An off-by-one error could cause tokens to be forgotten. Suppose
tokens := []string{"foo", "bar", "baz", "quux"}
i := 2
token := tokens[i] // token == "baz"
Then, after
copy(tokens[1:], tokens[:i+1])
tokens[0] = token
we have
tokens == []string{"baz", "foo", "bar", "baz"}
The short test actually relied on this bug.
Having a separate mutex for the three or four instructions needed to
fetch and increment nextID means the overhead exceeds the cost of this
operation. nextID is now handled inside the critical section for
awaiting instead, while the more expensive channel creation has been
moved outside it.
This is mostly a simplification, though it may have minor performance
benefits in some situations. The single-threaded sender benchmark shows
no significant difference:
name old speed new speed delta
RequestsRawTCP-8 55.3MB/s ± 7% 56.6MB/s ± 6% ~ (p=0.190 n=10+10)
RequestsTLSoTCP-8 20.5MB/s ±20% 20.8MB/s ± 8% ~ (p=0.604 n=10+9)
* lib/protocol: Require at least 3.125% savings from compression
The new lz4 library doesn't need its output buffer to be the maximum
size, unlike the old one (which would allocate if it weren't). It can
take a buffer that is of a smaller size and will report if compressed
data can fit inside the buffer (with a small chance of reporting a false
negative). Use that property to our advantage by requiring compressed
data to be at most n-n/32 = .96875*n bytes long for n input bytes.
* lib/protocol: Remove unused receivers
To make DeepSource happy.
* lib/protocol: Micro-optimize lz4Compress
Only write the length if compression was successful. This is a memory
write, so the compiler can't reorder it.
Only check the return value of lz4.CompressBlock. Length-zero inputs
are always expanded by LZ4 compression (the library documents this),
so the check on len(src) isn't needed.
When pathSep is a constant, the compiler precomputes pathSep+pathSep and
".."+pathSep instead of emitting function calls to compute "//" and
"../". Benchmark results in lib/osutil:
name old time/op new time/op delta
TraversesSymlink-8 8.86µs ± 3% 8.53µs ± 4% -3.79% (p=0.000 n=18+20)
name old alloc/op new alloc/op delta
TraversesSymlink-8 1.06kB ± 0% 1.06kB ± 0% ~ (all equal)
name old allocs/op new allocs/op delta
TraversesSymlink-8 15.0 ± 0% 15.0 ± 0% ~ (all equal)
The locking protocol in nat.Mapping was racy:
* Mapping.addressMap RLock'd, but then returned a map shared between
caller and Mapping, so the lock didn't do anything.
* Operations inside Service.{verifyExistingMappings,acquireNewMappings}
would lock the map for every update, but that means callers to
Mapping.ExternalAddresses can be looping over the map while the
Service methods are concurrently modifying it. When the Go runtime
detects that happening, it panics.
* Mapping.expires was read and updated without locking.
The Service methods now lock the map once and release the lock only when
done.
Also, subscribers no longer get the added and removed addresses, because
none of them were using the information. This was changed for a previous
attempt to retain the fine-grained locking and not reverted because it
simplifies the code.
By truncating time.Time to an int64 nanosecond count, we lose the
ability to precisely order timestamps before 1678 or after 2262, but we
gain (linux/amd64, Go 1.17.1):
name old time/op new time/op delta
JobQueuePushPopDone10k-8 2.85ms ± 5% 2.29ms ± 2% -19.80% (p=0.000 n=20+18)
JobQueueBump-8 34.0µs ± 1% 29.8µs ± 1% -12.35% (p=0.000 n=19+19)
name old alloc/op new alloc/op delta
JobQueuePushPopDone10k-8 2.56MB ± 0% 1.76MB ± 0% -31.31% (p=0.000 n=18+13)
name old allocs/op new allocs/op delta
JobQueuePushPopDone10k-8 23.0 ± 0% 23.0 ± 0% ~ (all equal)
Results for BenchmarkJobQueueBump are with the fixed version, which no
longer depends on b.N for the amount of work performed. rand.Rand.Intn
is cheap at ~10ns per iteration.
