[PATCH v2 2/5] lib: scatterlist: fix sg_split_phys() ->length clobber on !NEED_SG_DMA_LENGTH

[Charles Pellegrini <c4ffein.work@gmail.com>]

sg_split_phys() builds each output entry by copying the input entry and
then scrubbing the output's DMA view:

    *out_sg = *in_sg;
    sg_dma_address(out_sg) = 0;
    sg_dma_len(out_sg) = 0;

On configurations where CONFIG_NEED_SG_DMA_LENGTH is not selected,
sg_dma_len() is defined as ((sg)->length): there is no separate
dma_length member. The sg_dma_len(out_sg) = 0 line therefore zeroes the
CPU ->length that the function computes for the split. The trailing

    out_sg[-1].length = split->length_last_sg;

restores only the last entry, so for a multi-entry split every non-last
entry is left with length 0 and the split silently exposes fewer bytes
than requested. Where CONFIG_NEED_SG_DMA_LENGTH is selected (any config
with an IOMMU, and arches such as x86, arm64, powerpc, s390 or sparc)
sg_dma_len() touches the distinct dma_length field and ->length is
untouched, so the bug does not manifest.

Compute the length into a local and assign ->length after the DMA scrub,
so the value survives on both configurations:

    unsigned int len = in_sg->length;

    *out_sg = *in_sg;
    sg_dma_address(out_sg) = 0;
    sg_dma_len(out_sg) = 0;
    if (!j) {
        out_sg->offset += split->skip_sg0;
        len -= split->skip_sg0;
    }
    out_sg->length = len;

The sg_dma_len() = 0 scrub is kept deliberately: on NEED configurations it
clears the dma_length copied by *out_sg = *in_sg, which a consumer would
otherwise read as stale. Deleting it would fix the !NEED case but leak a
stale DMA length on NEED; reordering is correct on both.

No in-tree caller triggers this: the only caller that hands sg_split() an
unmapped list and requests a multi-entry partial split (the DTHEv2 AEAD
driver) builds for an arm64 platform, which selects
CONFIG_NEED_SG_DMA_LENGTH and is therefore immune. It was found while
correcting an in_mapped_nents argument in the KUnit suite (next patch),
which is what first exercised the unmapped multi-entry path on a !NEED
build.

Fixes: f8bcbe62acd0 ("lib: scatterlist: add sg splitting function")
Assisted-by: Claude:claude-opus-4-8 hegel-c
Signed-off-by: Charles Pellegrini <c4ffein.work@gmail.com>
---
 lib/sg_split.c | 9 ++++++---
 1 file changed, 6 insertions(+), 3 deletions(-)

diff --git a/lib/sg_split.c b/lib/sg_split.c
index 28fe7fa6c9ac..ab574dd26b03 100644
--- a/lib/sg_split.c
+++ b/lib/sg_split.c
@@ -84,13 +84,16 @@ static void sg_split_phys(struct sg_splitter *splitters, const int nb_splits)
 		in_sg = split->in_sg0;
 		out_sg = split->out_sg;
 		for (j = 0; j < split->nents; j++, out_sg++) {
+			unsigned int len = in_sg->length;
+
 			*out_sg = *in_sg;
+			sg_dma_address(out_sg) = 0;
+			sg_dma_len(out_sg) = 0;
 			if (!j) {
 				out_sg->offset += split->skip_sg0;
-				out_sg->length -= split->skip_sg0;
+				len -= split->skip_sg0;
 			}
-			sg_dma_address(out_sg) = 0;
-			sg_dma_len(out_sg) = 0;
+			out_sg->length = len;
 			in_sg = sg_next(in_sg);
 		}
 		out_sg[-1].length = split->length_last_sg;
-- 
2.47.3