Upon additional review, the new fast path in adiantum_finish() is
missing the call to flush_dcache_page() that scatterwalk_map_and_copy()
was doing. It's apparently debatable whether flush_dcache_page() is
actually needed, as per the discussion at
https://lore.kernel.org/lkml/YYP1lAq46NWzhOf0@casper.infradead.org/T/#u.
However, it appears that currently all the helper functions that write
to a page, such as scatterwalk_map_and_copy(), memcpy_to_page(), and
memzero_page(), do the dcache flush. So do it to be consistent.
Fixes: dadf5e56c967 ("crypto: adiantum - add fast path for single-page messages")
Signed-off-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
le128_sub(&rctx->rbuf.bignum, &rctx->rbuf.bignum, &rctx->header_hash);
if (dst_nents == 1 && dst->offset + req->cryptlen <= PAGE_SIZE) {
/* Fast path for single-page destination */
- void *virt = kmap_local_page(sg_page(dst)) + dst->offset;
+ struct page *page = sg_page(dst);
+ void *virt = kmap_local_page(page) + dst->offset;
err = crypto_shash_digest(&rctx->u.hash_desc, virt, bulk_len,
(u8 *)&digest);
}
le128_sub(&rctx->rbuf.bignum, &rctx->rbuf.bignum, &digest);
memcpy(virt + bulk_len, &rctx->rbuf.bignum, sizeof(le128));
+ flush_dcache_page(page);
kunmap_local(virt);
} else {
/* Slow path that works for any destination scatterlist */