Claude review: None

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

Overall Series Review

Subject: None
Author: None
Patches: 0
Reviewed: 2026-02-15T19:01:01.789602

---

This 8-patch series adds batch userptr allocation support to amdkfd, allowing multiple non-contiguous CPU virtual address ranges to be mapped to a single contiguous GPU virtual address space via the existing `AMDKFD_IOC_ALLOC_MEMORY_OF_GPU` ioctl. The design uses a single MMU interval notifier covering the VA span with an interval tree for efficient range lookup during invalidation callbacks, which is a sound architectural choice.

The series has several significant issues that need addressing before it can be merged. The most critical are: (1) missing upper bound on `num_ranges` from userspace, allowing excessive kernel memory allocation; (2) the return value of `set_user_pages_batch()` is silently ignored in `init_user_pages_batch()`, which could lead to `ttm_bo_validate` operating on a corrupt pages array; (3) the error paths in `amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu_batch()` leak the mem limit reservation when `drm_vma_node_allow()` or `drm_gem_handle_create()` fails; and (4) the `size` parameter of `init_user_pages_batch()` is `uint32_t` but receives a `uint64_t` value, silently truncating sizes above 4GB.

Additionally, the patch references functions (`amdgpu_ttm_tt_get_user_pages_done`, `amdgpu_ttm_tt_discard_user_pages`) and uses types (`struct hmm_range *` in `user_range_info`) that do not match the upstream kernel tree's API (`amdgpu_hmm_range_alloc`/`amdgpu_hmm_range_free`, `struct amdgpu_hmm_range *`). This suggests the series was developed against an AMD staging tree with a different internal API, and needs reconciliation for upstream submission.

---
Generated by Claude Code Patch Reviewer