Futureweb

Software

A Linux 7.4 pull closes files inline on close_range, exec, and exit. The tarball is not cut.

Christian Brauner's 7.4 pull says close_range, exec, and exit now close files inline. He cites a teardown of a million descriptors. Linux 7.4 is a merge window, not a release.

The short version

Christian Brauner's October 16, 2026 pull for the Linux 7.4 merge window says close_range, exec, and exit now put each final file close inline, and he reports a faster teardown of a million descriptors. Phoronix reported the pull. Linux 7.4 is not a released tarball.

A pull dated October 16, reported October 21

Christian Brauner's cover message for the Linux 7.4 vfs pulls is dated October 16, 2026. One item in that batch is that closing files is now synchronous for close_range, exec, and exit. Phoronix's October 21 story is the account on the software wire. The software desk already filed the UnixWare boot filesystem leaving this same window. This page files the close path, which is a different change. Asked directly, What new flags does close_range grow? The pull adds CLOSE_RANGE_EXCEPT, which acts on descriptors outside the range, and CLOSE_RANGE_CLOEXEC_ONLY, which closes only the close-on-exec descriptors in the range.Linux 7.3 without the rc suffix.

Why the old path was slow, in his words

Brauner writes that every final close of an exiting task used to be queued as task work and run later. For a large descriptor table that meant a compare-and-swap and a flag update per file, then a second pass over file objects that were no longer cache hot. close_files, close_range, and the close-on-exec pass now flush and put each file inline, the way a single close already did. The change is where the work runs, not a new system call name.UnixWare BFS leaves the 7.4 window.

The million-descriptor timing is his

The cover message says tearing down one million descriptors goes from 429 to 306 milliseconds on exit and from 471 to 325 milliseconds on execve. He calls that a 16 to 21 percent win, or about 40 to 150 nanoseconds per closed file. Phoronix repeats the direction of the change and does not add a lab of its own. A reader who wants the number wants his message, not a leaderboard.the 7.3 release candidate.

Two flags on close_range

The same batch adds CLOSE_RANGE_EXCEPT, so the call can apply to every descriptor outside a window, and CLOSE_RANGE_CLOEXEC_ONLY, so it closes only the close-on-exec descriptors inside a range. Brauner's note for that pull says a task sitting between clone and exec can drop everything except a window of descriptors in one call, and that descriptors in a skipped range are not copied. The flags are an interface change inside the merge window. They are not a userspace program this page is shipping.what a license obligates.

The merge window is not the tarball

kernel.org on October 21 still showed Linux 7.3, dated October 18, as mainline, and 7.2.12 as the stable update. A pull Linus has been asked to take is not the version a distribution can point at. The October 19 brief on 7.3 said a later document would be a 7.4 candidate or a stable patch. This pull is one of the documents inside the window. It does not cut 7.4. A separate October 18 sched/core pull carries short-slice latency changes and is not this close path. Asked directly, Does Phoronix add a benchmark? No. The October 21 item is an account of the pull. It does not publish a second machine's timing.the short-slice pull.

What else the cover message bundles

The cover note also describes coredump records, a flag on open_tree, and kernfs lock work. Those are in the same batch and they are not this page's question. The dependency here is the close path: inline puts, two close_range flags, and a timing the author measured on a million descriptors. A reader who needs the coredump protocol reads the pull that contains it. Asked directly, What got faster, in the pull's own numbers? Brauner says tearing down one million descriptors goes from 429 to 306 milliseconds on exit and from 471 to 325 milliseconds on execve, which he calls a 16 to 21 percent win. Those are his figures.the software desk.

Who is in scope

Processes that exit or exec with very large descriptor tables are the case the timing describes. Container runtimes that already use close_range are the callers the new flags address. A desktop that closes a handful of files will not see a 100 millisecond swing, and the pull does not claim it will. Phoronix's audience is the same kernel readers. It does not name a distribution that has packaged the change. Asked directly, Is this the UnixWare filesystem removal? No. That pull removed a boot filesystem driver. This pull changes how the kernel closes file descriptors. They are different documents in the same merge window.OpenBSD 8.0 is the 61st release.

What stays on the record

October 16 pull, October 21 retelling. Inline close on close_range, exec, and exit. His timings: 429 to 306 milliseconds on exit, 471 to 325 on execve, for one million descriptors. Two new close_range flags. Linux 7.4 not released. Distinct from the UnixWare filesystem removal. The short answer this page will stand on is this: Christian Brauner's October 16, 2026 pull for the Linux 7.4 merge window says close_range, exec, and exit now put each final file close inline, and he reports a faster teardown of a million descriptors. Phoronix reported the pull. Linux 7.4 is not a released tarball.Futureweb.

What would move the row

A Linux 7.4 tarball, or a stable patch that carries the same close path, would be the next document. Until then the record is the pull message and Phoronix's account of it. A distribution changelog that names the commit would be a third document, and it is not in these sources. Asked directly, Is Linux 7.4 released? No. The October 16 message is a pull request into the 7.4 merge window. kernel.org still listed 7.3 as mainline on October 21. A pull is not a tarball.BeamFS is still a request.

Sources

The reports this brief is filing. Futureweb did not republish them.

  1. vfs pull for Linux 7.4
  2. vfs sync_close pull
  3. Phoronix

Questions

Is Linux 7.4 released?

No. The October 16 message is a pull request into the 7.4 merge window. kernel.org still listed 7.3 as mainline on October 21. A pull is not a tarball.

What got faster, in the pull's own numbers?

Brauner says tearing down one million descriptors goes from 429 to 306 milliseconds on exit and from 471 to 325 milliseconds on execve, which he calls a 16 to 21 percent win. Those are his figures.

What new flags does close_range grow?

The pull adds CLOSE_RANGE_EXCEPT, which acts on descriptors outside the range, and CLOSE_RANGE_CLOEXEC_ONLY, which closes only the close-on-exec descriptors in the range.

Does Phoronix add a benchmark?

No. The October 21 item is an account of the pull. It does not publish a second machine's timing.

Is this the UnixWare filesystem removal?

No. That pull removed a boot filesystem driver. This pull changes how the kernel closes file descriptors. They are different documents in the same merge window.