Previously, we would split messages by line and call the logger
function for each line. We would hold a lock during this, to ensure
that multiple threads would not interleave their messages.
There are a few problems with this approach:
1) Using a lock is not efficient and is not fork safe
2) With APEX, there is one lock per instance of libbase, so we must
move the lock to a location where all instances can access it, or
perform the line splitting in a way that does not require the lock.
To solve these issues, we reimagine line splitting.
1) We move the lock out of the LogMessage::~LogMessage() and make it
the logger's responsibility to split lines, giving the logger the
option to lock or not.
2) We do not need any locks at all for StderrLogger.
Instead, we generate a single string that contains all of the lines
with their appropriate log header. A single write() call is used
to output this at once.
3) Logd handles log messages with newlines correctly, however it only
accepts up to a maximum size of log message. Therefore we
separate the incoming log message into chunks, delimited by new
lines, up to that maximum size, and send each of those to logd.
Note that this is the strategy used in
android.util.Log.printlns().
This should solve a majority of use cases, since the maximum size
that logd accepts is nearly 4K, while remaining lock free.
If interleaving messages absolutely must be avoided, a lock can
still be used given 1) above.
Bug: 65062446
Bug: 153824050
Test: logging, particularly multi-line stack traces, show correctly
Test: existing and new unit tests
Change-Id: Id0cb5669bee7f912da1e17f7010f0ee4c93be1e3
(cherry picked from commit 36d31c530d)
Some recent changes can have these logging functions potentially set
errno. This change places android::base::ErrnoRestorer at the entry
point of the public functions where we want to guarantee errno is
restored to ensure this will not happen again.
Test: build
Change-Id: Iab4170ab16b9c7301474a509ee42d38b370b91a4
See the previous commit moving SetLogger and SetAborter to liblog for
motivation.
This creates more harmony between the two mechanisms in libbase and
liblog for checking loggability.
Currently:
1) libbase filters all messages based on its minimum log priority. For
example, if minimum log priority in libbase remained at its
default, but a tag was specifically opted into DEBUG logs via
log.tag.<tag>, libbase would not print this log.
2) liblog ignores libbase's minimum log priority. For example if a
process called SetMinimumLogPriority(WARNING) but used a library
that logged via liblog's ALOGI macro, that log would still be
printed even though the process intends on filtering out those INFO
messages.
With this change:
1) If both a minimum log priority and a priority through log.tag.<tag>
are set, then the lower of the two values is used.
2) If only one or the other is set, then that value is used. This
fixes the two issues described above.
3) If neither of these values are set, then the default of using INFO
is unchanged.
Bug: 116329414
Bug: 119867234
Test: libbase and liblog minimum log priority tests
Change-Id: Icb49b30b9d93bf797470e23730ae9e537931bb6c
libbase is copied into each APEX module which requires it, meaning
that there may be multiple instances of libbase running within a
single process with their own copy of libbase's globals. This means
that SetLogger() and SetAborter() will only impact logs from the
instance of libbase that calls it. This change moves this state to
liblog, since it will only ever have one instance in a single
process.
One major side-effect here is that now both ALOGE style and LOG(...)
style logs will be handled through the same logger function. For
example, a logger specified through libbase's SetLogger() will now see
logs sent to liblog through ALOGE(). This is intended behavior.
A second side-effect is that libbase's stderr logger is used for all
host logging now. It's simply a better logging default than the
fake_log_device logger in liblog currently and makes ALOGE and
LOG(...) logs on host follow the same format.
Bug: 119867234
Test: libbase and liblog unit tests; logging works
Change-Id: Ib52cbfb4e43749e50910ed19a993dffae19ace86
This is for projects built with GCC that import parts of Android that,
despite not having Android-specific dependencies, still end up
depending on logging.h.
Also removes outdated notes.
Change-Id: I5a47b302bcaeeb935592d8fc7ad2fe5068d226c3
There were only two users of these and both of them were better off
setting up a default logger.
These macros are not particularly useful as it's not useful for a
single program to write to both the MAIN and SYSTEM logs. In fact,
the opposite of these macros would be more beneficial: having more
programs write to only the MAIN or only the SYSTEM buffer, so getting
rid of these macros removes a temptation for bad behavior.
Users that absolutely need to do this behavior can still use the
liblog macros or functions, but that should be an extreme edge case,
such as the few programs that write to the CRASH buffer and does not
need to exist in libbase.
Bug: 119867234
Test: build
Change-Id: I23369c3b48ed636b617220cab47f77fdd5559763
liblog will soon be required for all of libbase's logging. This
change proactively requires liblog in all configurations instead of
just Android.
Bug: 119867234
Test: build
Change-Id: I696162fbebc78d4ef23c6032412101ac51d397a4
android-base:
* Add NOLINT for expanding namespace std for std::string* ostream
overload
libdm:
* Fix missing parentesis around macro parameters
init:
* Fix missing CLOEXEC usage and add NOLINT for the intended
usages.
* Fix missing parentesis around macro parameters
* Fix erase() / remove_if() idiom
* Correctly specific unsigned char when intended
* 'namespace flags' should be signed, since 'flags' it signed for
clone()
* Add clear to property restore vector<string> to empty after move
* Explicit comparison against 0 for strcmp
Test: build
Change-Id: I8c31dafda2c43ebc5aa50124cbbd6e23ed2c4101
Allow the default tag (the program name) to be overwritten.
