Fix seqscan prefetch in pg17

[knizhnik]

Problem

See #9860

Neon doesn't have seqscan detection of its own, so stop read_stream from trying to utilize that readahead, and instead make it issue readahead of its own.

Summary of changes

Add loop calling PrefetchBuffer for next effective_io_concurrency blocks.

[github-actions]

6879 tests run: 6572 passed, 0 failed, 307 skipped (full report)

---

Flaky tests (2)

Postgres 17

• test_sharding_split_failures[failure12]: debug-x86-64

Postgres 15

• test_tenant_delete_stale_shards: release-arm64

Code coverage* (full report)

• functions: 31.0% (7969 of 25720 functions)
• lines: 48.8% (63276 of 129701 lines)

* collected from Rust tests only

---

_{The comment gets automatically updated with the latest test results
5ecc4b2 at 2024-11-26T08:15:44.358Z :recycle:}

[hlinnaka]

This addresses the same issue as #9860. I think PR #9860 is a better fix, read_stream.c should handle this in the new v17 model. I see your comment there that there's something wrong with the PR, so let's try to fix that.

[knizhnik]

I am not sure that #9860 is better fix.
Unless I missed something, current AIO approach is to pin all prefetched buffers.
I think it will not work well in Neon when size of shared_buffers is small enough.
I wrote about it in my comment in Slack.

I will try to investigate more why prefetch distance in pg17 is 1, while max_ios is 100 (effective_io_ooncurency).
I suspect it is case caused by max pinned buffer limit. But doesn't matter where it comes from, art seems to be not the best idea to pin prefetched buffers.

[hlinnaka]

I am not sure that #9860 is better fix. Unless I missed something, current AIO approach is to pin all prefetched buffers. I think it will not work well in Neon when size of shared_buffers is small enough. I wrote about it in my comment in Slack.

This is the way it's designed to work in upstream too. Our prefetching isn't that different from the upstream prefetching with posix_fadvise, nor the way async I/O will work in the future. So if there's a problem with pinning, we should address that.

We set shared_buffers to 16384. I think that leaves enough room for pinning buffers, even with pretty high effective_io_concurrency settings.

[MMeent]

I think this is obsolete once #9860 merges (now only a matter of CI passing) so I'll close this one.