| Message ID | 1531239793-11781-1-git-send-email-yibin.gong@nxp.com |
|---|---|
| Headers | show |
| Series | add memcpy support for sdma | expand |
Hi Robin, On 11-07-18, 00:23, Robin Gong wrote: > Add MEMCPY support, meanwhile, add SDMA_BD_MAX_CNT instead > of '0xffff'. latter part should be its own patch. Never mix things > +static struct dma_async_tx_descriptor *sdma_prep_memcpy( > + struct dma_chan *chan, dma_addr_t dma_dst, > + dma_addr_t dma_src, size_t len, unsigned long flags) > +{ > + struct sdma_channel *sdmac = to_sdma_chan(chan); > + struct sdma_engine *sdma = sdmac->sdma; > + int channel = sdmac->channel; > + size_t count; > + int i = 0, param; > + struct sdma_buffer_descriptor *bd; > + struct sdma_desc *desc; > + > + if (!chan || !len) > + return NULL; > + > + dev_dbg(sdma->dev, "memcpy: %pad->%pad, len=%zu, channel=%d.\n", > + &dma_src, &dma_dst, len, channel); > + > + desc = sdma_transfer_init(sdmac, DMA_MEM_TO_MEM, len / SDMA_BD_MAX_CNT > + + 1); this looks quite odd to read consider: esc = sdma_transfer_init(sdmac, DMA_MEM_TO_MEM, len / SDMA_BD_MAX_CNT + 1); > + if (!desc) > + goto err_out; > + > + do { > + count = min_t(size_t, len, SDMA_BD_MAX_CNT); > + bd = &desc->bd[i]; > + bd->buffer_addr = dma_src; > + bd->ext_buffer_addr = dma_dst; > + bd->mode.count = count; > + desc->chn_count += count; > + > + switch (sdmac->word_size) { > + case DMA_SLAVE_BUSWIDTH_4_BYTES: This looks wrong, we are in memcpy, there is no SLAVE so no SLAVE widths.. > static struct dma_async_tx_descriptor *sdma_prep_slave_sg( > struct dma_chan *chan, struct scatterlist *sgl, > unsigned int sg_len, enum dma_transfer_direction direction, > @@ -1344,9 +1431,9 @@ static struct dma_async_tx_descriptor *sdma_prep_slave_sg( > > count = sg_dma_len(sg); > > - if (count > 0xffff) { > + if (count > SDMA_BD_MAX_CNT) { > dev_err(sdma->dev, "SDMA channel %d: maximum bytes for sg entry exceeded: %d > %d\n", > - channel, count, 0xffff); > + channel, count, SDMA_BD_MAX_CNT); these changes dont belong to this patch > @@ -1486,6 +1573,8 @@ static int sdma_config(struct dma_chan *chan, > sdmac->watermark_level |= (dmaengine_cfg->dst_maxburst << 16) & > SDMA_WATERMARK_LEVEL_HWML; > sdmac->word_size = dmaengine_cfg->dst_addr_width; > + } else if (dmaengine_cfg->direction == DMA_MEM_TO_MEM) { > + sdmac->word_size = dmaengine_cfg->dst_addr_width; same here too, we are in .device_config which deals with slave. Not memcpy! > } else { > sdmac->per_address = dmaengine_cfg->dst_addr; > sdmac->watermark_level = dmaengine_cfg->dst_maxburst * > @@ -1902,6 +1991,7 @@ static int sdma_probe(struct platform_device *pdev) > > dma_cap_set(DMA_SLAVE, sdma->dma_device.cap_mask); > dma_cap_set(DMA_CYCLIC, sdma->dma_device.cap_mask); > + dma_cap_set(DMA_MEMCPY, sdma->dma_device.cap_mask); > > INIT_LIST_HEAD(&sdma->dma_device.channels); > /* Initialize channel parameters */ > @@ -1968,9 +2058,11 @@ static int sdma_probe(struct platform_device *pdev) > sdma->dma_device.dst_addr_widths = SDMA_DMA_BUSWIDTHS; > sdma->dma_device.directions = SDMA_DMA_DIRECTIONS; > sdma->dma_device.residue_granularity = DMA_RESIDUE_GRANULARITY_SEGMENT; > + sdma->dma_device.device_prep_dma_memcpy = sdma_prep_memcpy; > sdma->dma_device.device_issue_pending = sdma_issue_pending; > sdma->dma_device.dev->dma_parms = &sdma->dma_parms; > - dma_set_max_seg_size(sdma->dma_device.dev, 65535); > + sdma->dma_device.copy_align = DMAENGINE_ALIGN_4_BYTES; > + dma_set_max_seg_size(sdma->dma_device.dev, SDMA_BD_MAX_CNT); this line should not be part of this patch
On 11-07-18, 00:23, Robin Gong wrote: > Add check_bd_buswidth() to minimize the code size. this looks mostly fine and I think this should be first patch.. > > Signed-off-by: Robin Gong <yibin.gong@nxp.com> > --- > drivers/dma/imx-sdma.c | 64 +++++++++++++++++++++++--------------------------- > 1 file changed, 29 insertions(+), 35 deletions(-) > > diff --git a/drivers/dma/imx-sdma.c b/drivers/dma/imx-sdma.c > index 27ccabf..ed2267d 100644 > --- a/drivers/dma/imx-sdma.c > +++ b/drivers/dma/imx-sdma.c > @@ -1326,6 +1326,33 @@ static struct sdma_desc *sdma_transfer_init(struct sdma_channel *sdmac, > return NULL; > } > > +static int check_bd_buswidth(struct sdma_buffer_descriptor *bd, > + struct sdma_channel *sdmac, int count, > + dma_addr_t dma_dst, dma_addr_t dma_src) > +{ > + int ret = 0; > + > + switch (sdmac->word_size) { > + case DMA_SLAVE_BUSWIDTH_4_BYTES: > + bd->mode.command = 0; > + if ((count | dma_dst | dma_src) & 3) > + ret = -EINVAL; > + break; empty line after each break please > + case DMA_SLAVE_BUSWIDTH_2_BYTES: > + bd->mode.command = 2; > + if ((count | dma_dst | dma_src) & 1) > + ret = -EINVAL; > + break; > + case DMA_SLAVE_BUSWIDTH_1_BYTE: > + bd->mode.command = 1; > + break; > + default: > + return -EINVAL; > + } > + > + return ret; > +} > +
