This change fixes detailed UID stats to ensure network and battery stats
both take VPNs into account. NetworkStatsFactory is being made aware of
VPNs enabled, and the full set of underlying networks present.
Since traffic can only be migrated over a NetworkStats delta, NSF
maintains a NetworkStats snapshot across all UIDs/ifaces/tags.
This snapshot gets updated whenever NSF records a new snapshot
(based on various hooks such as VPN updating its underlying networks,
network getting lost, etc.), or NetworkStatsService's
getDetailedUidStats() method being called.
This change widens the scope of the existing mPersistentSnapshot lock,
renaming it to mPersistentDataLock, and ensures that TUN migrations are
not done in parallel. Additionally, mVpnInfos is updated via
pointer-swapping, to reduce the scope of the mPersistentDataLock.
The safety of this change is predicated on:
1. NetworkStatsFactory lock not held, so services cannot deadlock through
the cyclical lock.
2. The broadening of the scope of the lock in NetworkStatsFactory has no
threading implications, as it is always the last (leaf node) lock held,
and therefore is impossible to have lock inversion.
Additionally, to ensure VPNs work with 464xlat, the VPN info passed to
the NetworkStatsFactory includes all underlying interfaces, instead of
only passing the first one.
This (partially) re-applies changes from:
aosp/972848: Add one more test for VPN usage stats.
aosp/972847: Addressing comments for http://ag/7700679.
aosp/885338: NetworkStatsService: Fix getDetailedUidStats to take VPNs
into account.
Co-developed with: Varun Anand <vaanand@google.com>
Bug: 113122541
Bug: 120145746
Bug: 129264869
Bug: 134244752
Test: FrameworksNetTest passing
Test: Manual tests show data usage fixes maintained.
Change-Id: I6466ec1411fc5ed6954125d27d353b6cd1be719e
This change removes a redundant line in the xt_qtaguid_with_clat test
file.
Bug: 134244752
Test: FrameworksNetTest passing
Change-Id: I87deb82ba960102b617ab27362e8c5899fa478f8
Currently, 464xlat counts its ipv6 tx traffic into root uid.
When user is making ipv4 upload over ipv6-only network, ipv4
tx traffic may sometimes be counted faster then ipv6 tx
traffic.
Thus, NetworkStatsService may detect non-monotonic values due
to 464xlat adjustment.
So the solution is that: for clatd, make ipv6 tx traffic counts
into clat uid, and then ignore it in the framework side.
Bug: 118602783
Test: 1. manually verify clatd traffic on clat uid.
2. runtest frameworks-net
Change-Id: Ifb478b79e3e281918c70e16d1f90682c78f33db1
If the qtaguid proc file is no longer exist, the device is running new
eBPF module to do traffic accounting. So the NetworkStatsFactory need to
use the proc/net/dev interface to get the per interface traffic stats
summary. Also, adding a test to verify the helper function work properly
Bug: 30950746
Test: run NetworkStatsFactoryTest
Change-Id: Ia36808bf02f1637dd41a3e7c50917b91b1a77524
Follow-up of commit 4ff3a778ab.
Also add @SmallTest annotation to classes moved to tests/net.
Test: no functional change
Bug: 32561414
Change-Id: I0a0e8865d37ba7bae06ce352d3fc385989adc300
For 464xlat scenarios on IPv6 networks, the clatd interface setup
introduces double counting of apps ipv4 traffic. NetworkStatsFactory was
accounting for this on the tx path, but not on the rx path. Also it did
not accounted for the 20 bytes added by the IPv6 header.
This patch subtract correctly the rx and tx traffic from the root uid on
the underlying interface, and also adds correctly the 20 bytes cost per
packet on the stacked interface for 464xlat traffic.
Test: added several new unit tests, based on synthetic data and real
data also.
Bug: 33681750
Change-Id: I2675643b220acbc6110179fa937d4c313b6f5e32
This groups them together with the rest of the networking unit
tests. It also speeds up compile/test cycles ("runtest -x" of one
file goes from 1m15s to 30s).
Test: runtest frameworks-net passes on internal tree
Change-Id: I53cb0c51355fe4b4b30e451fa09fbbf58da39efd
