1d90) Flags: NORMAL
(standard input):10436025: XMIN: 596098791 XMAX: 596099954 CID|XVAC:
3 OID: 930322395
(standard input)-10436026- Block Id: 108700 linp Index: 19 Attributes:
33 Size: 32
(standard input)-10436027- infomask: 0x2909
(HASNULL|HASOID|XMIN_COMMITTED|XMAX_INVALID|UPDATED)
(standard input)-10436028- t_bits: [0]: 0xff [1]: 0xff [2]: 0xff [3]: 0x3f
(standard input)-10436029- [4]: 0x00
This error is reported in heap_prepare_freeze_tuple:
/*
* Process xmax. To thoroughly examine the current Xmax value we
need to
* resolve a MultiXactId to its member Xids, in case some of them are
* below the given cutoff for Xids. In that case, those values
might need
* freezing, too. Also, if a multi needs freezing, we cannot
simply take
* it out --- if there's a live updater Xid, it needs to be kept.
*
* Make sure to keep heap_tuple_needs_freeze in sync with this.
*/
xid = HeapTupleGetRawXmax(htup);
if (tuple->t_infomask & HEAP_XMAX_IS_MULTI)
{
...
}
else if (TransactionIdIsNormal(xid))
{
...
if (TransactionIdPrecedes(xid, cutoff_xid))
{
/*
* If we freeze xmax, make absolutely sure that it's not an XID
* that is important. (Note, a lock-only xmax can be removed
* independent of committedness, since a committed lock
holder has
* released the lock).
*/
if (!HEAP_XMAX_IS_LOCKED_ONLY(tuple->t_infomask) &&
TransactionIdDidCommit(xid))
ereport(ERROR,
(errcode(ERRCODE_DATA_CORRUPTED),
errmsg_internal("cannot freeze committed xmax
" XID_FMT,
xid)));
freeze_xmax = true;
}
else
freeze_xmax = false;
...
}
else if ((tuple->t_infomask & HEAP_XMAX_INVALID) ||
!TransactionIdIsValid(HeapTupleGetRawXmax(htup)))
{
freeze_xmax = false;
xmax_already_frozen = true;
}
So, as you can see, in all this records HEAP_XMAX_INVALID is set, but
xmax is normal transaction id.
This is why we produce error before check for HEAP_XMAX_INVALID in the
subsequent if.
I do not know value of cutoff_xid, because do not have access to the
debugger at customer site.
I will be please or any help how to localize the source of the problem.
Looks like there is no assumption that xmax should be set to
InvalidTransactionId when HEAP_XMAX_INVALID bit is set.
And I didn't find any check preventing cutoff_xid to be greater than
XID of some transaction which was aborted long time ago.
So is there some logical error that xmax is compared with cutoff_xid
before HEAP_XMAX_INVALID bit is checked?
Otherwise, where this constraint most likely be violated?
It is PG 11.7 version of Postgres.
Thanks is advance,
--
Konstantin Knizhnik
Postgres Professional: http://www.postgrespro.com
The Russian Postgres Company
> On Oct 28, 2020, at 6:44 AM, Konstantin Knizhnik <k.knizhnik@postgrespro.ru> wrote: > > Looks like there is no assumption that xmax should be set to InvalidTransactionId when HEAP_XMAX_INVALID bit is set. > And I didn't find any check preventing cutoff_xid to be greater than XID of some transaction which was aborted long timeago. Nothing in your example suggests that cutoff_xid is wrong, so I'd assume that part is probably working ok. Your data shows that HEAP_XMAX_INVALID and HEAP_UPDATED flags are both set. That should only happen if the updating transactionaborted. But the query of clog is saying that it committed. Something is wrong with that. How did the hint bitsget set to HEAP_XMAX_INVALID if the transaction did commit. Either some process is setting that hint bit when it shouldn't,or your clog is corrupted and returning a bogus answer about the xmax having been committed. Either way, you'vegot corruption. Your question "preventing cutoff_xid to be greater than XID of some transaction which was aborted long time ago" seems tobe ignoring that TransactionIdDidCommit(xid) is returning true, suggesting the transaction did not abort. — Mark Dilger EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company
