Linux 7.4 VFS Patch Speeds Up File Opens by 39% in One Test

A Linux Virtual File System optimization queued for the Linux 7.4 development cycle removes an unnecessary reference-counting round trip from a common file-open path. In developer Mateusz Guzik’s targeted will-it-scale test, opening the same file read-only on a 20-core virtual machine increased from 4,043,375 to 5,629,378 operations per second — a 39% gain.

The result is striking because the patch is not a new filesystem, scheduler or storage driver. It changes how an already-held terminal dentry reference is handed into the file-opening path, avoiding a second reference acquisition followed by dropping the first one. That cuts synchronization work from a very hot kernel operation.

During pathname lookup, the existing path acquires a reference to the terminal dentry in __legitimize_path(). The normal open path then takes another reference in do_dentry_open(), before terminate_walk() releases the original reference. Guzik’s patch changes the common case so do_dentry_open() can consume the reference that is already held instead of performing that redundant ref/unref cycle.

That distinction is useful when interpreting the benchmark. The optimization is reducing kernel bookkeeping around file opening; it is not making an SSD 39% faster, increasing filesystem read bandwidth by 39%, or reducing every application’s runtime by 39%. Workloads dominated by other costs can see a much smaller end-to-end difference.

Reference-count modifications can become disproportionately expensive when many CPU cores contend on the same cache line. Removing two unnecessary modifications from a common operation can therefore matter most when many threads are opening files concurrently. Build systems, package tooling, servers, container-heavy environments and metadata-intensive developer workloads are plausible beneficiaries, but the patch’s published benchmark does not establish a specific application-level gain for any of them.

For desktop users, the practical effect may be difficult to notice in ordinary launches where file opening is only one piece of the workload. The more important technical point is that the optimization sits in the VFS layer rather than being tied to one storage device or one specific on-disk filesystem.

The patch has gone through multiple revisions over more than two years and is now queued in the VFS vfs-7.4.lookup branch for the Linux 7.4 cycle. That is a meaningful upstream milestone, but it is not the same as the change already shipping in a stable Linux 7.4 release or in current distributions.

The next evidence to watch is the Linux 7.4 merge window and broader benchmarking once the change reaches mainline. Those tests will show whether removing the dentry reference cycle produces measurable improvements outside the deliberately file-open-heavy benchmark that motivated the patch.