[24.0 backport] pkg/fileutils: GetTotalUsedFds: reduce allocations

[thaJeztah]

• backport of pkg/fileutils: GetTotalUsedFds: reduce allocations #45848
• relates to dockerd eating too much memory and keeps growing #45842
• ⚠️ added one commit with a partial backport of the log changes

---

pkg/fileutils: GetTotalUsedFds(): don't pretend to support FreeBSD

Commit 8d56108 moved this function from
the generic (no build-tags) fileutils.go to a unix file, adding "freebsd"
to the build-tags.

This likely was a wrong assumption (as other files had freebsd build-tags).
FreeBSD's procfs does not mention /proc/<pid>/fd in the manpage, and
we don't test FreeBSD in CI, so let's drop it, and make this a Linux-only
file.

While updating also dropping the import-tag, as we're planning to move
this file internal to the daemon.

pkg/fileutils: add BenchmarkGetTotalUsedFds

go test -bench ^BenchmarkGetTotalUsedFds$ -run ^$ ./pkg/fileutils/
goos: linux
goarch: arm64
pkg: github.com/docker/docker/pkg/fileutils
BenchmarkGetTotalUsedFds-5   	  149272	      7896 ns/op	     945 B/op	      20 allocs/op

pkg/fileutils: GetTotalUsedFds: reduce allocations

Use File.Readdirnames instead of os.ReadDir, as we're only interested in
the number of files, and results don't have to be sorted.

Before:

BenchmarkGetTotalUsedFds-5            149272              7896 ns/op             945 B/op         20 allocs/op

After:

BenchmarkGetTotalUsedFds-5            153517              7644 ns/op             408 B/op         10 allocs/op

pkg/fileutils: GetTotalUsedFds(): use fast-path for Kernel 6.2 and up

Linux 6.2 and up (commit f1f1f2569901ec5b9d425f2e91c09a0e320768f3)
provides a fast path for the number of open files for the process.

From the Linux docs:

The number of open files for the process is stored in 'size' member of
stat() output for /proc//fd for fast access.

This patch adds a fast-path for Kernels that support this, and falls back
to the slow path if the Size fields is zero.

Comparing on a Fedora 38 (kernel 6.2.9-300.fc38.x86_64):

Before/After:

go test -bench ^BenchmarkGetTotalUsedFds$ -run ^$ ./pkg/fileutils/
BenchmarkGetTotalUsedFds        57264     18595 ns/op     408 B/op      10 allocs/op
BenchmarkGetTotalUsedFds       370392      3271 ns/op      40 B/op       3 allocs/op

Note that the slow path has 1 more file-descriptor, due to the open
file-handle for /proc//fd during the calculation.

- What I did

- How I did it

- How to verify it

- Description for the changelog

- A picture of a cute animal (not mandatory but encouraged)

[neersighted]

I just ran into the containerd log changes myself in #45857; personally I think it's more idiomatic to fix the conflict in the cherry-picked commit itself, rather than synthesize a single-file cherry-pick to make a downstream cherry-pick 'clean.'

[thaJeztah]

Yeah, always a bit on the fence; in this case; all the cherry-picks where squeaky clean after the first commit, and it may be more clear in history, but yeah YMMV I guess.

[cpuguy83]

Surprised you used Sprintf here :)

[thaJeztah]

Haha, yeah I was doubting, but in the total scheme of things, it was slightly more readable than strconv and path.Join combined