Check for invalid down link while prefetching B-Tree leave pages for index-only scan

[knizhnik]

Problem

See #9866

Index-only scan prefetch implementation doesn't take in account that down link may be invalid

Summary of changes

Check that downlink is valid block number

Correspondent Postgres PRs:
neondatabase/postgres#534
neondatabase/postgres#535
neondatabase/postgres#536
neondatabase/postgres#537

[github-actions]

7095 tests run: 6798 passed, 0 failed, 297 skipped (full report)

---

Code coverage* (full report)

• functions: 31.4% (8402 of 26778 functions)
• lines: 48.1% (66635 of 138650 lines)

* collected from Rust tests only

---

_{The comment gets automatically updated with the latest test results
af979af at 2024-12-13T20:37:05.642Z :recycle:}

[hlinnaka]

Please add a test case for this

[MMeent]

I don't think I understand how a downlink can contain InvalidBlockNumber, could you refer me to code which generates that?

[MMeent]

I think you have a mistaken understanding of the issue. In the issue you've linked the calculated value for skip is sometimes negative. That's bad, because we use that to read from the prefetch_blocks array, which thus causes out-of-bounds reads in that array, which finally would be the cause for InvalidBlockNumber results from that access.

I think you'll have to fix the skip calculation instead of papering over the invalid results from that skip calculation.

[MMeent]

This needs a different approach to fixing the issue, as this seems to just paper over it.

See #9867 (comment)

[knizhnik]

I think you have a mistaken understanding of the issue. In the issue you've linked the calculated value for skip is sometimes negative. That's bad, because we use that to read from the prefetch_blocks array, which thus causes out-of-bounds reads in that array, which finally would be the cause for InvalidBlockNumber results from that access.

I think you'll have to fix the skip calculation instead of papering over the invalid results from that skip calculation.

I think you are right. It seems to me that I have saw checks for InvalidBlockNumber but right now I have checked it once again ands that it is used only for parent link.

So I restored original code and added extra checks (including check that skip is not negative). But tests are now passed both CI and locally.
So still trying to reproduce the issue.

[MMeent]

LGTM