[PATCH v2] dma-buf: Fix silent overflow for phys vec to sgt

[PATCH v2] dma-buf: Fix silent overflow for phys vec to sgt

[David Hu]

In case MMIO size is bigger than 4G and peer2peer DMA goes
through host bridge, we trigger a code path that assigns the
total linked IOVA (which is greater than 4G) to mapped_len.

Previously, `mapped_len` was declared as 32-bit `unsigned int`.
When accumulating `size_t` lengths, this leads to a silent wrap-around.
This truncation causes truncated lengths to be passed to functions
like `fill_sg_entry()`.

Fix this by changing `mapped_len` to `size_t` (64-bit). While
at it, fix similar potential overflow issues in `calc_sg_nents`
by using `size_t` for `nents` and checking against `UINT_MAX`
and using `unsigned int` for the loop iterator in `fill_sg_entry`
to match.

to mapped_len, and leading to a silent overflow

Fixes: 3aa31a8bb11e ("dma-buf: provide phys_vec to scatter-gather mapping routine")
Cc: stable@vger.kernel.org
Cc: iommu@lists.linux.dev
Reviewed-by: Pranjal Shrivastava <praan@google.com>
Signed-off-by: David Hu <xuehaohu@google.com>
---
Changes in v2:
 Fixed 'IVOA' -> 'IOVA' typo and expanded commit message (Claude Bot).
 Added Reverse Xmas tree formatting (Pranjal).
 Folded in extra bounds checking for calc_sg_nents() (Pranjal).
 Folded in type consistency fix for fill_sg_entry() (Pranjal).
 Droped unnecessary `nents = 0` initialization (Claude Bot).

 drivers/dma-buf/dma-buf-mapping.c | 10 +++++++---
 1 file changed, 7 insertions(+), 3 deletions(-)

diff --git a/drivers/dma-buf/dma-buf-mapping.c b/drivers/dma-buf/dma-buf-mapping.c
index 794acff2546a..5bc769fc42ea 100644
--- a/drivers/dma-buf/dma-buf-mapping.c
+++ b/drivers/dma-buf/dma-buf-mapping.c
@@ -10,7 +10,7 @@ static struct scatterlist *fill_sg_entry(struct scatterlist *sgl, size_t length,
 					 dma_addr_t addr)
 {
 	unsigned int len, nents;
-	int i;
+	unsigned int i;
 
 	nents = DIV_ROUND_UP(length, UINT_MAX);
 	for (i = 0; i < nents; i++) {
@@ -36,7 +36,7 @@ static unsigned int calc_sg_nents(struct dma_iova_state *state,
 				  struct phys_vec *phys_vec, size_t nr_ranges,
 				  size_t size)
 {
-	unsigned int nents = 0;
+	size_t nents = 0;
 	size_t i;
 
 	if (!state || !dma_use_iova(state)) {
@@ -51,6 +51,9 @@ static unsigned int calc_sg_nents(struct dma_iova_state *state,
 		nents = DIV_ROUND_UP(size, UINT_MAX);
 	}
 
+	if (nents > UINT_MAX)
+		return 0;
+
 	return nents;
 }
 
@@ -95,9 +98,10 @@ struct sg_table *dma_buf_phys_vec_to_sgt(struct dma_buf_attachment *attach,
 					 size_t nr_ranges, size_t size,
 					 enum dma_data_direction dir)
 {
-	unsigned int nents, mapped_len = 0;
 	struct dma_buf_dma *dma;
 	struct scatterlist *sgl;
+	size_t mapped_len = 0;
+	unsigned int nents;
 	dma_addr_t addr;
 	size_t i;
 	int ret;
-- 
2.54.0.794.g4f17f83d09-goog

Claude review: dma-buf: Fix silent overflow for phys vec to sgt

[Claude Code Review Bot]

Overall Series Review

Subject: dma-buf: Fix silent overflow for phys vec to sgt
Author: David Hu <xuehaohu@google.com>
Patches: 1
Reviewed: 2026-05-28T12:05:55.710402

---

This is a single-patch fix for a real overflow bug in `dma_buf_phys_vec_to_sgt()` where MMIO regions larger than 4GB cause `mapped_len` (declared as `unsigned int`) to silently wrap around when accumulating `size_t` lengths. The fix is straightforward and correct: promote `mapped_len` to `size_t`. The secondary changes (loop iterator signedness in `fill_sg_entry`, overflow guard in `calc_sg_nents`) are reasonable hardening.

The patch is **mostly good** but has a couple of issues worth addressing.