Bug: 34867873
Test: m
Test: logging_test
Test: manual
Change-Id: I4ef32bad413a7cc82e46ce16a2f26212925964b1
If LOG_TAG was not defined, falling back to a default
behaviour (using binary name).
Bug: 35361699
Test: manual
Change-Id: I209a6ebaf0df882f98642f6d1831766cb296c951
This reverts commit 4ef5011a7b.
Reason for revert: Breaks an internal mocking case that cannot be changed.
Test: m
Change-Id: I83f9338bde02eb2b45b3e52b66ef78490ddbeeda
In most reasonable cases, this is actually a bug. So delete the
operator overload and let the compiler complain.
Test: m
Change-Id: I7d66ec2f33cc46588b6f549876241871f19ce995
::android::base::GetMinimumLogSeverity() is defined externally, so the
static analyzer was allowed to assume that we continue executing after a
LOG(FATAL).
I manually audited all of the code I have access to, and the only
"change the minimum log severity" statements I can see keep FATAL
enabled (...and continuing after a FATAL is highly sketchy to me
anyway).
(I'm sure I tested this at some point in making the previous patch. I
probably broke it in a refactor before sending it out for review; my
bad. :) )
Bug: None
Test: m without the static-analyzer builds; m with it yields fewer
false positives.
Change-Id: I216cd2034e1daa8d6f6c5e776f64b4cce88bb938
This helps us have less false-positives.
We do this instead of `#undef NDEBUG`, since undefing NDEBUG actually
gave us more false-positives (...and build breakages) than simply
leaving it defined.
Bug: None
Test: Ran the static analyzer across internal master. 213 fewer warnings
(15 Medium, 2 Low, the remainder are 'Analyzer'). All of the dropped
warnings I audited were false-positives. It adds ~3 Tidy warnings.
Change-Id: Ibedab60ca7e9d2b0772896222b83d2e7ab064afe
Two small changes in one:
- `foo || for (;;abort()) bar();` isn't valid C or C++, since for is a
statement. We need an expression instead.
- we'll now treat everything after LOG(foo) as unreachable in the
static analyzer, as long as we can prove at compile-time that
foo == FATAL.
The impact of this, running across internal master, is that we see ~50
fewer medium/high-severity false positives from clang-tidy. We see 15
new complaints about unreachable code (at the "tidy" severity), but all
of them are harmless AFAICT (e.g.
switch (foo) {
// ...
default:
LOG(FATAL) << "Unhandled case!"; // or CHECK(false);
break; // clang-tidy: unreachable break.
})
(Some of the macros were forcibly formatted by the clang-format hook)
Bug: None
Test: Ran
`DEFAULT_GLOBAL_TIDY_CHECKS=clang-analyzer*,-clang-analyzer-alpha* m`;
stared at warn.py output.
Change-Id: Ie984eda0481afad4274b9def7c61ba777cfa289a
Partially reverts commit 436f5a031f.
Remove the variants taking a fully qualified LogSeverity. Instead
use a lambda with "using" statements to translate both qualified
and unqualified names into valid expressions.
Compile-time regression was measured as 0.1s for a thousand LOG
statements on a z840.
Update tests.
Bug: 31338270
Test: m
Test: mmma system/core/base && $ANDROID_HOST_OUT/nativetest64/libbase_test/libbase_test64
Change-Id: I36fdf30a9d535b19543307b85d1b3c19a97f20dd
Add WOULD_LOG to determine whether a given severity would be logged.
Add LOG_STREAM to have direct access to a logging stream.
Add LOG_S variants that take a fully qualified severity. This allows
complex expressions as parameters, e.g., ternaries for conditional
severity levels.
Add tests.
Bug: 31338270
Test: m
Test: mmma system/core/base && $ANDROID_HOST_OUT/nativetest64/libbase_test/libbase_test64
Change-Id: I242b960594e68caff6db9cd8aaa4ce8aaf90474c
Add the INTERNAL_FATAL level. It will print 'F' like FATAL, but
does not abort after logging.
Add a test to logging_test.
Bug: 31338270
Test: m
Test: mmma system/core/base && adb sync && adb shell /data/nativetest/libbase_test/libbase_test32
Change-Id: Idf74c08e8516881efccaefc58fa3f41d57e56396
Assume C++11 support. logging.h already used constexpr before,
macros.h now also does no longer check the language level.
Use constexpr for the logging evaluator. This allows the usage
of CHECK and co in other constexpr expressions in C++14.
Test: m checkbuild (N9)
Change-Id: Ifdffa074271fff1f9949c48829a185800ec5e524
I've been meaning to do this for a while, and it came up on the bug below
that there have been conflicts with similar-named files in <base/*.h>,
so let's rule out one possible explanation.
Bug: http://b/27804373
Change-Id: I69e5d52b6260c573c308513420aee0e281426bd4
As logging macros uses `if xxx else yyy` style, it is reported as an
error when DCHECK() is compiled with -Wdangling-else option. Because
after preprocess, DCHECK(x) becomes:
if (EnableDChecks)
if (x)
;
else
LogMessage(FATAL) << yyy;
This CL avoids compilation error by replacing `if xxx else yyy`
with `xxx && yyy` or `!(xxx) || yyy`.
Bug: 26962895
Change-Id: Ib0bf242cc04a238ec31a1ab66b53fc8a5b5ed28f
