en: MT7601U driver performance note; VStarcam Hi3518EV200+GC2023 device page

README.md

@@ -138,6 +138,7 @@ OpenIPC Wiki
 - [Smartwares CIP-37210](en/device-smartwares-cip-37210.md)
 - [Foscam X5](en/device-foscam-x5.md)
 - [TP-Link Tapo C110](en/device-tapo-c110.md)
+- [VStarcam pan/tilt (Hi3518EV200 + GC2023)](en/device-vstarcam-hi3518ev200-gc2023.md)
 - [Xiaomi Mi Camera 2K (MJSXJ03HL)](https://github.com/OpenIPC/device-mjsxj03hl/)
 - [ATOM Cam / Hualai / Wyze](en/device-wyze-integration.md)
 - [Digitus DN-16048 Optizoom](en/device-digitus-dn16048.md)

en/adding-wifi-driver.md

@@ -72,6 +72,21 @@ BR2_PACKAGE_SSV635X_OPENIPC
 BR2_PACKAGE_MT7601U_OPENIPC
 ```
 
+Note: for MT7601U there are two different drivers. The mainline kernel driver
+(`mt7601u.ko`, enabled via the kernel config) and the vendor-lineage MediaTek
+DPO STA driver (`mt7601sta.ko`, built by `BR2_PACKAGE_MT7601U_OPENIPC`, with
+cfg80211 support so the interface is a plain `wlan0` and wpa_supplicant drives
+it as usual). They are not equivalent: on a Hi3518EV200 camera with a USB
+`148f:7601` dongle, measured on the same access point and channel, the mainline
+driver managed only 0.4–1.0 Mbps transmit (with repeated `mt7601u_mcu_wait_resp
+retrying` in dmesg and gateway RTT spikes to 800+ ms), while `mt7601sta` reached
+15–17 Mbps transmit with clean 9.7 ms RTT — the difference between an RTSP
+stream that lags by seconds and one that keeps its send queue empty. If your
+video stalls on this chip while download speed looks fine, measure the upload
+direction and try the STA driver. Do not swap the two drivers on a live system
+(rmmod/modprobe churn can wedge the dongle MCU until a power cycle) — pick one
+at boot.
+
 #### SigmaStar:
 ```
 BR2_PACKAGE_SSW101B

