Files
2026-11-18 01:39:26 +09:00

435 lines
26 KiB
Markdown
Raw Permalink Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# Cellular data experiments on platina
Status on **2026-11-17**: SDM660 IPA v2.6L BAM QMAP cellular data transport is
implemented in candidate kernel recipe **7.2.12-r25** (`0044-net-ipa-implement-sdm660-v2-6l-bam-transport.patch` + DT interconnects/SMMU in `0042` & ANOC2 IRQ polarity in `0045`). Interface-bound public IPv4 ICMP, idle single-packet UDP DNS, TCP DNS, and TCP+TLS HTTPS are verified on China Unicom LTE (`46001`, APN `3gnet`), alongside multi-wraparound packet delivery and disconnect/reconnect cycles. See the [2026-11-17 IPA v2.6L BAM section](#ipa-v26l-bam-transport-qualification-2026-11-17) below; historical **2026-11-12/13** AT/PPP and r23/r24 findings are preserved chronologically.
## What was verified on the phone
| Layer | Observed result |
|---|---|
| Modem control | QMI DMS, NAS and UIM respond over `qrtr://0`; USIM is ready |
| Radio | NAS reports China Unicom / LTE, registered, PS attached during tests; registration also times out in some attempts |
| Software transport | The unused modem GLINK `DS` channel responds to AT commands through `rpmsg_chrdev` |
| Firmware capabilities | `AT+CGDATA=?` advertises PPP; this does not establish a data call |
| Actual PPP call | With GLINK DTR/RTS signals set and registration complete, `ATD*99#` returns `CONNECT`; LCP and IPCP exchange packets in both directions |
| IPv4 | Multiple calls obtained an IPv4 address on `ppp0` |
| SIM subscription | The user confirmed this SIM can use mobile data under Android |
| Public data | Interface-bound ICMP to `1.1.1.1`/`223.5.5.5` and DNS queries to `223.5.5.5` received no replies |
| Packet integrity | Captured outgoing DNS frames passed PPP FCS, IPv4-header and UDP-checksum verification; their source matched IPCP |
| Compression | Disabling address/control, protocol and VJ compression did not restore DNS replies |
| Authentication | A diagnostic call with `unicom` credentials completed CHAP-MD5 authentication; DNS still timed out and a subsequent LCP echo reply arrived |
| PDP activation | `AT+CGACT=1,1` succeeded for `3gnet`; `CGCONTRDP` returned the APN, an IPv4 address and DNS `221.11.1.67`/`221.11.1.68`. Dialing PPP after this separate activation ended with `NO CARRIER` during IPCP |
| Alternative APN | A diagnostic `wonet` PPP call obtained IPv4 and passed CHAP; three queries to operator DNS `221.11.1.67` still timed out |
| Proven Internet uplink | A PPP-bound TCP SYN reached the authorized Tokyo server; its sequence number matched the phone's outgoing PPP frame exactly. The server transmitted SYN-ACKs, but no IP frame returned through PPP |
| Firmware stability | Kernel logs record repeated MPSS fatal errors at `ipa_util.c:1089`, asserting `ipa_util.imm_cmd.clients[cmd_handle].init_done`; remoteproc automatically restarts the modem, and ath10k also recovers |
| CGDATA alternative | `AT+CGDATA="PPP",1` returned `ERROR`, including after LTE registration |
Clearing both GLINK signals before dialing also returned `ERROR`. The current
RPMsg flow-control ioctl changes DTR and RTS together. Their interaction with
bulk data needs further qualification. A `CONNECT` speed string is a modem
response, **not a measured throughput**.
A separate diagnostic accepted `AT&D0` and cleared both signals after IPCP.
It did not restore downlink: the modem hit the same IPA assertion about
60 milliseconds after clearing the signals. This is not an operational
flow-control workaround. These diagnostic AT settings were volatile; no
profile-save command was sent, and the modem restarted after the assertion.
ModemManager was temporarily stopped during direct AT/PPP experiments and
restarted afterwards. At the start of investigation, Wi-Fi was disconnected
and the modem reported `shutting-down`; previously reported steady-state
network results must not be assumed to describe the current session.