On 11-07-18, 00:23, Robin Gong wrote: > dmatest(memcpy) will never call dmaengine_slave_config before prep, and that should have been a hint to you that you should not expect that > so jobs in dmaengine_slave_config need to be moved into somewhere > before device_prep_dma_memcpy. Besides, dmatest never setup chan > ->private as other common case like uart/audio/spi will always setup > chan->private. Here check it to judge if it's dmatest case and do > jobs in slave_config. and you should not do anything for dmatest. Supporting it means memcpy implementation is not correct :) > > Signed-off-by: Robin Gong <yibin.gong@nxp.com> > --- > drivers/dma/imx-sdma.c | 37 ++++++++++++++++++++++++++++--------- > 1 file changed, 28 insertions(+), 9 deletions(-) > > diff --git a/drivers/dma/imx-sdma.c b/drivers/dma/imx-sdma.c > index ed2267d..48f3749 100644 > --- a/drivers/dma/imx-sdma.c > +++ b/drivers/dma/imx-sdma.c > @@ -1222,10 +1222,36 @@ static int sdma_alloc_chan_resources(struct dma_chan *chan) > { > struct sdma_channel *sdmac = to_sdma_chan(chan); > struct imx_dma_data *data = chan->private; > + struct imx_dma_data default_data; > int prio, ret; > > - if (!data) > - return -EINVAL; > + ret = clk_enable(sdmac->sdma->clk_ipg); > + if (ret) > + return ret; > + ret = clk_enable(sdmac->sdma->clk_ahb); > + if (ret) > + goto disable_clk_ipg; > + /* > + * dmatest(memcpy) will never call dmaengine_slave_config before prep, > + * so jobs in dmaengine_slave_config need to be moved into somewhere > + * before device_prep_dma_memcpy. Besides, dmatest never setup chan > + * ->private as other common cases like uart/audio/spi will setup > + * chan->private always. Here check it to judge if it's dmatest case > + * and do jobs in slave_config. > + */ > + if (!data) { > + dev_warn(sdmac->sdma->dev, "dmatest is running?\n"); why is that a warning! > + sdmac->word_size = sdmac->sdma->dma_device.copy_align; > + default_data.priority = 2; > + default_data.peripheral_type = IMX_DMATYPE_MEMORY; > + default_data.dma_request = 0; > + default_data.dma_request2 = 0; > + data = &default_data; > + > + sdma_config_ownership(sdmac, false, true, false); > + sdma_get_pc(sdmac, IMX_DMATYPE_MEMORY); > + sdma_load_context(sdmac); > + } this needs to be default for memcpy
> Hi Robin, > > On 11-07-18, 00:23, Robin Gong wrote: > > Add MEMCPY support, meanwhile, add SDMA_BD_MAX_CNT instead of > > '0xffff'. > > latter part should be its own patch. Never mix things Okay, I will split it even for this minor change. > > > +static struct dma_async_tx_descriptor *sdma_prep_memcpy( > > + struct dma_chan *chan, dma_addr_t dma_dst, > > + dma_addr_t dma_src, size_t len, unsigned long flags) { > > + struct sdma_channel *sdmac = to_sdma_chan(chan); > > + struct sdma_engine *sdma = sdmac->sdma; > > + int channel = sdmac->channel; > > + size_t count; > > + int i = 0, param; > > + struct sdma_buffer_descriptor *bd; > > + struct sdma_desc *desc; > > + > > + if (!chan || !len) > > + return NULL; > > + > > + dev_dbg(sdma->dev, "memcpy: %pad->%pad, len=%zu, channel=%d.\n", > > + &dma_src, &dma_dst, len, channel); > > + > > + desc = sdma_transfer_init(sdmac, DMA_MEM_TO_MEM, len / > SDMA_BD_MAX_CNT > > + + 1); > > this looks quite odd to read consider: > > esc = sdma_transfer_init(sdmac, DMA_MEM_TO_MEM, > len / SDMA_BD_MAX_CNT + 1); > Okay, will fix on v2. > > + if (!desc) > > + goto err_out; > > + > > + do { > > + count = min_t(size_t, len, SDMA_BD_MAX_CNT); > > + bd = &desc->bd[i]; > > + bd->buffer_addr = dma_src; > > + bd->ext_buffer_addr = dma_dst; > > + bd->mode.count = count; > > + desc->chn_count += count; > > + > > + switch (sdmac->word_size) { > > + case DMA_SLAVE_BUSWIDTH_4_BYTES: > > This looks wrong, we are in memcpy, there is no SLAVE so no SLAVE widths.. > Okay, will remove check bus width. > > static struct dma_async_tx_descriptor *sdma_prep_slave_sg( > > struct dma_chan *chan, struct scatterlist *sgl, > > unsigned int sg_len, enum dma_transfer_direction direction, @@ > > -1344,9 +1431,9 @@ static struct dma_async_tx_descriptor > > *sdma_prep_slave_sg( > > > > count = sg_dma_len(sg); > > > > - if (count > 0xffff) { > > + if (count > SDMA_BD_MAX_CNT) { > > dev_err(sdma->dev, "SDMA channel %d: maximum bytes for sg > entry exceeded: %d > %d\n", > > - channel, count, 0xffff); > > + channel, count, SDMA_BD_MAX_CNT); > > these changes dont belong to this patch Will split in v2. > > > @@ -1486,6 +1573,8 @@ static int sdma_config(struct dma_chan *chan, > > sdmac->watermark_level |= (dmaengine_cfg->dst_maxburst << 16) > & > > SDMA_WATERMARK_LEVEL_HWML; > > sdmac->word_size = dmaengine_cfg->dst_addr_width; > > + } else if (dmaengine_cfg->direction == DMA_MEM_TO_MEM) { > > + sdmac->word_size = dmaengine_cfg->dst_addr_width; > > same here too, we are in .device_config which deals with slave. Not memcpy! Will remove it. > > > } else { > > sdmac->per_address = dmaengine_cfg->dst_addr; > > sdmac->watermark_level = dmaengine_cfg->dst_maxburst * @@ > -1902,6 > > +1991,7 @@ static int sdma_probe(struct platform_device *pdev) > > > > dma_cap_set(DMA_SLAVE, sdma->dma_device.cap_mask); > > dma_cap_set(DMA_CYCLIC, sdma->dma_device.cap_mask); > > + dma_cap_set(DMA_MEMCPY, sdma->dma_device.cap_mask); > > > > INIT_LIST_HEAD(&sdma->dma_device.channels); > > /* Initialize channel parameters */ > > @@ -1968,9 +2058,11 @@ static int sdma_probe(struct platform_device > *pdev) > > sdma->dma_device.dst_addr_widths = SDMA_DMA_BUSWIDTHS; > > sdma->dma_device.directions = SDMA_DMA_DIRECTIONS; > > sdma->dma_device.residue_granularity = > > DMA_RESIDUE_GRANULARITY_SEGMENT; > > + sdma->dma_device.device_prep_dma_memcpy = sdma_prep_memcpy; > > sdma->dma_device.device_issue_pending = sdma_issue_pending; > > sdma->dma_device.dev->dma_parms = &sdma->dma_parms; > > - dma_set_max_seg_size(sdma->dma_device.dev, 65535); > > + sdma->dma_device.copy_align = DMAENGINE_ALIGN_4_BYTES; > > + dma_set_max_seg_size(sdma->dma_device.dev, SDMA_BD_MAX_CNT); > > this line should not be part of this patch > > -- > ~Vinod
> -----Original Message----- > From: Vinod [mailto:vkoul@kernel.org] > Sent: 2018年7月10日 23:31 > To: Robin Gong <yibin.gong@nxp.com> > Cc: dan.j.williams@intel.com; shawnguo@kernel.org; > s.hauer@pengutronix.de; Fabio Estevam <fabio.estevam@nxp.com>; > linux@armlinux.org.uk; linux-arm-kernel@lists.infradead.org; > kernel@pengutronix.de; dmaengine@vger.kernel.org; > linux-kernel@vger.kernel.org; dl-linux-imx <linux-imx@nxp.com> > Subject: Re: [PATCH v1 2/4] dmaengine: imx-sdma: add check_bd_buswidth() to > kill the dulicated code > > On 11-07-18, 00:23, Robin Gong wrote: > > Add check_bd_buswidth() to minimize the code size. > > this looks mostly fine and I think this should be first patch.. Since no need to check bus width in memcpy case, I'll remove this patch too. > > > > > Signed-off-by: Robin Gong <yibin.gong@nxp.com> > > --- > > drivers/dma/imx-sdma.c | 64 > > +++++++++++++++++++++++--------------------------- > > 1 file changed, 29 insertions(+), 35 deletions(-) > > > > diff --git a/drivers/dma/imx-sdma.c b/drivers/dma/imx-sdma.c index > > 27ccabf..ed2267d 100644 > > --- a/drivers/dma/imx-sdma.c > > +++ b/drivers/dma/imx-sdma.c > > @@ -1326,6 +1326,33 @@ static struct sdma_desc > *sdma_transfer_init(struct sdma_channel *sdmac, > > return NULL; > > } > > > > +static int check_bd_buswidth(struct sdma_buffer_descriptor *bd, > > + struct sdma_channel *sdmac, int count, > > + dma_addr_t dma_dst, dma_addr_t dma_src) { > > + int ret = 0; > > + > > + switch (sdmac->word_size) { > > + case DMA_SLAVE_BUSWIDTH_4_BYTES: > > + bd->mode.command = 0; > > + if ((count | dma_dst | dma_src) & 3) > > + ret = -EINVAL; > > + break; > > empty line after each break please > > > + case DMA_SLAVE_BUSWIDTH_2_BYTES: > > + bd->mode.command = 2; > > + if ((count | dma_dst | dma_src) & 1) > > + ret = -EINVAL; > > + break; > > + case DMA_SLAVE_BUSWIDTH_1_BYTE: > > + bd->mode.command = 1; > > + break; > > + default: > > + return -EINVAL; > > + } > > + > > + return ret; > > +} > > + > -- > ~Vinod
On Wed, Jul 11, 2018 at 12:23:10AM +0800, Robin Gong wrote: > Add MEMCPY support, meanwhile, add SDMA_BD_MAX_CNT instead > of '0xffff'. > > Signed-off-by: Robin Gong <yibin.gong@nxp.com> > --- > +static struct dma_async_tx_descriptor *sdma_prep_memcpy( > + struct dma_chan *chan, dma_addr_t dma_dst, > + dma_addr_t dma_src, size_t len, unsigned long flags) > +{ > + struct sdma_channel *sdmac = to_sdma_chan(chan); > + struct sdma_engine *sdma = sdmac->sdma; > + int channel = sdmac->channel; > + size_t count; > + int i = 0, param; > + struct sdma_buffer_descriptor *bd; > + struct sdma_desc *desc; > + > + if (!chan || !len) > + return NULL; > + > + dev_dbg(sdma->dev, "memcpy: %pad->%pad, len=%zu, channel=%d.\n", > + &dma_src, &dma_dst, len, channel); > + > + desc = sdma_transfer_init(sdmac, DMA_MEM_TO_MEM, len / SDMA_BD_MAX_CNT > + + 1); > + if (!desc) > + goto err_out; > + > + do { > + count = min_t(size_t, len, SDMA_BD_MAX_CNT); When len is bigger than 0xffff you initialize count to 0xffff... > + bd = &desc->bd[i]; > + bd->buffer_addr = dma_src; > + bd->ext_buffer_addr = dma_dst; > + bd->mode.count = count; > + desc->chn_count += count; > + > + switch (sdmac->word_size) { > + case DMA_SLAVE_BUSWIDTH_4_BYTES: > + bd->mode.command = 0; > + if ((count | dma_src | dma_dst) & 3) > + goto err_bd_out; > + break; ...In which case you bail out here with an error. Please make sure bigger transfers are working. > + case DMA_SLAVE_BUSWIDTH_2_BYTES: > + bd->mode.command = 2; > + if ((count | dma_src | dma_dst) & 1) > + goto err_bd_out; > + break; > + case DMA_SLAVE_BUSWIDTH_1_BYTE: > + bd->mode.command = 1; > + break; > + default: > + goto err_bd_out; > + } > + > + dma_src += count; > + dma_dst += count; > + len -= count; > + i++; > + > + param = BD_DONE | BD_EXTD | BD_CONT; Probably better readable if you drop BD_CONT here and do a if (len) { param |= BD_CONT; } else { ... } Sascha
