Claude review: drm/msm: Add PERFCNTR_CONFIG ioctl

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

Overall Series Review

Subject: drm/msm: Add PERFCNTR_CONFIG ioctl
Author: Rob Clark <robin.clark@oss.qualcomm.com>
Patches: 17
Reviewed: 2026-05-16T10:50:07.754640

---

This is a well-structured 16-patch series by Rob Clark that adds a new `PERFCNTR_CONFIG` ioctl to the MSM/Adreno DRM driver, enabling both global GPU performance counter streaming (for system profilers, restricted to `perfmon_capable()`) and local per-context counter reservation (for UMD per-draw profiling). The series is at v6, has accumulated reviews from Dmitry Baryshkov and Anna Maniscalco, and shows a mature design that has evolved through multiple iterations addressing real issues (null pointer deref on older gens, FIFO size limits, loop counter fixes, IFPC support).

**Strengths:**
- Clean separation: removes old dead perf infrastructure first, then builds new subsystem incrementally
- Sensible UABI design: fd-based stream lifetime, group-name-based counter identification, extensible via flags/stride
- Thoughtful concurrency: global vs. local counter conflict detection, proper ordering of IFPC pwrup reglist updates, per-context reservation model
- Good use of `__free(kfree)` cleanup guards in the ioctl path

**Concerns (detailed per-patch below):**
- A few locking/ordering issues in the ioctl error paths
- The `msm_perfcntrs_stream_read()` does not handle wrap-around properly (only returns up to end of buffer)
- The `regs[62]` -> `regs[]` flex array change in `cpu_gpu_lock` may need coordinated size validation
- Minor UABI nit: `group_stride` could allow future extension but also allows reading less than the struct size

Overall this is solid work from an experienced kernel developer. The issues I found are minor and the design is sound. With the few items below addressed, this looks ready.

---
Generated by Claude Code Patch Reviewer