Packages P1 and P2 can define identical struct types T1 and T2, and one
can convert from type T1 to T2 or vice versa.
The spec defines two identical struct types as:
Two struct types are identical if they have the same sequence of
fields, and if corresponding fields have the same names, and
identical types, and identical tags. Non-exported field names
from different packages are always different.
Unfortunately, garble broke this: since we obfuscated field names
differently depending on the package, cross-package conversions like the
case above would result in typechecking errors.
To fix this, implement Joe Tsai's idea: hash struct field names with the
string representation of the entire struct. This way, identical struct
types will have their field names obfuscated in the same way in all
packages across a build.
Note that we had to refactor "reverse" a bit to start using transformer,
since now it needs to keep track of struct types as well.
This failure was affecting the build of google.golang.org/protobuf,
since it makes regular use of cross-package struct conversions.
Note that the protobuf module still fails to build, but for other
reasons. The package that used to fail now succeeds, so the build gets a
bit further than before. #240 tracks adding relevant third-party Go
modules to CI, so we'll track the other remaining failures there.
Fixes#310.
Previously, the local name part of linkname directives were never
obfuscated. Now that we can obfuscate function names in Go assembly,
that limitation is required no longer.
Fixes#309
Obfuscating some std packages for windows/arm triggered a bug; when
encountering a call to runtime·memmove, we'd hash "memmove" with the
current package's action ID.
This is wrong on two levels: First, we aren't obfuscating the runtime
package yet. And second, if we did, we would have to hash the symbol
appropriately, with that package's action ID.
For now, only hashing the local names does the trick. That's all that
the code currently supports, anyway.
Since 1a8e32227f we have been setting line numbers on call sites, and
that line number is always the minimum value for -tiny: 1.
Still not as zero-overhead as the old mechanism which entirely deleted
line numbers, but a useless and small line number is still pretty good.
THey don't show up in stack traces, and if they show up in regular
program output, we should in theory not obfuscate those names via the
detection of reflection.
However, there's one relatively common scenario where obfuscated field
names can appear: in build error messages, when obfuscation fails a
build.
Our previous logic only took care of fairly simple types, such as a
simple struct or a pointer to a struct. If we had a struct embedding
another struct, we'd fail to record the objects for the fields in the
inner struct, and that would lead to miscompilation:
> garble build
[stderr]
# test/main
LZmt64Nm.go:7: outer.InnerField undefined (type *CcUt1wkQ.EmbeddingOuter has no field or method InnerField)
To fix this issue, make the function that records all objects under a
types.Type smarter. Since it now does more than just dealing with
structs, it's also renamed.
Since the function now walks types properly, we get to remove the extra
ast.Inspect in recordReflectArgs, which is nice.
We also make it a method, to avoid the map parameter. A boolean
parameter is also added, since we need this feature to only look at the
current package when looking at reflect calls.
Finally, we add a test case, a simplified version of the original bug
report.
Fixes#315.
The docs for go/importer.ForCompiler say:
The lookup function is called each time the resulting importer
needs to resolve an import path. In this mode the importer can
only be invoked with canonical import paths (not relative or
absolute ones); it is assumed that the translation to canonical
import paths is being done by the client of the importer.
We use a lookup func for two reasons: first, to support modules, and
second, to be able to use our information from "go list -json -export".
However, go/types does not canonicalize import paths before calling
ImportFrom. This is somewhat understandable; it doesn't know whether an
importer was created with a lookup func, and ImportFrom only requires
the input path to be canonicalized in that scenario. When the lookup
func is nil, the importer canonicalizes by itself via go/build.Import.
Before this change, the added crossbuild test would fail:
> garble build net/http
[stderr]
# vendor/golang.org/x/crypto/chacha20
typecheck error: /usr/lib/go/src/vendor/golang.org/x/crypto/chacha20/chacha_generic.go:10:2: could not import crypto/cipher (can't find import: "crypto/cipher")
# vendor/golang.org/x/text/secure/bidirule
typecheck error: /usr/lib/go/src/vendor/golang.org/x/text/secure/bidirule/bidirule.go:12:2: could not import errors (can't find import: "errors")
# vendor/golang.org/x/crypto/cryptobyte
typecheck error: /usr/lib/go/src/vendor/golang.org/x/crypto/cryptobyte/asn1.go:8:16: could not import encoding/asn1 (can't find import: "encoding/asn1")
# vendor/golang.org/x/text/unicode/norm
typecheck error: /usr/lib/go/src/vendor/golang.org/x/text/unicode/norm/composition.go:7:8: could not import unicode/utf8 (can't find import: "unicode/utf8")
This is because we'd fall back to importer.Default, which only knows how
to find packages in $GOROOT/pkg. Those are missing for cross-builds,
unsurprisingly, as those built archives end up in the build cache.
