s3lvol: rename COS to S3, add MinIO support, portable ISA and py3.8 RPC client

[fslongjin]

Summary

Rename the s3lvol backend from Tencent COS to generic S3 and add first-class MinIO support, across the target, scripts, tests and deploy tooling.

Highlights

• COS → S3 config: the per-node config moves from cos.cfg to s3.cfg with S3-style field names (access_key_id, secret_access_key, endpoint, buckets, path_style, no_tls). There is no fallback from the old file or the old keys. The rcow_add_cos_config RPC is renamed to rcow_add_s3_config.
• MinIO support: path_style / no_tls flags are threaded through the startup scripts, the purge path and an address-flags helper, so a path-style plain-HTTP backend works out of the box.
• Ensure-bucket tool: a stdlib SigV4 s3_bucket.py creates/confirms the dedicated s3lvol bucket (kept separate from the volume plugin's) before rcow_start.sh; no awscli or boto3 needed.
• CPU-portable releases: SPDK is built with --target-arch=haswell (x86_64) / armv8.2-a+crypto (aarch64) instead of native; make_release.sh refuses to package a native tree, and the one-click preflight checks for AVX2.
• Python 3.8 RPC client: a rpc.py launcher plus rpc_compat.py backfills argparse.BooleanOptionalAction so SPDK's unmodified client runs on Ubuntu 20.04; SPDK's own rpc.py ships as spdk_rpc.py.
• One-click deploy: install.sh renders s3.cfg from CUBE_S3_*, installs nvme-cli when missing, validates the packaged RPC client, and the supervisor preserves rcow_start's exit status so systemd Restart=on-failure retries.

Tests

• New offline suites: ISA baseline (test_isa_baseline.sh) and rpc.py py3.8 compat (test_rpc_py38_compat.sh), plus the s3_bucket.py self-test, all wired into run_all.sh.
• Extended test_minio_s3_guard.sh (s3.cfg rendering, bucket separation, nvme-cli installer, RPC client validation, supervisor exit code) and test_package_layout.sh (packaged bucket tool and RPC launcher).

[cubesandboxbot]

Review: s3lvol COS→S3 rename, MinIO support, portable ISA, py3.8 RPC client (PR #1524)

AI-generated review. Reviewed against the base branch (master) using the PR diff and the local base-branch workspace. No live test run was performed.

Verdict

Approve. This is a large PR (1398+/219−, 37 files) but internally consistent: the COS→S3 rename is complete, the new S3/MinIO wiring is correct, the Python 3.8 RPC launcher is a sound design, and the accompanying tests genuinely exercise the new code paths. I found no correctness bugs. Three low-severity observations follow.

What I verified

• Rename completeness. Every cos.cfg / RCOW_COS_CFG / cos_endpoint / cos_bucket_name / secretid / secretkey / rcow_add_cos_config reference is converted in the diff; no stale additions remain. The module's rcow_add_s3_config RPC, struct, decoders, handler and SPDK_RPC_REGISTER are renamed coherently, and the scripts/tests that call it are updated in lockstep.
• Supervisor exit-code fix is a real fix. The base if ! "${RCOW_START}"; then rc=$? captures the negated status (always 0), so systemd Restart=on-failure never retried. The new rc=0; "${RCOW_START}" || rc=$? is correct, and test_supervise_preserves_rcow_start_exit_code guards it.
• rcow_s3_addr_flags eval is guarded. The identifier whitelist (''|*[!A-Za-z0-9_]*|[0-9]*) plus fixed eval strings make the eval safe for the intended callers; every call site passes a hardcoded identifier and pre-initializes the array (so set -u is satisfied). The three expansion sites in rcow_purge.sh match the grep -c assertion in the test.
• Python 3.8 launcher works. argparse does not pass nargs/const to a custom Action class unless the caller supplied them explicitly (the store_true/store_const built-ins are proof), so rpc_compat.py's __init__ signature is sufficient. runpy.run_path(..., run_name='__main__') preserves sys.argv and stdin, so -s/-t arguments and batch-mode stdin both pass through. The launcher's search order mirrors rcow_common.sh's own SPDK_ROOT fallback chain, so the two stay consistent.
• s3_bucket.py SigV4 is correct. Canonical request shape, header ordering (host < x-amz-content-sha256 < x-amz-date), path-style vs virtual-hosted addressing, the us-east-1 LocationConstraint omission, and default HTTPS certificate verification are all right. Treating HEAD 403 as "bucket exists" is a safe heuristic here because the caller treats failure as non-fatal.
• ISA gate. The haswell/armv8.2-a+crypto pin and the native-refusal already existed in the base (setup_dep.sh / mk/s3lvol_isa_gate.sh / make_release.sh); the PR documents it and adds the AVX2 preflight plus tests. The hardcoded "14 passed" matches the 14 checks in test_isa_baseline.sh.
• One-click s3.cfg generation. The sentinel check correctly preserves hand-written files; quote rejection, endpoint scheme-stripping, and the path_style/no_tls derivation all match rcow_cfg_get / rcow_s3_buckets parsing.

