@@ -2677,6 +2677,52 @@ static inline int get_burst_len(struct dma_pl330_desc *desc, unsigned int brst_s
return burst_len;
}
+/*
+ * Returns burst size to be used to copy data from/to memory during a
+ * peripheral transfer
+ */
+static unsigned int get_periph_mem_brst_sz(dma_addr_t addr, size_t len,
+ struct dma_pl330_chan *pch)
+{
+ unsigned int burst;
+
+ /* Select max possible burst size */
+ burst = pch->dmac->pcfg.data_bus_width / 8;
+
+ /*
+ * Make sure we use a burst size that aligns with the memory and length.
+ */
+ while ((addr | len) & (burst - 1))
+ burst /= 2;
+
+ return __ffs(burst);
+}
+
+/*
+ * Returns burst length to be used to copy data from/to memory during a
+ * peripheral transfer
+ */
+static unsigned int get_periph_mem_brst_len(struct dma_pl330_desc *desc,
+ struct dma_pl330_chan *pch,
+ unsigned int burst_size)
+{
+ unsigned int burst_len = pch->burst_len;
+
+ if (burst_size != pch->burst_sz) {
+ /* Select max possible burst len */
+ burst_len = get_burst_len(desc, burst_size);
+
+ /*
+ * Adjust AxLen to keep number of bytes same in Load/Store
+ */
+ if (burst_size > pch->burst_sz)
+ burst_len = pch->burst_len >> (burst_size - pch->burst_sz);
+ else
+ pch->burst_len = burst_len >> (pch->burst_sz - burst_size);
+ }
+ return burst_len;
+}
+
static struct dma_async_tx_descriptor *pl330_prep_dma_cyclic(
struct dma_chan *chan, dma_addr_t dma_addr, size_t len,
size_t period_len, enum dma_transfer_direction direction,
@@ -2684,8 +2730,8 @@ static struct dma_async_tx_descriptor *pl330_prep_dma_cyclic(
{
struct dma_pl330_desc *desc = NULL, *first = NULL;
struct dma_pl330_chan *pch = to_pchan(chan);
+ unsigned int i, burst_size, burst_len;
struct pl330_dmac *pl330 = pch->dmac;
- unsigned int i;
dma_addr_t dst;
dma_addr_t src;
@@ -2729,28 +2775,35 @@ static struct dma_async_tx_descriptor *pl330_prep_dma_cyclic(
return NULL;
}
+ burst_size = get_periph_mem_brst_sz(dma_addr, period_len, pch);
+ burst_len = get_periph_mem_brst_len(desc, pch, burst_size);
+
switch (direction) {
case DMA_MEM_TO_DEV:
desc->rqcfg.src_inc = 1;
desc->rqcfg.dst_inc = 0;
src = dma_addr;
dst = pch->fifo_dma;
+ desc->rqcfg.src_brst_size = burst_size;
+ desc->rqcfg.src_brst_len = burst_len;
+ desc->rqcfg.dst_brst_size = pch->burst_sz;
+ desc->rqcfg.dst_brst_len = pch->burst_len;
break;
case DMA_DEV_TO_MEM:
desc->rqcfg.src_inc = 0;
desc->rqcfg.dst_inc = 1;
src = pch->fifo_dma;
dst = dma_addr;
+ desc->rqcfg.src_brst_size = pch->burst_sz;
+ desc->rqcfg.src_brst_len = pch->burst_len;
+ desc->rqcfg.dst_brst_size = burst_size;
+ desc->rqcfg.dst_brst_len = burst_len;
break;
default:
break;
}
desc->rqtype = direction;
- desc->rqcfg.src_brst_size = pch->burst_sz;
- desc->rqcfg.src_brst_len = pch->burst_len;
- desc->rqcfg.dst_brst_size = pch->burst_sz;
- desc->rqcfg.dst_brst_len = pch->burst_len;
desc->bytes_requested = period_len;
fill_px(&desc->px, dst, src, period_len);
@@ -2850,7 +2903,10 @@ pl330_prep_slave_sg(struct dma_chan *chan, struct scatterlist *sgl,
