fortify: Provide a memcpy trap door for sharp corners

As we continue to narrow the scope of what the FORTIFY memcpy() will
accept and build alternative APIs that give the compiler appropriate
visibility into more complex memcpy scenarios, there is a need for
"unfortified" memcpy use in rare cases where combinations of compiler
behaviors, source code layout, etc, result in cases where the stricter
memcpy checks need to be bypassed until appropriate solutions can be
developed (i.e. fix compiler bugs, code refactoring, new API, etc). The
intention is for this to be used only if there's no other reasonable
solution, for its use to include a justification that can be used
to assess future solutions, and for it to be temporary.

Example usage included, based on analysis and discussion from:
https://lore.kernel.org/netdev/CANn89iLS_2cshtuXPyNUGDPaic=sJiYfvTb_wNLgWrZRyBxZ_g@mail.gmail.com

Cc: Jakub Kicinski <kuba@kernel.org>
Cc: Eric Dumazet <edumazet@google.com>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Paolo Abeni <pabeni@redhat.com>
Cc: Coco Li <lixiaoyan@google.com>
Cc: Tariq Toukan <tariqt@nvidia.com>
Cc: Saeed Mahameed <saeedm@nvidia.com>
Cc: Leon Romanovsky <leon@kernel.org>
Cc: netdev@vger.kernel.org
Cc: linux-hardening@vger.kernel.org
Signed-off-by: Kees Cook <keescook@chromium.org>
Link: https://lore.kernel.org/r/20220511025301.3636666-1-keescook@chromium.orgSigned-off-by: Paolo Abeni <pabeni@redhat.com>

drivers/net/ethernet/mellanox/mlx5/core/en_tx.c

@@ -386,7 +386,13 @@ mlx5e_sq_xmit_wqe(struct mlx5e_txqsq *sq, struct sk_buff *skb,
 			stats->added_vlan_packets++;
 		} else {
 			eseg->inline_hdr.sz |= cpu_to_be16(attr->ihs);
-			memcpy(eseg->inline_hdr.start, skb->data, attr->ihs);
+			unsafe_memcpy(eseg->inline_hdr.start, skb->data, attr->ihs,
+				/* This copy has been bounds-checked earlier in
+				 * mlx5i_sq_calc_wqe_attr() and intentionally
+				 * crosses a flex array boundary. Since it is
+				 * performance sensitive, splitting the copy is
+				 * undesirable.
+				 */);
 		}
 		dseg += wqe_attr->ds_cnt_inl;
 	} else if (skb_vlan_tag_present(skb)) {

include/linux/fortify-string.h

@@ -52,6 +52,22 @@ extern char *__underlying_strncpy(char *p, const char *q, __kernel_size_t size)
 #define __underlying_strncpy	__builtin_strncpy
 #endif
 
+/**
+ * unsafe_memcpy - memcpy implementation with no FORTIFY bounds checking
+ *
+ * @dst: Destination memory address to write to
+ * @src: Source memory address to read from
+ * @bytes: How many bytes to write to @dst from @src
+ * @justification: Free-form text or comment describing why the use is needed
+ *
+ * This should be used for corner cases where the compiler cannot do the
+ * right thing, or during transitions between APIs, etc. It should be used
+ * very rarely, and includes a place for justification detailing where bounds
+ * checking has happened, and why existing solutions cannot be employed.
+ */
+#define unsafe_memcpy(dst, src, bytes, justification)		\
+	__underlying_memcpy(dst, src, bytes)
+
 /*
  * Clang's use of __builtin_object_size() within inlines needs hinting via
  * __pass_object_size(). The preference is to only ever use type 1 (member

include/linux/string.h

@@ -252,6 +252,10 @@ static inline const char *kbasename(const char *path)
 #if !defined(__NO_FORTIFY) && defined(__OPTIMIZE__) && defined(CONFIG_FORTIFY_SOURCE)
 #include <linux/fortify-string.h>
 #endif
+#ifndef unsafe_memcpy
+#define unsafe_memcpy(dst, src, bytes, justification)		\
+	memcpy(dst, src, bytes)
+#endif
 
 void memcpy_and_pad(void *dest, size_t dest_len, const void *src, size_t count,
 		    int pad);