Providing a hash.Block512 is an implementation detail of how deephash
works today, but providing an opaque type with mostly equivalent API
(i.e., HashUint8, HashBytes, etc. methods) is still sensible.
Thus, define a public Hasher type that exposes exactly the API
that an implementation of SelfHasher would want to call.
This gives us freedom to change the hashing algorithm of deephash
at some point in the future.
Also, this type is likely going to be called by types that are
going to memoize their own hash results, we additionally add
a HashSum method to simplify this use case.
Add documentation to SelfHasher on how a type might implement it.
Updates: corp#16409
Signed-off-by: Joe Tsai <joetsai@digital-static.net>
This updates all source files to use a new standard header for copyright
and license declaration. Notably, copyright no longer includes a date,
and we now use the standard SPDX-License-Identifier header.
This commit was done almost entirely mechanically with perl, and then
some minimal manual fixes.
Updates #6865
Signed-off-by: Will Norris <will@tailscale.com>
It is unclear whether the lack of checking nil-ness of slices
was an oversight or a deliberate feature.
Lacking a comment, the assumption is that this was an oversight.
Also, expand the logic to perform cycle detection for recursive slices.
We do this on a per-element basis since a slice is semantically
equivalent to a list of pointers.
Signed-off-by: Joe Tsai <joetsai@digital-static.net>
Add a new lookupTypeHasher function that is just a cached front-end
around the makeTypeHasher function.
We do not need to worry about the recursive type cycle issue that
made getTypeInfo more complicated since makeTypeHasher
is not directly recursive. All calls to itself happen lazily
through a sync.Once upon first use.
Signed-off-by: Joe Tsai <joetsai@digital-static.net>
The entry logic of Hash has extra complexity to make sure
we always have an addressable value on hand.
If not, we heap allocate the input.
For this reason we document that there are performance benefits
to always providing a pointer.
Rather than documenting this, just enforce it through generics.
Also, delete the unused HasherForType function.
It's an interesting use of generics, but not well tested.
We can resurrect it from code history if there's a need for it.
Signed-off-by: Joe Tsai <joetsai@digital-static.net>
This helps pprof better identify which Go kinds take the most time
since the kind is always in the function name.
Signed-off-by: Joe Tsai <joetsai@digital-static.net>
This helps pprof better identify which Go kinds take the most time
since the kind is always in the function name.
There is a minor adjustment where we hash the length of the map
to be more on the cautious side.
Signed-off-by: Joe Tsai <joetsai@digital-static.net>
Rather than having two copies []fieldInfo,
just maintain one and perform merging in the same pass.
Signed-off-by: Joe Tsai <joetsai@digital-static.net>
This helps pprof better identify which Go kinds take the most time
since the kind is always in the function name.
Signed-off-by: Joe Tsai <joetsai@digital-static.net>
Use of reflect.Value.SetXXX panics if the provided argument was
obtained from an unexported struct field.
Instead, pass an unsafe.Pointer around and convert to a
reflect.Value when necessary (i.e., for maps and interfaces).
Converting from unsafe.Pointer to reflect.Value guarantees that
none of the read-only bits will be populated.
When running in race mode, we attach type information to the pointer
so that we can type check every pointer operation.
This also type-checks that direct memory hashing is within
the valid range of a struct value.
We add test cases that previously caused deephash to panic,
but now pass.
Performance:
name old time/op new time/op delta
Hash 14.1µs ± 1% 14.1µs ± 1% ~ (p=0.590 n=10+9)
HashPacketFilter 2.53µs ± 2% 2.44µs ± 1% -3.79% (p=0.000 n=9+10)
TailcfgNode 1.45µs ± 1% 1.43µs ± 0% -1.36% (p=0.000 n=9+9)
HashArray 318ns ± 2% 318ns ± 2% ~ (p=0.541 n=10+10)
HashMapAcyclic 32.9µs ± 1% 31.6µs ± 1% -4.16% (p=0.000 n=10+9)
There is a slight performance gain due to the use of unsafe.Pointer
over reflect.Value methods. Also, passing an unsafe.Pointer (1 word)
on the stack is cheaper than passing a reflect.Value (3 words).
Performance gains are diminishing since SHA-256 hashing now dominates the runtime.
Signed-off-by: Joe Tsai <joetsai@digital-static.net>
Rather than separate functions to hash each kind,
just rely on the fact that these are direct memory hashable,
thus simplifying the code.
Signed-off-by: Joe Tsai <joetsai@digital-static.net>
Every implementation of typeHasherFunc always returns true,
which implies that the slow path is no longer executed.
Delete it.
h.hashValueWithType(v, ti, ...) is deleted as it is equivalent to:
ti.hasher()(h, v)
h.hashValue(v, ...) is deleted as it is equivalent to:
ti := getTypeInfo(v.Type())
ti.hasher()(h, v)
Signed-off-by: Joe Tsai <joetsai@digital-static.net>
Add support for maps and interfaces to the fast path.