---
Generated by Claude Code Patch Reviewer

Claude review: dma-buf: Fix silent overflow for phys vec to sgt

[Claude Code Review Bot]

Patch Review

**Commit message nit:** There's a dangling fragment in the commit message body:

```
to mapped_len, and leading to a silent overflow
```

This line (line 103 of the mbox) looks like leftover text from an earlier draft that wasn't cleaned up. It should be removed before merging.

**`mapped_len` promotion (main fix) — correct:**

```c
-	unsigned int nents, mapped_len = 0;
...
+	size_t mapped_len = 0;
+	unsigned int nents;
```

This is the core fix. `mapped_len` accumulates `phys_vec[i].len` values (which are `size_t`) at line 154 of the source:

```c
mapped_len += phys_vec[i].len;
```

With >4GB total MMIO, the old `unsigned int` wraps silently. Then `dma_iova_sync()` and `fill_sg_entry()` receive a truncated length, which is a data-corruption-grade bug. The fix is correct and minimal.

**`fill_sg_entry` loop iterator — correct but cosmetic:**

```c
-	int i;
+	unsigned int i;
```

The loop compares `i < nents` where `nents` is `unsigned int`. Using a signed `int` for the iterator means the comparison is well-defined (the signed value gets implicitly converted), so this is not a functional bug, but matching signedness is cleaner and avoids compiler warnings with `-Wsign-compare`. Fine to include.

**`calc_sg_nents` overflow guard — needs attention:**

```c
-	unsigned int nents = 0;
+	size_t nents = 0;
...
+	if (nents > UINT_MAX)
+		return 0;
```

The function signature still returns `unsigned int`:

```c
static unsigned int calc_sg_nents(struct dma_iova_state *state, ...)
```

Promoting `nents` to `size_t` locally and clamping to `UINT_MAX` before returning is logically correct — it prevents the intermediate accumulation from overflowing before the check. However, there are concerns:

1. **Returning 0 on overflow is a silent failure.** The caller at line 135-137 does:

   ```c
   nents = calc_sg_nents(dma->state, phys_vec, nr_ranges, size);
   ret = sg_alloc_table(&dma->sgt, nents, GFP_KERNEL | __GFP_ZERO);
   ```

   Passing `nents = 0` to `sg_alloc_table()` — which takes `unsigned int nents` — will cause `sg_alloc_table` to fail (it returns `-EINVAL` for 0 entries per the kernel implementation). So the error *will* propagate, but through `sg_alloc_table` failing rather than `calc_sg_nents` explicitly signaling the problem. This works but is fragile — it relies on `sg_alloc_table`'s behavior for a degenerate input rather than the caller explicitly handling the overflow. A `WARN_ON_ONCE` or explicit error return (e.g., returning `UINT_MAX` with a warning, or having the caller check for 0) would make the failure mode more obvious and debuggable.

2. **The non-IOVA path can also overflow.** In the loop:

   ```c
   for (i = 0; i < nr_ranges; i++)
       nents += DIV_ROUND_UP(phys_vec[i].len, UINT_MAX);
   ```

   Each `phys_vec[i].len` produces at least 1 entry, and `nr_ranges` is `size_t`. If `nr_ranges` is extremely large (though unlikely in practice for MMIO regions), `nents` could still overflow even as `size_t` on 64-bit. The `> UINT_MAX` check after the loop catches this, so it's fine. But the IOVA path with `DIV_ROUND_UP(size, UINT_MAX)` will never exceed `UINT_MAX` on 64-bit since `size` is `size_t` (at most 2^64-1 / (2^32-1) ≈ 2^32), so the check is mostly relevant for the non-IOVA accumulation path.

**Reverse Christmas tree ordering — correct:**

The variable declarations in `dma_buf_phys_vec_to_sgt` are properly reordered by line length:

```c
	struct dma_buf_dma *dma;
	struct scatterlist *sgl;
	size_t mapped_len = 0;
	unsigned int nents;
	dma_addr_t addr;
	size_t i;
	int ret;
```

This follows the kernel coding style convention.

**Summary:** The core `mapped_len` overflow fix is correct and important. The `calc_sg_nents` overflow guard works but returning 0 silently is not ideal — consider adding a `WARN_ON_ONCE(nents > UINT_MAX)` before the return so the failure is visible in dmesg rather than silently propagating through `sg_alloc_table`. The dangling text fragment in the commit message ("to mapped_len, and leading to a silent overflow") must be removed.

---
Generated by Claude Code Patch Reviewer