How Xavier Xia’s Persistent Optimizations Made contpte_ptep_get Faster in All Scenarios

The Linux kernel function contpte_ptep_get() originally scans CONT_PTES page‑table entries unconditionally to collect the dirty and young flags for the whole contpte, resulting in a full loop regardless of the actual state of the entries.

Xavier Xia’s first patch introduced an early‑exit: as soon as both a dirty and a young entry were found, the loop would break. This could reduce work when such entries appear early, but it also risked a negative impact if none of the entries were dirty or young, because the break would never be taken and the extra conditional checks could add overhead.

The community identified three representative scenarios to evaluate the patch:

All entries are both young and dirty.

No entry is young or dirty.

Entries are not dirty, but a single entry in the middle is young.

Most observers doubted that a single patch could improve performance across all three cases; some feared regression in the second scenario.

In response, Xavier produced several iterative versions, each adding more complex logic to handle edge cases. The code grew increasingly difficult to read, testing community patience.

After multiple revisions, Xavier released a final version (v6) that incorporated additional checks and branch predictions designed to keep the fast path short while preserving correctness. The final patch is shown below.

Benchmark data collected by Xavier demonstrated that the new implementation outperformed the original in every scenario, and even in the worst‑case (no dirty or young entries) it showed no measurable regression.

Despite the data, some community members remained skeptical. Ryan Roberts publicly questioned whether the added complexity justified the gains. He later ran his own tests, reproducing Xavier’s results and publishing the numbers.

The episode illustrates how persistent, data‑driven engineering can overcome initial skepticism. Xavier’s willingness to iterate, measure, and publish concrete results ultimately convinced the kernel community that the optimization was both safe and beneficial.