* changes:
adbd: avoid compiling more code in the daemon.
adb: don't run all of the tests again over TCP in coverage.
adbd: respect ADB_TRACE on host adbd.
adb: mark kMaxProcessNameLength as constexpr.
When parsing an RTM_NEWADDR or RTM_DELADDR, ifaddr is always
present (unless the message is invalid). So ifaddr->ifa_flags is
always known before any attributes are parsed.
Bug: 155005831
Test: atest NetworkStackNextIntegrationTests:IpClientIntegrationTest continues to apss
Change-Id: Id1998faccca7d81c1b7f3e85e4912aa22919e94a
* changes:
libnetutils/packet.c - create socket with close-on-exec
libnetutils/packet.c - fix a raw socket reception race
libnetutils/packet.c - fix a socket leak on bind error
In RTM_NEWADDR messages, the first 8 flags are reported in the
ifa_flags field in struct ifaddrmsg, but flags above 0x80 are
reported in the IFA_FLAGS attribute. NetlinkEvent currently only
looks at ifa_flags, so it cannot see higher-value flags such as
IFA_F_STABLE_PRIVACY. Fix this by parsing the IFA_FLAGS
attribute.
Bug: 155005831
Test: makes new test in aosp/1295495 pass
Change-Id: I723f1106cbcea2186fc6452305942a0f8301fd2a
A socket created bound to a specific protocol can receive packets before
it's bound to a specific interface/mac.
It's best to rely on the bind().
This replicates the way packet sockets are created in external/android-clat.
See: ring.c ring_create() and clatd.c configure_packet_socket()
Test: builds, atest
Bug: 155297277
Signed-off-by: Maciej Żenczykowski <maze@google.com>
Change-Id: I569d2b81a396f352ab5d1b71bd5ad0d96217aab1
and clean up the code a little bit while at it.
Test: builds, atest
Bug: 155297277
Signed-off-by: Maciej Żenczykowski <maze@google.com>
Change-Id: I01a10e36f852cde1b93a91f95b51294a434885ab
The exit of init panics the system *after* process context (mm, stack,
...etc.) are recycled, according to Linux kernel's 'do_exit'
implementation. To preserve most init process context for debugging,
triggers the panic via proc-sysrq explicitly.
Note: after this change, there will be no "Attempt to kill init" panic
when androidboot.init_fatal_panic is set.
Test: Insert data abort fault in init, the full process context is
preserved in memory dump captured after panic.
Bug: 155940351
Change-Id: I3393bd00f99b8cb432cfa19a105b7d636b411764
It's impractical to test the contents of the stack trace, but we
should at least test that *a* stack trace is present, which would
have caught the bug fixed by r.android.com/1306754 .
Bug: 135772972
Change-Id: Ic5e0b997caa53c7eeec4e5185df5c043c9d4fe3d
The code was using an off64_t but calling mmap. This caused the code
to abort on 32 bit.
Add a unit test that would abort on the previous version.
Bug: 155662887
Test: New unit test passes.
Change-Id: I7a6efbc0d4227403c3d08a08deea56f239382157
LogListener uses the libsysutils SocketListener class to read incoming
log messages, which means that it poll()'s to wait for new data then
calls recvmsg() when this data is available. That is unnecessary,
since this is a blocking socket, it already has a dedicated thread,
and we're not using any other SocketListner features, so a tight loop
of recvmsg() is sufficient.
Below are the stats from simpleperf during a framework restart. To
get an apples to apples comparison, these stats calculate the overhead
of receiving the message as 1 - (<number of perf events dispatching
the logs in logd>/<total number of perf events for the listener thread>).
With SocketListener / poll()
378563153 total events for the listener thread - SocketListener::runListener()
245799364 events for reading and dispatching logs + recvmsg() - LogListener::onDataAvailable()
67622952 events for recvmsg() inside LogListener::onDataAvailable()
=> (378563153 - 245799364) = 132763789 SocketListener + poll() overhead
=> (132763789 + 67622952) = 200386741 total overhead to for recvmsg() + poll() + SocketListener
=> (200386741 / 378563153) = 52.9% overhead to read each log message.
