From 9632d2a7f63ce72a914c478da828965a9f6492b6 Mon Sep 17 00:00:00 2001
From: Gregory Nutt <gnutt@nuttx.org>
Date: Wed, 4 Dec 2013 07:41:29 -0600
Subject: [PATCH] SAMA5 NAND: Do not perform DMA on small transfers

---
 arch/arm/src/sama5/Kconfig    | 22 ++++++++++++++++++++--
 arch/arm/src/sama5/sam_nand.c | 18 +++++++++++-------
 arch/arm/src/sama5/sam_nand.h | 18 ++++++++++++++++++
 3 files changed, 49 insertions(+), 9 deletions(-)

diff --git a/arch/arm/src/sama5/Kconfig b/arch/arm/src/sama5/Kconfig
index 12513c5c0b..ea4b0b7295 100644
--- a/arch/arm/src/sama5/Kconfig
+++ b/arch/arm/src/sama5/Kconfig
@@ -3348,8 +3348,26 @@ config SAMA5_NAND_DMA
 	default y
 	depends on SAMA5_DMAC0
 	---help---
-		Use DMA to perform NAND data transfers.  NOTE that DMAC0 must be
-		selected (DMAC1 cannot access NFC SRAM). (highly recommended)
+		Use memory-to-memory DMA to perform NAND data transfers.  NOTE that
+		DMAC0 must be selected (DMAC1 cannot access NFC SRAM).
+
+config SAMA5_NAND_DMA_THRESHOLD
+	int "DMA threshold"
+	default 784
+	depends on SAMA5_NAND_DMA
+	---help---
+		Defines a threshold value for performing memory-to-memory DMA.
+
+		If memory-to-memory DMAs are used, then two context switches will
+		occur:  (1) when the NAND logic waits for the DMA to complete, and
+		(2) again when the DMA completes and the NAND logic is re-awakened.
+		Each context switch will required saving and restoring a set of
+		registers defining the task state.  Those register include the PSR,
+		16 general purpose registers, and 32 floating point registers or
+		about 196 bytes per task state.  That is then 392*2 bytes per
+		context and 784 bytes for both.  Plus there is processing overhead.
+		So certainly, there is no reason to use a memory-to-memory DMA
+		transfer for much smaller blocks of data.
 
 config SAMA5_NAND_READYBUSY
 	bool "NAND Ready/Busy"
diff --git a/arch/arm/src/sama5/sam_nand.c b/arch/arm/src/sama5/sam_nand.c
index 6fecea58f3..b5865be8cd 100644
--- a/arch/arm/src/sama5/sam_nand.c
+++ b/arch/arm/src/sama5/sam_nand.c
@@ -1599,11 +1599,13 @@ static int nand_read(struct sam_nandcs_s *priv, bool nfcsram,
     }
 
 #ifdef CONFIG_SAMA5_NAND_DMA
-  /* Then perform the transfer via DMA or not, depending on if we have
-   * a DMA channel assigned.
+  /* Then perform the transfer via memory-to-memory DMA or not, depending
+   * on if we have a DMA channel assigned and if the transfer is
+   * sufficiently large.  Small DMAs (e.g., for spare data) are not peformed
+   * because the DMA context switch can take more time that the DMA itself.
    */
 
-  if (priv->dma)
+  if (priv->dma && buflen > CONFIG_SAMA5_NAND_DMA_THRESHOLD)
     {
       /* Transfer using DMA */
 
@@ -1909,12 +1911,14 @@ static int nand_write(struct sam_nandcs_s *priv, bool nfcsram,
 
   dest += offset;
 
-  /* Then perform the transfer via DMA or not, depending on if we have
-   * a DMA channel assigned.
+#ifdef CONFIG_SAMA5_NAND_DMA
+  /* Then perform the transfer via memory-to-memory DMA or not, depending
+   * on if we have a DMA channel assigned and if the transfer is
+   * sufficiently large.  Small DMAs (e.g., for spare data) are not peformed
+   * because the DMA context switch can take more time that the DMA itself.
    */
 
-#ifdef CONFIG_SAMA5_NAND_DMA
-  if (priv->dma)
+  if (priv->dma && buflen > CONFIG_SAMA5_NAND_DMA_THRESHOLD)
     {
       /* Transfer using DMA */
 
diff --git a/arch/arm/src/sama5/sam_nand.h b/arch/arm/src/sama5/sam_nand.h
index cd282c6a9a..ca1d8fd818 100644
--- a/arch/arm/src/sama5/sam_nand.h
+++ b/arch/arm/src/sama5/sam_nand.h
@@ -74,6 +74,24 @@
 #  endif
 #endif
 
+/* If memory-to-memory DMAs are used, then two context switches will occur:
+ * (1) when the NAND logic waits for the DMA to complete, and (2) again when
+ * the DMA completes and the NAND logic is re-awakened.  Each context switch
+ * will required saving and restoring a set of registers defining the task
+ * state.  Those register include the PSR, 16 general purpose registers, and
+ * 32 floating point registers or about 196 bytes per task state.  That is
+ * then 392*2 bytes per context and 784 bytes for both.  Plus there is
+ * processing overhead. So certainly, there is no reason to use a memory-to-
+ * memory DMA transfer for much smaller blocks of data.
+ */
+
+#ifdef CONFIG_SAMA5_NAND_DMA
+#  ifndef CONFIG_SAMA5_NAND_DMA_THRESHOLD
+#    define CONFIG_SAMA5_NAND_DMA_THRESHOLD 784
+#  endif
+#endif
+
+
 /* Hardware ECC types.  These are extensions to the NANDECC_HWECC value
  * defined in include/nuttx/mtd/nand_raw.h.
  *