shim: gracefully skip optional telemetry queries when the ioctl is unsupported (#1447 follow-up)

[Scottcjn]

Follow-up to #1447 (in-tree amdxdna ≤6.17 missing GET_ARRAY/AIE2 telemetry ioctls).

Problem. When the loaded kernel driver predates DRM_AMDXDNA_QUERY_TELEMETRY, the ioctl returns EOPNOTSUPP/EINVAL. The SHIM's telemetry query handlers call it unconditionally, so shim_err throws and the entire xrt-smi report aborts — even though basic BO/exec paths work fine on that driver.

Fix. A small helper (telemetry_handler::telemetry_query) wraps the optional telemetry ioctl: on EOPNOTSUPP/ENOTSUP/EINVAL/ENOTTY it logs a one-time warning and returns false, so the caller returns an empty result and the rest of the report continues. Any other (real) error is rethrown. Applied to all 9 optional telemetry query handlers (aie/misc/opcode/rtos/stream_buffer, incl. the aie4 overrides). Required paths are untouched.

This turns #1447 from "this ioctl is missing" into "the SHIM handles it safely on older drivers."

Testing.

• Compiles cleanly under the repo's -Werror release build (cmake --build ... --target xrt_driver_xdna). The -Werror=maybe-uninitialized check caught that misc_telemetry's scalar result needed value-initialization on the new early-return path — fixed (result_type output{}).
• Surgical/no-regression by construction: on a telemetry-capable driver the try succeeds and behavior is unchanged.
• Disclosure: I could not live-reproduce the unsupported-ioctl path on my test box — the in-tree 6.17 driver had been replaced by the DKMS staging build, and the other installed in-tree modules are for different kernels (vermagic mismatch). The runtime degradation is straightforward (standard errno catch) but is best confirmed against an older in-tree driver / in CI. Happy to iterate.

Surfaced during an open-source XDNA1 (Phoenix) Linux bring-up.

[maxzhen]

ok to test

[Copilot]

Pull request overview

This PR updates the XDNA SHIM telemetry query path to gracefully handle older kernel drivers that don’t support DRM_AMDXDNA_QUERY_TELEMETRY, preventing xrt-smi from aborting the entire report when optional telemetry ioctls are unsupported.

Changes:

• Add a telemetry_handler::telemetry_query() helper that treats EOPNOTSUPP/ENOTSUP/EINVAL/ENOTTY as “telemetry unsupported” and returns an empty telemetry result instead of throwing.
• Apply the helper across the telemetry query handlers (AIE2 + AIE4), and value-initialize misc_telemetry outputs to support early returns safely.
• Add required headers for errno handling, std::system_error, and one-time logging (std::once_flag).

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[maxzhen]

It seems that the fix should be in xrt-smi, not in SHIM. The SHIM should faithfully signaling the caller and return whatever error it sees. It is the caller to decide if this is optional or required.

[Scottcjn]

You're right, thanks. Moved it out of the shim.

The shim goes back to just returning whatever errno the ioctl gives, no swallowing. The optional decision now lives in the caller in XRT: Xilinx/XRT#9907.

It skips telemetry only when the driver reports it unsupported (EOPNOTSUPP, or ENOTTY if the ioctl is not registered), which is what aie2_get_info returns from its default case on an older driver. Everything else including EINVAL still surfaces so real errors are not hidden. I gave the preemption path the same handling since it hits the same rtos_telemetry query.

That leaves nothing in this shim PR, so closing it in favor of the XRT one.