en/device-vstarcam-hi3518ev200-gc2023.md

@@ -0,0 +1,107 @@
+# OpenIPC Wiki
+[Table of Content](../README.md)
+
+VStarcam pan/tilt cameras (Hi3518EV200 + GC2023 + MT7601U)
+---
+
+Notes from a full conversion of a VStarcam indoor pan/tilt camera (P2P UID
+prefix `VSTA...`, close sibling of the C8892WIP in the builder device list).
+Everything below was verified on real hardware.
+
+### Hardware
+
+| Component | Detail |
+|-----------|--------|
+| SoC       | HiSilicon Hi3518EV200 |
+| Sensor    | GalaxyCore GC2023, 1080p, wired to **MIPI** (2-lane RAW10), SDK I2C address `0x6e` |
+| Flash     | 16 MB SPI NOR (GD25Q128, `0xc8 0x40 0x18`) |
+| Wi-Fi     | MediaTek MT7601U on USB (`148f:7601`) |
+| PTZ       | Two 28BYJ-48-style steppers on GPIOs, pan `31 32 33 34`, tilt `58 59 60 61` |
+| Serial    | ttyAMA0 at 115200 8N1; U-Boot 2010.06, prompt `hisilicon #`, `bootdelay=1` |
+
+Vendor firmware serial login is `vstarcam2017` / `20170912` (uid 0) — the
+username is not `root`; `root` is always rejected.
+
+### Installing while keeping the vendor U-Boot
+
+The vendor U-Boot stores its environment at offset `0x80000`, which sits inside
+the kernel region of OpenIPC's standard 16 MB layout — with standard offsets a
+later `saveenv` would write the environment straight through the kernel. Keeping
+the vendor U-Boot (and therefore the vendor recovery path) works fine with a
+shifted layout:
+
+```
+setenv bootargs 'mem=32M console=ttyAMA0,115200 panic=20 root=/dev/mtdblock2 rootfstype=squashfs init=/init mtdparts=hi_sfc:1024k(boot),3072k(kernel),8192k(rootfs),-(rootfs_data)'
+```
+
+- Kernel at `0x100000`, rootfs at `0x400000`. The vendor `bootcmd`
+  (`sf read 0x82000000 0x100000 0x300000; bootm`) needs no change at all —
+  OpenIPC's uImage uses the same `0x80008000` load address as the vendor kernel.
+- `root=/dev/mtdblock2`, not the `mtdblock3` from the upstream install guide —
+  this layout has no separate `env` partition, so every index shifts down one.
+- The overlay partition must be named exactly `rootfs_data` and boot needs
+  `init=/init`, or you get a read-only system with no persistent config.
+
+Two U-Boot traps worth knowing before you flash:
+
+- **This U-Boot has no `tftpdstp`** (`strings` on the boot partition shows only
+  `tftpblocksize`/`tftptimeout`), so it silently retries against anything that
+  is not port 69. Run the TFTP server on the standard port.
+- The stripped build has no `boot` or `run` commands — invoke the `bootcmd`
+  steps by hand.
+
+### Sensor: GC2023 over MIPI
+
+Use the `gc2023_mipi` sensor identity and `gc2023_mipi_1080p.ini` (2-lane
+RAW10, `Isp_Bayer=0`/RGGB). The plain `gc2023` entry in `load_hisilicon` muxes
+the parallel VI pads, which disables the MIPI PHY on this board — the tell is
+`/proc/umap/vi` showing `IntCnt=0` with `TmgErr=0` and `ccErrN=0` (no VSYNC at
+all; bad ini timings would raise the error counters instead).
+
+Caveat: on the tested unit the closed `libsns_gc2023.so` shipped with the
+firmware did not produce frames (VENC timeout — it appears to program the
+sensor for DVP output). The working setup used the *vendor's* MIPI sensor
+library plus a replay of its 123-write init table, because that library leaves
+digital gain (page 0 register `0xb1`) at 0 — which multiplies every pixel to
+zero: perfectly valid MIPI framing carrying an all-black raster. If you see
+valid framing but black frames, check `0xb1` before suspecting optics or IQ.
+
+The sensor answers only on the SDK's own I2C (via `ipctool i2cget`, 8-bit
+address `0x6e`) — sweeping `/dev/i2c-*` finds nothing and proves nothing.
+
+### Wi-Fi
+
+Build with `BR2_PACKAGE_MT7601U_OPENIPC=y` and load `mt7601sta`, not the
+mainline `mt7601u` — see the driver note in
+[Adding a wifi driver](adding-wifi-driver.md): mainline managed 0.4–1.0 Mbps
+transmit on this hardware, the STA driver 15–17 Mbps on the same dongle and AP.
+
+If both eth0 and wlan0 are up, make sure they do not share a DHCP hostname:
+busybox `ifup` sends the same hostname on every interface, and a DHCP server
+that keys reservations on hostname will hand both interfaces the same address,
+which blackholes the camera (ARP answered by the wrong interface).
+
+### PTZ
+
+`fw_setenv gpio_motors "31 32 33 34 58 59 60 61"` drives the head with the
+`gpio-motors` package. The pins came from the vendor's `motogpio.ko`: it never
+calls `ioremap` and uses static-mapping virtual addresses, so subtracting the
+fixed `phys 0x20000000 → virt 0xFE100000` offset turns its literal pool back
+into real GPIO banks. There are no working limit switches — the vendor homes by
+driving into the mechanical stop and counting steps back out.
+
+### IR-cut and IR LED
+
+The IR-cut solenoid is bi-stable: pulse sysfs GPIO **64** for ~200 ms to switch
+to day (filter in), pulse GPIO **65** for night (filter out). The IR illuminator
+is a plain level on sysfs GPIO **1**. Both filter pins power up muxed as
+`SAR_ADC` channels, so nothing responds until the pinmux is set:
+
+```
+devmem 0x200F0100 32 1   # MUXCTRL_REG64 -> GPIO8_0
+devmem 0x200F0104 32 1   # MUXCTRL_REG65 -> GPIO8_1
+```
+
+Do not brute-force sweep GPIOs looking for pins on this board — some lines are
+sensor power/reset, and a sweep that leaves one asserted breaks sensor I2C until
+a power cycle.