On Wed, Jul 11, 2018 at 12:23:11AM +0800, Robin Gong wrote: > Add check_bd_buswidth() to minimize the code size. > > Signed-off-by: Robin Gong <yibin.gong@nxp.com> > --- > drivers/dma/imx-sdma.c | 64 +++++++++++++++++++++++--------------------------- > 1 file changed, 29 insertions(+), 35 deletions(-) > > diff --git a/drivers/dma/imx-sdma.c b/drivers/dma/imx-sdma.c > index 27ccabf..ed2267d 100644 > --- a/drivers/dma/imx-sdma.c > +++ b/drivers/dma/imx-sdma.c > @@ -1326,6 +1326,33 @@ static struct sdma_desc *sdma_transfer_init(struct sdma_channel *sdmac, > return NULL; > } > > +static int check_bd_buswidth(struct sdma_buffer_descriptor *bd, > + struct sdma_channel *sdmac, int count, > + dma_addr_t dma_dst, dma_addr_t dma_src) > +{ Better name this function set_bd_buswidth. I would assume that a function named check_foo actually checks something, but doesn't set anything. Sascha
> -----Original Message----- > From: Vinod [mailto:vkoul@kernel.org] > Sent: 2018年7月10日 23:33 > To: Robin Gong <yibin.gong@nxp.com> > Cc: dan.j.williams@intel.com; shawnguo@kernel.org; > s.hauer@pengutronix.de; Fabio Estevam <fabio.estevam@nxp.com>; > linux@armlinux.org.uk; linux-arm-kernel@lists.infradead.org; > kernel@pengutronix.de; dmaengine@vger.kernel.org; > linux-kernel@vger.kernel.org; dl-linux-imx <linux-imx@nxp.com> > Subject: Re: [PATCH v1 3/4] dmaengine: imx-sdma: support dmatest > > On 11-07-18, 00:23, Robin Gong wrote: > > dmatest(memcpy) will never call dmaengine_slave_config before prep, > > and that should have been a hint to you that you should not expect that > > > so jobs in dmaengine_slave_config need to be moved into somewhere > > before device_prep_dma_memcpy. Besides, dmatest never setup chan > > ->private as other common case like uart/audio/spi will always setup > > chan->private. Here check it to judge if it's dmatest case and do > > jobs in slave_config. > > and you should not do anything for dmatest. Supporting it means memcpy > implementation is not correct :) Okay, I will any word about dmatest here since memcpy assume no calling slave_config. > > > > > Signed-off-by: Robin Gong <yibin.gong@nxp.com> > > --- > > drivers/dma/imx-sdma.c | 37 ++++++++++++++++++++++++++++--------- > > 1 file changed, 28 insertions(+), 9 deletions(-) > > > > diff --git a/drivers/dma/imx-sdma.c b/drivers/dma/imx-sdma.c index > > ed2267d..48f3749 100644 > > --- a/drivers/dma/imx-sdma.c > > +++ b/drivers/dma/imx-sdma.c > > @@ -1222,10 +1222,36 @@ static int sdma_alloc_chan_resources(struct > > dma_chan *chan) { > > struct sdma_channel *sdmac = to_sdma_chan(chan); > > struct imx_dma_data *data = chan->private; > > + struct imx_dma_data default_data; > > int prio, ret; > > > > - if (!data) > > - return -EINVAL; > > + ret = clk_enable(sdmac->sdma->clk_ipg); > > + if (ret) > > + return ret; > > + ret = clk_enable(sdmac->sdma->clk_ahb); > > + if (ret) > > + goto disable_clk_ipg; > > + /* > > + * dmatest(memcpy) will never call dmaengine_slave_config before prep, > > + * so jobs in dmaengine_slave_config need to be moved into somewhere > > + * before device_prep_dma_memcpy. Besides, dmatest never setup chan > > + * ->private as other common cases like uart/audio/spi will setup > > + * chan->private always. Here check it to judge if it's dmatest case > > + * and do jobs in slave_config. > > + */ > > + if (!data) { > > + dev_warn(sdmac->sdma->dev, "dmatest is running?\n"); > > why is that a warning! Current SDMA driver assume filter function to set chan->private with specific data (struct imx_dma_data dma_data)like below (sound/soc/fsl/fsl_asrc_dma.c): static bool filter(struct dma_chan *chan, void *param) { if (!imx_dma_is_general_purpose(chan)) return false; chan->private = param; return true; } But in memcpy case, at lease dmatest case, no chan->private set in its filter function. So here take dmatest a special case and do some prepare jobs for memcpy. But if the Upper device driver call dma_request_channel() with their specific filter without 'chan->private' setting in the future. The warning message is a useful hint to them to Add 'chan->private' in filter function. Or doc it somewhere? > > > + sdmac->word_size = sdmac->sdma->dma_device.copy_align; > > + default_data.priority = 2; > > + default_data.peripheral_type = IMX_DMATYPE_MEMORY; > > + default_data.dma_request = 0; > > + default_data.dma_request2 = 0; > > + data = &default_data; > > + > > + sdma_config_ownership(sdmac, false, true, false); > > + sdma_get_pc(sdmac, IMX_DMATYPE_MEMORY); > > + sdma_load_context(sdmac); > > + } > > this needs to be default for memcpy > > -- > ~Vinod