On 28.10.2020 18:25, Mark Dilger wrote: > >> On Oct 28, 2020, at 6:44 AM, Konstantin Knizhnik <k.knizhnik@postgrespro.ru> wrote: >> >> Looks like there is no assumption that xmax should be set to InvalidTransactionId when HEAP_XMAX_INVALID bit is set. >> And I didn't find any check preventing cutoff_xid to be greater than XID of some transaction which was aborted long timeago. > Nothing in your example suggests that cutoff_xid is wrong, so I'd assume that part is probably working ok. > > Your data shows that HEAP_XMAX_INVALID and HEAP_UPDATED flags are both set. That should only happen if the updating transactionaborted. But the query of clog is saying that it committed. Something is wrong with that. How did the hint bitsget set to HEAP_XMAX_INVALID if the transaction did commit. Either some process is setting that hint bit when it shouldn't,or your clog is corrupted and returning a bogus answer about the xmax having been committed. Either way, you'vegot corruption. > > Your question "preventing cutoff_xid to be greater than XID of some transaction which was aborted long time ago" seemsto be ignoring that TransactionIdDidCommit(xid) is returning true, suggesting the transaction did not abort. > > — > Mark Dilger > EnterpriseDB: http://www.enterprisedb.com > The Enterprise PostgreSQL Company Yes, I forgot to say that transaction is treated as committed (txid_status() returns "committed"). Also database was previously upgraded from 11.5 to 11.7 Certainly the hypothesis of CLOG corruption explains everything. I wonder if there can be some other scenario (upgrade, multixacts, previous freeze attempt...) which can cause such combination of flags? I have inspected all cases where HEAP_XMAX_INVALID is set, but have not found any one which can explain it. -- Konstantin Knizhnik Postgres Professional: http://www.postgrespro.com The Russian Postgres Company
> On Oct 28, 2020, at 8:56 AM, Konstantin Knizhnik <k.knizhnik@postgrespro.ru> wrote: > > > > On 28.10.2020 18:25, Mark Dilger wrote: >> >>> On Oct 28, 2020, at 6:44 AM, Konstantin Knizhnik <k.knizhnik@postgrespro.ru> wrote: >>> >>> Looks like there is no assumption that xmax should be set to InvalidTransactionId when HEAP_XMAX_INVALID bit is set. >>> And I didn't find any check preventing cutoff_xid to be greater than XID of some transaction which was aborted longtime ago. >> Nothing in your example suggests that cutoff_xid is wrong, so I'd assume that part is probably working ok. >> >> Your data shows that HEAP_XMAX_INVALID and HEAP_UPDATED flags are both set. That should only happen if the updating transactionaborted. But the query of clog is saying that it committed. Something is wrong with that. How did the hint bitsget set to HEAP_XMAX_INVALID if the transaction did commit. Either some process is setting that hint bit when it shouldn't,or your clog is corrupted and returning a bogus answer about the xmax having been committed. Either way, you'vegot corruption. >> >> Your question "preventing cutoff_xid to be greater than XID of some transaction which was aborted long time ago" seemsto be ignoring that TransactionIdDidCommit(xid) is returning true, suggesting the transaction did not abort. >> >> — >> Mark Dilger >> EnterpriseDB: http://www.enterprisedb.com >> The Enterprise PostgreSQL Company > Yes, I forgot to say that transaction is treated as committed (txid_status() returns "committed"). > Also database was previously upgraded from 11.5 to 11.7 > Certainly the hypothesis of CLOG corruption explains everything. > I wonder if there can be some other scenario (upgrade, multixacts, previous freeze attempt...) which can cause such combinationof flags? > I have inspected all cases where HEAP_XMAX_INVALID is set, but have not found any one which can explain it. The other possibillity is that this tuple is erroneously marked as HEAP_UPDATED. heap_update() sets that, which makes sense. rewrite_heap_tuple() copies the old tuple's bits to the new tuple and then does some work to resolve update chains. I guess you could look at whether that logic might leave things in an invalid state. I don't have any theory aboutthat. Looking at the git logs, it seems 699bf7d05c68734f800052829427c20674eb2c6b introduced the check that is ereporting, and didso along with commit 9c2f0a6c3cc8bb85b78191579760dbe9fb7814ec, which cleaned up some corruption bugs. I wonder if you'rejust unlucky enough to have had one of these corruptions, and now you're bumping into the ereport which is intendedto prevent the corruption from spreading further? — Mark Dilger EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company