Findings (low severity)

1. Repo-layout preflight no longer verifies the upstream SPDK rpc.py exists (CubeS3lvol/scripts/rcow_common.sh). Pointing RCOW_SPDK_RPC_PY at the always-present launcher means rcow_start.sh's [ -x "${RCOW_SPDK_RPC_PY}" ] passes even when SPDK_ROOT is wrong and no fallback exists; the failure then surfaces only on the first RPC call, i.e. up to ~60 s into rcow_wait_rpc (after the target has already been started and will be torn down). The new rcow_wait_rpc last-error capture does report the launcher's clear "cannot find SPDK rpc.py" message at that point, so this is a fail-fast regression with a much better late error. Consider a preflight that proves the launcher resolves its upstream (e.g. run the launcher with --help, like validate_s3lvol_rpc_client does for the package).
2. Dataplane test scripts bypass the py3.8 launcher (CubeS3lvol/test/dataplane/*.sh). They still invoke ${SPDK_ROOT}/scripts/rpc.py directly, so on Python 3.8 — the exact platform this PR enables — they would hit the BooleanOptionalAction error the launcher fixes. Not currently breaking (the tests run on the dev/CI host, which is 3.9+), but the runtime scripts and the test scripts now use different RPC entry points.
3. Nit: cos_config.err log filenames are retained in several dataplane tests after the rename. Cosmetic.

Notes

• This is an intentional breaking change: the config file and every key are renamed with no fallback, and the rcow_add_cos_config RPC is renamed rcow_add_s3_config. The target binary, the scripts, and the config must upgrade together — an old target with new scripts (or vice versa) fails loudly. The README / env.example document this clearly.
• The supervise bucket-ensure is correctly non-fatal and runs before rcow_start.sh, matching the "HEAD 403 = exists, MinIO still coming up" retry model.

[fslongjin]

On review finding 3 (ensure_nvme_cli / apt-get update -qq with no || die): install.sh is set -e, so a failed apt-get update already aborts the install — the gap is only a contextual die around apt's raw stderr. The same bare apt-get update -qq is used in install_docker(). Not changing that in this PR; follow-up.

[cubesandboxbot]

Low-severity observation: pointing RCOW_SPDK_RPC_PY at the always-present launcher changes the failure behavior of rcow_start.sh's preflight [ -x "${RCOW_SPDK_RPC_PY}" ]. That check now passes even when SPDK_ROOT is wrong and no fallback exists, so the problem surfaces only on the first rcow_srpc call — i.e. after the target is already started and rcow_wait_rpc polls for up to 60 s before bail tears it down. The new rcow_wait_rpc last-error capture does surface the launcher's clear "cannot find SPDK rpc.py" message at that point (better than the old silent timeout), but the fail-fast preflight is gone. A preflight that proves the launcher resolves its upstream — e.g. run it with --help, as validate_s3lvol_rpc_client already does for the package — would close the gap.

[cubesandboxbot]

Low-severity consistency note: this file (like the other test/dataplane/*.sh scripts) still invokes ${SPDK_ROOT}/scripts/rpc.py directly, so it bypasses the new scripts/rpc.py launcher. On Python 3.8 — the exact platform this PR makes the deployment support — that direct call hits the argparse.BooleanOptionalAction error the launcher was added to fix, so the dataplane suite would break on a 20.04 host even though the runtime scripts are 3.8-safe. Not currently breaking (the tests run on the dev/CI host, which is 3.9+), but the runtime and test paths now use different RPC entry points. Consider having these scripts resolve SPDK_RPC_PY through rcow_common.sh (or RCOW_SPDK_RPC_PY) so both stay in sync.