The code previously assumed that write() took a non-zero number of clock
ticks. This is mostly true, but can't be guaranteed. Add a check to
avoid crashing if it's 0.
Bug: 233008287
Test: Compile
Change-Id: Idec0052f534dc4abb81a414ca76f6c11f4a4dded
SharedRefBase objects can only be created via
SharedRefBase::make, not std::make_shared. Otherwise
there'll be a double-ownership problem.
Test: TH
Fixes: 225448524
Change-Id: I856677626cb146eb52a5b4a4562148e899d5a57b
proto_loaded is not thread safe, so we must protect it with
a mutex proto_lock.
Signed-off-by: hanhongli@360.cn
Signed-off-by: gongguang@360.cn
Signed-off-by: Mark Salyzyn <salyzyn@google.com>
Bug: 170732441
Test: compile
Change-Id: I93416598b6047e890c66b922b360093aa42db8c5
(cherry picked from commit d179cbcdc2)
Storaged always writes to the SYSTEM log, so instead of using
LOG_TO(), it can use the normal LOG() macros as long as it calls
InitLogging() with a default to the SYSTEM log.
Test: storaged writes to the SYSTEM log still
Change-Id: I47e63bc92d55f9b1262a67e3508601ddbd9edc87
After a few years of being available, there only ended up being one
user of this, so it is clear that logd's duplicate message mechanism
is the favored solution.
The one user of this rate limiting is questionable as is, since due to
the nature of storaged, the mainloop only runs once per minute by
default as is, so there is essentially nothing to be gained by rate
limiting any further.
Test: build
Change-Id: I0610d11efda1ce8b581b939bad11ff295cb2daa6
It is possible for storaged to load the same saved proto file multiple
times, for example, if system_server crashes. In this case we do not
want to fill io_history with duplicate entries. This patch elides
records for which an app+userid record already exists for the same
timestamp.
Bug: 111578975
Test: gtest storaged_test.load_uid_io_proto
Change-Id: I87bc3e37d6464079cece03b5852285d79067b935
Merged-In: I87bc3e37d6464079cece03b5852285d79067b935
Add missing libbatteryservice_headers dependency.
It used to use global header include dirs, which is bad.
Add old transitive dependency from libbatteryservice_headers
as well.
Test: builds
Bug: 68724651
Change-Id: Idcc4452160dce70b4f8b48e2d809dcd5744b93b4
Merged-In: Idcc4452160dce70b4f8b48e2d809dcd5744b93b4
(cherry picked from commit 0b31908912)
storaged_t has more fields that depend on the health service;
they should be initialized in the new init() function.
Test: storaged unit tests
Change-Id: I70d41e5d0a0ef20c39c41c7539a284937bd21ad5
protobuf is only needed when serializing/deserializing data. Instead of
maintaining a permanent buffer in storaged object, move the container to
stack so that the buffer is released when we don't need it. In addition,
we don't need to clear the buffer before updating it.
Also added a function to clear user io history when the user is removed.
Bug: 63740245
Change-Id: Ia5d19b9a0c3f92a93b061a56be89bb0b958a2a29
Use user_id (from app uid) to determine file location.
/data/misc_ce/<user_id>/storaged/storaged.proto
Vold notifies storaged when a user's CE area becomes available.
Then storaged restores data from the proto in that area and
combines them into IO history.
Vold also notifies storaged when the CE area is being deleted.
Storaged clears internal history about this user and deletes the
proto file.
IO perf is stored in user_0 area since it's not user related.
Test: dumpsys storaged before/after multiple users' unlock
Bug: 63740245
Change-Id: I39f923f6b09e9f2a29e9286ce02b3b3bcbfb9f94
Use proto_stat to indicate status of CE area. Before user_0 unlocks,
storaged detects CE NOT_AVAILABLE when attempting to read the proto
file. It then skips reading/writting the proto.
When user_0 logs in, vold calls onUserStart in storaged, which sets
proto_stat to AVAILABLE. At next event, storaged tries to read the
proto. If it's a success, proto_stat is changed to LOADED. After that,
storaged reads and writes proto normally.
Test: adb shell storaged -u -p
Bug: 63740245
Change-Id: I1fdd42c430e91682f6cc07497fcad5be52489b4e
Record the bandwidth when writing storaged.proto to monitor system
storage performance.
The history is maintained in 3 catergories.
1. samples of last 24 hours.
2. daily average of last 7 days.
3. weekly average of last 52 weeks.
Sampling frequency of the first catergory is set to once every hour.
Storaged also flushes proto at same frequency.
The proto file is padded to reach a minimal size of 128KB. Storaged
writes 16KB at a time. Bandwidth calculation ignores time spent in
the first write and writes smaller than 16KB.
bandwidth = total size of 16KB writes / time spent in 16KB writes
Restructured the code so that storaged.cpp handles proto load/flush.
