[PATCH 1/2] dmaengine: fsl-edma: add dma memcpy support
From: stefan@agner.ch (Stefan Agner)
Date: 2015-01-30 12:39:25
Also in:
lkml
On 2014-12-30 09:41, Jingchang Lu wrote:
This adds the memory copy capability support, the memcpy functionality needs to configure the DMAMUX of the channel with the always on slave id.
Hi Jingchang, I run some tests on v3.19-rc6 with this patches applied on Vybrid SoC VF500 and VF610. On VF500 clocked at 400MHz I get this numbers: # insmod dmatest.ko max_channels=1 iterations=100 run=1 [ 616.809594] dmatest: Started 1 threads using dma0chan0 [ 617.293498] dmatest: dma0chan0-copy0: summary 100 tests, 0 failures 209 iops 1613 KB/s (0) In contrast, on a VF610 clocked at 500MHz I get this: # insmod dmatest.ko max_channels=1 iterations=100 run=1 [ 154.203290] dmatest: Started 1 threads using dma0chan4 [ 154.614225] dmatest: dma0chan4-copy0: summary 100 tests, 0 failures 246 iops 2002 KB/s (0) Not that huge numbers, but I guess this is what the hardware can deliver... Did you run some performance tests? But generally, looks good. Tested-by: Stefan Agner <stefan@agner.ch>
quoted hunk ↗ jump to hunk
Signed-off-by: Jingchang Lu <redacted> --- drivers/dma/fsl-edma.c | 63 ++++++++++++++++++++++++++++++++++++++++++++++++-- 1 file changed, 61 insertions(+), 2 deletions(-)diff --git a/drivers/dma/fsl-edma.c b/drivers/dma/fsl-edma.c index 09e2842..e0bd517 100644 --- a/drivers/dma/fsl-edma.c +++ b/drivers/dma/fsl-edma.c@@ -110,6 +110,8 @@ #define EDMAMUX_CHCFG_ENBL 0x80 #define EDMAMUX_CHCFG_SOURCE(n) ((n) & 0x3F) +#define SLAVE_ID_ALWAYSON 63 /* the always on slave id */ + #define DMAMUX_NR 2 #define FSL_EDMA_BUSWIDTHS BIT(DMA_SLAVE_BUSWIDTH_1_BYTE) | \@@ -147,6 +149,7 @@ struct fsl_edma_slave_config { struct fsl_edma_chan { struct virt_dma_chan vchan; enum dma_status status; + u32 slave_id; struct fsl_edma_engine *edma; struct fsl_edma_desc *edesc; struct fsl_edma_slave_config fsc;@@ -637,6 +640,45 @@ static struct dma_async_tx_descriptor*fsl_edma_prep_slave_sg( return vchan_tx_prep(&fsl_chan->vchan, &fsl_desc->vdesc, flags); } +static struct dma_async_tx_descriptor * +fsl_edma_prep_memcpy(struct dma_chan *chan, dma_addr_t dst, + dma_addr_t src, size_t len, unsigned long tx_flags) +{ + struct fsl_edma_chan *fsl_chan = to_fsl_edma_chan(chan); + struct fsl_edma_desc *fsl_desc; + u16 soff, doff, attr; + + /* + * use 4-bytes data transfer size if all is 4-bytes aligned, + * else 2-bytes data transfer size of all is 2-bytes aligned, + * otherwise 1-byte tranfer size. + */ + if (src & 0x1 || dst & 0x1 || len & 0x1) { + attr = EDMA_TCD_ATTR_SSIZE_8BIT | EDMA_TCD_ATTR_DSIZE_8BIT; + soff = 0x1; + doff = 0x1; + } else if (src & 0x2 || dst & 0x2 || len & 0x2) { + attr = EDMA_TCD_ATTR_SSIZE_16BIT | EDMA_TCD_ATTR_DSIZE_16BIT; + soff = 0x2; + doff = 0x2; + } else { + attr = EDMA_TCD_ATTR_SSIZE_32BIT | EDMA_TCD_ATTR_DSIZE_32BIT; + soff = 0x4; + doff = 0x4; + } + + fsl_desc = fsl_edma_alloc_desc(fsl_chan, 