Add CI on linux-riscv64

[luhenry]

Proposed Changes

Types of Changes

Verification

Testing

Linked Issues

User-Facing Change

Further Comments

cc @pgonin

[luhenry]

Fixing missing gh cli at riseproject-dev/riscv-runner-images#17

[luhenry]

Looking at the timing differences between build-arm64 and build-riscv64:

Overall: riscv64 is ~10-11x slower than arm64 across the board.

Key metrics:

• Median per-package slowdown: 10.2x (very consistent)
• p90: 15.5x, p99: 25.6x, max: 42x
• Minimum: 1.5x, even the fastest riscv64 compiles are still slower

Top bottlenecks on riscv64:

Package	arm64	riscv64	Ratio
runtime	6.2s	57.1s	9.2x
k8s.io/api/core/v1	4.5s	53.5s	12.0x
reflect	3.2s	36.6s	11.6x
net	2.5s	26.3s	10.3x
net/http	1.8s	22.9s	12.7x

Wall-clock for key build steps:

• scripts/build: arm64=3m55s vs riscv64=40m28s (10.4x)
• scripts/package-cli: arm64=22s vs riscv64=2m31s (6.7x)
• docker buildx build Dockerfile.local: arm64=5m46s vs riscv64=48m32s (8.4x) — note these are cumulative buildkit step times

The slowdown is remarkably uniform at ~10-12x for compilation. The outliers (25-42x) tend to be tiny packages where the ratio is amplified by fixed overhead. The user time tracks closely with real time, confirming this is CPU-bound work, not I/O.

[luhenry]

I'm also adding kubectl to the base image at riseproject-dev/riscv-runner-images#24

[luhenry]

riscv64 CI Failure Classification

Run: https://github.com/luhenry/k3s/actions/runs/23554260675

Every failure is riscv64-only — all amd64 and arm64 jobs passed.

Category 1: Missing rancher/systemd-node:v0.0.8 image for riscv64 (7 jobs)

These fail because the rancher/systemd-node:v0.0.8 Docker image doesn't exist for riscv64:

Job	Error
hardened	Unable to find image 'rancher/systemd-node:v0.0.8'
autoimport	Unable to find image 'rancher/systemd-node:v0.0.8'
snapshotrestore	Unable to find image 'rancher/systemd-node:v0.0.8'
dualstack	Unable to find image 'rancher/systemd-node:v0.0.8'
svcpoliciesandfirewall	Unable to find image 'rancher/systemd-node:v0.0.8'
secretsencryption	Unable to find image 'rancher/systemd-node:v0.0.8'
token	Unable to find image 'rancher/systemd-node:v0.0.8'

Category 2: Missing rancher/k3s release images for riscv64 (2 jobs)

Job	Error
skew	Unable to find image 'rancher/k3s:v1.34.5-k3s1'
upgrade	Unable to find image 'rancher/k3s:v1.35.2-k3s1'

These tests need prior K3s release images which aren't published for riscv64.

Category 3: Missing rancher/mirrored-pause:3.6 image for riscv64 (4 jobs)

The rancher/mirrored-pause:3.6 manifest has no riscv64 platform entry (no match for platform in manifest: not found). This prevents any pod sandbox from starting, which causes coredns (and all other pods) to never come up, ultimately hitting the test timeout.

Job	Error
etcd	Timed out after 120.000s — failed to pull image "rancher/mirrored-pause:3.6": no match for platform in manifest: not found
basics	Timed out after 180.001s — failed to pull image "rancher/mirrored-pause:3.6": no match for platform in manifest: not found
bootstraptoken	Timed out after 120.000s — failed to pull image "rancher/mirrored-pause:3.6": no match for platform in manifest: not found
cacerts	failed to run command: docker run ... rancher/mirrored-pause:3.6, exit status 125

Category 5: Nix not supported on riscv64 (1 job)

Job	Error
nixsnapshotter	ArchOs (RISCV64-Linux) doesn't map to a supported Nix platform

Workaround: Disable nix jobs like on arm64

Category 6: Test setup issue — bind mount path doesn't exist (1 job)

Job	Error
lazypull	bind source path does not exist: /tmp/k3s-test-4264788357/server-0.yaml

Fix: riseproject-dev/riscv-runner-app@1fea2f7

Summary

Root Cause	Jobs
Missing rancher/systemd-node:v0.0.8 riscv64 image	7
Missing rancher/mirrored-pause:3.6 riscv64 platform in manifest	4
Missing rancher/k3s prior release images for riscv64	2
Nix doesn't support riscv64	1
Test setup bug (config file not created before bind mount)	1

[shanduur]

From my observation, compiling riscv64 on arm64 is little bit faster than riscv64 on amd64.

[luhenry]

From my observation, compiling riscv64 on arm64 is little bit faster than riscv64 on amd64.

@shanduur this is building on native RISC-V hardware, so no QEMU involved.