It calls individual module to sync proto with internal data
structures.
Added a cmdline argument to dump perf history.
adb shell storaged -p
I/O perf history (KB/s) : most_recent <--------- least_recent
last 24 hours : 5315 3775 3659 2042 3890 5652 3649 3696 6023
last 7 days : 4648 3532 3828 3567 3841 3359 4676
last 52 weeks : 3817 4275 3836 3766 4280 0 0 0 0 0 ...
Test: adb shell storaged -p
atp:asit/perf/jank_systemui_test
atp:asit/perf/appstartup_hermetic_cyclic_dropcache_test
atp:asit/perf/appstartup_non_hermetic_cyclic_dropcache_test
Bug: 63629306
Change-Id: Ie7051e7a8df883d4a6818ea9a54a10f4dccb4843
Merged disk_stats_publisher into disk_stats_monitor class.
Moved proc file check to disk_stats class instead of doing so in
storaged, same for uid_io.
Added is_zero function to check zero disk_stats.
Added operators to compute disk_stats.
Change-Id: I277ae51f6fe18464f92e45c03471343784e67093
Convert storaged internal io_history to protobuf format and serialize
it to userdata partition. Also load this file during storaged startup
to reconstruct io history.
Bug: 63740245
Change-Id: I0697525df1c31fdec20f5ed4e3e9363e2dde244f
Update disk space utilization together with diskstats since we will
use free space size to co-relate disk performance.
Bug: 63629306
Change-Id: I4e5694aaff3b71aa56db451f1bc92ccfb07e5086
storaged relies on batteryproperties service (healthd) to provide
charger stats. If healthd is killed, kill storaged as well. Both
will be restarted by init process anyway to restore a correct state.
Test: kill healthd when storaged is running
Bug: 36652060
Merged-In: Ia785bb6f5ea259aa43cd1efab9505ebefaf3db12
Change-Id: Ia785bb6f5ea259aa43cd1efab9505ebefaf3db12
uid_monitor is the only user of batteryproperties, no need to register
a callback if uid_io not enabled.
Test: run storaged without uid_io, plug/unplug usb cable
Merged-In: If890a93fec155b67c450341c54b44bb18d3aae3d
Change-Id: If890a93fec155b67c450341c54b44bb18d3aae3d
Sysfs data is straightforward so we don't need parsing anymore.
Also removed periodical check since data is set only once during
driver initialization. Checking at every device boot or storaged
restart should be sufficient to monitor long term status change.
Test: adb logcat -d -b events | grep storaged_emmc_info
Bug: 36228467
Merged-In: I2a181f52c9f19de1e679a3a905aaebafe4d08227
Change-Id: Ic05e353f0af9363f3bcbe793ba0c351082e446ca
Register a listener to batteryproperties service for charger stats
change.
Aggregate IO usage based on charger stats in a collection window.
Bug: 33086174
Bug: 34198239
Change-Id: Ibe306c9c3ff8b8ada6be034aa8511268cb9a9b1c
Skips event time check if failed to get start time.
Use ratelimited logging when gettime fails.
Test: adb shell setprop ro.storaged.event.perf_check 10000
Bug: 34630954
Bug: 34198239
Change-Id: I168ef3b4e4ed27257ec76373219c6214a7b7b2ed
Add properties to control event intervals.
Add a property to check time spent in event loop.
Bug: 34612341
Bug: 34198239
Change-Id: I01f64c84e17153377ece7ae530be106e3f55287e
Add uid_monitor class to query /proc/uid_io/stats periodically.
Add a log tag to record any UID that exceeds IO threshold.
Test: adb shell storaged -u
Bug: 34198239
Change-Id: I53568c30dbefe2f4bdb18054d3dedb30b4133d8b
Don't log disk_stats and emmc_info to system buffer since they
are already uploaded to events buffer.
Use DEBUG flag to control disk perf logging.
Bug: 34331490
Change-Id: I5f628a1a46e1f72588da064523da69737bd67dcb
Storaged periodically checks the devices' disk stats and eMMC lifetime.
This information is logged to both the kernel and event log. By event
logging, the data can be aggregated by the checkin service and then
sent to Sawmill.
It also periodically traverses /proc/[pid], maintaining the I/O usage of
all tasks (all sampled tasks). The task I/O data can be reported using
the --dump option.
Storaged is booted when the device boots up and requires the permission
for reading /proc/diskstats, /proc/[pid]/io and debugfs(eMMC ext_csd).
For detailed description, please refer to go/storaged.
Bug: 28826771
Change-Id: I774b1a44cc30092bac1bfcbc08bf487295e7cfed
