When NDC delivers limitReached() event, NMS re-arms the alert, which
makes another call back to NDC. If NDC already started processing
another call, this causes deadlock.
Fix this by re-arming on a handler thread.
Bug: 5879848
Change-Id: I46fa78fe1f1122c5f31dd6020483e5ae89a9f39d
Split existing network stats into two separate classes: a recorder
which generates historical data based on periodic counter snapshots,
and a collection of historical data with persistance logic.
Recorder keeps a pending history in memory until outstanding data
crosses a specific threshold. Persisting is handled through a given
FileRotator. This pattern significantly reduces disk churn and
memory overhead. Separate UID data from UID tag data, enabling a
shorter rotation cycle. Migrate existing stats into new structure.
Remove "xt" stats until iptables hooks are ready. Avoid consuming
Entry values when recording into NetworkStatsHistory. Assign
operation counts to default route interface.
Introduce "Rewriter" interface in FileRotator with methods to enable
rewriteAll(). Introduce IndentingPrintWriter to handle indenting in
dump() methods.
Bug: 5386531
Change-Id: Ibe086230a17999a197206ca62d45f266225fdff1
Create a LocalLog class for logging within a service for dumping in dumps.
Use it in the NativeDaemonConnector so we can get some insight into what
is happening in these lockups.
bug:5864209
Change-Id: I68ddc58847f3c8de613be9528570f8c3157d8274
Report non-monotonic NetworkStats through an observer interface
instead of throwing, since those events are still recoverable.
Change-Id: Ic0749f4634b0ac05dbe90e95ca490957ec8b2f23
Uses argument escaping inside NativeDaemonConnector, using varargs
to separate boundaries. Also introduces Command object to help build
argument lists.
Bug: 5472606
Change-Id: I357979fc19bb0171a056e690064e01b5a7119501
Perform uniform argument escaping inside NativeDaemonConnector, using
varargs to separate boundaries. Also move to parsed NativeDaemonEvent
instances instead of raw Strings.
Bug: 5472606
Change-Id: I1270733e2b2eeb2f6b810240df82ab24d38ebf40
Move NTP updates outside locks to avoid blocking ShutdownThread. Add
logging around trim events that remove substantial history. Include
history start in sample events.
Bug: 5627247, 5584564
Change-Id: If9bbd93842c710efcdec94e68da87e9e42a5f961
Instead of trusting NTP time alone, use the most-conservative of
system clock and NTP.
Bug: 5584564
Change-Id: I5dd87fc009959b1cf0a7d660e385a0b1a8be238b
When encountering non-monotonic stats rows, recover remaining data by
clamping to 0. In particular, this avoids edge-case where persisting
threshold checks would never trigger. Also recover when tethering
snapshots are missing.
Bug: 5600785, 5433871, 5600678
Change-Id: I1871954ce3955cc4ac8846f9841bae0066176ffe
Moved away from BufferedReader, which only reads the first 8KB of
some proc files because it aggresively fills its buffer. Optimized
proc parsing, now double the speed. Tests to cover.
Log when NetworkStats counters roll backwards when subtracting, and
optimizations around findIndex(). When system removes UID, also
remove from last stats snapshot to avoid xt counters from rolling
backwards.
Bug: 5472949, 5458380
Change-Id: I07c08fe5233156fac2b84450f6291868bf9bfaf2
Offer parameters to pick out specific list of packages, and to change
duration and total data.
Bug: 5416742
Change-Id: I6970a5e76f3198ef23a400c079037212b000ca6b
Begin tracking xtables summary of data usage to compare with values
reported from /proc/net/dev. Roll tethering directly into UID stats
to trigger UID stats persisting when crossing threshold.
Include xtables summary and authoritative time in samples.
Bug: 5373561, 5397882, 5381980
Change-Id: Ib7945522caadfbe0864fdf391582dc820f4f371e
Watch for changes to telephony network subtype, and update iface
mapping to persist stats under correct type. Update network stats
before removing tethering NAT rules.
Skip recording that would create empty historical buckets. Query UID
stats before iface stats to always skew positive when counters are
actively rolling forward.
Bug: 5360042, 5359860, 5335674, 5334448
Change-Id: I8aa37b568e8ffb70647218aa1aff5195d3e44d5a
When sampling network stats, always use atomic buckets instead of
interpolating. Always poll iface and UID together so we distribute
into buckets equally. Move stale bucket trimming to just before
writing stats.
Bug: 5321340
Change-Id: I78a2226778a79c875f3668336e39ea24a7b4d5c4
Use new "gettetherstats" netd command to retrieve statistics for
active tethering connections. Keep tethering poll events separate
from UID poll, even though they end up same historical structures.
Bug: 5244846
Change-Id: Ia0c5165f6712c12b51586f86c331a2aad4ad6afb
Devices currently delay collecting UID stats during common operations
like global alerts to reduce parsing load. Here we introduce a flag
to always collect UID and iface stats together to aid debugging.
Bug: 5321340
Change-Id: Ia8e1daf1a6f9d5f5cf1f64a04b6b6766ee965ab6
When ifaces change, poll UID stats without persisting, since they
depend on knowing active iface to store correctly. Log dropped UID
stats when iface is unknown. Switch to using flags when calling
performPoll().
Enforce that "idx" values are consistent from xt_qtaguid. Transition
to using Log.wtf() for important checks, mostly around file I/O,
kernel stats parsing, and kernel module control. Increase stats
persist threshold to 2MB to reduce churn.
Bug: 5269476, 5270106
Change-Id: I721215bfb65127f95775c71cf135e907cd567e92
New broadcast that is dispatched immediately after connectivity
changes have been handled by ConnectivityService, bypassing any
applicable CONNECTIVITY_CHANGE_DELAY.
