Don't call elog() while holding spinlock.
authorFujii Masao <fujii@postgresql.org>
Tue, 2 Jun 2020 10:18:13 +0000 (19:18 +0900)
committerFujii Masao <fujii@postgresql.org>
Tue, 2 Jun 2020 10:21:04 +0000 (19:21 +0900)
Previously UpdateSpillStats() called elog(DEBUG2) while holding
the spinlock even though the local variables that the elog() accesses
don't need to be protected by the lock. Since spinlocks are intended
for very short-term locks, they should not be used when calling
elog(DEBUG2). So this commit moves that elog() out of spinlock period.

Author: Kyotaro Horiguchi
Reviewed-by: Amit Kapila and Fujii Masao
Discussion: https://postgr.es/m/20200602.161518.1399689010416646074.horikyota.ntt@gmail.com

src/backend/replication/walsender.c

index 86847cbb54f54bf5ce0bcc5f0da86b053bb759b6..2364cbfc61b568d85fcb8c551b03324c47ed34f1 100644 (file)
@@ -3685,17 +3685,15 @@ UpdateSpillStats(LogicalDecodingContext *ctx)
 {
        ReorderBuffer *rb = ctx->reorder;
 
-       SpinLockAcquire(&MyWalSnd->mutex);
-
-       MyWalSnd->spillTxns = rb->spillTxns;
-       MyWalSnd->spillCount = rb->spillCount;
-       MyWalSnd->spillBytes = rb->spillBytes;
-
        elog(DEBUG2, "UpdateSpillStats: updating stats %p %lld %lld %lld",
                 rb,
                 (long long) rb->spillTxns,
                 (long long) rb->spillCount,
                 (long long) rb->spillBytes);
 
+       SpinLockAcquire(&MyWalSnd->mutex);
+       MyWalSnd->spillTxns = rb->spillTxns;
+       MyWalSnd->spillCount = rb->spillCount;
+       MyWalSnd->spillBytes = rb->spillBytes;
        SpinLockRelease(&MyWalSnd->mutex);
 }