RE: [Patch] Optimize dropping of relation buffers using dlist

Hi,

I revised the patch again. Attached is V19.
The previous patch's algorithm missed entering the optimization loop.
So I corrected that and removed the extra function I added
in the previous versions.

The revised patch goes something like this:
	for (forks of rel)
	{
		if (smgrcachednblocks() == InvalidBlockNumber) 
			break; //go to full scan
		if (nBlocksToInvalidate < buf_full_scan_threshold)
			for (blocks of the fork)
		else
			break; //go to full scan
	}
	<execute full scan>

Recovery performance measurement results below.
But it seems there are overhead even with large shared buffers.

| s_b   | master | patched | %reg  | 
|-------|--------|---------|-------| 
| 128MB | 36.052 | 39.451  | 8.62% | 
| 1GB   | 21.731 | 21.73   | 0.00% | 
| 20GB  | 24.534 | 25.137  | 2.40% | 
| 100GB | 30.54  | 31.541  | 3.17% |

I'll investigate further. Or if you have any feedback or advice, I'd appreciate it.

Machine specs used for testing:
RHEL7, 8 core, 256 GB RAM, xfs

Configuration:
wal_level = replica
autovacuum = off
full_page_writes = off

# For streaming replication from primary. 
synchronous_commit = remote_write
synchronous_standby_names = ''

# For Standby.
#hot_standby = on
#primary_conninfo

shared_buffers = 128MB
# 1GB, 20GB, 100GB

Just in case it helps for some understanding,
I also attached the recovery log 018_wal_optimize_node_replica.log
with some ereport that prints whether we enter the optimization loop or do full scan.

Regards,
Kirk Jamison