Add cycle-detection to the pointer handling logic.
This logic is mostly copied from the slow path.
A future commit will delete the slow path once
the fast path never falls back to the slow path.
Performance:
name old time/op new time/op delta
Hash-24 18.5µs ± 1% 14.9µs ± 2% -19.52% (p=0.000 n=10+10)
HashPacketFilter-24 2.54µs ± 1% 2.60µs ± 1% +2.19% (p=0.000 n=10+10)
HashMapAcyclic-24 31.6µs ± 1% 30.5µs ± 1% -3.42% (p=0.000 n=9+8)
TailcfgNode-24 1.44µs ± 2% 1.43µs ± 1% ~ (p=0.171 n=10+10)
HashArray-24 324ns ± 1% 324ns ± 2% ~ (p=0.425 n=9+9)
The additional cycle detection logic doesn't incur much slow down
since it only activates if a type is recursive, which does not apply
for any of the types that we care about.
There is a notable performance boost since we switch from the fath path
to the slow path less often. Most notably, a struct with a field that
could not be handled by the fast path would previously cause
the entire struct to go through the slow path.
Signed-off-by: Joe Tsai <joetsai@digital-static.net>
There are 5 types that we care about that implement AppendTo:
key.DiscoPublic
key.NodePublic
netip.Prefix
netipx.IPRange
netip.Addr
The key types are thin wrappers around [32]byte and are memory hashable.
The netip.Prefix and netipx.IPRange types are thin wrappers over netip.Addr
and are hashable by default if netip.Addr is hashable.
The netip.Addr type is the only one with a complex structure where
the default behavior of deephash does not hash it correctly due to the presence
of the intern.Value type.
Drop support for AppendTo and instead add specialized hashing for netip.Addr
that would be semantically equivalent to == on the netip.Addr values.
The AppendTo support was already broken prior to this change.
It was fully removed (intentionally or not) in #4870.
It was partially restored in #4858 for the fast path,
but still broken in the slow path.
Just drop support for it altogether.
This does mean we lack any ability for types to self-hash themselves.
In the future we can add support for types that implement:
interface { DeepHash() Sum }
Test and fuzz cases were added for the relevant types that
used to rely on the AppendTo method.
FuzzAddr has been executed on 1 billion samples without issues.
Signed-off-by: Joe Tsai joetsai@digital-static.net
Rename Hash as Block512 to indicate that this is a general-purpose
hash.Hash for any algorithm that operates on 512-bit block sizes.
While we rename the package as hashx in this commit,
a subsequent commit will move the sha256x package to hashx.
This is done separately to avoid confusing git.
Signed-off-by: Joe Tsai <joetsai@digital-static.net>
Also, rename canMemHash to typeIsMemHashable to be consistent.
There are zero changes to the semantics.
Signed-off-by: Joe Tsai <joetsai@digital-static.net>
Any type that is memory hashable must not be recursive since
there are definitely no pointers involved to make a cycle.
Signed-off-by: Joe Tsai <joetsai@digital-static.net>
Put the t.Size() == 0 check first since this is applicable in all cases.
Drop the last struct field conditional since this is covered by the
sumFieldSize check at the end.
Signed-off-by: Joe Tsai <joetsai@digital-static.net>
Hashing []any is slow since hashing of interfaces is slow.
Hashing of interfaces is slow since we pessimistically assume
that cycles can occur through them and start cycle tracking.
Drop the variadic signature of Update and fix callers to pass in
an anonymous struct so that we are hashing concrete types
near the root of the value tree.
Signed-off-by: Joe Tsai <joetsai@digital-static.net>
Signed-off-by: Joe Tsai <joetsai@digital-static.net>
Formatting a time.Time as RFC3339 is slow.
See https://go.dev/issue/54093
Now that we have efficient hashing of fixed-width integers,
just hash the time.Time as a binary value.
Performance:
Hash-24 19.0µs ± 1% 18.6µs ± 1% -2.03% (p=0.000 n=10+9)
TailcfgNode-24 1.79µs ± 1% 1.40µs ± 1% -21.74% (p=0.000 n=10+9)
Signed-off-by: Joe Tsai <joetsai@digital-static.net>
Switch deephash to use sha256x.Hash.
We add sha256x.HashString to efficiently hash a string.
It uses unsafe under the hood to convert a string to a []byte.
We also modify sha256x.Hash to export the underlying hash.Hash
for testing purposes so that we can intercept all hash.Hash calls.