After this change, we properly support importing std-vendored packages,
so we can get rid of the importer.Default workaround. And, by extension,
cross-builds now work as well.
Note that, in the added test script, the full build of the binary fails,
as there seems to be some sort of linker problem:
> garble build
[stderr]
# test/main
d9rqJyxo.uoqIiDs5: relocation target runtime.os9A16A3 not defined
We leave that as a TODO for now, as this change is subtle enough as it
is.
In the worst case scenario, when GOPRIVATE isn't set at all, we would
run these three commands:
* "go env GOPRIVATE", to fetch GOPRIVATE itself
* "go list -m", for GOPRIVATE's fallback
* "go version", to check the version of Go being used
Now that we support Go 1.16 and later, all these three can be obtained
via "go env -json":
$ go env -json GOPRIVATE GOMOD GOVERSION
{
"GOMOD": "/home/mvdan/src/garble/go.mod",
"GOPRIVATE": "",
"GOVERSION": "go1.16.3"
}
Note that we don't get the module path directly, but we can use the
x/mod/modfile Go API to parse it from the GOMOD file cheaply.
Notably, this also simplifies our Go version checking logic, as now we
get just the version string without the "go version" prefix and
"GOOS/GOARCH" suffix we don't care about.
This makes our code a bit more maintainable and robust. When running a
short incremental build, we can also see a small speed-up, as saving two
"go" invocations can save a few milliseconds:
name old time/op new time/op delta
Build/Cache-8 168ms ± 0% 166ms ± 1% -1.26% (p=0.009 n=6+6)
name old bin-B new bin-B delta
Build/Cache-8 6.36M ± 0% 6.36M ± 0% +0.12% (p=0.002 n=6+6)
name old sys-time/op new sys-time/op delta
Build/Cache-8 222ms ± 2% 219ms ± 3% ~ (p=0.589 n=6+6)
name old user-time/op new user-time/op delta
Build/Cache-8 857ms ± 1% 846ms ± 1% -1.31% (p=0.041 n=6+6)
In ce2c45440a, we simplified the code a bit and removed one call to
obfuscatedTypesPackage.
Unfortunately, we introduced a regression; if an exported function is
linknamed to another symbol name, and it's called from an importer
package, we would have a build failure now:
> garble build
[stderr]
# test/main
ZiOACuw7.go:1: undefined: ODC0xN52.BaDqbhkj
This is because the imported package would not hash the original name,
via its ignoreObjects logic. And, since the importer package has no
access to that knowledge, it would hash the same name, and fail to find
it in the final build.
The regression happened because we used to have a types.Scope Lookup
that saved us in this scenario. Add the test, and re-add the Lookup,
this time only for this particular scenario with function names.
Thanks to Andrew LeFevre for reporting and describing the test case.
While at it, replace more uses of "garbled" to "obfuscated".
Mention two more bugfixes. Not the initial Go 1.17 support just yet,
because coincidentally it broke in the past 24h due to upstream changes.
Also tweak some of the other wording to be clearer.
Since it changes rapidly, especially during merge cycles, and we don't
want CI to surprisingly blow up in our faces from one day to another.
Pin this to a commit from yesterday which works, since some changes
merged today moved where the Go build version is recorded and broke
garble.
While at it, replace "git clone" with a wget of a source archive. This
is much, much faster, mainly because a tarball is significantly smaller.
We now download about 20MiB instead of over 350MiB.
One downside is that, without git, make.bash can't construct a devel
version on its own. For that reason, add a pretty basic manual version
via the VERSION file.
This means that we must not reject custom devel version strings. This is
a good thing anyway, because custom devel strings are already common
when building Go in custom ways. Those people tend to be advanced users,
such as CI, so fall back to assuming they know what they are doing and
don't error.
Plus, starting last week, devel versions in Go master now contain the
major Go version like in build tags, such as "go1.17-commit...", so we
will soon start relying on that instead of parsing dates:
$ go version
go version devel go1.17-a7e16abb22 Thu Apr 8 07:33:58 2021 +0000 linux/amd64
Just two minor tweaks were necessary to get "go test" to pass on:
go version devel go1.17-a25c584629 Tue Apr 6 04:48:09 2021 +0000 linux/amd64
Re-enable the CI for it, too. The config needed changing since the
set-env and add-path commands now use special files instead, due to some
security issues uncovered last winter.
