Btrfs Gets Up to 5x Faster in Some Linux 7.3 Paths, and Phoronix Benchmarks Show a 25% Average Gain

Updated Oct. 26, 2026, 3:20 a.m. ET: Adds that Linux 7.3 shipped as a stable kernel on Oct. 18.

Btrfs picks up some of its largest performance gains in years with Linux 7.3. SUSE’s David Sterba sent the file system’s feature updates for the merge window on Aug. 19, and Phoronix summarized the pull request the same day: several changes bring wins of roughly 3x to 5x in specific code paths.

The biggest change is to direct I/O. Btrfs now uses an IOmap bounce buffer instead of falling back to buffered I/O. According to the merge request, earlier trade-offs between correctness and speed held direct I/O to about 50% of its theoretical maximum, and the new path reaches about 95%, so it roughly doubles.

The pull also includes smaller fixes, and some of them have big effects:

• A local LRU list replaces XArray for tracking inhibited extent buffers, worth about 3x in that path.
• Dropping an unnecessary one-jiffy delay in non-SSD mode with multiple logging tasks cut latency and raised throughput about 5x on a sample workload.
• Skipping hole detection during a full fsync, for files with many extents and no holes, gave another 5x.
• Locking around extent readahead was reduced and extent buffer allocation was improved.

Btrfs also turns off its old free-space v1 code by default. Free-space v2 has been the default for years.

What the benchmarks showed

Michael Larabel benchmarked Linux 7.2 stable against Linux 7.3 Git (as of Aug. 23) on an HP Z4 G6i workstation with an Intel Xeon 678X Granite Rapids processor and a 2TB Sandisk SN8000S PCIe 4.0 SSD. He formatted Btrfs fresh for each run and used stock mount options.

FIO random reads were mostly flat, and SQLite was only slightly faster. The large gains showed up in FIO random and sequential writes with direct I/O, which Larabel attributes to the new bounce-buffer path. Dbench improved substantially with multiple clients but stayed flat with a single client. TigerBeetle, the financial transactions database, saw much higher throughput and lower latency, ClickHouse and PostgreSQL improved modestly, and MariaDB was one of the few tests that regressed.

Across the whole suite, the geometric mean came out about 25% better on 7.3. That result comes from one SSD-backed workstation, and Larabel said EXT4, XFS and F2FS comparisons on the same kernels would follow.

Linux 7.3 was released as stable on Oct. 18, so distributions that track recent kernels now get these changes. Teams running write-heavy databases or virtual machines on Btrfs with NVMe storage should see the biggest difference.