Without SocketListener / poll()
324802401 total events for the new listener thread - LogListener::ThreadFunction()
133208105 events for recvmsg()
=> (133208105 / 324802401) = 41% overhead to read each log message.
This shows a 22.5% = (41 - 52.9)/52.9 reduction in cost spent for each log message.
Test: logging unit tests
Test: simpleperf
Change-Id: I1f93b7a06fac6adce8e64e727bf6d8c5935c77c9
This is the last isMonotonic() user and can go away. This timestamp
is set in the kernel source in either the audit_get_stamp() or
__audit_syscall_entry() functions. In both cases, the value used is
from ktime_get_coarse_real_ts64(), which is a realtime timestamp.
Test: audit messages show in the log correctly.
Change-Id: Ife6c09dd97fccdfc7a8f07ee63161079ae2eccc4
This has been around for ~5 years but there has been roughly no
adoption, so remove this as we clean up the logging code.
Future efforts may track the monotonic timestamp in all cases.
Test: logging unit tests
Change-Id: I55ed565669f923988e741f6b384141bba893630d
libsnapshot used to assume CoW images are always mapped
with device mapper, and call GetDeviceString directly.
However, in tests, it is possible that it is mapped as loop
devices. Fix the incorrect assumption and calls
IImageManager::GetMappedImageDevice instead.
Test: libsnapshot_fuzzer with corpus
Bug: 154633114
Change-Id: I1ec39744480e514ae7a7902643ed863e0ca0a340
libfiemap functionality is extensively tested via libsnapshot, gsid, and
fiemap_writer_test. However those tests rely on the configuration of the
device, and libfiemap has fairly different behavior based on whether FBE
or metadata encryption is in use. (The former will use loop devices, the
latter uses dm-linear).
To expand coverage, we added tests with very complicated nesting,
creating fiemaps on a rw partition mounted off a device backed by a
fiemap. Unsurprisingly, this was brittle to begin with, and now
consistently fails due to the intermediate partition not having
sepolicy.
Given the preference toward metadata encryption, and the fact that the
FBE mode will still be properly tested via VTS, let's just remove these
tests.
Bug: 148874852
Test: fiemap_image_test
Change-Id: I7be7d44dea81508505c050ed995b3b78086ad35f
We were getting stale service ip addresses because we weren't
destroying the sdref correctly.
Also, we were leaking the ResolvedServices when removing it from the
ServiceRegistry. Converted them to smart pointers to fix that.
Bug: 153343580
Test: test_adb.py
Change-Id: Ib7c1dbf54937d4ac6d9885cb5f7289bef616d12e
1) We can use real member functions with std::thread and
std::function, so use those instead of the 'me' pointer.
2) Don't expose member variables directly.
3) Rename and document member variables, since all of their references
are being touched anyway.
Test: logging unit tests
Change-Id: I9a357a3ea8691433d58687c95356b984b83e9c36
LogTimes has evolved from being simply a store of the last timestamp
that each reader has read to being a class representing an individual
reader thread, including the thread function, so name it
appropriately.
Test: logging unit tests
Change-Id: I6914824376a6ff1f7509e657fa4dc044ead62954
This code has evolved to the point that it LogReader::notifyNewLog()
executes FlushCommand on every socket. FlushCommand then iterates
over all log reader threads in LogTimes and flushes triggers them to
flush logs if they match the client. This can be simplified to simply
looping over all of the reads in LogTimes.
Code readability was the primary motivation here, but note that 64% of
LogReader::notifyNewLog() was spent looping over the sockets, and an
additional 78% of FlushCommand::runSocketCommand() was spent
repeatedly locking and unlocking the LogTimes lock.
Overall, this reduces the cost of LogReader::notifyNewLog() in
LogListener::onDataAvailable() from 5.91% to 2.93%. This is the
critical path for handling incoming log messages, so it's a
non-trivial savings.
Test: logging unit tests
Test: unprivileged clients still cannot view privileged logs
Change-Id: Ic7620978a6c23e5e2cb179ff0c42b7cea52fc011
