The patch looks OK. I think as Thomas-san suggested, we can remove the modification to smgrnblocks() and don't care wheter the size is cached or not. But I think the current patch is good too, so I'd like to leave it up to a committer to decide which to choose.
I measured performance in a different angle -- the time DropRelFileNodeBuffers() and DropRelFileNodeAllBuffers() took. That reveals the direct improvement and degradation. I used 1,000 tables, each of which is 1 MB. I used shared_buffers = 128 MB for the case where the traditional full buffer scan is done, and shared_buffers = 100 GB for the case where the optimization path takes effect. The results are almost good as follows: A. UNPATCHED 128 MB shared_buffers 1. VACUUM = 0.04 seconds 2. TRUNCATE = 0.04 seconds 100 GB shared_buffers 3. VACUUM = 69.4 seconds 4. TRUNCATE = 69.1 seconds B. PATCHED 128 MB shared_buffers (full scan) 5. VACUUM = 0.04 seconds 6. TRUNCATE = 0.07 seconds 100 GB shared_buffers (optimized path) 7. VACUUM = 0.02 seconds 8. TRUNCATE = 0.08 seconds So, I'd like to mark this as ready for committer. Regards Takayuki Tsunakawa