Merge branch 'xsk-the-lost-bits-from-chapter-iii'

Alexander Lobakin says:

====================
xsk: the lost bits from Chapter III

Before introducing libeth_xdp, we need to add a couple more generic
helpers. Notably:

* 01: add generic loop unrolling hint helpers;
* 04: add helper to get both xdp_desc's DMA address and metadata
  pointer in one go, saving several cycles and hotpath object
  code size in drivers (especially when unrolling).

Bonus:

* 02, 03: convert two drivers which were using custom macros to
  generic unrolled_count() (trivial, no object code changes).
====================

Link: https://patch.msgid.link/20250206182630.3914318-1-aleksander.lobakin@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

8 files changed, 141 insertions, 26 deletions

diff --git a/drivers/net/ethernet/intel/i40e/i40e_xsk.c b/drivers/net/ethernet/intel/i40e/i40e_xsk.c
index e28f1905a4a0..9f47388eaba5 100644
--- a/drivers/net/ethernet/intel/i40e/i40e_xsk.c
+++ b/drivers/net/ethernet/intel/i40e/i40e_xsk.c
@@ -2,6 +2,7 @@
 /* Copyright(c) 2018 Intel Corporation. */
 
 #include <linux/bpf_trace.h>
+#include <linux/unroll.h>
 #include <net/xdp_sock_drv.h>
 #include "i40e_txrx_common.h"
 #include "i40e_xsk.h"
@@ -529,7 +530,8 @@ static void i40e_xmit_pkt_batch(struct i40e_ring *xdp_ring, struct xdp_desc *des
 	dma_addr_t dma;
 	u32 i;
 
-	loop_unrolled_for(i = 0; i < PKTS_PER_BATCH; i++) {
+	unrolled_count(PKTS_PER_BATCH)
+	for (i = 0; i < PKTS_PER_BATCH; i++) {
 		u32 cmd = I40E_TX_DESC_CMD_ICRC | xsk_is_eop_desc(&desc[i]);
 
 		dma = xsk_buff_raw_get_dma(xdp_ring->xsk_pool, desc[i].addr);
diff --git a/drivers/net/ethernet/intel/i40e/i40e_xsk.h b/drivers/net/ethernet/intel/i40e/i40e_xsk.h
index ef156fad52f2..dd16351a7af8 100644
--- a/drivers/net/ethernet/intel/i40e/i40e_xsk.h
+++ b/drivers/net/ethernet/intel/i40e/i40e_xsk.h
@@ -6,7 +6,7 @@
 
 #include <linux/types.h>
 
-/* This value should match the pragma in the loop_unrolled_for
+/* This value should match the pragma in the unrolled_count()
  * macro. Why 4? It is strictly empirical. It seems to be a good
  * compromise between the advantage of having simultaneous outstanding
  * reads to the DMA array that can hide each others latency and the
@@ -14,14 +14,6 @@
  */
 #define PKTS_PER_BATCH 4
 
-#ifdef __clang__
-#define loop_unrolled_for _Pragma("clang loop unroll_count(4)") for
-#elif __GNUC__ >= 8
-#define loop_unrolled_for _Pragma("GCC unroll 4") for
-#else
-#define loop_unrolled_for for
-#endif
-
 struct i40e_ring;
 struct i40e_vsi;
 struct net_device;
diff --git a/drivers/net/ethernet/intel/ice/ice_xsk.c b/drivers/net/ethernet/intel/ice/ice_xsk.c
index 8975d2971bc3..a3a4eaa17739 100644
--- a/drivers/net/ethernet/intel/ice/ice_xsk.c
+++ b/drivers/net/ethernet/intel/ice/ice_xsk.c
@@ -2,6 +2,7 @@
 /* Copyright (c) 2019, Intel Corporation. */
 