Registry.Get used a full sort to get the minimum of a list, and the sort
was broken because util.AddressUnspecifiedLess assumed it could find out
whether an address is IPv4 or IPv6 from its Network method. However,
net.(TCP|UDP)Addr.Network always returns "tcp"/"udp".
Benchmark results on Linux/amd64, using updated benchmark for old and
new:
name old time/op new time/op delta
HashFile-8 88.6ms ± 1% 88.3ms ± 1% -0.33% (p=0.046 n=19+19)
name old speed new speed delta
HashFile-8 201MB/s ± 1% 202MB/s ± 1% +0.33% (p=0.044 n=19+19)
name old alloc/op new alloc/op delta
HashFile-8 59.4kB ± 0% 46.1kB ± 0% -22.47% (p=0.000 n=14+20)
name old allocs/op new allocs/op delta
HashFile-8 29.0 ± 0% 27.0 ± 0% -6.90% (p=0.000 n=20+20)
Co-authored-by: greatroar <@>
Most notably, it now detects all-lowercase files and returns these
as-is. The tests have been expanded with two cases and are now used
as a benchmark (admittedly a rather trivial one).
name old time/op new time/op delta
UnicodeLowercaseMaybeChange-8 4.59µs ± 2% 4.57µs ± 1% ~ (p=0.197 n=10+10)
UnicodeLowercaseNoChange-8 3.26µs ± 1% 3.09µs ± 1% -5.27% (p=0.000 n=9+10)
When cap(permanentRelays) >= len(permanentRelays) + len(knownRelays),
append(permanentRelays, knownRelays...)
returns a slice of the array underlying permanentRelays. The subsequent
rand.Shuffle then mixes the permanent and known relays. Sequential
requests may cause strelaypoolsrv to forget its permanent relays. Worse,
concurrent requests may cause shuffling of the same slice on multiple
processors concurrently.
Co-authored-by: greatroar <@>
If the GC finds a key k that it wants to keep, it records that in a
Bloom filter. If a key k' can be removed but its hash collides with k,
it will be kept. Since the old Bloom filter code was completely
deterministic, the next run would encounter the same collision, assuming
k must still be kept.
A randomized hash function that uses all the SHA-256 bits solves this
problem: the second run has a non-zero probability of removing k', as
long as the Bloom filter is not completely full.
This is shorter, skips two allocations, makes the function inlineable
and is safer, since the compiler now check whether
DeviceIDLength == sha256.Size.
crypto/rand output is cryptographically secure by the Go library
documentation's promise. That, rather than strength (= passes randomness
tests) is the property that Syncthing needs).
This matches the convention of the stdlib and avoids ambiguity: when
customErr{} and &customErr{} both implement error, client code needs to
check for both.
Memory use should remain the same, since storing a non-pointer type in
an interface value still copies the value to the heap.
Storing assets as []byte requires every compiled-in asset to be copied
into writable memory at program startup. That currently takes up 1.6MB
per syncthing process. Strings stay in the RODATA section and should be
shared between processes running the same binary.
Successful LRU cache lookups modify the cache's recency list, so
RWMutex.RLock isn't enough protection.
Secondarily, multiple concurrent lookups with the same key should not
create separate rate limiters, so release the lock only when presence
of the key in the cache has been ascertained.
Co-authored-by: greatroar <@>
The previous implementation was very generic; its tests didn't cover the
actual alphabet for device IDs.
Benchmark results on amd64:
name old time/op new time/op delta
Luhnify-8 1.00µs ± 1% 0.28µs ± 4% -72.38% (p=0.000 n=9+10)
Unluhnify-8 992ns ± 2% 274ns ± 1% -72.39% (p=0.000 n=10+9)
greatroarandJakob Borg
