nvidia-drm: propagate errno from semaphore surface fence wait ioctl

[VaggelisGian]

Found while investigating #1297.

Problem

nv_drm_semsurf_fence_wait_ioctl() initializes ret = -EINVAL, sets it to 0 on success paths, but then ends in a bare return 0. Every validation failure logs an error and reports success to userspace even though no wait was registered with RM:

• non-monotonic pre_wait_value >= post_wait_value
• unknown fence context handle
• wrong fence context type
• callback data allocation failure
• invalid sync FD (nv_drm_sync_file_get_fence() returned NULL)

A caller that trusts the return value believes its wait is armed. The semaphore it waits on is never signaled, and the caller hangs. This is exactly the sequence in #1297's log: Attempt to wait on invalid sync FD: 1511 followed by a hung render thread.

Every sibling ioctl in the file (SEMSURF_FENCE_CREATE, SEMSURF_FENCE_ATTACH, the prime fence ioctls) propagates its errno the same way this one already computes but discards it, and the first check in this very function (pDevice == NULL) already returns -EOPNOTSUPP.

Change

One line: done: now does return ret;.

• Success paths still return 0, including the deliberate proceed-as-signaled path where dma_fence_add_callback() fails after the wait was registered with RM (there the wait IS registered, so reporting success is correct).
• Error paths return -EINVAL; no path changes from nonzero to zero.
• Ownership is unchanged: on error paths wait_data is NULL or freshly allocated and unshared when it is freed; the inline work-callback path resets ret = 0 before the label.
• The ioctl number and params struct are untouched.

Test Plan

Prepared on Windows without Linux kernel toolchain or NVIDIA GPU, so no build and no runtime test were possible; verification was static:

git diff --stat
 kernel-open/nvidia-drm/nvidia-drm-fence.c | 2 +-

git diff --check        # clean

Traced all eight paths through the function for ret/wait_data state at the label, confirmed the sole caller is the DRM ioctl dispatch table, and searched open-source userspace (Mesa, libdrm, Proton ecosystem) for callers of DRM_IOCTL_NVIDIA_SEMSURF_FENCE_WAIT: none found outside header vendoring, so the only caller affected is NVIDIA's own userspace, which already handles nonzero returns from this same function. A repro of #1297 with the patched module should show the game receiving an error at wait time instead of hanging.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}