The full command line run inside sbox can be very long, and printing it
after the errors printed by the failing command can hide the error
messages. Buffer the output of the command and print the failing command
line before the output if it fails.
Bug: 185516277
Test: m out/soong/.intermediates/frameworks/base/system-api-stubs-docs-non-updatable/android_common/metalava/api_lint.timestamp with lint error
Change-Id: I893f3dd01f1baf195e182111c5c49e92eb82f3b0
Error messages printed by failing commands may reference output files
that were created by the command, for example printing a command line
to copy and paste to update a baseline file. Copy output files in the
sandbox to their final locations, ignoring missing files, so that the
messages are valid.
Bug: 185516277
Test: m out/soong/.intermediates/frameworks/base/system-api-stubs-docs-non-updatable/android_common/metalava/api_lint.timestamp with lint error
Change-Id: I604a11c9b54e409ca5bc5c016cd04b3133f74a60
Sbox preserves the permissions of input files when copying them into the
sandbox. A read-only file copied into the sandbox multiple times causes
a permission denied error on the second write. Building in Bazel results
in more read-only files, which triggers the issue on existing sbox rules
with duplicate input files. Remove the destination file when copying if
it exists.
Bug: 184113103
Test: m USE_BAZEL=true
Change-Id: I7edf92d82b766100e3cbbd90d22428269d7d0167
The current implementation causes inputs listed in an rsp file used with
sbox to be duplicated 3 times in the build.ninja file; once as a
dependency of the rule, once in the rspfile_content field of the rule
with the paths rewritten to be relative to the sandbox, and once in the
rule to write the sbox manifest. When RBE is enabled it also gets a
fourth copy in the list of files to be treated as inputs by rewrapper.
Reduce this to a single copy by using "$in" for the rspfile_content so
that the files only have to be listed in the input dependencies of the
rule, and then add support to sbox to rewrite the rsp file while copying
it into the sandbox, and pass it to rewrapper as well.
Test: m lint-check
Change-Id: I3f46f61119508d39a8bb231c99fc130153fb6f04
When running commands in chdir mode sbox needs to translate the
directory placeholders to paths relative to the sandbox directory
instead of relative to the top of the source tree.
Also translate relative PATH entries into absolute paths so they
are still valid when the current directory is changed.
Bug: 182612695
Test: manual
Change-Id: Idcbe20466888909d423d62788bc9c35f4e03b398
Add a boolean flag to the sbox proto to request the executable bit
to be set after copying a file. This will be used for sandboxing
tools when copying a file in from the source tree that would
normally get its executable bit set during installation.
Bug: 124313442
Test: sbox_test.go
Change-Id: Ie2c197bb5183ffc1bf63fd6effd175143cd324d4
The sbox tool is used to wrap genrule actions, and run them in a
temporary directory. Afterwards, the outputs are moved into their
proper location in the output tree. However, some tools embed the file
name of the output file (as passed to the tool) into the output file.
For example, the perfetto code generator script uses the output file
name to generate a C header guard when it generates
perfetto_src_base_version_gen_h/gen/perfetto_version.gen.h.
When using remote execution / remote caching, these genrule actions are
run locally whenever the output file doesn't exist, effectively making
them unique (across time and between users).
They then cause cache misses on all actions depending on these output
files as well as on transitive actions. In the above example, this
causes libperfetto.so to differ, which then causes all actions
depending on libperfetto.so to be rerun unnecessarily.
As a fix, this commit changes the sbox tool to use the sha1 hash of the
manifest path. The tool already seems to assume that there are no
concurrent runs of the same tool with the same manifest (otherwise
it would fail flakily depending on exact timing). It seems therefore
safe to use a temporary path deterministically derived from the
manifest path.
Test: ran multiple builds w/ remote execution; observe proper caching
Change-Id: I5b73ffd3b7f85cbb0336dfa1675de7ac0e2fd1a8
The hardlink optimization attempts to save I/O by hardlinking input
files into the sandbox instead of copying them, but it never works on
a Linux build due to nsjail always making out and the source tree into
separate bind-mounted filesystems. If the optimization did work it
would actually cause build failures when it hardlinked the relative
bionic/libc/fs_config_generator.py symlink into the sandbox directory
without the target. Just remove the optimization for now since it
never works.
Test: m checkbuild
Change-Id: I96e8c0c145e7c99958639594edf8a93b69ae90eb
This relands I3b918a6643cea77199fd39577ef71e34cdeacdb1 with a fix
to create the directory for the output depfile if doesn't exist.
In preparation for more complicated sandboxing that copies tools
and/or inputs into the sandbox directory, make sbox use a textproto
input that describes the commands to be run and the files to copy
in or out of the sandbox.
Bug: 124313442
Test: m checkbuild
Test: rule_builder_test.go
Test: genrule_test.go
Change-Id: I8af00c8c0b25a92f55a5032fcb525715ae8297c2
In preparation for more complicated sandboxing that copies tools
and/or inputs into the sandbox directory, make sbox use a textproto
input that describes the commands to be run and the files to copy
in or out of the sandbox.
