dma: avoid redundant calls for sync operations

Quite often, devices do not need dma_sync operations on x86_64 at least.
Indeed, when dev_is_dma_coherent(dev) is true and
dev_use_swiotlb(dev) is false, iommu_dma_sync_single_for_cpu()
and friends do nothing.

However, indirectly calling them when CONFIG_RETPOLINE=y consumes about
10% of cycles on a cpu receiving packets from softirq at ~100Gbit rate.
Even if/when CONFIG_RETPOLINE is not set, there is a cost of about 3%.

Add dev->need_dma_sync boolean and turn it off during the device
initialization (dma_set_mask()) depending on the setup:
dev_is_dma_coherent() for the direct DMA, !(sync_single_for_device ||
sync_single_for_cpu) or the new dma_map_ops flag, %DMA_F_CAN_SKIP_SYNC,
advertised for non-NULL DMA ops.
Then later, if/when swiotlb is used for the first time, the flag
is reset back to on, from swiotlb_tbl_map_single().

On iavf, the UDP trafficgen with XDP_DROP in skb mode test shows
+3-5% increase for direct DMA.

Suggested-by: Christoph Hellwig <[email protected]> # direct DMA shortcut
Co-developed-by: Eric Dumazet <[email protected]>
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: Alexander Lobakin <[email protected]>
Signed-off-by: Christoph Hellwig <[email protected]>

Author: alobakin

include/linux/device.h

@@ -691,6 +691,7 @@ struct device_physical_location {
  *		and optionall (if the coherent mask is large enough) also
  *		for dma allocations.  This flag is managed by the dma ops
  *		instance from ->dma_supported.
+ * @dma_need_sync: The device needs performing DMA sync operations.
  *
  * At the lowest level, every device in a Linux system is represented by an
  * instance of struct device. The device structure contains the information
@@ -803,6 +804,9 @@ struct device {
 #ifdef CONFIG_DMA_OPS_BYPASS
 	bool			dma_ops_bypass : 1;
 #endif
+#ifdef CONFIG_DMA_NEED_SYNC
+	bool			dma_need_sync:1;
+#endif
 };
 
 /**

include/linux/dma-map-ops.h

@@ -18,8 +18,11 @@ struct iommu_ops;
  *
  * DMA_F_PCI_P2PDMA_SUPPORTED: Indicates the dma_map_ops implementation can
  * handle PCI P2PDMA pages in the map_sg/unmap_sg operation.
+ * DMA_F_CAN_SKIP_SYNC: DMA sync operations can be skipped if the device is
+ * coherent and it's not an SWIOTLB buffer.
  */
 #define DMA_F_PCI_P2PDMA_SUPPORTED     (1 << 0)
+#define DMA_F_CAN_SKIP_SYNC            (1 << 1)
 
 struct dma_map_ops {
 	unsigned int flags;
@@ -273,6 +276,15 @@ static inline bool dev_is_dma_coherent(struct device *dev)
 }
 #endif /* CONFIG_ARCH_HAS_DMA_COHERENCE_H */
 
+static inline void dma_reset_need_sync(struct device *dev)
+{
+#ifdef CONFIG_DMA_NEED_SYNC
+	/* Reset it only once so that the function can be called on hotpath */
+	if (unlikely(!dev->dma_need_sync))
+		dev->dma_need_sync = true;
+#endif
+}
+
 /*
  * Check whether potential kmalloc() buffers are safe for non-coherent DMA.
  */

include/linux/dma-mapping.h

@@ -282,16 +282,59 @@ static inline int dma_mmap_noncontiguous(struct device *dev,
 #endif /* CONFIG_HAS_DMA */
 