On Wed, Jul 11, 2018 at 06:37:02AM +0000, Robin Gong wrote: > > > -----Original Message----- > > From: Vinod [mailto:vkoul@kernel.org] > > Sent: 2018年7月10日 23:33 > > To: Robin Gong <yibin.gong@nxp.com> > > Cc: dan.j.williams@intel.com; shawnguo@kernel.org; > > s.hauer@pengutronix.de; Fabio Estevam <fabio.estevam@nxp.com>; > > linux@armlinux.org.uk; linux-arm-kernel@lists.infradead.org; > > kernel@pengutronix.de; dmaengine@vger.kernel.org; > > linux-kernel@vger.kernel.org; dl-linux-imx <linux-imx@nxp.com> > > Subject: Re: [PATCH v1 3/4] dmaengine: imx-sdma: support dmatest > > > > On 11-07-18, 00:23, Robin Gong wrote: > > > dmatest(memcpy) will never call dmaengine_slave_config before prep, > > > > and that should have been a hint to you that you should not expect that > > > > > so jobs in dmaengine_slave_config need to be moved into somewhere > > > before device_prep_dma_memcpy. Besides, dmatest never setup chan > > > ->private as other common case like uart/audio/spi will always setup > > > chan->private. Here check it to judge if it's dmatest case and do > > > jobs in slave_config. > > > > and you should not do anything for dmatest. Supporting it means memcpy > > implementation is not correct :) > Okay, I will any word about dmatest here since memcpy assume no calling > slave_config. > > > > > > > > Signed-off-by: Robin Gong <yibin.gong@nxp.com> > > > --- > > > drivers/dma/imx-sdma.c | 37 ++++++++++++++++++++++++++++--------- > > > 1 file changed, 28 insertions(+), 9 deletions(-) > > > > > > diff --git a/drivers/dma/imx-sdma.c b/drivers/dma/imx-sdma.c index > > > ed2267d..48f3749 100644 > > > --- a/drivers/dma/imx-sdma.c > > > +++ b/drivers/dma/imx-sdma.c > > > @@ -1222,10 +1222,36 @@ static int sdma_alloc_chan_resources(struct > > > dma_chan *chan) { > > > struct sdma_channel *sdmac = to_sdma_chan(chan); > > > struct imx_dma_data *data = chan->private; > > > + struct imx_dma_data default_data; > > > int prio, ret; > > > > > > - if (!data) > > > - return -EINVAL; > > > + ret = clk_enable(sdmac->sdma->clk_ipg); > > > + if (ret) > > > + return ret; > > > + ret = clk_enable(sdmac->sdma->clk_ahb); > > > + if (ret) > > > + goto disable_clk_ipg; > > > + /* > > > + * dmatest(memcpy) will never call dmaengine_slave_config before prep, > > > + * so jobs in dmaengine_slave_config need to be moved into somewhere > > > + * before device_prep_dma_memcpy. Besides, dmatest never setup chan > > > + * ->private as other common cases like uart/audio/spi will setup > > > + * chan->private always. Here check it to judge if it's dmatest case > > > + * and do jobs in slave_config. > > > + */ > > > + if (!data) { > > > + dev_warn(sdmac->sdma->dev, "dmatest is running?\n"); > > > > why is that a warning! > Current SDMA driver assume filter function to set chan->private with specific data > (struct imx_dma_data dma_data)like below (sound/soc/fsl/fsl_asrc_dma.c): > static bool filter(struct dma_chan *chan, void *param) > { > if (!imx_dma_is_general_purpose(chan)) > return false; > chan->private = param; > return true; > } > > But in memcpy case, at lease dmatest case, no chan->private set in its filter function. > So here take dmatest a special case and do some prepare jobs for memcpy. But if the > Upper device driver call dma_request_channel() with their specific filter without > 'chan->private' setting in the future. The warning message is a useful hint to them to > Add 'chan->private' in filter function. Or doc it somewhere? Instead of doing heuristics to guess whether we are doing memcpy you could instead make memcpy the default when slave_config is not called, i.e. drop the if (!data) check completely. > > > > > + sdmac->word_size = sdmac->sdma->dma_device.copy_align; > > > + default_data.priority = 2; > > > + default_data.peripheral_type = IMX_DMATYPE_MEMORY; > > > + default_data.dma_request = 0; > > > + default_data.dma_request2 = 0; > > > + data = &default_data; > > > + > > > + sdma_config_ownership(sdmac, false, true, false); > > > + sdma_get_pc(sdmac, IMX_DMATYPE_MEMORY); > > > + sdma_load_context(sdmac); > > > + } > > > > this needs to be default for memcpy The problem seems to be that we do not know whether we are doing memcpy or not. Normally we get the information how a channel is to be configured in dma_device->device_config, but this function is not called in the memcpy case. An alternative might also be to do the setup in dma_device->device_prep_dma_memcpy. Sascha