The firmware assertions and missing downlink make modem-side IPA
initialization a concrete dependency to investigate. They do not prove that
implementing a particular QMI message alone will fix it. Do not repeatedly
redial this experiment as an
automatic failover service: the observed modem restart also affects Wi-Fi.
USB SSH and the battery guard remained available during these tests.
In the correlated TCP test, the phone sent sequence `2844831695` through
PPP at monotonic time `17402.104`. The Tokyo capture received the same
sequence after carrier NAT and sent SYN-ACK sequence `797834871`, ACK
`2844831696`. The modem's IPA assertion occurred at `17414.509`, about
12.4 seconds after the first SYN. Active PPP shutdown began at `17437.118`.
Thus this assertion preceded local teardown; it cannot be attributed solely
to closing the bridge. Server SSH was listening and its existing firewall
allowed this port; no firewall rule was changed. This establishes an
Internet uplink and a missing PPP downlink, not a completed TCP connection.
## Run the experimental software path
`scripts/phone-cellular-rpmsg.py` is packaged as
`/usr/local/sbin/platina-cellular-rpmsg` in newly generated runtime overlays.
It is opt-in and has no automatically enabled OpenRC service.
The updated helper was also installed on the test phone on 2026-11-12;
its SHA-256 matched the local source and its AT-only smoke test returned
`OK`. No automatic PPP service was installed.
Dependencies are `python3`, `ppp-daemon`, matching kernel `slhc`, `ppp_generic`,
`ppp_async` and CRC modules, and the existing `platina-cellular` battery gate.
On this installation, rebuilding the module dependency index with `depmod -a`
resolved missing `slhc_*` symbols when loading `ppp_generic`. This was an index
problem; no incompatible module was forced into the kernel.
The default command sends only `AT`; `--query` reads SIM, radio and PDP state:
```sh
sudo platina-cellular-rpmsg
sudo platina-cellular-rpmsg --query
```
The following foreground experiment checks the battery, temporarily stops a
running ModemManager, requests QMI online mode and restores the manager's
initial service state on exit:
```sh
sudo platina-cellular-rpmsg --connect --apn 3gnet \
--dial-method atd --glink-signals 1 \
--registration-timeout 180 --duration 90 \
--verify-target 223.5.5.5 --verify-kind dns
```
The helper refuses already bound channels and existing PPP interfaces. It
uses an exclusive session lock, restores the channel override and closes
the temporary endpoint on exit. During PPP shutdown it continues forwarding
the termination handshake before closing the transport.
GLINK may remove and re-announce the channel while closing it. The helper
supports only the modem's `DS`, `DATA1`, `DATA2`, `DATA3`, `DATA4` and `DATA11`
names; only `DS` has been verified as an AT transport here.
`--connect` configures PDP context 1 for the specified APN and IPv4. It changes
the modem's AT echo/error settings. It does not set a default route or DNS.
`--verify-target` adds an exact temporary host route, binds its probe to the
PPP device, and removes that route after testing. It never falls back to a
Wi-Fi or USB probe. `--verify-kind dns` sends a DNS query for `example.com` to
the specified public resolver; `ping` uses four ICMP packets. A failed probe
returns nonzero even if IPCP assigned an address. A successful test closes
the experimental call on exit; continuous operation needs separate qualification.
The result's `ppp_frames.tx` and `ppp_frames.rx` count FCS-validated PPP
frames by protocol, IPv4 bytes, corrupt frames and oversized frames. These
counters distinguish LCP/IPCP/CHAP replies from actual received IP packets;
they do not retain payloads. In the `wonet` diagnostic, TX contained three
IPv4 frames (171 bytes) and RX contained only LCP/CHAP/IPCP, with zero bad
FCS frames. Those counts alone demonstrate valid frames entering the local
RPMsg send path. The separate sequence-correlated server capture establishes
Internet uplink delivery. A carrier-side or modem-side downlink capture is
still needed to locate the loss of replies.