{
struct dma_pl330_desc *first, *desc = NULL;
struct dma_pl330_chan *pch = to_pchan(chan);
+ unsigned int burst_size, burst_len;
struct scatterlist *sg;
+ dma_addr_t mem_addr;
+ size_t len;
int i;
if (unlikely(!pch || !sgl || !sg_len))
@@ -2882,29 +2938,37 @@ pl330_prep_slave_sg(struct dma_chan *chan, struct scatterlist *sgl,
else
list_add_tail(&desc->node, &first->node);
+ mem_addr = sg_dma_address(sg);
+ len = sg_dma_len(sg);
+
+ burst_size = get_periph_mem_brst_sz(mem_addr, len, pch);
+ burst_len = get_periph_mem_brst_len(desc, pch, burst_size);
+
switch (direction) {
case DMA_MEM_TO_DEV:
desc->rqcfg.src_inc = 1;
desc->rqcfg.dst_inc = 0;
- fill_px(&desc->px, pch->fifo_dma, sg_dma_address(sg),
- sg_dma_len(sg));
+ desc->rqcfg.src_brst_size = burst_size;
+ desc->rqcfg.src_brst_len = burst_len;
+ desc->rqcfg.dst_brst_size = pch->burst_sz;
+ desc->rqcfg.dst_brst_len = pch->burst_len;
+ fill_px(&desc->px, pch->fifo_dma, mem_addr, len);
break;
case DMA_DEV_TO_MEM:
desc->rqcfg.src_inc = 0;
desc->rqcfg.dst_inc = 1;
- fill_px(&desc->px, sg_dma_address(sg), pch->fifo_dma,
- sg_dma_len(sg));
+ desc->rqcfg.src_brst_size = pch->burst_sz;
+ desc->rqcfg.src_brst_len = pch->burst_len;
+ desc->rqcfg.dst_brst_size = burst_size;
+ desc->rqcfg.dst_brst_len = burst_len;
+ fill_px(&desc->px, mem_addr, pch->fifo_dma, len);
break;
default:
break;
}
- desc->rqcfg.src_brst_size = pch->burst_sz;
- desc->rqcfg.src_brst_len = pch->burst_len;
- desc->rqcfg.dst_brst_size = pch->burst_sz;
- desc->rqcfg.dst_brst_len = pch->burst_len;
desc->rqtype = direction;
- desc->bytes_requested = sg_dma_len(sg);
+ desc->bytes_requested = len;
}
/* Return the last desc in the chain */
Add logic to choose the maximum possible AxSize for transactions towards memory during usecases which copy data between memory and peripherals. Currently PL330 driver chooses equal AxLen and AxSize for both loads and stores to/from memory and peripherals which is inefficient towards memory as the whole bus width is not used for transfers as a peripheral might be limited to use only a narrow size of the buswidth available. Example scenario: A peripheral might require data byte by byte which would make AxSize = 1 byte and AxLen = 16 for both load from memory and store to Peripheral. This can be optimized for memory by using maximum AxSize (say 16bytes) then load from memory can be done with AxSize = 16byte, AxLen = 1 and store to peripheral with AxSize = 1byte, AxLen = 16 beats. Instruction setup post changes : 512bytes copy from Memory(16bytes * 4beats) to Peripheral(4bytes * 16 beats) --- DMAMOV CCR 0xbd0239 DMAMOV SAR 0xffffe000 DMAMOV DAR 0xffffc860 DMALP_1 7 DMAFLUSHP 0 DMAWFPB 0 DMALDB DMASTPB 0 DMALPENDA_1 bjmpto_7 DMASEV 3 DMAEND --- Signed-off-by: Joy Chakraborty <joychakr@google.com> --- drivers/dma/pl330.c | 92 ++++++++++++++++++++++++++++++++++++++------- 1 file changed, 78 insertions(+), 14 deletions(-)