Bug: 124313442
Test: m checkbuild
Test: rule_builder_test.go
Test: genrule_test.go
Change-Id: I3b918a6643cea77199fd39577ef71e34cdeacdb1
This change automates one of the best practices for
custom tools. The best practice says "Prefer to list
input files on the command line, otherwise we may not
know to re-run your command when a new input file is
added." [1]
Normally you'd reference these inputs with $(in) or
one of the forms of the $(location*) substituions on
the command line for the custom tool. However, if the
custom tool does not accept the list of inputs on the
command line, the build system could fail to re-run
the custom tool if the list changes.
This change adds a hash of the list of input names to
the sbox command that wraps the custom tool. If the
list of inputs change, the hash will change, and
therefore the sbox command will change, causing ninja
to re-run the custom tool. The hash is visible to
(but ignored by) the sbox command, and hidden from
your custom tool.
[1] https://android.googlesource.com/platform/build/soong/+/refs/heads/master/docs/best_practices.md#custom-build-tools
Test: TestGenruleHashInputs
Bug: 149397658
Change-Id: I18b547ea3c4296ee15bd6150a4778a8f376d80b7
sbox will generate a random directory for the output root, and most
tools will encode that directory name in the output target of the
depfile.
So embed the library from dep_fixer into sbox so that it can rewrite the
output filename to a static (reproducible) value. Ninja doesn't care
what that value is, so it's just "outputfile".
Also fix up rule_builder to actually tell sbox about the depfile.
Test: mmma system/iorap; check the contents of:
out/soong/.intermediates/system/iorap/libiorap-binder/android_arm_armv7-a-neon_core_static/gen/aidl/system/iorap/binder/com/google/android/startop/iorap/IIorap.cpp.d
Change-Id: I3640a2e8b0c034f143a35e398a8418a6d621b265
This essentially allows you to declare that everything in a directory
will be created by the rule, and we'll ensure that your command actually
writes out all of the claimed outputs, and remove any other files that
previously existed in that directory.
Test: built-in tests
Change-Id: I990dce2b3a0d89ebd2736ac1a0cadfb5864c6e73
Some tools used by genrules set the timestamp of the generated file
to an old value. For example, when using tar to extract files it
will set the timestamp of the extracted file to the value stored in
the tar file. This can cause unnecessary rebuilds due to output files
that are older than their inputs. Touch output files in sbox before
moving them into the output location.
Test: m checkbuild && m checkbuild
Change-Id: Idb0e71d2bf0ac3618338371132cc71be74bf9d9d
The path to the output directory may be arbitrarily long, use
__SBOX_OUT_DIR__ in the list of output files passed to sbox
to avoid expanding it multiple times in the command line.
Fixes:
ninja: fatal: posix_spawn: Argument list too long
09:40:14 ninja failed with: exit status 1
when building libchrome with a long OUT or OUT_DIR_COMMON_BASE.
Bug: 73726635
Test: m checkbuild
Change-Id: I59024b2164287c8e531711afd9273b692ce9c28a
which should make debugging faster.
Bug: 66921582
Test: ./out/soong/host/linux-x86/bin/sbox --output-root outs --sandbox-path out/.temp/sbox-work -c "cd __SBOX_OUT_DIR__ && mkdir asubdir && mkdir b && touch asubdir/child a b c d e f g h i j k l m n o p" outs/a outs/b outs/z # and observe the below output:
mismatch between declared and actual outputs
in sbox command(cd out/.temp/sbox-work/sbox343858828 && mkdir asubdir && mkdir b && touch asubdir/child a b c d e f g h i j k l m n o p)
in sandbox out/.temp/sbox-work/sbox343858828,
failed to create 2 files:
b: not a file
z: does not exist
did create 16 files:
a
asubdir/child
c
d
e
f
g
h
i
j
...6 more
Change-Id: I75e37834c44d4279dec874701d67ce8bb01b872c
Also removeTempDir -> !keepOutDir
Also shorten the help comment
Bug: 68336760
Test: rm out -rf && m -j
Test: sbox -c # and observe that the output does not have a stacktrace
Change-Id: I6fbb385132128ab1273c8398ff2425191049751e
Most notably, the sandbox depfile path should be passed into the tool.
Bug: 68336760
Test: m -j checkbuild
Change-Id: I22f944a3f57d613fda26de0ea777a915cafcd020
Cleanup of some failure messages
Also, this leaves the temp directory untouched if a declared output was not created
Bug: 35562758
Test: make
Change-Id: I8ef1315af80eb327752501f12a331dbdf52ba3e9
This causes Soong to put the outputs of each genrule into a temporary
location and copy the declared outputs back to the output directory.
This gets the process closer to having an actual sandbox.
Bug: 35562758
Test: make
Change-Id: I8048fbf1a3899a86fb99d71b60669b6633b07b3e
Colin Cross