Claude review: gpu: nova-core: use the Generic Bootloader to boot FWSEC on Turing

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

Patch Review

The largest new-code patch. Well structured.

**`BootloaderDmemDescV2` is `repr(C, packed)`**: The `AsBytes` safety comment says "doesn't contain uninitialized bytes", which is correct for a packed struct with no padding. However, the `packed` attribute means accessing fields creates unaligned references. The code only accesses fields during construction (by value) and through `as_bytes()`, so this is fine in practice.

**Mirror image logic**: The comment and code correctly explain the DMA firmware object layout:
```rust
// `BootloaderDmemDescV2` expects the source to be a mirror image of the destination
// and uses the same offset parameter for both.
```
The padding computation `imem_sec.dst_start.checked_sub(imem_sec.src_start)` ensures the DMA object layout mirrors IMEM. This is well-documented and handles the alignment requirement.

**`_firmware_dma` field**: The leading underscore is appropriate — this DMA object must be kept alive for the bootloader to DMA from, but the driver doesn't directly access it after creation. Good pattern.

**Boot address**: The boot vector computation `u32::from(self.start_tag) << 8` is correct — tags are 256-byte aligned (matching `MEM_BLOCK_ALIGNMENT`).

**Error path in `run`**: The `mbox0 != 0` check and `EIO` return is correct for FWSEC error reporting.

**`firmware.ucode.0` access**: In `FwsecFirmwareWithBl::new`, the code accesses `firmware.ucode.0` directly:
```rust
firmware_obj.extend_from_slice(firmware.ucode.0.as_slice(), GFP_KERNEL)?;
```
This reaches into the private `FirmwareObject` tuple struct field. This works because `bootloader.rs` is a submodule of `fwsec`, but it does bypass the abstraction boundary. Using `firmware.as_slice()` (from the `FalconDmaLoadable` trait) would be cleaner.

**`gsp/boot.rs` changes**: The `chipset` parameter is properly threaded through to `run_fwsec_frts`, and the branching logic is clear and correct.

Overall this is a solid series. The main actionable items are:
1. Unchecked addition in `try_as_pio_loadable` (patch 6)
2. `saturating_sub`/`saturating_add` vs `checked_sub`/`checked_add` in V2 header offsets (patch 9)
3. Mbox patch ordering (7 and 8 are swapped)

---
Generated by Claude Code Patch Reviewer