dmaengine: stm32-dma3: refactor HW linked-list to optimize memory accesses

Current behavior splits the buffer/sg in n * STM32_DMA3_MAX_BLOCK_SIZE + 1
for the remainder without optimization.

New behavior splits the buffer/sg in n * STM32_DMA3_MAX_BLOCK_SIZE + 1 for
(x * chan->max_burst) + 1 for the remainder.
Depending on channel FIFO size, optimal double-word (word if only 8-byte
FIFO size) bursts can be programmed before managing the very last remainder
with lower data width.
In case of _prep_slave_sg, and depending on the channel Transfer Complete
event configuration, the user is warned about the refactored linked-list,
not having the same items count than the initial sg_list. This warning is
shown only if the configuration is successful.

Signed-off-by: Amelie Delaunay <amelie.delaunay@foss.st.com>
Link: https://lore.kernel.org/r/20241016-dma3-mp25-updates-v3-3-8311fe6f228d@foss.st.com
Signed-off-by: Vinod Koul <vkoul@kernel.org>

Author: ADESTM

drivers/dma/stm32/stm32-dma3.c

@@ -1126,6 +1126,25 @@ static void stm32_dma3_free_chan_resources(struct dma_chan *c)
 	chan->config_set = 0;
 }
 
+static u32 stm32_dma3_get_ll_count(struct stm32_dma3_chan *chan, size_t len)
+{
+	u32 count;
+
+	count = len / STM32_DMA3_MAX_BLOCK_SIZE;
+	len -= (len / STM32_DMA3_MAX_BLOCK_SIZE) * STM32_DMA3_MAX_BLOCK_SIZE;
+
+	if (len >= chan->max_burst) {
+		count += 1; /* len < STM32_DMA3_MAX_BLOCK_SIZE here, so it fits in one item */
+		len -= (len / chan->max_burst) * chan->max_burst;
+	}
+
+	/* Unaligned remainder fits in one extra item */
+	if (len > 0)
+		count += 1;
+
+	return count;
+}
+
 static void stm32_dma3_init_chan_config_for_memcpy(struct stm32_dma3_chan *chan,
 						   dma_addr_t dst, dma_addr_t src)
 {
@@ -1161,7 +1180,7 @@ static struct dma_async_tx_descriptor *stm32_dma3_prep_dma_memcpy(struct dma_cha
 	size_t next_size, offset;
 	u32 count, i, ctr1, ctr2;
 
-	count = DIV_ROUND_UP(len, STM32_DMA3_MAX_BLOCK_SIZE);
+	count = stm32_dma3_get_ll_count(chan, len);
 
 	swdesc = stm32_dma3_chan_desc_alloc(chan, count);
 	if (!swdesc)
@@ -1177,6 +1196,9 @@ static struct dma_async_tx_descriptor *stm32_dma3_prep_dma_memcpy(struct dma_cha
 		remaining = len - offset;
 		next_size = min_t(size_t, remaining, STM32_DMA3_MAX_BLOCK_SIZE);
 
+		if (next_size < STM32_DMA3_MAX_BLOCK_SIZE && next_size >= chan->max_burst)
+			next_size = chan->max_burst * (remaining / chan->max_burst);
+
 		ret = stm32_dma3_chan_prep_hw(chan, DMA_MEM_TO_MEM, &swdesc->ccr, &ctr1, &ctr2,
 					      src + offset, dst + offset, next_size);
 		if (ret)
@@ -1215,12 +1237,9 @@ static struct dma_async_tx_descriptor *stm32_dma3_prep_slave_sg(struct dma_chan
 	u32 i, j, count, ctr1, ctr2;
 	int ret;
 
-	count = sg_len;
-	for_each_sg(sgl, sg, sg_len, i) {
-		len = sg_dma_len(sg);
-		if (len > STM32_DMA3_MAX_BLOCK_SIZE)
-			count += DIV_ROUND_UP(len, STM32_DMA3_MAX_BLOCK_SIZE) - 1;
-	}
+	count = 0;
+	for_each_sg(sgl, sg, sg_len, i)
+		count += stm32_dma3_get_ll_count(chan, sg_dma_len(sg));
 
 	swdesc = stm32_dma3_chan_desc_alloc(chan, count);
 	if (!swdesc)
@@ -1237,6 +1256,9 @@ static struct dma_async_tx_descriptor *stm32_dma3_prep_slave_sg(struct dma_chan
 		do {
 			size_t chunk = min_t(size_t, len, STM32_DMA3_MAX_BLOCK_SIZE);
 
+			if (chunk < STM32_DMA3_MAX_BLOCK_SIZE && chunk >= chan->max_burst)
+				chunk = chan->max_burst * (len / chan->max_burst);
+
 			if (dir == DMA_MEM_TO_DEV) {
 				src = sg_addr;
 				dst = dev_addr;
@@ -1269,6 +1291,10 @@ static struct dma_async_tx_descriptor *stm32_dma3_prep_slave_sg(struct dma_chan
 		} while (len);
 	}
 
+	if (count != sg_len && chan->tcem != CTR2_TCEM_CHANNEL)
+		dev_warn(chan2dev(chan), "Linked-list refactored, %d items instead of %d\n",
+			 count, sg_len);
+
 	/* Enable Error interrupts */
 	swdesc->ccr |= CCR_USEIE | CCR_ULEIE | CCR_DTEIE;
 	/* Enable Transfer state interrupts */