The foreground duration starts after dial/registration. Opening a GLINK
endpoint and requesting a remote RX intent can also wait inside the kernel;
`--timeout` bounds AT response polling, not every kernel operation.
Local evidence from this investigation is retained outside the source
release under `reports/cellular-20261112/`: correlated phone/server TCP logs,
the alternative-APN DNS result, firmware assertions and regression logs.
## IPA v2.6L kernel findings
The original experimental `0043` patch is a scaffold, not an implemented
data transport. It skips BAM, endpoint initialization, memory and route-table
setup. Its TX path frees packets while reporting successful transmission.
Creating `rmnet_ipa0`, exposing endpoint numbers or surviving a short idle
test does not qualify this implementation.
The new `0044` guard rejects this hardware with `-EOPNOTSUPP` before mapping
registers, and fixes both false-success TX branches. The candidate recipe
and lock file are **7.2.12-r24**. This is an explicit unsupported-transport
guard, not a BAM driver. The earlier GCC check of three affected objects used
an isolated **7.2.3-sdm660+** source copy. That historical result does not
qualify the candidate against the phone's **7.2.12-sdm660** kernel. The current
r23 build uses the locked archive, all recipe patches, Alpine Clang 22.1.3,
and the phone's exact configuration, including non-permissive KCFI. The
configuration SHA256 is
`7f816caaea4cb3889b57e564c28a2384c7ddcc4ec21cd7f38473e06c643c23ad`.
The r23 full Image/modules build and sequential finalization completed successfully,
confirmed on 2026-11-13 (Asia/Tokyo). Both regenerated panel DTBs contain the
corrected IPA/BAM resources and AP SMMU stream `0x19c0`. The latest diagnostic
also passed normal modpost against the completed symbol table, with split BTF
omitted. Finalization evidence is retained in
`reports/cellular-20261112/build-r23-final.log`. No r23 kernel has been flashed
or qualified for boot. The finalized diagnostic was
subsequently loaded, sampled and unloaded normally on 2026-11-13; its log is
`reports/cellular-20261112/probe-live-final-r23.log`.
| Final artifact | SHA256 |
|---|---|
| `Image.gz` | `41facfb07f2d8d573d322e45f82c0b7b47790b78a77cbeba313085954be1e4ad` |
| `sdm660-xiaomi-platina.dtb` | `6ae7d0d10fc99a60d8c51ef60805eb55113b616e1a44800288ddf664f043df74` |
| `sdm660-xiaomi-platina-boe.dtb` | `4d82fcd7bcef5c7a4ab98d168994ff4e3ec438a0e0e1b3acb87c5802cb4223f2` |
| `platina_ipa_v2_probe.ko` | `1dbc19916a0adf96f7751020d419c5554e84ceb411d8dd23a11c3211c8a0def7` |
The user-space hardware activation check also rejects a v2.6 shim, its RMNET
children, a bare driver binding, or an interface name without backing hardware.
A future implemented v2 backend must expose `modem/datapath_ready=1` before
this integration permits activation. Even that marker does not prove public
connectivity. The independent PPP experiment does not use this NM hardware gate.
The Xiaomi `platina-o-oss` sources, commit
`0bb2780bccb9c7e05f0edf5db8860e779179b5f9`, establish these concrete differences:
| Item | Original hardware/source definition |
|---|---|
| Hardware generation | `IPA_HW_v2_6L` (`qcom,ipa-hw-ver = <6>`) |
| Raw IPA resource | `0x14780000`, size `0x4effc` |
| BAM resource | `0x14784000`, size `0x26934`; **not shared SRAM** |
| IPA register offset | `IPA_REG_BASE_OFST_v2_6L = 0x40000` from the raw IPA base |
| SRAM direct access | v2 direct-access offsets and memory partitioning; not the v3.1 register table |
| AP command producer | Endpoint **3** |
| AP LAN/WAN producer | Endpoint **4** |
| AP LAN consumer | Endpoint **2** |
| AP WAN consumer | Endpoint **5** |
| Q6 LAN producer / command producer | Endpoints **6 / 7** |
| Q6 LAN consumer / WAN consumer | Endpoints **8 / 9** |
| SMMU | AP context stream ID `0x19c0`; separate WLAN/uC contexts |
These differ from the GSI-style endpoint assignments copied into `0043`.