Also protect CONNECTIVITY_CHANGE broadcasts, since they should only
be sent by system.
Bug: 5198167
Change-Id: I75f1fb44b21da1879f0ab960bcaa481126d70fde
When reading network counters, always splice in xt_qtaguid values
to avoid counting backwards. Test to verify.
Remove verbose logging around global alerts, and add dumpsys debug
info for NMS and MDST. Also fix subtle bug around stats persisting
and dumping.
Bug: 5212893, 5211028
Change-Id: I783e5286637a67ee2dd2b09878198711a926d0cb
Instead of polling every 15 minutes, register for alerts that trigger
when system-wide traffic passes a threshold. Still mixed with polling
to persist UID stats, but relaxed to 30 minutes. Currently watches
for every 512kB.
Make persistence decision separately for network versus UID, and use
total delta bytes when making decision. Use light bootstrap during
systemReady() instead of heavy poll, which had been force-loading all
UID data unnecessarily.
Bug: 5023631
Change-Id: I04b723d6c4bf872fb1028071122dba66a8e1b576
Teach NetworkStats about "counter sets" coming from kernel, and use
them to track usage in foreground/background. Add AID_NET_BW_ACCT to
system_server so it can control counter sets.
Move to composite key of NetworkIdentitySet, UID, set, and tag when
recording historical usage. Persisting still clusters by identity,
since that is heaviest object.
Request async stats poll during systemReady() to bootstrap later
delta calculations. Reset kernel counters when UID removed. Update
various tests.
Bug: 5105592, 5146067
Change-Id: Idabec9e3ffcaf212879821515602ecde0a03de8c
To catch instances where NativeDaemonConnector hangs while executing
commands, add users to system_server watchdog. Also moved qtaguid
enabler to separate thread to prevent blocking systemReady().
Bug: 5145437
Change-Id: I4c6c1dc09864a1c05f9486a31faa1b366548e911
Persist NetworkStatsHistory using variable-length encoding; since
most buckets have small numbers, we can encode them tighter. Initial
test showed 44% space savings. Also persist packet and operation
counters.
Let NetworkStatsHistory consumers request which fields they actually
need to reduce parcel overhead.
Tests for verify varint and history field requests, also verify end-
to-end by persisting history into byte[] and restoring. Expose
bandwidth control enabled state. Extend random generation to create
packet and operation counts. Moved operation counts to long.
Fix bug that miscalculated bytes since last persist, which would
cause partial stats loss when battery pulled.
Bug: 4581977, 5023706, 5023635, 5096903
Change-Id: If61e89f681ffa11fe5711471fd9f7c238d3d37b0
Now callbacks can communicate to the same daemon without causing a
deadlock. This also improves the latency of calls because they no
longer need to wait for the callbacks for the pending events.
Change-Id: I153fcf16bd64de79ee1c1a57d3cfdb12b354cf47
Provide API to increment "operation counts" for a UID and tag, used
to eventually derive bytes/operation stats. Internally is stored at
network layer, but should belong at data layer. Switch profiling
to use data layer stats, which are emulated by summarizing network
layer details.
Read packet counts from new /proc/ columns and collect them into
NetworkStatsHistory. Prevent double-counting by ignoring values from
first snapshot. Watch for duplicate /proc/ entries. Update tests
to verify packet and operation counters.
Bug: 5052136, 5097392
Change-Id: I1832f65a2b8a9188f8088f253474a667c21a2f09
Move template matching to builder methods instead of exposing the
internal constants. Also rule to match Ethernet usage. Buffer
reading and writing of network stats, making operations 5x faster.
Change-Id: Iedb2d0ab3b26a976811c050f84a164e909eb74b6
When requesting historical values with time ranges, always include
full values for buckets that span current time. (It doesn't make
sense to interpolate partial data.) Move getTotalData() to return
full Entry objects to prepare for packet counts.
Bug: 4691901
Change-Id: I717bd721be9f1d4a47c4121e46e07a56cb15bbf1
Instead of exposing internal storage details, offer getValues()
accessor which populates values into a recycled structure.
Change-Id: I5228d7099a8f61b2018a435a813435a67dae49b4
Uses NTP server and timeout from secure settings, or fallback to
defaults in resources. Update various system services to use cached
NTP time when fresh enough, or force updates as needed.
Bug: 4517273
Change-Id: Ie1c4c4883836013d02ca0bbd850cf8949f93b34b
Both stats and policy make NMS calls that depend on bandwidth control
being enabled, so move enable/disable into NMS and drop calls when
disabled. This avoids throwing heavy ISE exceptions when disabled.
Only include recent data when writing NetworkStatsHistory as part of
dumpsys call. Introduce manual poll event for Settings UI.
Bug: 4982115, 4770435, 4515856
Change-Id: I257820b057af2f0f99c736fb4f61e55b9fdc3e66
Change NMS parsing to handle extended /proc/ stats formats by pairing
values with header keys. Move TrafficStats to integer tags to match
kernel internals, and offer well-known tags for system services.
Async policy event dispatch from NPMS, and update tests to block for
event dispatch. Narrow app policy to exclude apps signed with system
key, which are usually critical.
Bug: 4948913, 4903489, 4585280
Change-Id: Idb357227ccaa617906411f309371cea18d7bc519
For now, there is still too much breakage around the kernel's new
qtaguid netfilter.
So no netstats up front.
But enable on the device until wiped:
./vendor/google/tools/override-gservices secure:netstats_enabled=1
adb reboot
Bug: 4903513
Change-Id: I230bd2d4efab85cac5c9abaa7d638c94f5e0c7d6
Signed-off-by: JP Abgrall <jpa@google.com>