It's possible that CI on master could suddenly break, if Go master
changes in some substantial way that requires more tweaks. If that turns
out to be an issue pretty often, we could always pin a specific git repo
commit and update it every few weeks.
Two bugs were remaining which made the build with -literals of std fail.
First, we were ignoring too many objects in constant expressions,
including type names. This resulted in type names declared in
dependencies which were incorrectly not obfuscated in the current
package:
# go/constant
O1ku7TCe.go:1: undefined: alzLJ5Fd.Word
b0ieEGVQ.go:1: undefined: alzLJ5Fd.Word
LEpgYKdb.go:4: undefined: alzLJ5Fd.Word
FkhHJCfm.go:1: undefined: alzLJ5Fd.Word
This edge case is easy to reproduce, so a test case is added to
literals.txt.
The second issue is trickier; in some packages like os/user, we would
get syntax errors because of comments printed out of place:
../tip/os/user/getgrouplist_unix.go:35:130: syntax error: unexpected newline, expecting comma or )
This is a similar kind of error that we tried to fix with e2f06cce94. In
particular, it's fixed by also setting CallExpr.Rparen in withPos. We
also add many other missing Pos fields for good measure, even though
we're not sure they help just yet.
Unfortunately, all my attempts to minimize this into a reproducible
failure have failed. We can't just copy the failing file from os/user,
as it only builds on some OSs. It seems like it was the perfect mix of
cgo (which adds line directive comments) plus unlucky positioning of
literals.
For that last reason, as well as for ensuring that -literals works well
with a wide variety of software, we add a build of all of std with
-literals when not testing with -short. This is akin to what we do in
goprivate.txt, but with the -literals flag. This does make "go test"
more expensive, but also more thorough.
Fixes#285, hopefully for good this time.
The regular obfuscation process simply modifies some simple nodes, such
as identifiers and strings. In those cases, we modify the nodes
in-place, meaning that their positions remain the same. This hasn't
caused any problems.
Literal obfuscation is trickier. Since we replace one expression with an
entirely different one, we use cursor.Replace. The new expression is
entirely made up on the spot, so it lacks position information.
This was causing problems. For example, in the added test input:
> garble -literals build
[stderr]
# test/main
dgcm4t6w.go:3: misplaced compiler directive
dgcm4t6w.go:4: misplaced compiler directive
dgcm4t6w.go:3: misplaced compiler directive
dgcm4t6w.go:6: misplaced compiler directive
dgcm4t6w.go:7: misplaced compiler directive
dgcm4t6w.go:3: misplaced compiler directive
dgcm4t6w.go:9: misplaced compiler directive
dgcm4t6w.go:3: misplaced compiler directive
dgcm4t6w.go:3: too many errors
The build errors are because we'd move the compiler directives, which
makes the compiler unhappy as they must be directly followed by a
function declaration.
The root cause there seems to be that, since the replacement nodes lack
position information, go/printer would try to estimate its printing
position by adding to the last known position. Since -literals adds
code, this would result in the printer position increasing rapidly, and
potentially printing directive comments earlier than needed.
For now, making the replacement nodes have the same position as the
original node seems to stop go/printer from making this mistake.
It's possible that this workaround won't be bulletproof forever, but it
works well for now, and I don't see a simpler workaround right now.
It would be possible to use fancier mechanisms like go/ast.CommentMap or
dave/dst, but those are a significant amount of added complexity as well.
Fixes#285.
This is easier to understand, since now the modification of the
*ast.File is all within a single chunk of code. We can also simplify
literals.Obfuscate to work on a single file, as transformGo runs in a
loop.
We also remove the "use receiver" TODOs, since the code is now in a
different package and it can't declare methods on a type here.
Walking the entire Go file again seems unnecessary, when we have an
astutil.Apply just next to it.
We could add to its "pre" function, but it's a bit easier to use the
"post" instead, which is empty.
Another advantage here is that astutil.Apply is more powerful than
ast.Inspect, which can come in handy in the future.
Reduces "go test -short" with a warm build cache from ~9s to ~4s. The
main offender was the use of "-a" in the "garble test" call; I think I
added that flag to rebuild all packages when debugging an error, and
forgot to remove it.
The sections are a bit more organized now. We add better docs for flags,
a new section on build speed, and we also remove the "withgo" helper as
it seems a bit over the top.
Fixes#275.
Just the exit status code is used. Printing an error would be confusing,
since we're printing the input too. Plus, the error is binary anyway -
either we changed something, or we did not.