Correctly implementing BAM requires v2 register definitions and resource
ownership, verified DMA/IOMMU mappings, descriptor/completion handling, RX
replenishment, immediate commands and memory/table initialization, modem QMI
initialization, and restart/power cleanup. Do not enable the old DT node and
simply substitute a BAM IRQ for GSI.
The r23 source description corrects the IPA register aperture to
`0x147c0000`, labels the complete `0x14784000` BAM range as BAM, and uses
the original SPS driver's level-high BAM interrupt. It removes the GSI
loader property and declares the original AP stream `0x19c0` on ANOC2 SMMU.
The guard remains in place; these corrections do not implement BAM DMA or
its buffer mappings. The running phone uses strict IOMMU translation with
unassigned-stream bypass disabled, so an unmapped physical-address DMA
fallback is not an acceptable transport implementation. An existing DTB-list
hunk in patch `0040` was also corrected to match the locked source's tabs and
surrounding context.
The Android initialization dependencies and opt-in read-only register
diagnostic are documented in
[`project/cellular/ipa-v2`](../project/cellular/ipa-v2/README.md). A key
transport distinction is `SPS_IOVEC_FLAG_IMME` (`0x0100`), with the opcode
in the descriptor's size field. Mainline BAM's `DMA_PREP_CMD` sets ordinary
BAM command bit `0x0800` and cannot substitute for this IPA descriptor.
### Read-only hardware verification, 2026-11-12
The opt-in diagnostic loaded through normal module checks on the current
7.2.12 phone and was unloaded after each snapshot. It retained KCFI and did
not use forced loading. The early diagnostic build preceded the full kernel
symbol table, so modpost emitted unresolved-symbol warnings. All 28 required
symbols were separately checked against the running kernel's actual exports;
This export check applied to the earlier live-tested binary. The finalized
diagnostic subsequently passed normal modpost and was loaded, sampled and
unloaded on the phone on 2026-11-13.
The diagnostic omitted split BTF so it did not depend on a rebuilt kernel's
BTF type IDs.
| Register/state | Observed value | Interpretation |
|---|---|---|
| IPA hardware version | `0x20060001` | Hardware registers are readable at the corrected aperture |
| IPA version | `0x01030300` | Raw version register, separate from the hardware version |
| IPA configuration | `0x10` | IPA enable bit 0 is clear at the snapshot |
| SRAM size / restricted bytes | `8192 / 0` | Matches the original v2.6L layout's 8 KiB end offset |
| IPA / BAM pipe counts | `20 / 20` | Available pipe counts, not active-pipe bitmaps |
| BAM revision | `0x00674427` | BAM revision field is `0x27` |
| BAM EE0 pipe attributes | `0x8009f83f` | EE0 is enabled and owns AP pipes 2 through 5 |
| AP LAN RX / command TX / TX / WAN RX | control and offsets all `0` | These four BAM pipes are not enabled |
The first snapshot exposed an additional clock-ordering error in the
diagnostic: this kernel's RPM provider returns success without recording a
rate change while unprepared. The corrected diagnostic follows the vendor's
prepare-before-set sequence. Its second snapshot reported a 75 MHz vote and
75 MHz cached request, then restored the original cached request and zero
prepare/enable counts. Those values are requests, not measured frequencies.
Neither snapshot reset IPA/BAM, acknowledged IRQs, submitted DMA or sent QMI.
No new firmware assertion appeared during the snapshots; USB, routes,
ModemManager and battery guard retained their prior state. Loading the
external module set the kernel's normal out-of-tree taint bit (`4096`), which
remains until reboot even after unloading. The complete logs and JSON are in
`reports/cellular-20261112/probe-live-r23.log` and
`probe-live-clock-r23.log`.
These observations establish an accessible but uninitialized AP transport.
