xdp, libeth: make the xdp_init_buff() micro-optimization generic

Often times the compilers are not able to expand two consecutive 32-bit
writes into one 64-bit on the corresponding architectures. This applies
to xdp_init_buff() called for every received frame (or at least once
per each 64 frames when the frag size is fixed).
Move the not-so-pretty hack from libeth_xdp straight to xdp_init_buff(),
but using a proper union around ::frame_sz and ::flags.
The optimization is limited to LE architectures due to the structure
layout.

One simple example from idpf with the XDP series applied (Clang 22-git,
CONFIG_CC_OPTIMIZE_FOR_PERFORMANCE => -O2):

add/remove: 0/0 grow/shrink: 0/1 up/down: 0/-27 (-27)
Function                                     old     new   delta
idpf_vport_splitq_napi_poll                 5076    5049     -27

The perf difference with XDP_DROP is around +0.8-1% which I see as more
than satisfying.

Suggested-by: Simon Horman <horms@kernel.org>
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Tested-by: Ramu R <ramu.r@intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>

2 files changed, 26 insertions, 13 deletions

diff --git a/include/net/libeth/xdp.h b/include/net/libeth/xdp.h
index f4880b50e804..bc3507edd589 100644
--- a/include/net/libeth/xdp.h
+++ b/include/net/libeth/xdp.h
@@ -1274,7 +1274,6 @@ bool libeth_xdp_buff_add_frag(struct libeth_xdp_buff *xdp,
  * Internal, use libeth_xdp_process_buff() instead. Initializes XDP buffer
  * head with the Rx buffer data: data pointer, length, headroom, and
  * truesize/tailroom. Zeroes the flags.
- * Uses faster single u64 write instead of per-field access.
  */
 static inline void libeth_xdp_prepare_buff(struct libeth_xdp_buff *xdp,
 					   const struct libeth_fqe *fqe,
@@ -1282,17 +1281,9 @@ static inline void libeth_xdp_prepare_buff(struct libeth_xdp_buff *xdp,
 {
 	const struct page *page = __netmem_to_page(fqe->netmem);
 
-#ifdef __LIBETH_WORD_ACCESS
-	static_assert(offsetofend(typeof(xdp->base), flags) -
-		      offsetof(typeof(xdp->base), frame_sz) ==
-		      sizeof(u64));
-
-	*(u64 *)&xdp->base.frame_sz = fqe->truesize;
-#else
-	xdp_init_buff(&xdp->base, fqe->truesize, xdp->base.rxq);
-#endif
 	xdp_prepare_buff(&xdp->base, page_address(page) + fqe->offset,
 			 pp_page_to_nmdesc(page)->pp->p.offset, len, true);
+	xdp_init_buff(&xdp->base, fqe->truesize, xdp->base.rxq);
 }
 
 /**
diff --git a/include/net/xdp.h b/include/net/xdp.h
index b40f1f96cb11..af60e11b336c 100644
--- a/include/net/xdp.h
+++ b/include/net/xdp.h
@@ -85,8 +85,20 @@ struct xdp_buff {
 	void *data_hard_start;
 	struct xdp_rxq_info *rxq;
 	struct xdp_txq_info *txq;
-	u32 frame_sz; /* frame size to deduce data_hard_end/reserved tailroom*/
-	u32 flags; /* supported values defined in xdp_buff_flags */
+
+	union {
+		struct {
+			/* frame size to deduce data_hard_end/tailroom */
+			u32 frame_sz;
+			/* supported values defined in xdp_buff_flags */
+			u32 flags;
+		};
+
+#ifdef __LITTLE_ENDIAN
+		/* Used to micro-optimize xdp_init_buff(), don't use directly */
+		u64 frame_sz_flags_init;
+#endif
+	};
 };
 
 static __always_inline bool xdp_buff_has_frags(const struct xdp_buff *xdp)
@@ -118,9 +130,19 @@ static __always_inline void xdp_buff_set_frag_pfmemalloc(struct xdp_buff *xdp)
 static __always_inline void
 xdp_init_buff(struct xdp_buff *xdp, u32 frame_sz, struct xdp_rxq_info *rxq)
 {
-	xdp->frame_sz = frame_sz;
 	xdp->rxq = rxq;
+
+#ifdef __LITTLE_ENDIAN
+	/*
+	 * Force the compilers to initialize ::flags and assign ::frame_sz with
+	 * one write on 64-bit LE architectures as they're often unable to do
+	 * it themselves.
+	 */
+	xdp->frame_sz_flags_init = frame_sz;
+#else
+	xdp->frame_sz = frame_sz;
 	xdp->flags = 0;
+#endif
 }
 
 static __always_inline void