Bisecting shows that the performance drop is caused by
this commit:
commit ca9e912a28 (refs/bisect/bad)
Author: Jakob Unterwurzacher <jakobunt@gmail.com>
Date: Sat Feb 29 19:58:08 2020 +0100
fusefrontend: drop xattr user namespace restriction
Adding flags allows to use inomap in reverse mode,
replacing the clunky inoBaseDirIV/inoBaseNameFile
logic that causes problems with high underlying
inode numbers ( https://github.com/rfjakob/gocryptfs/issues/457 )
Microbenchmarks (values below) show that the "SingleDev"
case is now much slower due to an extra map lookup,
but this has no visible effects in ./test.bash results,
so there was no time spent optimizing the case further.
$ go test -bench=.
goos: linux
goarch: amd64
pkg: github.com/rfjakob/gocryptfs/internal/inomap
BenchmarkTranslateSingleDev-4 18757510 61.5 ns/op
BenchmarkTranslateManyDevs-4 18061515 64.5 ns/op
PASS
ok github.com/rfjakob/gocryptfs/internal/inomap 2.467s
The case of a git repo without any tags used to fail
with:
fatal: No names found, cannot describe anything.
Now we continue, using "[no_tags_found]" as the
version string.
The comment still mentioned CBC, which has been removed
a long time ago.
The test definition can be rewritten using slice literals,
saving sume stuttering.
We used to prefer openssl in this situation, which
used to make sense, but now Go gained an optimized
assembly implementation for aes-gcm on arm64 with
aes instructions:
root@q1:~/go/src/github.com/rfjakob/gocryptfs# ./gocryptfs -speed
gocryptfs v1.7.1-46-g73436d9; go-fuse v1.0.1-0.20190319092520-161a16484456; 2020-04-13 go1.14.2 linux/arm64
AES-GCM-256-OpenSSL 212.30 MB/s (selected in auto mode)
AES-GCM-256-Go 452.30 MB/s
AES-SIV-512-Go 100.25 MB/s
XChaCha20-Poly1305-Go 137.35 MB/s
https://github.com/rfjakob/gocryptfs/issues/452
Output now looks like this
$ gocryptfs -speed
gocryptfs v1.7.1-38-gbe3b9df-dirty; go-fuse v2.0.2-57-gd1cfa17; 2020-04-13 go1.13.6 linux/amd64
AES-GCM-256-OpenSSL 607.90 MB/s
AES-GCM-256-Go 920.75 MB/s (selected in auto mode)
AES-SIV-512-Go 169.85 MB/s
XChaCha20-Poly1305-Go 794.30 MB/s
and has go version and arch information, which is important
when comparing results.
From https://github.com/golang/go/wiki/GoArm :
In cross compilation situations, it is recommended
that you always set an appropriate GOARM value
along with GOARCH.
The value seems to default to GOARM=5 if not set
during cross-compilation.
We used to restrict setting xattrs to the "user."
namespace. I don't see a real reason for this
anymore, and it causes trouble for users who are using
acls.
Tests will be added in the next commit.
https://github.com/rfjakob/gocryptfs/issues/453
Now that I have discovered golang.org/x/sys/cpu and that Go
versions below 1.6 are uncommon, there was not much useful
code left in prefer_openssl.
Merge the remains into stupidgcm.