Performance:
name old time/op new time/op delta
Hash-24 19.8µs ± 1% 19.2µs ± 1% -3.01% (p=0.000 n=10+10)
HashPacketFilter-24 2.61µs ± 0% 2.53µs ± 1% -3.01% (p=0.000 n=8+10)
HashMapAcyclic-24 31.3µs ± 1% 29.8µs ± 0% -4.80% (p=0.000 n=10+9)
TailcfgNode-24 1.83µs ± 1% 1.82µs ± 2% ~ (p=0.305 n=10+10)
HashArray-24 344ns ± 2% 323ns ± 1% -6.02% (p=0.000 n=9+10)
The performance gains is not as dramatic as sha256x over sha256 due to:
1. most of the hashing already occurring through the direct memory hashing logic, and
2. what does not go through direct memory hashing is slowed down by reflect.
Signed-off-by: Joe Tsai <joetsai@digital-static.net>
In Go 1.19, the reflect.Value.MapRange method uses "function outlining"
so that the allocation of reflect.MapIter is inlinable by the caller.
If the iterator doesn't escape the caller, it can be stack allocated.
See https://go.dev/cl/400675
Performance:
name old time/op new time/op delta
HashMapAcyclic-24 31.9µs ± 2% 32.1µs ± 1% ~ (p=0.075 n=10+10)
name old alloc/op new alloc/op delta
HashMapAcyclic-24 0.00B 0.00B ~ (all equal)
Signed-off-by: Joe Tsai <joetsai@digital-static.net>
The logic of deephash is both simpler and easier to reason about
if values are always addressable.
In Go, the composite kinds are slices, arrays, maps, structs,
interfaces, pointers, channels, and functions,
where we define "composite" as a Go value that encapsulates
some other Go value (e.g., a map is a collection of key-value entries).
In the cases of pointers and slices, the sub-values are always addressable.
In the cases of arrays and structs, the sub-values are always addressable
if and only if the parent value is addressable.
In the case of maps and interfaces, the sub-values are never addressable.
To make them addressable, we need to copy them onto the heap.
For the purposes of deephash, we do not care about channels and functions.
For all non-composite kinds (e.g., strings and ints), they are only addressable
if obtained from one of the composite kinds that produce addressable values
(i.e., pointers, slices, addressable arrays, and addressable structs).
A non-addressible, non-composite kind can be made addressable by
allocating it on the heap, obtaining a pointer to it, and dereferencing it.
Thus, if we can ensure that values are addressable at the entry points,
and shallow copy sub-values whenever we encounter an interface or map,
then we can ensure that all values are always addressable and
assume such property throughout all the logic.
Performance:
name old time/op new time/op delta
Hash-24 21.5µs ± 1% 19.7µs ± 1% -8.29% (p=0.000 n=9+9)
HashPacketFilter-24 2.61µs ± 1% 2.62µs ± 0% +0.29% (p=0.037 n=10+9)
HashMapAcyclic-24 30.8µs ± 1% 30.9µs ± 1% ~ (p=0.400 n=9+10)
TailcfgNode-24 1.84µs ± 1% 1.84µs ± 2% ~ (p=0.928 n=10+10)
HashArray-24 324ns ± 2% 332ns ± 2% +2.45% (p=0.000 n=10+10)
Signed-off-by: Joe Tsai <joetsai@digital-static.net>
Regression from 09afb8e35b, in which the
same reflect.Value scratch value was being used as the map iterator
copy destination.
Also: make nil and empty maps hash differently, add test.
Fixes#4871
Co-authored-by: Josh Bleecher Snyder <josharian@gmail.com>
Change-Id: I67f42524bc81f694c1b7259d6682200125ea4a66
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
The index for every struct field or slice element and
the number of fields for the struct is unncessary.
The hashing of Go values is unambiguous because every type (except maps)
encodes in a parsable manner. So long as we know the type information,
we could theoretically decode every value (except for maps).
At a high level:
* numbers are encoded as fixed-width records according to precision.
* strings (and AppendTo output) are encoded with a fixed-width length,
followed by the contents of the buffer.
* slices are prefixed by a fixed-width length, followed by the encoding
of each value. So long as we know the type of each element, we could
theoretically decode each element.
* arrays are encoded just like slices, but elide the length
since it is determined from the Go type.
* maps are encoded first with a byte indicating whether it is a cycle.
If a cycle, it is followed by a fixed-width index for the pointer,
otherwise followed by the SHA-256 hash of its contents. The encoding of maps
is not decodeable, but a SHA-256 hash is sufficient to avoid ambiguities.
* interfaces are encoded first with a byte indicating whether it is nil.
If not nil, it is followed by a fixed-width index for the type,
and then the encoding for the underlying value. Having the type be encoded
first ensures that the value could theoretically be decoded next.
* pointers are encoded first with a byte indicating whether it is
1) nil, 2) a cycle, or 3) newly seen. If a cycle, it is followed by
a fixed-width index for the pointer. If newly seen, it is followed by
the encoding for the pointed-at value.