 #include <linux/bpf_trace.h>
+#include <linux/unroll.h>
 #include <net/xdp_sock_drv.h>
 #include <net/xdp.h>
 #include "ice.h"
@@ -989,7 +990,8 @@ static void ice_xmit_pkt_batch(struct ice_tx_ring *xdp_ring,
 	struct ice_tx_desc *tx_desc;
 	u32 i;
 
-	loop_unrolled_for(i = 0; i < PKTS_PER_BATCH; i++) {
+	unrolled_count(PKTS_PER_BATCH)
+	for (i = 0; i < PKTS_PER_BATCH; i++) {
 		dma_addr_t dma;
 
 		dma = xsk_buff_raw_get_dma(xsk_pool, descs[i].addr);
diff --git a/drivers/net/ethernet/intel/ice/ice_xsk.h b/drivers/net/ethernet/intel/ice/ice_xsk.h
index 45adeb513253..8dc5d55e26c5 100644
--- a/drivers/net/ethernet/intel/ice/ice_xsk.h
+++ b/drivers/net/ethernet/intel/ice/ice_xsk.h
@@ -7,14 +7,6 @@
 
 #define PKTS_PER_BATCH 8
 
-#ifdef __clang__
-#define loop_unrolled_for _Pragma("clang loop unroll_count(8)") for
-#elif __GNUC__ >= 8
-#define loop_unrolled_for _Pragma("GCC unroll 8") for
-#else
-#define loop_unrolled_for for
-#endif
-
 struct ice_vsi;
 
 #ifdef CONFIG_XDP_SOCKETS
diff --git a/include/linux/unroll.h b/include/linux/unroll.h
index d42fd6366373..863fb69f6a7e 100644
--- a/include/linux/unroll.h
+++ b/include/linux/unroll.h
@@ -9,6 +9,50 @@
 
 #include <linux/args.h>
 
+#ifdef CONFIG_CC_IS_CLANG
+#define __pick_unrolled(x, y)		_Pragma(#x)
+#elif CONFIG_GCC_VERSION >= 80000
+#define __pick_unrolled(x, y)		_Pragma(#y)
+#else
+#define __pick_unrolled(x, y)		/* not supported */
+#endif
+
+/**
+ * unrolled - loop attributes to ask the compiler to unroll it
+ *
+ * Usage:
+ *
+ * #define BATCH 8
+ *
+ *	unrolled_count(BATCH)
+ *	for (u32 i = 0; i < BATCH; i++)
+ *		// loop body without cross-iteration dependencies
+ *
+ * This is only a hint and the compiler is free to disable unrolling if it
+ * thinks the count is suboptimal and may hurt performance and/or hugely
+ * increase object code size.
+ * Not having any cross-iteration dependencies (i.e. when iter x + 1 depends
+ * on what iter x will do with variables) is not a strict requirement, but
+ * provides best performance and object code size.
+ * Available only on Clang and GCC 8.x onwards.
+ */
+
+/* Ask the compiler to pick an optimal unroll count, Clang only */
+#define unrolled							\
+	__pick_unrolled(clang loop unroll(enable), /* nothing */)
+
+/* Unroll each @n iterations of the loop */
+#define unrolled_count(n)						\
+	__pick_unrolled(clang loop unroll_count(n), GCC unroll n)
+
+/* Unroll the whole loop */
+#define unrolled_full							\
+	__pick_unrolled(clang loop unroll(full), GCC unroll 65534)
+
+/* Never unroll the loop */
+#define unrolled_none							\
+	__pick_unrolled(clang loop unroll(disable), GCC unroll 1)
+
 #define UNROLL(N, MACRO, args...) CONCATENATE(__UNROLL_, N)(MACRO, args)
 
 #define __UNROLL_0(MACRO, args...)
diff --git a/include/net/xdp_sock_drv.h b/include/net/xdp_sock_drv.h
index 784cd34f5bba..15086dcf51d8 100644
--- a/include/net/xdp_sock_drv.h
+++ b/include/net/xdp_sock_drv.h
@@ -196,6 +196,23 @@ static inline void *xsk_buff_raw_get_data(struct xsk_buff_pool *pool, u64 addr)
 	return xp_raw_get_data(pool, addr);
 }
 
