I found autovacuum can be canceled by blocked backends even if the vacuum
is for preventing XID wraparound in 8.3.0 and HEAD. Autovacuum sets
PROC_VACUUM_FOR_WRAPAROUND flag just before vacuum, but the flag will be
cleared at the beginning of vacuum; PROC_VACUUM_FOR_WRAPAROUND is not set
during the vacuum.
The sequence is below:
vacuum()
-> CommitTransactionCommand()
-> ProcArrayEndTransaction()
-> proc->vacuumFlags &= ~PROC_VACUUM_STATE_MASK;
-> vacuum_rel()
PROC_VACUUM_STATE_MASK is defined as (0x0E), that is including
PROC_VACUUM_FOR_WRAPAROUND (0x08). The wraparound flag is cleared
before vacuum tasks.
I tried to make a patch to exclude PROC_VACUUM_FOR_WRAPAROUND
from PROC_VACUUM_STATE_MASK and make autovacuum workers to clear
PROC_VACUUM_FOR_WRAPAROUND by themselves. Is it a reasonable solution?
Index: src/backend/postmaster/autovacuum.c
===================================================================
--- src/backend/postmaster/autovacuum.c (HEAD)
+++ src/backend/postmaster/autovacuum.c (working copy)
@@ -2163,6 +2163,12 @@
PG_END_TRY();
/* the PGPROC flags are reset at the next end of transaction */
+ if (tab->at_wraparound)
+ {
+ LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+ MyProc->vacuumFlags &= ~PROC_VACUUM_FOR_WRAPAROUND;
+ LWLockRelease(ProcArrayLock);
+ }
/* be tidy */
pfree(tab);
Index: src/include/storage/proc.h
===================================================================
--- src/include/storage/proc.h (HEAD)
+++ src/include/storage/proc.h (working copy)
@@ -45,7 +45,7 @@
#define PROC_VACUUM_FOR_WRAPAROUND 0x08 /* set by
autovac only */
/* flags reset at EOXact */
-#define PROC_VACUUM_STATE_MASK (0x0E)
+#define PROC_VACUUM_STATE_MASK (PROC_IN_VACUUM |
PROC_IN_ANALYZE)
/*
* Each backend has a PGPROC struct in shared memory. There is also a list of
Regards,
---
ITAGAKI Takahiro
NTT Open Source Software Center
--
Sent via pgsql-hackers mailing list ([email protected])
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-hackers