One could make a case for only printing if we changed anything, and
showing an error otherwise. But that would mean buffering the entire
input, which could be problematic or slow when e.g. reversing large log
files.
Fix up a few TODOs, and simplify the way we handle comments.
We now add whitespace around inline /*line*/ directives, to ensure we
don't break programs. A test case is added too.
We now add line directives to call sites, not function declarations,
since those are what actually shows up in stack traces.
It's unclear if we care about any other lines inside functions at all.
This also fixes reversing with -literals, since that feature adds a
significant amount of code which shuffles line numbers around.
Finally, we extend the tests with types, methods, and anonymous
functions, and we make all of them work well.
Updates #5.
In 6898d61637, we switched from using action IDs from "go list
-toolexec=garble" to those from the original "go list". We still wanted
the obfuscation and hashing to change if the version of garble changes,
so we hashed that "original action ID" with garble's own content ID, and
called the new hash "garble action ID".
While working on a different patch, I noticed something weird: with the
new mechanism, adding or removing flags like -literals did not alter
those hashes, unlike the old method. This is because the old method used
ownContentID, which includes such bits of information, but the new
method does not.
Change that, and add a test that locks in the behavior we want. In
seed.txt, we check that a single function name gets hashed in particular
ways in different scenarios.
Note that we use a mix of "cmp" and "! bincmp", since the former has no
negated form.
While at it, the seed.txt test is revamped a bit. Now, we only run with
-literals once, as this test is mainly about -seed. We also declare seed
strings once, as environment variables, which makes it easier to track
what each step is doing.
In particular, the positions within function declarations, including the
positions of call sites to other functions.
Note that this isn't well tested just yet, particularly not with other
features like -literals. We can extend the tests and code over time.
This gets us the core basics.
The issue will be closed once the feature is documented for users, in a
follow-up PR.
Updates #5.
obfuscatedTypesPackage is now only useful for one scenario: a type which
is used with reflection in a dependency, so neither its name nor any of
its potential struct field names are obfuscated.
This is because we can only detect the use of reflection with go/ast,
which we don't have for dependencies.
As such, we only need obfuscatedTypesPackage in two places - when
considering to obfuscate a field or a type name.
There were two other calls to the function, which we remove.
The first was used for linkname directives. Those directives only work
for variables and functions, neither of which is affected by the
reflection detection.
The second was used for all identifiers, at the very end of the
transformGo inner func. It's entirely unnecessary right now, as it never
triggers anymore. It's possible it was necessary some time ago when we
still didn't obfuscate assembly functions.
While at it, improve some comments and add a few TODOs for edge cases
which do not have code coverage.
First, stop writing binaries into the current directory, which pollutes
the git clone.
Second, split the benchmark into two. The old benchmark always used the
build cache after the first iteration, meaning that we weren't really
measuring the cost of cold fresh builds.
The new benchmarks show a build with an always-warm cache, and one
without any cache.
Note that NoCache with the main package importing "fmt" took about 4s
wall time, which makes benchmarking too slow. For that reason, the new
bench-nocache program has no std dependencies other than runtime, which
already pulls in half a dozen dependencies we recompile at every
iteration. This reduces the wall time to 2s, which is bearable.
On the other hand, Cache is already fast, so we add a second and
slightly heavier dependency, net/http. The build still takes under 300ms
of wall time. This also helps the Cache benchmark imitate larger rebuilds
with a warm cache.
Longer term, both benchmarks will be useful, because we want both
scenarios to be as efficient as possible.
name time/op
Build/Cache-8 161ms ± 1%
Build/NoCache-8 1.21s ± 1%
name bin-B
Build/Cache-8 6.35M ± 0%
Build/NoCache-8 6.35M ± 0%
name sys-time/op
Build/Cache-8 218ms ± 7%
Build/NoCache-8 522ms ± 4%
name user-time/op
Build/Cache-8 825ms ± 1%
Build/NoCache-8 8.17s ± 1%
We had two nearly identical copies of the code to open a temp file with
CreateTemp, write to it, and handle closing properly. Unify them.
With the added docs this isn't exactly a net win, but we want the long
funcs such as transformCompile to be easy to follow, and this helps.
Mainly comments. "garble" refers to the tool, but the verb and adjective
is more intuitive as "obfuscate" and "obfuscated" instead of "garble"
and "garbled".
First, rename line_obfuscator.go to position.go. We obfuscate filenames,
not just line numbers, and "obfuscator" is a bit redundant.
Second, use "/*line :x*/" comments rather than the "//line :x" form, as
the former allows us to insert them in any position without adding
unnecessary newlines. This will be important for changing the position
of call sites, which will be important for "garble reverse".