+/**
+ * xsk_buff_raw_get_ctx - get &xdp_desc context
+ * @pool: XSk buff pool desc address belongs to
+ * @addr: desc address (from userspace)
+ *
+ * Wrapper for xp_raw_get_ctx() to be used in drivers, see its kdoc for
+ * details.
+ *
+ * Return: new &xdp_desc_ctx struct containing desc's DMA address and metadata
+ * pointer, if it is present and valid (initialized to %NULL otherwise).
+ */
+static inline struct xdp_desc_ctx
+xsk_buff_raw_get_ctx(const struct xsk_buff_pool *pool, u64 addr)
+{
+	return xp_raw_get_ctx(pool, addr);
+}
+
 #define XDP_TXMD_FLAGS_VALID ( \
 		XDP_TXMD_FLAGS_TIMESTAMP | \
 		XDP_TXMD_FLAGS_CHECKSUM | \
@@ -207,20 +224,27 @@ xsk_buff_valid_tx_metadata(const struct xsk_tx_metadata *meta)
 	return !(meta->flags & ~XDP_TXMD_FLAGS_VALID);
 }
 
-static inline struct xsk_tx_metadata *xsk_buff_get_metadata(struct xsk_buff_pool *pool, u64 addr)
+static inline struct xsk_tx_metadata *
+__xsk_buff_get_metadata(const struct xsk_buff_pool *pool, void *data)
 {
 	struct xsk_tx_metadata *meta;
 
 	if (!pool->tx_metadata_len)
 		return NULL;
 
-	meta = xp_raw_get_data(pool, addr) - pool->tx_metadata_len;
+	meta = data - pool->tx_metadata_len;
 	if (unlikely(!xsk_buff_valid_tx_metadata(meta)))
 		return NULL; /* no way to signal the error to the user */
 
 	return meta;
 }
 
+static inline struct xsk_tx_metadata *
+xsk_buff_get_metadata(struct xsk_buff_pool *pool, u64 addr)
+{
+	return __xsk_buff_get_metadata(pool, xp_raw_get_data(pool, addr));
+}
+
 static inline void xsk_buff_dma_sync_for_cpu(struct xdp_buff *xdp)
 {
 	struct xdp_buff_xsk *xskb = container_of(xdp, struct xdp_buff_xsk, xdp);
@@ -388,12 +412,25 @@ static inline void *xsk_buff_raw_get_data(struct xsk_buff_pool *pool, u64 addr)
 	return NULL;
 }
 
+static inline struct xdp_desc_ctx
+xsk_buff_raw_get_ctx(const struct xsk_buff_pool *pool, u64 addr)
+{
+	return (struct xdp_desc_ctx){ };
+}
+
 static inline bool xsk_buff_valid_tx_metadata(struct xsk_tx_metadata *meta)
 {
 	return false;
 }
 
