libmpi.so xxx symbol not found issue #1621

Activity

1. widgetii

   added this to firmwareon Apr 13, 2026
2. widgetii

   added

   platform:hisi

   platform:HiSiliconHiSilicon SoC issuesHiSilicon SoC issues

   and removed

   platform:hisi

   on Apr 13, 2026
3. widgetii

   added a commit that references this issue on Oct 16, 2026

   hisilicon-osdrv-hi3536dv100: take the modules from openhisilicon, and…

   ...

   fc30c88
4. widgetii commented on Oct 17, 2026

   

   widgetii

   on Oct 17, 2026

   Member

   More actions

   Fixed on master by #2614 (f182c17), with #2603 and OpenIPC/msposd#107.

   What was wrong. The package shipped userspace libraries and kernel modules from different MPP releases. libmpi.so came from 1.0.1.0 and the .ko files from 1.0.4.2, so the ioctl payloads disagree. For example, RGN_ATTR_S grew a u32CanvasNum field, so libmpi sends 0xc0185200 while region.ko only accepts 0xc01c5200 and answers -1. That is the RGN_Create failed with 0xffffffff reported in OpenIPC/msposd#18. The two kinds of error in the original log have different causes:

   • The jpeg_* / HI_JPEG_* relocations are libmpi's dependency on HiSilicon's own libjpeg.so (a jpeg6b build carrying the HI_JPEG_* extensions). The binary has to link -ljpeg, and the image has to ship the matching library.
   • MI_RGN_UpdateCanvas and unsupported relocation type 3 mean that binary was not a hi3536 build at all.

   What changed

   • One SDK release throughout: the userspace set is now from Hi3536DV100 SDK 2.0.2.0 (libmpi, libjpeg, libhdmi, libdnvqe, libupvqe, libVoiceEngine), and so are the kernel modules.
   • The modules are built by OpenIPC/openhisilicon, against the board's kernel (added radio_port, udp_port and fec settings to wfb.conf #266, Add mavlink-router package to fpv profile #267, rewrite config read, added guard_interval and ldpc #268 there).
     • OSAL is built from the SDK's source. The vendor's prebuilt hi_osal.ko registers its devices without a dev_t on OpenIPC's kernel, so /dev/sys, /dev/rgn and the rest never appeared.
     • The closed MPP modules are relinked from the same SDK's .ko files.
   • load_hisilicon loads them as open_* with modprobe. It also unloads vdec before vfmw, which avoids an oops on -r.
   • Display buffers: the 2.0.2.0 VO driver refuses fewer than 3. vdec falls back to 3 since vdec: fall back to 3 display buffers when the VO driver refuses 2 research#57.
   • msposd: firmware builds its hi3536 target on this family (msposd: build the hi3536 target on hi3536dv100 #2603). On an NVR it draws on the hifb layer (/dev/fb0) rather than an RGN overlay (hi3536: draw the OSD on the hifb layer instead of an RGN overlay msposd#107), because the NVR has no encoder channel to attach an overlay to.

   Tested on a lab Hi3536DV100 NVR:

   • all 22 modules load, and the MPI probe returns 0 on every call;
   • vdec configures HDMI at 720p60;
   • msposd's OSD is drawn, read back from /dev/fb0.

   That covers the board in our lab. If yours behaves differently on an image built from master after f182c17, please reopen with dmesg and the output of the failing command.
5. widgetii

   closed this as completedon Oct 17, 2026