 #if defined(CONFIG_HAS_DMA) && defined(CONFIG_DMA_NEED_SYNC)
-void dma_sync_single_for_cpu(struct device *dev, dma_addr_t addr, size_t size,
+void __dma_sync_single_for_cpu(struct device *dev, dma_addr_t addr, size_t size,
 		enum dma_data_direction dir);
-void dma_sync_single_for_device(struct device *dev, dma_addr_t addr,
+void __dma_sync_single_for_device(struct device *dev, dma_addr_t addr,
 		size_t size, enum dma_data_direction dir);
-void dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
+void __dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
 		int nelems, enum dma_data_direction dir);
-void dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
+void __dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
 		int nelems, enum dma_data_direction dir);
-bool dma_need_sync(struct device *dev, dma_addr_t dma_addr);
+bool __dma_need_sync(struct device *dev, dma_addr_t dma_addr);
+
+static inline bool dma_dev_need_sync(const struct device *dev)
+{
+	/* Always call DMA sync operations when debugging is enabled */
+	return dev->dma_need_sync || IS_ENABLED(CONFIG_DMA_API_DEBUG);
+}
+
+static inline void dma_sync_single_for_cpu(struct device *dev, dma_addr_t addr,
+		size_t size, enum dma_data_direction dir)
+{
+	if (dma_dev_need_sync(dev))
+		__dma_sync_single_for_cpu(dev, addr, size, dir);
+}
+
+static inline void dma_sync_single_for_device(struct device *dev,
+		dma_addr_t addr, size_t size, enum dma_data_direction dir)
+{
+	if (dma_dev_need_sync(dev))
+		__dma_sync_single_for_device(dev, addr, size, dir);
+}
+
+static inline void dma_sync_sg_for_cpu(struct device *dev,
+		struct scatterlist *sg, int nelems, enum dma_data_direction dir)
+{
+	if (dma_dev_need_sync(dev))
+		__dma_sync_sg_for_cpu(dev, sg, nelems, dir);
+}
+
+static inline void dma_sync_sg_for_device(struct device *dev,
+		struct scatterlist *sg, int nelems, enum dma_data_direction dir)
+{
+	if (dma_dev_need_sync(dev))
+		__dma_sync_sg_for_device(dev, sg, nelems, dir);
+}
+
+static inline bool dma_need_sync(struct device *dev, dma_addr_t dma_addr)
+{
+	return dma_dev_need_sync(dev) ? __dma_need_sync(dev, dma_addr) : false;
+}
 #else /* !CONFIG_HAS_DMA || !CONFIG_DMA_NEED_SYNC */
+static inline bool dma_dev_need_sync(const struct device *dev)
+{
+	return false;
+}
 static inline void dma_sync_single_for_cpu(struct device *dev, dma_addr_t addr,
 		size_t size, enum dma_data_direction dir)
 {

kernel/dma/mapping.c

@@ -330,7 +330,7 @@ void dma_unmap_resource(struct device *dev, dma_addr_t addr, size_t size,
 EXPORT_SYMBOL(dma_unmap_resource);
 
 #ifdef CONFIG_DMA_NEED_SYNC
-void dma_sync_single_for_cpu(struct device *dev, dma_addr_t addr, size_t size,
+void __dma_sync_single_for_cpu(struct device *dev, dma_addr_t addr, size_t size,
 		enum dma_data_direction dir)
 {
 	const struct dma_map_ops *ops = get_dma_ops(dev);
@@ -342,9 +342,9 @@ void dma_sync_single_for_cpu(struct device *dev, dma_addr_t addr, size_t size,
 		ops->sync_single_for_cpu(dev, addr, size, dir);
 	debug_dma_sync_single_for_cpu(dev, addr, size, dir);
 }
-EXPORT_SYMBOL(dma_sync_single_for_cpu);
+EXPORT_SYMBOL(__dma_sync_single_for_cpu);
 
-void dma_sync_single_for_device(struct device *dev, dma_addr_t addr,
+void __dma_sync_single_for_device(struct device *dev, dma_addr_t addr,
 		size_t size, enum dma_data_direction dir)
 {
 	const struct dma_map_ops *ops = get_dma_ops(dev);
@@ -356,9 +356,9 @@ void dma_sync_single_for_device(struct device *dev, dma_addr_t addr,
 		ops->sync_single_for_device(dev, addr, size, dir);
 	debug_dma_sync_single_for_device(dev, addr, size, dir);
 }
-EXPORT_SYMBOL(dma_sync_single_for_device);
+EXPORT_SYMBOL(__dma_sync_single_for_device);
 