-static inline struct xsk_tx_metadata *xsk_buff_get_metadata(struct xsk_buff_pool *pool, u64 addr)
+static inline struct xsk_tx_metadata *
+__xsk_buff_get_metadata(const struct xsk_buff_pool *pool, void *data)
+{
+	return NULL;
+}
+
+static inline struct xsk_tx_metadata *
+xsk_buff_get_metadata(struct xsk_buff_pool *pool, u64 addr)
 {
 	return NULL;
 }
diff --git a/include/net/xsk_buff_pool.h b/include/net/xsk_buff_pool.h
index 50779406bc2d..1dcd4d71468a 100644
--- a/include/net/xsk_buff_pool.h
+++ b/include/net/xsk_buff_pool.h
@@ -141,6 +141,14 @@ u32 xp_alloc_batch(struct xsk_buff_pool *pool, struct xdp_buff **xdp, u32 max);
 bool xp_can_alloc(struct xsk_buff_pool *pool, u32 count);
 void *xp_raw_get_data(struct xsk_buff_pool *pool, u64 addr);
 dma_addr_t xp_raw_get_dma(struct xsk_buff_pool *pool, u64 addr);
+
+struct xdp_desc_ctx {
+	dma_addr_t dma;
+	struct xsk_tx_metadata *meta;
+};
+
+struct xdp_desc_ctx xp_raw_get_ctx(const struct xsk_buff_pool *pool, u64 addr);
+
 static inline dma_addr_t xp_get_dma(struct xdp_buff_xsk *xskb)
 {
 	return xskb->dma;
diff --git a/net/xdp/xsk_buff_pool.c b/net/xdp/xsk_buff_pool.c
index 1f7975b49657..c263fb7a68dc 100644
--- a/net/xdp/xsk_buff_pool.c
+++ b/net/xdp/xsk_buff_pool.c
@@ -699,18 +699,56 @@ void xp_free(struct xdp_buff_xsk *xskb)
 }
 EXPORT_SYMBOL(xp_free);
 
-void *xp_raw_get_data(struct xsk_buff_pool *pool, u64 addr)
+static u64 __xp_raw_get_addr(const struct xsk_buff_pool *pool, u64 addr)
+{
+	return pool->unaligned ? xp_unaligned_add_offset_to_addr(addr) : addr;
+}
+
+static void *__xp_raw_get_data(const struct xsk_buff_pool *pool, u64 addr)
 {
-	addr = pool->unaligned ? xp_unaligned_add_offset_to_addr(addr) : addr;
 	return pool->addrs + addr;
 }
+
+void *xp_raw_get_data(struct xsk_buff_pool *pool, u64 addr)
+{
+	return __xp_raw_get_data(pool, __xp_raw_get_addr(pool, addr));
+}
 EXPORT_SYMBOL(xp_raw_get_data);
 
-dma_addr_t xp_raw_get_dma(struct xsk_buff_pool *pool, u64 addr)
+static dma_addr_t __xp_raw_get_dma(const struct xsk_buff_pool *pool, u64 addr)
 {
-	addr = pool->unaligned ? xp_unaligned_add_offset_to_addr(addr) : addr;
 	return (pool->dma_pages[addr >> PAGE_SHIFT] &
 		~XSK_NEXT_PG_CONTIG_MASK) +
 		(addr & ~PAGE_MASK);
 }
+
+dma_addr_t xp_raw_get_dma(struct xsk_buff_pool *pool, u64 addr)
+{
+	return __xp_raw_get_dma(pool, __xp_raw_get_addr(pool, addr));
+}
 EXPORT_SYMBOL(xp_raw_get_dma);
+
+/**
+ * xp_raw_get_ctx - get &xdp_desc context
+ * @pool: XSk buff pool desc address belongs to
+ * @addr: desc address (from userspace)
+ *
+ * Helper for getting desc's DMA address and metadata pointer, if present.
+ * Saves one call on hotpath, double calculation of the actual address,
+ * and inline checks for metadata presence and sanity.
+ *
+ * Return: new &xdp_desc_ctx struct containing desc's DMA address and metadata
+ * pointer, if it is present and valid (initialized to %NULL otherwise).
+ */
+struct xdp_desc_ctx xp_raw_get_ctx(const struct xsk_buff_pool *pool, u64 addr)
+{
+	struct xdp_desc_ctx ret;
+
+	addr = __xp_raw_get_addr(pool, addr);
+
+	ret.dma = __xp_raw_get_dma(pool, addr);
+	ret.meta = __xsk_buff_get_metadata(pool, __xp_raw_get_data(pool, addr));
+
+	return ret;
+}
+EXPORT_SYMBOL(xp_raw_get_ctx);