> -----Original Message----- > From: Sascha Hauer [mailto:s.hauer@pengutronix.de] > Sent: 2018年7月11日 14:25 > To: Robin Gong <yibin.gong@nxp.com> > Cc: vkoul@kernel.org; dan.j.williams@intel.com; shawnguo@kernel.org; Fabio > Estevam <fabio.estevam@nxp.com>; linux@armlinux.org.uk; > linux-arm-kernel@lists.infradead.org; kernel@pengutronix.de; > dmaengine@vger.kernel.org; linux-kernel@vger.kernel.org; dl-linux-imx > <linux-imx@nxp.com> > Subject: Re: [PATCH v1 1/4] dmaengine: imx-sdma: add memcpy interface > > On Wed, Jul 11, 2018 at 12:23:10AM +0800, Robin Gong wrote: > > Add MEMCPY support, meanwhile, add SDMA_BD_MAX_CNT instead of > > '0xffff'. > > > > Signed-off-by: Robin Gong <yibin.gong@nxp.com> > > --- > > +static struct dma_async_tx_descriptor *sdma_prep_memcpy( > > + struct dma_chan *chan, dma_addr_t dma_dst, > > + dma_addr_t dma_src, size_t len, unsigned long flags) { > > + struct sdma_channel *sdmac = to_sdma_chan(chan); > > + struct sdma_engine *sdma = sdmac->sdma; > > + int channel = sdmac->channel; > > + size_t count; > > + int i = 0, param; > > + struct sdma_buffer_descriptor *bd; > > + struct sdma_desc *desc; > > + > > + if (!chan || !len) > > + return NULL; > > + > > + dev_dbg(sdma->dev, "memcpy: %pad->%pad, len=%zu, channel=%d.\n", > > + &dma_src, &dma_dst, len, channel); > > + > > + desc = sdma_transfer_init(sdmac, DMA_MEM_TO_MEM, len / > SDMA_BD_MAX_CNT > > + + 1); > > + if (!desc) > > + goto err_out; > > + > > + do { > > + count = min_t(size_t, len, SDMA_BD_MAX_CNT); > > When len is bigger than 0xffff you initialize count to 0xffff... In this case, the data will be split into several bds, for example, If the total count is 0x10000, two bd used then. One is for 0xffff, Another is for the last 1 > > > + bd = &desc->bd[i]; > > + bd->buffer_addr = dma_src; > > + bd->ext_buffer_addr = dma_dst; > > + bd->mode.count = count; > > + desc->chn_count += count; > > + > > + switch (sdmac->word_size) { > > + case DMA_SLAVE_BUSWIDTH_4_BYTES: > > + bd->mode.command = 0; > > + if ((count | dma_src | dma_dst) & 3) > > + goto err_bd_out; > > + break; > > ...In which case you bail out here with an error. In case dma_src/dma_dst is not align with bus width. But I'll remove such bus width in v2 as Vinod comments. > > Please make sure bigger transfers are working. > > > + case DMA_SLAVE_BUSWIDTH_2_BYTES: > > + bd->mode.command = 2; > > + if ((count | dma_src | dma_dst) & 1) > > + goto err_bd_out; > > + break; > > + case DMA_SLAVE_BUSWIDTH_1_BYTE: > > + bd->mode.command = 1; > > + break; > > + default: > > + goto err_bd_out; > > + } > > + > > + dma_src += count; > > + dma_dst += count; > > + len -= count; > > + i++; > > + > > + param = BD_DONE | BD_EXTD | BD_CONT; > > Probably better readable if you drop BD_CONT here and do a > > if (len) { > param |= BD_CONT; > } else { > ... > } Okay. Will improve it. > > Sascha > > -- > Pengutronix e.K. | > | > Industrial Linux Solutions | > https://emea01.safelinks.protection.outlook.com/?url=http%3A%2F%2Fwww. > pengutronix.de%2F&data=02%7C01%7Cyibin.gong%40nxp.com%7C0aa96 > 135702b4414dbe708d5e6f709fe%7C686ea1d3bc2b4c6fa92cd99c5c301635%7C > 0%7C0%7C636668871060103585&sdata=k2h0RIWlujaCs8ioduJsnJAfGW0 > ZS7uzlHMMtjpQ%2Fw4%3D&reserved=0 | > Peiner Str. 6-8, 31137 Hildesheim, Germany | Phone: +49-5121-206917-0 | > Amtsgericht Hildesheim, HRA 2686 | Fax: > +49-5121-206917-5555 |
On Wed, Jul 11, 2018 at 06:56:18AM +0000, Robin Gong wrote: > > > -----Original Message----- > > From: Sascha Hauer [mailto:s.hauer@pengutronix.de] > > Sent: 2018年7月11日 14:25 > > To: Robin Gong <yibin.gong@nxp.com> > > Cc: vkoul@kernel.org; dan.j.williams@intel.com; shawnguo@kernel.org; Fabio > > Estevam <fabio.estevam@nxp.com>; linux@armlinux.org.uk; > > linux-arm-kernel@lists.infradead.org; kernel@pengutronix.de; > > dmaengine@vger.kernel.org; linux-kernel@vger.kernel.org; dl-linux-imx > > <linux-imx@nxp.com> > > Subject: Re: [PATCH v1 1/4] dmaengine: imx-sdma: add memcpy interface > > > > On Wed, Jul 11, 2018 at 12:23:10AM +0800, Robin Gong wrote: > > > Add MEMCPY support, meanwhile, add SDMA_BD_MAX_CNT instead of > > > '0xffff'. > > > > > > Signed-off-by: Robin Gong <yibin.gong@nxp.com> > > > --- > > > +static struct dma_async_tx_descriptor *sdma_prep_memcpy( > > > + struct dma_chan *chan, dma_addr_t dma_dst, > > > + dma_addr_t dma_src, size_t len, unsigned long flags) { > > > + struct sdma_channel *sdmac = to_sdma_chan(chan); > > > + struct sdma_engine *sdma = sdmac->sdma; > > > + int channel = sdmac->channel; > > > + size_t count; > > > + int i = 0, param; > > > + struct sdma_buffer_descriptor *bd; > > > + struct sdma_desc *desc; > > > + > > > + if (!chan || !len) > > > + return NULL; > > > + > > > + dev_dbg(sdma->dev, "memcpy: %pad->%pad, len=%zu, channel=%d.