Add unit tests for the native API to log atoms to statsd.
(This CL is being cherrypicked into AOSP in order to prevent automerger
issues with the rename from libstatssocket to libstatspush.)
Test: bit libstatssocket_test:*
Bug: 145231901
Change-Id: If427c17319787200260cbe3b71075ca556c9a82b
Merged-In: If427c17319787200260cbe3b71075ca556c9a82b
Now, the number of elements in each simulated stats_event is correctly
sampled from a uniform distribution.
Rerunning the benchmarking tests indicates that truncating the buffer
increases the number of cycles require to log an event by only 1.6%.
Test: m libstatssocket_benchmark
Test: bit libstatssocket_benchmark:*
Change-Id: I8d5c1734861038beec8eeb7ce1f84a24bde25a75
We create a .map.txt file that lists all the stable entry points into
libstatssocket. This should allow other APEXes to link to libstatssocket
without having to copy the library within the APEX.
Test: m -j libstatssocket
Bug: 146377784
Change-Id: I9f77a0c380b6884d9ca60807a8974380420cfe0a
This cleanup makes it clear that there are only two acceptable ways of
logging to the socket: either through libstatspush_compat or
libstatssocket (which will soon be renamed to libstatspush).
Test: m -j
Change-Id: I467b08bc300c2ef394b22422c6acf55a978a380a
Merged-In: I6a49b4995b56df462e1be901ea5e54acfe2e8bac
We now truncate the buffer to the appropriate length when clients call
stats_event_build().
Benchmarking tests indicate that truncating the buffer to the
appropriate length increases the cost clients pay to write to the socket
by 2%. This is negligible enough that I decided to truncate the buffer
for both pushed and pulled atoms in order to simplify the API.
Test: m libstatssocket
Test: bit libstatssocket_benchmark:*
Bug: 144126231
Change-Id: I35dec748ff87c0821d0d06779a406997e6e64966
Merged-In: Ife976bb383ecff8de5064730692a95e2a3a82c9d
StatsEventCompat is a wrapper library that toggles between the old
logging scheme and the new logging scheme. It was designed particularly
for the DNS Resolver module and should not be used by others.
We will merge libstatspush_compat and libstatssocket_q in a future CL.
Test: m libstatspush_compat
Test: bit libstatspush_compat_test:* (passes on Q and R)
Bug: 145534143
Change-Id: Ib355031f3573101ea90bd8694861fbfc33b0c788
Merged-In: Idf35ccb6669798166475f08b2fbab40534b5db19
Move iovec* construction and noteLogLoss from stats_event_list to
stats_buffer_writer.h.
Expose write_buffer_to_statsd that takes in a byte array and writes
it to the statsd socket. Currently exposed for StatsLog.write JNI call.
This change also allows getting rid of stats_event_list once all clients
have migrated to stats_event
Bug: 145619049
Test: m -j libstatssocket
Change-Id: I73f914985421fc1a4f094982c826bf364604b788
Merged-In: I0048e392c2f5039eb70dacf4e91a43d3f32e8749
Simplifications/changes:
(1) We remove the bufPos field and track the end of the buffer using
only the size field.
(2) We rename put_<type>() functions to append_<type>() to signify that the
value will be placed at the end of the buffer. We also move the
increment of event->size to within the append_<type>() functions;
this improves readability of the write_<type>() functions.
(3) We immediately write the timestamp and atom id to the buffer in
order to simplify stats_event_build().
(4) We never check for null pointers; checking for null pointers
delays errors and obfuscates the root problem.
(5) We change the the annotationId and numPairs parameters to be
uint8_t's. This helps signify to clients that these values must fit
in 1 byte.
(6) Clients no longer have to pass in the length of strings. Instead, we
expect them to pass in null-terminated strings, and we will
calculate the length outselves using strnlen.
Test: m -j libstatssocket
Change-Id: I6d192768876a23d7016173bcdaf59f8b7e92b182
Merged-In: I0173f8bc76ef25118379dce5d2481f5f7a9b7519
Encoding of key-value pairs follows strategy described in
go/statsd-socket-doc.
Fixed bug where timestamp and atom id were not contributing to the
number of fields in the buffer.
Note: Similar to ag/9603344, which was abandoned because of git rebasing
issues.
Test: m -j libstatssocket compiles
Bug: 143079361
Change-Id: I15abdf13ad8debf83ae1f6c372676db4fe390289
Merged-In: I8d23ddf2dbfee518ad970894ac09ba3ad5b5b4ad
This allows the parsing code within statsd to access the type ids.
Also, we move stats_event.h to the include directory for consistency.
Test: m -j libstatssocket
Change-Id: I7030416872677f3a523c2baf7bf6bcb4ca0ae6a4
Merged-In: I1db7fc9e9d0efd2225612353a8fc49bb13220ff0
Cherry pick to aosp.
stats_event_list changes for creating a generic api for statsd mainline
logging.
Test: builds
Test: existing logs continue to flow
Change-Id: I4163edc80ffcca61db7ab2a1e14ae8c14aab2347
Merged-In: I4163edc80ffcca61db7ab2a1e14ae8c14aab2347
(cherry picked from commit 4c2f079dad)
In addition to the dgram queue size, the sender buffer size also matters.
Set it to a higher value (default is 224KB).
Test: manual
Change-Id: Ib610c1fa1eaf3c5e36f51854a2b0720f387dc1b8
stats_event_list changes for creating a generic api for statsd mainline
logging.
Test: builds
Test: existing logs continue to flow
Change-Id: I4163edc80ffcca61db7ab2a1e14ae8c14aab2347
readv() isn't used by anyone, writev() has one easily replaced user.
uio.h can be left as a private header for windows compatibility with
struct iovec.
Test: build
Change-Id: I33d4c6bdee6fd818271f78ae06abdd2aa09430f2
The existing APIs will truncate the byte array in case of '\0' in between.
Test: test_drive with atom 103
Change-Id: I104d35b330c6b52e2fdce1f5d4305dca08228f1b
Previous submission of this change, aosp/790068, caused a failure in
sdk_mac for CLOCK_REALTIME and <endian.h> not being available.
Bug: 110537511
Test: compiles without failures and verified the correct metric
Change-Id: Iba1dc920ad82e88a4bcdd2feaee9a06202a440c2
+ Record all types of failures which lead to log loss
+ Only record eventual failure after the retries to get a accurate log loss count.
Bug: 80538532
Test: manually tested
Change-Id: I23a55e62e140bf22ae5aa9e6c40539d51149cd66
Merged-In: I23a55e62e140bf22ae5aa9e6c40539d51149cd66
(cherry picked from commit ad3a40bf76)
The original liblog code has this additional check because there may be multiple writers registered,
there is a for-loop and the retVal is set to be the FIRST error. Statsd doesn't need it because it
only has one writer.
Test: locally tested.
Change-Id: I03f0015d0c29eba19eb2b5ea145380cb98fb2509
So that lmkd can build on PDK.
Bug: 79349329
Test: builds locally
Merged-In: I981e6ef9f9769b873640e5f169a9495ccea2f25c
Change-Id: I981e6ef9f9769b873640e5f169a9495ccea2f25c
(cherry picked from commit b13a102c0a)