They support implementing the real BAM command and data paths; they do not
prove that missing initialization is the sole cause of PPP downlink loss.
### Strict DMA mapping and r24 DT qualification, 2026-11-13
The opt-in DMA mapping module created a separate test platform device through
an OF overlay, attached AP stream `0x19c0` to the strict default DMA domain,
and allocated three coherent pages. For each page it compared the actual CPU
backing pages from `dma_get_sgtable()` with `iommu_iova_to_phys()` at both ends.
All comparisons passed. This tested translation setup, with no hardware DMA
or IPA/BAM MMIO. The result explicitly reported `mapping_verified=true` and
`hardware_dma_verified=false`. The overlay, allocations and device were
removed afterward. The old-DT experiment also emitted context-IRQ and
device-link warnings; mapping success did not excuse those warnings.
The ANOC2 context fault SPIs 374–378 were incorrectly described as level-low.
The GIC rejected that trigger type, and SMMU setup continued without the
requested context fault handler. Candidate patch `0045` changes these five
entries to level-high. Both panel DTBs were rebuilt and the standard panel
DTB was temporarily RAM-booted with the existing release kernel and ramdisk.
The IPA device joined IOMMU group 1 and its context fault handler registered
successfully, without the earlier trigger/request errors.
| r24 DT artifact | SHA256 |
|---|---|
| `sdm660-xiaomi-platina.dtb` | `f36404a07e0cadba7bf1d41edfa3b8e9b9acfe65f7de15f364a77427952e2a25` |
| `sdm660-xiaomi-platina-boe.dtb` | `4a95727fd071ed7f68c501447b8a082074d78426d8ae08035759e31f51baf15f` |
This qualifies the DT change on the existing kernel. It does not qualify a
new r24 kernel Image or APK: the complete Image/modules build remains the
r23 build above. No boot partition was flashed.
### BAM command experiment: descriptor consumed, SRAM DMA unverified
Stopping a running modem during the first experiment was followed by an
unexpected reboot; the available record cannot identify its exact cause.
Subsequent tests used an opt-in `platina.ipa_bringup=1` RAM boot instead.
The DSP service held MSS offline from startup, while ADSP/CDSP ran normally,
and RMTFS served files without its modem start/stop option. This avoided
repeating a live modem stop.
A second reboot was localized by logging each MMIO read. Reads of AP/USB
pipe controls returned zero, but the last recorded operation was the read
of Q6 DECOMP pipe 11 at BAM offset `0x1e000`. Its bit was present in
`PIPE_ATTR_EE0`; this attribute alone therefore cannot authorize an AP read
of every advertised pipe while MSS is offline. The diagnostic now restricts
inspection to the vendor's AP/USB endpoints 0–5, requires IPA initially
disabled and MSS locked offline, and never resets global BAM or touches
Q6/WLAN pipe registers. This checks the AP/USB scope, not all hardware pipes.
The revised read-only inspection passed and unloaded normally.
The command test prepares one translated descriptor and a 16-byte
`DMA_SHARED_MEM` read at SRAM offset `0x1ff0`, with the DMA target initialized
to the inverse of the CPU SRAM sample. IPA enable precedes 32-bit CPU SRAM
access, and BCR uses the pinned vendor value `0x001fff7f`. Actual observations:
| Check | Result |
|---|---|
| Strict page mapping comparisons | Passed |
| AP/USB idle inspection and command-pipe register checks | Passed |
| BAM descriptor offset | Advanced to 8 bytes |
| BAM pipe/global status | `0x2c / 0`, including EOT without reported BAM errors |
| CPU SRAM before/after | Stable |
| DMA target after EOT | Still the initial inverse pattern |
| Low IOVA comparison, below 1 GiB | Same failure, including a further 1-second data wait |
| Command endpoint control | `0`, no delay/suspend bits set |
These observations qualify BAM descriptor consumption only. They do not
prove IPA command execution or SRAM-to-system DMA. The latest diagnostic
reports `bam_eot_seen=true`, `command_completed=false`,
`sram_bytes_match=false`, `result=-5`, and `datapath_ready=false`.