1); + if (!fsl_desc) + return NULL; + fsl_desc->iscyclic = false; + + fsl_edma_fill_tcd(fsl_desc->tcd[0].vtcd, src, dst, attr, soff, len, + 0, 1, 1, doff, 0, true, true, false); + + return vchan_tx_prep(&fsl_chan->vchan, &fsl_desc->vdesc, tx_flags); + +} + static void fsl_edma_xfer_desc(struct fsl_edma_chan *fsl_chan) { struct virt_dma_desc *vdesc;@@ -735,6 +777,7 @@ static struct dma_chan *fsl_edma_xlate(structof_phandle_args *dma_spec, { struct fsl_edma_engine *fsl_edma = ofdma->of_dma_data; struct dma_chan *chan, *_chan; + struct fsl_edma_chan *fsl_chan; unsigned long chans_per_mux = fsl_edma->n_chans / DMAMUX_NR; if (dma_spec->args_count != 2)@@ -748,8 +791,10 @@ static struct dma_chan *fsl_edma_xlate(structof_phandle_args *dma_spec, chan = dma_get_slave_channel(chan); if (chan) { chan->device->privatecnt++; - fsl_edma_chan_mux(to_fsl_edma_chan(chan), - dma_spec->args[1], true); + fsl_chan = to_fsl_edma_chan(chan); + fsl_chan->slave_id = dma_spec->args[1]; + fsl_edma_chan_mux(fsl_chan, fsl_chan->slave_id, + true); mutex_unlock(&fsl_edma->fsl_edma_mutex); return chan; }@@ -763,6 +808,17 @@ static int fsl_edma_alloc_chan_resources(structdma_chan *chan) { struct fsl_edma_chan *fsl_chan = to_fsl_edma_chan(chan); + /* + * If the slave id of the channel DMAMUX is not set yet, + * this could happy when the channel is requested by the + * dma_request_channel() for memory copy purpose instead + * of by dts binding, then configure the DMAMUX with the + * always on slave id. + */ + if (fsl_chan->slave_id == 0) { + fsl_chan->slave_id = SLAVE_ID_ALWAYSON; + fsl_edma_chan_mux(fsl_chan, fsl_chan->slave_id, true); + } fsl_chan->tcd_pool = dma_pool_create("tcd_pool", chan->device->dev, sizeof(struct fsl_edma_hw_tcd), 32, 0);@@ -778,6 +834,7 @@ static void fsl_edma_free_chan_resources(structdma_chan *chan) spin_lock_irqsave(&fsl_chan->vchan.lock, flags); fsl_edma_disable_request(fsl_chan); fsl_edma_chan_mux(fsl_chan, 0, false); + fsl_chan->slave_id = 0; fsl_chan->edesc = NULL; vchan_get_all_descriptors(&fsl_chan->vchan, &head); spin_unlock_irqrestore(&fsl_chan->vchan.lock, flags);@@ -903,6 +960,7 @@ static int fsl_edma_probe(struct platform_device *pdev) dma_cap_set(DMA_PRIVATE, fsl_edma->dma_dev.cap_mask); dma_cap_set(DMA_SLAVE, fsl_edma->dma_dev.cap_mask); dma_cap_set(DMA_CYCLIC, fsl_edma->dma_dev.cap_mask); + dma_cap_set(DMA_MEMCPY, fsl_edma->dma_dev.cap_mask); fsl_edma->dma_dev.dev = &pdev->dev; fsl_edma->dma_dev.device_alloc_chan_resources@@ -912,6 +970,7 @@ static int fsl_edma_probe(struct platform_device *pdev) fsl_edma->dma_dev.device_tx_status = fsl_edma_tx_status; fsl_edma->dma_dev.device_prep_slave_sg = fsl_edma_prep_slave_sg; fsl_edma->dma_dev.device_prep_dma_cyclic = fsl_edma_prep_dma_cyclic; + fsl_edma->dma_dev.device_prep_dma_memcpy = fsl_edma_prep_memcpy; fsl_edma->dma_dev.device_config = fsl_edma_slave_config; fsl_edma->dma_dev.device_pause = fsl_edma_pause; fsl_edma->dma_dev.device_resume = fsl_edma_resume;