Claude review: drm/msm/a6xx: Allow IFPC with perfcntr stream

[Claude Code Review Bot <claude-review@example.com>]

Patch Review

Clean final patch. The `sysprof_setup(gpu, true/false)` bracketing in `sample_worker` temporarily inhibits IFPC only during register reads, which is the correct approach. The `msm_gpu_sysprof_no_ifpc()` simplification to only check the legacy SYSPROF refcount (no longer checking for active perfcntr stream) is correct because the kernel now handles IFPC itself via the force_on mechanism.

---

**Summary of actionable items**:

1. **Patch 13 (CRITICAL)**: `copy_from_user(&g, userptr, args->group_stride)` — no upper bound on `group_stride`, stack buffer overflow possible. Should use `min(args->group_stride, sizeof(g))`.

2. **Patch 13 (CRITICAL)**: `1 << args->bufsz_shift` — undefined behavior for large shift values. Validate `bufsz_shift` early (e.g., `bufsz_shift > 27` → EINVAL).

3. **Patch 13 (bug)**: Non-blocking read on empty FIFO returns 0 instead of `-EAGAIN`.

4. **Patch 10 (minor)**: `WARN_ON_ONCE(n > stream->nr_groups)` should be `>=`.

5. **Patch 13 (minor)**: Include style should use angle brackets for kernel headers.

6. **Patch 13 (minor)**: UAPI doc typo: "read from the has" → "read from the fd has".

7. **Patch 13 (minor)**: Counter allocation state not rolled back on partial failure in the ioctl loop.

---
Generated by Claude Code Patch Reviewer