Removing unnecessary details speeds up hashing:
name old time/op new time/op delta
Hash-8 76.0µs ± 1% 55.8µs ± 2% -26.62% (p=0.000 n=10+10)
HashMapAcyclic-8 61.9µs ± 0% 62.0µs ± 0% ~ (p=0.666 n=9+9)
TailcfgNode-8 10.2µs ± 1% 7.5µs ± 1% -26.90% (p=0.000 n=10+9)
HashArray-8 1.07µs ± 1% 0.70µs ± 1% -34.67% (p=0.000 n=10+9)
Signed-off-by: Joe Tsai <joetsai@digital-static.net>
Instead of hashing the humanly formatted forms of a number,
hash the native machine bits of the integers themselves.
There is a small performance gain for this:
name old time/op new time/op delta
Hash-8 75.7µs ± 1% 76.0µs ± 2% ~ (p=0.315 n=10+9)
HashMapAcyclic-8 63.1µs ± 3% 61.3µs ± 1% -2.77% (p=0.000 n=10+10)
TailcfgNode-8 10.3µs ± 1% 10.2µs ± 1% -1.48% (p=0.000 n=10+10)
HashArray-8 1.07µs ± 1% 1.05µs ± 1% -1.79% (p=0.000 n=10+10)
Signed-off-by: Joe Tsai <joetsai@digital-static.net>
The swapping of bufio.Writer between hasher and mapHasher is subtle.
Just embed a hasher in mapHasher to avoid complexity here.
No notable change in performance:
name old time/op new time/op delta
Hash-8 76.7µs ± 1% 77.0µs ± 1% ~ (p=0.182 n=9+10)
HashMapAcyclic-8 62.4µs ± 1% 62.5µs ± 1% ~ (p=0.315 n=10+9)
TailcfgNode-8 10.3µs ± 1% 10.3µs ± 1% -0.62% (p=0.004 n=10+9)
HashArray-8 1.07µs ± 1% 1.06µs ± 1% -0.98% (p=0.001 n=8+9)
Signed-off-by: Joe Tsai <joetsai@digital-static.net>
The previous algorithm used a map of all visited pointers.
The strength of this approach is that it quickly prunes any nodes
that we have ever visited before. The detriment of the approach
is that pruning is heavily dependent on the order that pointers
were visited. This is especially relevant for hashing a map
where map entries are visited in a non-deterministic manner,
which would cause the map hash to be non-deterministic
(which defeats the point of a hash).
This new algorithm uses a stack of all visited pointers,
similar to how github.com/google/go-cmp performs cycle detection.
When we visit a pointer, we push it onto the stack, and when
we leave a pointer, we pop it from the stack.
Before visiting a pointer, we first check whether the pointer exists
anywhere in the stack. If yes, then we prune the node.
The detriment of this approach is that we may hash a node more often
than before since we do not prune as aggressively.
The set of visited pointers up until any node is only the
path of nodes up to that node and not any other pointers
that may have been visited elsewhere. This provides us
deterministic hashing regardless of visit order.
We can now delete hashMapFallback and associated complexity,
which only exists because the previous approach was non-deterministic
in the presence of cycles.
This fixes a failure of the old algorithm where obviously different
values are treated as equal because the pruning was too aggresive.
See https://github.com/tailscale/tailscale/issues/2443#issuecomment-883653534
The new algorithm is slightly slower since it prunes less aggresively:
name old time/op new time/op delta
Hash-8 66.1µs ± 1% 68.8µs ± 1% +4.09% (p=0.000 n=19+19)
HashMapAcyclic-8 63.0µs ± 1% 62.5µs ± 1% -0.76% (p=0.000 n=18+19)
TailcfgNode-8 9.79µs ± 2% 9.88µs ± 1% +0.95% (p=0.000 n=19+17)
HashArray-8 643ns ± 1% 653ns ± 1% +1.64% (p=0.000 n=19+19)
However, a slower but more correct algorithm seems
more favorable than a faster but incorrect algorithm.
Signed-off-by: Joe Tsai <joetsai@digital-static.net>
A Go interface may hold any number of different concrete types.
Just because two underlying values hash to the same thing
does not mean the two values are identical if they have different
concrete types. As such, include the type in the hash.
Seed the hash upon first use with the current time.
This ensures that the stability of the hash is bounded within
the lifetime of one program execution.
Hopefully, this prevents future bugs where someone assumes that
this hash is stable.
Signed-off-by: Joe Tsai <joetsai@digital-static.net>
Brad Fitzpatrick
Joe Tsai
Tom DNetto
Will Norris
Josh Soref
Andrew Dunham
Josh Bleecher Snyder
slowy07