EOT without the expected bytes retains the module, driver, clock and DMA
mappings until reboot; neither forced unloading nor modem startup is an
acceptable cleanup for that state. The retained-resource path was exercised
and followed by a normal reboot into the original persistent image.
The temporary DSP/RMTFS files were restored from hash-checked backups.
ModemManager, RMTFS, modem startup, USB SSH, Mesh connectivity and battery
guard were checked after recovery. Hardware diagnostics remain opt-in and
the production v2 transport guard remains enabled. Kernel module builds
passed normal modpost with matching Clang/KCFI. The final source also restores
the device DMA masks on normal cleanup and checks BAM errors during the
post-EOT data wait; that final cleanup revision is compile-checked, not
live-qualified. The 191-test regression suite
passed with one skip; project/public-index checks still have their existing
tracked-path/private-pattern failures and are not reported as passing.
Wi-Fi scanning works after recovery, but the saved hotspot was absent from
the scan results and reconnect timed out; automatic connection remains set.
Evidence is retained outside the public source tree in
`reports/cellular-20261112/`: `dma-test-live-path-fix.log`,
`ram-boot-r24-live.log`, `bam-inspect-reboot-pstore.log`,
`bam-ap-inspect-live.log`, `bam-command-bytes-live.log`,
`bam-command-low-iova-live.log`, and `final-device-recovery.log`.
The next hardware milestone is actual command-engine execution and DMA
writeback. Memory/table initialization, WAN RX/TX, modem QMI initialization,
RMNET/MAP delivery and restart recovery remain incomplete.
The subsequent source-only investigation adds explicit observations of BAM
control, configuration bits, descriptor summing threshold and the cached
core-clock request. The Android comparison is recorded in
[`INITIALIZATION.md`](../project/cellular/ipa-v2/INITIALIZATION.md), including
the v2.6L-specific configuration branch and IPA's explicit threshold of 16.
These added reads have not been compiled or live-tested. USB SSH stopped
responding during the follow-up connectivity check, and neither the direct
nor phone-proxied build connection was available. No new module was loaded
and no service or boot change was made in that follow-up; prior recovery
checks above describe the last verified state, not a new health check.
QMI configures the modem and its bearers. A WDS start-network response alone
does not provide an IP packet transport. The software experiment above uses
GLINK/PPP for payloads and QMI for modem control.
Primary references in the pinned Xiaomi source are
`arch/arm/boot/dts/qcom/sdm660.dtsi`,
`drivers/platform/msm/ipa/ipa_v2/ipa_reg.h`, `ipa.c` and `ipa_utils.c`.
The running kernel's `drivers/rpmsg/rpmsg_char.c`, `rpmsg_ctrl.c` and
`qcom_glink_native.c`, plus `include/uapi/linux/rpmsg.h`, define the RPMsg
binding, receive-intent and signal APIs used by this experiment.
## IPA v2.6L BAM transport qualification, 2026-11-17
Candidate recipe **7.2.12-r25** replaces the `r24` unsupported-transport guard
with an implemented SDM660 IPA v2.6L BAM transport (`ipa_v2_bam.c`) derived from
Xiaomi `platina-o-oss` (`0bb2780bccb9c7e05f0edf5db8860e779179b5f9`):
- **Device tree & clock/bus votes (`0042`, `0045`)**: IPA registers at
`0x147c0000`, BAM at `0x14784000`, level-high BAM IRQ (SPI 432), ANOC2 SMMU
stream `0x19c0` with level-high context fault IRQs, core clock (`RPM_SMD_IPA_CLK`,
75 MHz), and `a2noc`/`gnoc` interconnect bandwidth votes (`memory` and `config`).
- **AP/Modem ownership & XPU boundaries**: AP never asserts global IPA or BAM
SW resets while MSS (`4080000.remoteproc`) is running and never accesses
Q6/restricted IPA MMIO apertures (`0x2000..0x4fff`). Pipe 3 (`CMD_PROD`) drives
128-byte-aligned DDR/SRAM routing, header, and filter table initialization via
`BAM_DESC_FLAG_IMME` immediate commands before starting the QMI service `0x31`
handshake with MSS.
