load_hisilicon (cv300): insmod cma_osal.ko but only hi_osal.ko shipped

[widgetii]

Summary

load_hisilicon on hi3516cv300 references a cma_osal.ko that the firmware never builds or ships. When the kernel cmdline contains mmz=… (which the qemu-hisilicon machine emulator and most production U-Boots do), insert_osal takes the CMA branch:

insmod: can't insert 'cma_osal.ko': No such file or directory
******* Error: There's something wrong, please check! *****

load_hisilicon then report_error → exit 1 and zero modules load.

The cv300 rootfs ships hi_osal.ko only (verified in /lib/modules/4.9.37/hisilicon/ of openipc.hi3516cv300-nor-lite.tgz); there is no cma_osal.ko anywhere in the build.

This bug pre-dates #2059 — it was hidden on cv300 because load_hisilicon used to bail at the os_mem= guard before ever reaching insert_osal. Post-#2060, the guard is correct and the cma_osal call now runs.

Where

general/package/hisilicon-osdrv-hi3516cv300/files/script/load_hisilicon, lines 58 / 60:

insert_osal() {
    MMZ=$(awk -F '=' '$1=="mmz"{print $2}' RS=" " /proc/cmdline)
    if [ -z "$MMZ" ]; then
        insmod hi_osal.ko anony=1 mmz_allocator=hisi mmz=anonymous,0,$mmz_start,$mmz_size || report_error
    else
        insmod cma_osal.ko anony=1 mmz_allocator=cma mmz=$MMZ || report_error    # <-- file doesn't exist
    fi
}

Suggested fix

Two options, pick one:

1. Use hi_osal.ko in both branches, passing the right args. The cv300 OSAL module already has mmz_allocator= parameter support; passing mmz_allocator=cma mmz=$MMZ to hi_osal.ko should drive the same code path the (hypothetical) cma_osal.ko would. Tracking issue mmz: fail allocator init when every configured zone is rejected openhisilicon#73 made the V3 cma allocator init fail fast on register error, so this path is exercised correctly upstream.

2. Build and ship cma_osal.ko as a separate module from openhisilicon for V3 / cv300, and add it to hisilicon-osdrv-hi3516cv300's install list.

Option 1 is the smaller change and keeps cv300's rootfs footprint identical.

CI hardcoding to undo once fixed

OpenIPC/openhisilicon/.github/workflows/build.yml qemu-boot matrix has:

- machine: hi3516cv300
  firmware: hi3516cv300
  mem: 128M
  allow-failure: true                # <-- remove this line
  append: "console=ttyAMA0,115200 mem=32M root=/dev/ram0 rootfstype=squashfs"

After this issue is fixed and a new latest firmware release is published:

1. Drop the allow-failure: true line from the cv300 matrix entry.
2. Optionally add min_modules: N to gate the lsmod count assertion.
3. Drop the explanatory comment block above the cv300 row.

Refs

• Discovered while landing ci: add hi3516ev300_neo qemu-boot + lsmod-count regression probe openhisilicon#86.
• Same root cause as load_hisilicon (cv500): insmod cma_osal.ko but only hi_osal.ko shipped #2062 (cv500 also insmod cma_osal.ko, also ships only hi_osal.ko).