From mboxrd@z Thu Jan  1 00:00:00 1970
From: Claude Code Review Bot <claude-review@example.com>
To: dri-devel-reviews@example.com
Subject: Claude review: drm/gpusvm: Use dma-map IOVA alloc, link, and sync API in GPU SVM
Date: Sun, 12 Apr 2026 12:09:35 +1000
Message-ID: <review-patch1-20260408201537.3580549-2-matthew.brost@intel.com>
In-Reply-To: <20260408201537.3580549-2-matthew.brost@intel.com>
References: <20260408201537.3580549-1-matthew.brost@intel.com> <20260408201537.3580549-2-matthew.brost@intel.com>
X-Mailer: Claude Code Patch Reviewer
Content-Type: text/plain; charset="utf-8"
Content-Transfer-Encoding: 7bit
MIME-Version: 1.0

Patch Review

**Reviewed-by: Thomas Hellstrom** -- good to see.

**The try/fallback pattern is correct.** The code calls `dma_iova_try_alloc` on the first iteration (`!i`), then checks `dma_use_iova(state)` to decide the path. This matches the documented API usage where `dma_iova_try_alloc` returns false if the device doesn't support the IOVA API.

**Nit: Misleading `phys` parameter:**
```c
+			if (!i)
+				dma_iova_try_alloc(gpusvm->drm->dev, state,
+						   npages * PAGE_SIZE >=
+						   HPAGE_PMD_SIZE ?
+						   HPAGE_PMD_SIZE : 0,
+						   npages * PAGE_SIZE);
```
The `phys` parameter to `dma_iova_try_alloc` is documented as "only used to calculate the IOVA alignment. Callers that always do PAGE_SIZE aligned transfers can safely pass 0 here." Since all transfers here are PAGE_SIZE-aligned, `HPAGE_PMD_SIZE` and `0` both produce `iova_offset() == 0` -- the conditional is dead logic. Other callers (HMM, mlx5, dma-buf) just pass `0` or the actual physical address of the first page. Passing `HPAGE_PMD_SIZE` as a fake physical address is confusing. Consider just passing `0`.

**Unmap path looks correct:**
```c
+		bool use_iova = dma_use_iova(&svm_pages->state);
+
+		if (use_iova) {
+			dma_iova_unlink(dev, &svm_pages->state, 0,
+					svm_pages->state_offset,
+					svm_pages->dma_addr[0].dir, 0);
+			dma_iova_free(dev, &svm_pages->state);
+		}
```
The bulk unlink + free before the per-page loop, with the per-page `dma_unmap_page` skipped via `!use_iova`, is clean.

**Minor: Over-allocation of IOVA space.** `dma_iova_try_alloc` allocates `npages * PAGE_SIZE`, but some entries in the pfn array may be device-private pages (handled differently) or NULL pages, meaning the IOVA space won't be fully used. The `state_offset` correctly tracks only the actually-linked portion. Not a bug, just a minor waste of IOVA address space.

**Error path:** On `err_unmap`, `__drm_gpusvm_unmap_pages` is called which handles the IOVA unlink/free. On `retry`, `state` is re-zeroed before `map_pages:`. Looks correct.

---
Generated by Claude Code Patch Reviewer