-void dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
+void __dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
 		    int nelems, enum dma_data_direction dir)
 {
 	const struct dma_map_ops *ops = get_dma_ops(dev);
@@ -370,9 +370,9 @@ void dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
 		ops->sync_sg_for_cpu(dev, sg, nelems, dir);
 	debug_dma_sync_sg_for_cpu(dev, sg, nelems, dir);
 }
-EXPORT_SYMBOL(dma_sync_sg_for_cpu);
+EXPORT_SYMBOL(__dma_sync_sg_for_cpu);
 
-void dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
+void __dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
 		       int nelems, enum dma_data_direction dir)
 {
 	const struct dma_map_ops *ops = get_dma_ops(dev);
@@ -384,18 +384,47 @@ void dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
 		ops->sync_sg_for_device(dev, sg, nelems, dir);
 	debug_dma_sync_sg_for_device(dev, sg, nelems, dir);
 }
-EXPORT_SYMBOL(dma_sync_sg_for_device);
+EXPORT_SYMBOL(__dma_sync_sg_for_device);
 
-bool dma_need_sync(struct device *dev, dma_addr_t dma_addr)
+bool __dma_need_sync(struct device *dev, dma_addr_t dma_addr)
 {
 	const struct dma_map_ops *ops = get_dma_ops(dev);
 
 	if (dma_map_direct(dev, ops))
+		/*
+		 * dma_need_sync could've been reset on first SWIOTLB buffer
+		 * mapping, but @dma_addr is not necessary an SWIOTLB buffer.
+		 * In this case, fall back to more granular check.
+		 */
 		return dma_direct_need_sync(dev, dma_addr);
-	return ops->sync_single_for_cpu || ops->sync_single_for_device;
+	return true;
 }
-EXPORT_SYMBOL_GPL(dma_need_sync);
-#endif /* CONFIG_DMA_NEED_SYNC */
+EXPORT_SYMBOL_GPL(__dma_need_sync);
+
+static void dma_setup_need_sync(struct device *dev)
+{
+	const struct dma_map_ops *ops = get_dma_ops(dev);
+
+	if (dma_map_direct(dev, ops) || (ops->flags & DMA_F_CAN_SKIP_SYNC))
+		/*
+		 * dma_need_sync will be reset to %true on first SWIOTLB buffer
+		 * mapping, if any. During the device initialization, it's
+		 * enough to check only for the DMA coherence.
+		 */
+		dev->dma_need_sync = !dev_is_dma_coherent(dev);
+	else if (!ops->sync_single_for_device && !ops->sync_single_for_cpu &&
+		 !ops->sync_sg_for_device && !ops->sync_sg_for_cpu)
+		/*
+		 * Synchronization is not possible when none of DMA sync ops
+		 * is set.
+		 */
+		dev->dma_need_sync = false;
+	else
+		dev->dma_need_sync = true;
+}
+#else /* !CONFIG_DMA_NEED_SYNC */
+static inline void dma_setup_need_sync(struct device *dev) { }
+#endif /* !CONFIG_DMA_NEED_SYNC */
 
 /*
  * The whole dma_get_sgtable() idea is fundamentally unsafe - it seems
@@ -785,6 +814,8 @@ int dma_set_mask(struct device *dev, u64 mask)
 
 	arch_dma_set_mask(dev, mask);
 	*dev->dma_mask = mask;
+	dma_setup_need_sync(dev);
+
 	return 0;
 }
 EXPORT_SYMBOL(dma_set_mask);

kernel/dma/swiotlb.c

@@ -1408,6 +1408,12 @@ phys_addr_t swiotlb_tbl_map_single(struct device *dev, phys_addr_t orig_addr,
 		return (phys_addr_t)DMA_MAPPING_ERROR;
 	}
 
+	/*
+	 * If dma_need_sync wasn't set, reset it on first SWIOTLB buffer
+	 * mapping to always sync SWIOTLB buffers.
+	 */
+	dma_reset_need_sync(dev);
+
 	/*
 	 * Save away the mapping from the original address to the DMA address.
 	 * This is needed when we sync the memory.  Then we sync the buffer if