Claude review: accel/neutron: Add driver for NXP Neutron NPU

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

Patch Review

**Firmware ELF loading lacks bounds checking (security issue):** In `neutron_load_firmware()`, the ELF header fields are read from firmware data without validating them against `fw->size`:

```c
ehdr = (struct elf32_hdr *)fw->data;
if (memcmp(ehdr->e_ident, ELFMAG, SELFMAG) != 0) {
    ...
}

phdr = (struct elf32_phdr *)(fw->data + ehdr->e_phoff);
for (i = 0; i < ehdr->e_phnum; i++) {
    seg = &phdr[i];
    ...
    memcpy_toio(dest, fw->data + seg->p_offset, seg->p_filesz);
```

- `ehdr->e_phoff` is not validated against `fw->size`, so `phdr` could point past the firmware buffer.
- `ehdr->e_phnum` is not checked, so the loop could read past the buffer.
- `seg->p_offset + seg->p_filesz` is not validated against `fw->size`, so `memcpy_toio` could read OOB.
- There's no check that the firmware is even large enough for the ELF header (`fw->size >= sizeof(struct elf32_hdr)`).

A corrupt or maliciously crafted firmware file could trigger kernel memory reads. You need bounds checks before accessing every firmware-derived offset.

**`of_match_ptr()` usage:**

```c
.of_match_table	= of_match_ptr(neutron_match_table),
```

`of_match_ptr()` should not be used here. The driver `depends on ARCH_MXC` which always has OF, but `of_match_ptr` is still bad practice -- it causes an unused variable warning when `CONFIG_OF=n`. Just use `.of_match_table = neutron_match_table` directly. No other accel driver uses `of_match_ptr`.

**Return value of `dma_set_mask_and_coherent()` is not checked:**

```c
dma_set_mask_and_coherent(dev, DMA_BIT_MASK(48));
```

This can fail and the return value should be checked.

**`neutron_irq_handler_thread`:** Using a purely threaded IRQ handler (hardirq = NULL) with `IRQF_ONESHOT` is fine for a level-triggered interrupt, but means interrupt latency is subject to scheduling. This is probably acceptable for an NPU.

---
Generated by Claude Code Patch Reviewer