Claude review: drm/tyr: add GPU virtual memory module

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

Patch Review

**Author:** Boris Brezillon / Daniel Almeida / Deborah Brouwer

This is the largest patch. The VM module integrates GPUVM with the hardware page table.

Observations:

1. `reserve_range` is set to `0..0u64` (empty range):

```rust
+        let reserve_range = 0..0u64;
```

This means no VA space is reserved for the kernel. For the MCU firmware VM this is fine since the driver controls all mappings, but for user VMs this will need to be revisited.

2. In `sm_step_map`, when a mapping fails partway through SGT entries, the cleanup unmaps from `start_iova`:

```rust
+                Err(e) => {
+                    let total_mapped = iova - start_iova;
+                    if total_mapped > 0 {
+                        pt_unmap(context.pt, start_iova..(start_iova + total_mapped)).ok();
+                    }
+                    return Err(e);
+                }
```

This is correct — it rolls back successful mappings from previous SGT entries. However, the `.ok()` silently ignores unmap failures during error recovery. A `pr_warn!` or similar would help debug issues.

3. The `get_pgsize()` function handles 4K and 2M pages:

```rust
+fn get_pgsize(addr: u64, size: u64) -> (u64, u64) {
+    let blk_offset_2m = addr.wrapping_neg() % (SZ_2M as u64);
```

Using `wrapping_neg()` to compute distance to next boundary is idiomatic. The function also caps the number of 2M pages at the 1G boundary, which is correct for AArch64 LPAE page tables where the next level table covers 1G.

4. The `PtUpdateContext::drop()` calls both `end_vm_update()` and `flush_vm()`, logging errors but not propagating them. This is unavoidable in `Drop`, but means that if the flush fails, GPU state may be inconsistent.

5. In `pt_map()`:

```rust
+        let (mapped, result) = unsafe {
+            pt.map_pages(
+                curr_iova as usize,
+                (curr_paddr as usize).try_into().unwrap(),
```

The `unwrap()` on `try_into()` could panic if `curr_paddr` doesn't fit in the target type of `PhysAddr`. On 32-bit systems this could be an issue, but the driver requires ARM64, so this is safe in practice. Worth a comment.

6. The SAFETY comments for `pt.map_pages()` and `pt.unmap_pages()` reference "PtUpdateContext" ensuring exclusive access. This is sound because the MMU lock + start_vm_update() sequence prevents concurrent page table modifications.

---
Generated by Claude Code Patch Reviewer