Third, do not rely on go/ast to remove and add comments. Since they are
free-floating, we can very easily end up with misplaced comments,
especially as the literal obfuscator heavily modifies the AST.
The new method prints and re-parses the file, to ensure all node
positions are consistent with a buffer, buf1. Then, we copy the contents
into a new buffer, buf2, while inserting the comments that we need.
The new method also modifies line numbers at the very end of obfuscating
a Go file, instead of at the very beginning. That's going to be more
robust long-term, as we will also obfuscate line numbers for any
additions or modifications to the AST.
Fourth, detachedDirectives is unnecessary, as we can accomplish the same
with two simple prefix matches.
Finally, this means we can stop using detachedComments entirely, as
printFile already inserts the comments we need.
For #5.
Position information was obfuscated with math/rand manually, which meant
that the resulting positions were pretty small like "x.go:34", but they
were also very hard to reverse due to their short length and difficulty
to reproduce.
We now hash them with hashWith and the package's GarbleActionID:
"main.go:203" hashed with 933ad1c700755b7c3a9913c55cade1 to "mwu1xuNz.go"
The input to the hash is the base filename and the byte offset of the
declaration within the file, meaning that it's unique within a package.
The output filename is long enough to allow easy reversal.
The line number is always 1, since the information needed for reversing
is contained entirely within the filename. It doesn't really matter if
we encode data in the filename or line number, but it's easier for us to
use a string.
For #5.
Explain why we print the random seed on failure, and why we accept
base64 padding when we don't use it.
While at it, the flagOptions.Random field is unused, so remove it. It
also seems wrong for it to exist; the random value is only for the seed
flag, which already has a field of its own.
We were still printing files into a buffer, which got left behind from
when we used to store the obfuscated source in object files.
All that code is gone, so this buffer is now just wasting CPU cycles and
memory.
obfuscatedTypesPackage can't be outright removed, since we still do
require knowledge of what was obfuscated for one reason: types used with
reflection. Since we can only figure that out with go/ast and not just
go/types, we rely on the original compilation to tell us that
information.
IgnoreFuncBodies=true for typechecking the original source code would be
nice, as we would save time, but ultimately it doesn't work. When we
rename top-level declarations such as functions and types, we also need
to amend their references in func bodies. We depend on type information
for that.
Finally, we've been randomizing filenames for a while now. Randomizing
the order of the files doesn't seem to be useful.
Historically, it was impossible to rename those funcs as the
implementation was in assembly files, and we only transformed Go code.
Now that transformAsm exists, it's only about fifty lines to do some
very basic parsing and rewriting of assembly files.
This fixes the obfuscated builds of multiple std packages, including a
few dependencies of net/http, since they included assembly funcs which
called pure Go functions. Those pure Go functions had their names
obfuscated, breaking the call sites in assembly.
Fixes#258.
Fixes#261.
We no longer cache the obfuscated source code for -debugdir, so
obfSrcArchive is entirely useless at this point. We likely didn't notice
because the Go compiler doesn't realise it's unused.
symAbis is also unnecessary. It used to be necessary as we could only
collect the action ID from transformCompile in the second asm
invocation. Since we no longer need the temporary file dance with
transformCompile, we can simplify that code too.
With the many refactors building up to v0.1.0, we broke "garble test" as
we no longer dealt with test packages well.
Luckily, now that we can depend on TOOLEXEC_IMPORTPATH, we can support
the test command again, as we can always figure out what package we're
currently compiling, without having to track a "main" package.
Note that one major pitfall there is test packages, where
TOOLEXEC_IMPORTPATH does not agree with ImportPath from "go list -json".
However, we can still work around that with a bit of glue code, which is
also copiously documented.
The second change necessary is to consider test packages private
depending on whether their non-test package is private or not. This can
be done via the ForTest field in "go list -json".
The third change is to obfuscate "_testmain.go" files, which are the
code-generated main functions which actually run tests. We used to not
need to obfuscate them, since test function names are never obfuscated
and we used to not obfuscate import paths at compilation time. Now we do
rewrite import paths, so we must do that for "_testmain.go" too.
The fourth change is to re-enable test.txt, and expand it with more
sanity checks and edge cases.
Finally, document "garble test" again.
Fixes#241.
Our use of go-internal's gotooltest.Setup means that "go" is set up as a
top-level command in the scripts, so we can simply "go build" instead of
having to "exec go build". The result is practically the same, but the
scripts are simpler.
While at it, I had left an "exec cat" behind; remove it.
Daniel Martí
Andrew LeFevre
lu4p