- **Filter polarity (`IPA_V2_FILTER`)**: On IPA v2.6L (`ipa_reg.h`), bit 0 of
`0x0048` (`IPA_V2_FILTER`) is `FILTER_DIS` (`0` enables filtering, `1`
disables it). Writing `0` ensures modem UL filter entries route QMAP traffic
to Q6 LAN consumer pipe 8 rather than falling back to AP pipe 2.
- **Idle UL aggregation flush (`ep[8]`)**: Q6 configures `ep[8]` (`Q6_LAN_CONS`)
with HW aggregation (`aggr = 0x1907f1`, `AGGR_SW_EOF_ACTIVE = 0`). After TX
ring completion leaves no pending TX slots, a workqueue waits for `ep[8]` to
exit suspend (`ENDP_INIT_CTRL(8)` bit 0 clear) and submits `IPA_REGISTER_WRITE`
(`opcode = 12`) on pipe 3 (`CMD_PROD`) to pulse `BIT(22)` (`AGGR_FORCE_CLOSE`)
on `IPA_ENDP_INIT_AGGR(8)` (`0x0340`), flushing isolated single UDP/SYN
packets immediately.
- **Single-frame QMAP RX & ring indexing**: Pipe 5 (`APPS_WAN_CONS`) uses
`IPA_BYPASS_AGGR` (`0`) to match `rmnet_ipa.c` non-deaggregated mode and
mainline `rmnet`'s default `qmapmux0.0` format, with exact 1:1 repost descriptor
buffer indexing in both `LAN_CONS` and `WAN_CONS`.
### Live verification on China Unicom 4G LTE (`46001`, APN `3gnet`)
Tested on the phone over `qmapmux0.0` (strict `SO_BINDTODEVICE` / `-I qmapmux0.0`):
- **Isolated single-packet UDP DNS from idle** (2.5 s sleep before each query,
confirming `ep[8]` wake and flush): `221.11.1.67` (`36.3 ms`), `223.5.5.5`
(`70.4 ms`), `119.29.29.29` (`63.6 ms`), all `rcode=0` with 3 answers on the
first packet.
- **TCP DNS & TCP+TLS HTTPS**: TCP DNS to `223.5.5.5:53` connected in `68.4 ms`
(`total 126.5 ms`, `rcode=0`); three HTTPS `GET /` requests to `www.baidu.com`
connected in `38.7–45.6 ms` and returned `HTTP/1.1 200 OK` (`18160` bytes each).
- **Sustained RX wraparound**: 100-packet ICMP ping (`-i 0.1 -c 100 223.5.5.5`)
completed with `100/100` received (`0%` loss, `min/avg/max = 34.5/35.5/52.2 ms`),
`lan_rx = 0`, and increasing TX/RX counters on both `rmnet_ipa0` and `qmapmux0.0`.
- **Disconnect/reconnect & coexistence**: Three consecutive
`platina-cellular --down` / `--up` cycles reassigned fresh IPv4 leases, passed
ICMP + idle single-packet UDP DNS, emitted zero MPSS/SMMU/IPA errors in
`dmesg`, and kept USB NCM SSH (`usb0`), EasyTier (`tun0`), `ath10k_snoc`
(`wlan0`), and `platina-battery-guard` operational. Route priority follows
Wi-Fi (`metric 600`) > `platina-cellular` (`ipv4.route-metric 700`) >
`platina-usb` (`route-metric=2000`), ensuring unbound traffic uses 4G LTE
rather than USB host sharing when Wi-Fi is disconnected.
## Completion criteria
Keep cellular Internet status unverified until an interface-bound public
DNS or TCP exchange succeeds, repeated calls recover reliably, and a
sustained test confirms actual IP receive counters increase. Only then add
an automatic service, route priority, DNS policy and EasyTier failover tests.
SIM-ready, LTE registration, IPCP and PPP control-byte counters are separate
milestones and cannot substitute for that final packet-delivery test.