\n", > > > + &dma_src, &dma_dst, len, channel); > > > + > > > + desc = sdma_transfer_init(sdmac, DMA_MEM_TO_MEM, len / > > SDMA_BD_MAX_CNT > > > + + 1); > > > + if (!desc) > > > + goto err_out; > > > + > > > + do { > > > + count = min_t(size_t, len, SDMA_BD_MAX_CNT); > > > > When len is bigger than 0xffff you initialize count to 0xffff... > In this case, the data will be split into several bds, for example, > If the total count is 0x10000, two bd used then. One is for 0xffff, > Another is for the last 1 And you are doing byte size DMA? Wouldn't word size accesses be more optimal? Sascha
> -----Original Message----- > From: Sascha Hauer [mailto:s.hauer@pengutronix.de] > Sent: 2018年7月11日 15:01 > To: Robin Gong <yibin.gong@nxp.com> > Cc: vkoul@kernel.org; dan.j.williams@intel.com; shawnguo@kernel.org; Fabio > Estevam <fabio.estevam@nxp.com>; linux@armlinux.org.uk; > linux-arm-kernel@lists.infradead.org; kernel@pengutronix.de; > dmaengine@vger.kernel.org; linux-kernel@vger.kernel.org; dl-linux-imx > <linux-imx@nxp.com> > Subject: Re: [PATCH v1 1/4] dmaengine: imx-sdma: add memcpy interface > > On Wed, Jul 11, 2018 at 06:56:18AM +0000, Robin Gong wrote: > > > > > -----Original Message----- > > > From: Sascha Hauer [mailto:s.hauer@pengutronix.de] > > > Sent: 2018年7月11日 14:25 > > > To: Robin Gong <yibin.gong@nxp.com> > > > Cc: vkoul@kernel.org; dan.j.williams@intel.com; shawnguo@kernel.org; > > > Fabio Estevam <fabio.estevam@nxp.com>; linux@armlinux.org.uk; > > > linux-arm-kernel@lists.infradead.org; kernel@pengutronix.de; > > > dmaengine@vger.kernel.org; linux-kernel@vger.kernel.org; > > > dl-linux-imx <linux-imx@nxp.com> > > > Subject: Re: [PATCH v1 1/4] dmaengine: imx-sdma: add memcpy > > > interface > > > > > > On Wed, Jul 11, 2018 at 12:23:10AM +0800, Robin Gong wrote: > > > > Add MEMCPY support, meanwhile, add SDMA_BD_MAX_CNT instead of > > > > '0xffff'. > > > > > > > > Signed-off-by: Robin Gong <yibin.gong@nxp.com> > > > > --- > > > > +static struct dma_async_tx_descriptor *sdma_prep_memcpy( > > > > + struct dma_chan *chan, dma_addr_t dma_dst, > > > > + dma_addr_t dma_src, size_t len, unsigned long flags) { > > > > + struct sdma_channel *sdmac = to_sdma_chan(chan); > > > > + struct sdma_engine *sdma = sdmac->sdma; > > > > + int channel = sdmac->channel; > > > > + size_t count; > > > > + int i = 0, param; > > > > + struct sdma_buffer_descriptor *bd; > > > > + struct sdma_desc *desc; > > > > + > > > > + if (!chan || !len) > > > > + return NULL; > > > > + > > > > + dev_dbg(sdma->dev, "memcpy: %pad->%pad, len=%zu, > channel=%d.\n", > > > > + &dma_src, &dma_dst, len, channel); > > > > + > > > > + desc = sdma_transfer_init(sdmac, DMA_MEM_TO_MEM, len / > > > SDMA_BD_MAX_CNT > > > > + + 1); > > > > + if (!desc) > > > > + goto err_out; > > > > + > > > > + do { > > > > + count = min_t(size_t, len, SDMA_BD_MAX_CNT); > > > > > > When len is bigger than 0xffff you initialize count to 0xffff... > > In this case, the data will be split into several bds, for example, If > > the total count is 0x10000, two bd used then. One is for 0xffff, > > Another is for the last 1 > > And you are doing byte size DMA? Wouldn't word size accesses be more > optimal? > > Sascha Default is words, and I'll force the buswidth to word and set into BD. sdma->dma_device.copy_align = DMAENGINE_ALIGN_4_BYTES; > > > -- > Pengutronix e.K. | > | > Industrial Linux Solutions | > https://emea01.safelinks.protection.outlook.com/?url=http%3A%2F%2Fwww. > pengutronix.de%2F&data=02%7C01%7Cyibin.gong%40nxp.com%7C41fbd > 3765b10424a7f1e08d5e6fc196d%7C686ea1d3bc2b4c6fa92cd99c5c301635%7C > 0%7C0%7C636668892793035822&sdata=mLKjTKaojuO1Zv%2F4ohwkkzeK > FDFbmqYturqh6eSblIM%3D&reserved=0 | > Peiner Str. 6-8, 31137 Hildesheim, Germany | Phone: +49-5121-206917-0 | > Amtsgericht Hildesheim, HRA 2686 | Fax: > +49-5121-206917-5555 |
