Unlike the FUSE implementation on Linux, macFUSE doesn't cache the file
attributes from the `LOOKUP` call, so it calls `GETATTR` prior to
accessing a file.
In the case of the `VirtualConfNode` (reverse config file passthrough),
this resulted in the default `GETATTR` implementation returning an empty
result, ultimately resulting in a "permission denied" error.
14:44:14.095207 rx 3: GETATTR n2
14:44:14.095229 tx 3: OK, {tA=1s {M0100000 SZ=0 L=0 0:0 0 0:8954996 A 0.000000 M 0.000000 C 0.000000}}
14:44:14.099943 rx 4: ACCESS n2 {u=501 g=20 r}
14:44:14.099990 tx 4: 13=permission denied
By impementing `Getattr` (from `fs.NodeGetattrer`) on `VirtualConfNode`
this solves the issue.
$ ./crossbuild.bash
[...]
+ GOOS=darwin
+ GOARCH=amd64
+ build
+ go build -tags without_openssl -o /dev/null
internal/fusefrontend/node.go:397:2: duplicate case syscallcompat.RENAME_NOREPLACE (value 0) in switch
previous case at internal/fusefrontend/node.go:397:7
internal/fusefrontend/node.go:397:2: duplicate case syscallcompat.RENAME_EXCHANGE (value 0) in switch
previous case at internal/fusefrontend/node.go:397:7
internal/fusefrontend/node.go:397:2: duplicate case syscallcompat.RENAME_WHITEOUT (value 0) in switch
previous case at internal/fusefrontend/node.go:397:7
internal/fusefrontend/node.go:399:38: duplicate case syscallcompat.RENAME_NOREPLACE | syscallcompat.RENAME_WHITEOUT (value 0) in switch
previous case at internal/fusefrontend/node.go:397:7
Quoting fusefrontend_reverse/node_helpers.go :
// File names are padded to 16-byte multiples, encrypted and
// base64-encoded. We can encode at most 176 bytes to stay below the 255
// bytes limit:
// * base64(176 bytes) = 235 bytes
// * base64(192 bytes) = 256 bytes (over 255!)
// But the PKCS#7 padding is at least one byte. This means we can only use
// 175 bytes for the file name.
Noticed by @bailey27 at https://github.com/rfjakob/gocryptfs/issues/499#issuecomment-955790427
Failure is:
# github.com/rfjakob/gocryptfs/v2/internal/nametransform
internal/nametransform/names.go:47:33: undefined: math.MaxInt
math.MaxInt was only introduced in Go 1.17. Use MaxInt32 instead,
which is good enough, even on amd64. It only has to be larger than
any name we might encounter.
Determines when to start hashing long names instead
of hardcoded 255. Will be used to alleviate "name too long"
issues some users see on cloud storage.
https://github.com/rfjakob/gocryptfs/issues/499
Reported by codacity:
internal/cryptocore/cryptocore.go
Minor icon MINOR
Code Style
should omit type AEADTypeEnum from declaration of var BackendAESSIV; it will be inferred from the right-hand side
var BackendAESSIV AEADTypeEnum = AEADTypeEnum{"AES-SIV-512", "Go", siv_aead.NonceSize}
Minor icon MINOR
Code Style
should omit type AEADTypeEnum from declaration of var BackendXChaCha20Poly1305; it will be inferred from the right-hand side
var BackendXChaCha20Poly1305 AEADTypeEnum = AEADTypeEnum{"XChaCha20-Poly1305", "Go", chacha20poly1305.NonceSizeX}
Minor icon MINOR
Code Style
should omit type AEADTypeEnum from declaration of var BackendXChaCha20Poly1305OpenSSL; it will be inferred from the right-hand side
var BackendXChaCha20Poly1305OpenSSL AEADTypeEnum = AEADTypeEnum{"XChaCha20-Poly1305", "OpenSSL", chacha20poly1305.NonceSizeX}
Found 2 possible new issues
internal/cryptocore/cryptocore.go
Minor icon MINOR
Code Style
should omit type AEADTypeEnum from declaration of var BackendOpenSSL; it will be inferred from the right-hand side
var BackendOpenSSL AEADTypeEnum = AEADTypeEnum{"AES-GCM-256", "OpenSSL", 16}
Minor icon MINOR
Code Style
should omit type AEADTypeEnum from declaration of var BackendGoGCM; it will be inferred from the right-hand side
var BackendGoGCM AEADTypeEnum = AEADTypeEnum{"AES-GCM-256", "Go", 16}
When somebody posts "gocryptfs -speed" results, they are
most helpful together with the CPU model. Add the cpu model
to the output.
Example:
$ ./gocryptfs -speed
gocryptfs v2.2.0-beta1-5-g52b0444-dirty; go-fuse v2.1.1-0.20210825171523-3ab5d95a30ae; 2021-09-14 go1.17.1 linux/amd64
cpu: Intel(R) Core(TM) i5-3470 CPU @ 3.20GHz; with AES acceleration
AES-GCM-256-OpenSSL 862.79 MB/s
AES-GCM-256-Go 997.71 MB/s (selected in auto mode)
AES-SIV-512-Go 159.58 MB/s
XChaCha20-Poly1305-OpenSSL 729.65 MB/s
XChaCha20-Poly1305-Go 843.97 MB/s (selected in auto mode)
We used to have "first Translate() wins". This is not deterministic,
as the LOOKUP for the root directory does not seem to reach us, so
the first user LOOKUP would win, which may be on a mountpoint.
I noticed that growslice() shows up in the cpuprofile.
Avoiding slice append for the private jey copy gives a 0.6% speedup:
gocryptfs/internal/speed$ benchstat old new
name old time/op new time/op delta
StupidXchacha-4 5.68µs ± 0% 5.65µs ± 0% -0.63% (p=0.008 n=5+5)
name old speed new speed delta
StupidXchacha-4 721MB/s ± 0% 725MB/s ± 0% +0.63% (p=0.008 n=5+5)
2% performance improvement, almost for free.
gocryptfs/internal/speed$ benchstat old new
name old time/op new time/op delta
StupidXchacha-4 5.82µs ± 0% 5.68µs ± 0% -2.37% (p=0.008 n=5+5)
name old speed new speed delta
StupidXchacha-4 704MB/s ± 0% 721MB/s ± 0% +2.43% (p=0.008 n=5+5)
Gets the decryption speed to the same level as the
encryption speed.
internal/speed$ benchstat old.txt new.txt
name old time/op new time/op delta
StupidXchacha-4 732MB/s ± 0% 740MB/s ± 0% ~ (p=1.000 n=1+1)
StupidXchachaDecrypt-4 602MB/s ± 0% 741MB/s ± 0% ~ (p=1.000 n=1+1)
Go has a high overhead for each C call, so batch
all openssl operations in the new C function chacha20poly1305_seal.
Benchmark results:
internal/speed$ go test -bench BenchmarkStupidXchacha -count 10 > old.txt
internal/speed$ go test -bench BenchmarkStupidXchacha -count 10 > new.txt
internal/speed$ benchstat old.txt new.txt
name old time/op new time/op delta
StupidXchacha-4 8.79µs ± 1% 7.25µs ± 1% -17.54% (p=0.000 n=10+10)
name old speed new speed delta
StupidXchacha-4 466MB/s ± 1% 565MB/s ± 1% +21.27% (p=0.000 n=10+10)