On Wed, Jul 11, 2018 at 07:05:23AM +0000, Robin Gong wrote: > > > > -----Original Message----- > > From: Sascha Hauer [mailto:s.hauer@pengutronix.de] > > Sent: 2018年7月11日 15:01 > > To: Robin Gong <yibin.gong@nxp.com> > > Cc: vkoul@kernel.org; dan.j.williams@intel.com; shawnguo@kernel.org; Fabio > > Estevam <fabio.estevam@nxp.com>; linux@armlinux.org.uk; > > linux-arm-kernel@lists.infradead.org; kernel@pengutronix.de; > > dmaengine@vger.kernel.org; linux-kernel@vger.kernel.org; dl-linux-imx > > <linux-imx@nxp.com> > > Subject: Re: [PATCH v1 1/4] dmaengine: imx-sdma: add memcpy interface > > > > On Wed, Jul 11, 2018 at 06:56:18AM +0000, Robin Gong wrote: > > > > > > > -----Original Message----- > > > > From: Sascha Hauer [mailto:s.hauer@pengutronix.de] > > > > Sent: 2018年7月11日 14:25 > > > > To: Robin Gong <yibin.gong@nxp.com> > > > > Cc: vkoul@kernel.org; dan.j.williams@intel.com; shawnguo@kernel.org; > > > > Fabio Estevam <fabio.estevam@nxp.com>; linux@armlinux.org.uk; > > > > linux-arm-kernel@lists.infradead.org; kernel@pengutronix.de; > > > > dmaengine@vger.kernel.org; linux-kernel@vger.kernel.org; > > > > dl-linux-imx <linux-imx@nxp.com> > > > > Subject: Re: [PATCH v1 1/4] dmaengine: imx-sdma: add memcpy > > > > interface > > > > > > > > On Wed, Jul 11, 2018 at 12:23:10AM +0800, Robin Gong wrote: > > > > > Add MEMCPY support, meanwhile, add SDMA_BD_MAX_CNT instead of > > > > > '0xffff'. > > > > > > > > > > Signed-off-by: Robin Gong <yibin.gong@nxp.com> > > > > > --- > > > > > +static struct dma_async_tx_descriptor *sdma_prep_memcpy( > > > > > + struct dma_chan *chan, dma_addr_t dma_dst, > > > > > + dma_addr_t dma_src, size_t len, unsigned long flags) { > > > > > + struct sdma_channel *sdmac = to_sdma_chan(chan); > > > > > + struct sdma_engine *sdma = sdmac->sdma; > > > > > + int channel = sdmac->channel; > > > > > + size_t count; > > > > > + int i = 0, param; > > > > > + struct sdma_buffer_descriptor *bd; > > > > > + struct sdma_desc *desc; > > > > > + > > > > > + if (!chan || !len) > > > > > + return NULL; > > > > > + > > > > > + dev_dbg(sdma->dev, "memcpy: %pad->%pad, len=%zu, > > channel=%d.\n", > > > > > + &dma_src, &dma_dst, len, channel); > > > > > + > > > > > + desc = sdma_transfer_init(sdmac, DMA_MEM_TO_MEM, len / > > > > SDMA_BD_MAX_CNT > > > > > + + 1); > > > > > + if (!desc) > > > > > + goto err_out; > > > > > + > > > > > + do { > > > > > + count = min_t(size_t, len, SDMA_BD_MAX_CNT); > > > > > > > > When len is bigger than 0xffff you initialize count to 0xffff... > > > In this case, the data will be split into several bds, for example, If > > > the total count is 0x10000, two bd used then. One is for 0xffff, > > > Another is for the last 1 > > > > And you are doing byte size DMA? Wouldn't word size accesses be more > > optimal? > > > > Sascha > Default is words, and I'll force the buswidth to word and set into BD. > sdma->dma_device.copy_align = DMAENGINE_ALIGN_4_BYTES; So guess which alignment the second transfer has when the first has the size 0xffff. Sascha
On 11-07-18, 05:34, Robin Gong wrote: > > On 11-07-18, 00:23, Robin Gong wrote: > > > Add MEMCPY support, meanwhile, add SDMA_BD_MAX_CNT instead of > > > '0xffff'. > > > > latter part should be its own patch. Never mix things > Okay, I will split it even for this minor change. Yes, a patch should represent _one_ thing, describe that one thing and do that one thing. Anything else, doesn't matter how simple or complex should be an individual patch > > > + if (!desc) > > > + goto err_out; > > > + > > > + do { > > > + count = min_t(size_t, len, SDMA_BD_MAX_CNT); > > > + bd = &desc->bd[i]; > > > + bd->buffer_addr = dma_src; > > > + bd->ext_buffer_addr = dma_dst; > > > + bd->mode.count = count; > > > + desc->chn_count += count; > > > + > > > + switch (sdmac->word_size) { > > > + case DMA_SLAVE_BUSWIDTH_4_BYTES: > > > > This looks wrong, we are in memcpy, there is no SLAVE so no SLAVE widths.. > > > Okay, will remove check bus width. it is not about bus_width but the fact that you are using slave concepts. In memcpy we have _no_ slave, hence do not use anything related to slave including dma_slave_config
> -----Original Message----- > From: s.hauer@pengutronix.de [mailto:s.hauer@pengutronix.de] > Sent: 2018年7月11日 14:54 > To: Robin Gong <yibin.gong@nxp.com> > Cc: Vinod <vkoul@kernel.org>; dan.j.williams@intel.com; > shawnguo@kernel.org; Fabio Estevam <fabio.estevam@nxp.com>; > linux@armlinux.org.uk; linux-arm-kernel@lists.infradead.org; > kernel@pengutronix.de; dmaengine@vger.kernel.org; > linux-kernel@vger.kernel.org; dl-linux-imx <linux-imx@nxp.com> > Subject: Re: [PATCH v1 3/4] dmaengine: imx-sdma: support dmatest > > On Wed, Jul 11, 2018 at 06:37:02AM +0000, Robin Gong wrote: > > > > > -----Original Message----- > > > From: Vinod [mailto:vkoul@kernel.org] > > > Sent: 2018年7月10日 23:33 > > > To: Robin Gong <yibin.gong@nxp.com> > > > Cc: dan.j.williams@intel.com; shawnguo@kernel.org; > > > s.hauer@pengutronix.de; Fabio Estevam <fabio.estevam@nxp.com>; > > > linux@armlinux.org.uk; linux-arm-kernel@lists.infradead.org; > > > kernel@pengutronix.de; dmaengine@vger.kernel.org; > > > linux-kernel@vger.kernel.org; dl-linux-imx <linux-imx@nxp.com> > > > Subject: Re: [PATCH v1 3/4] dmaengine: imx-sdma: support dmatest > > > > > > On 11-07-18, 00:23, Robin Gong wrote: > > > > dmatest(memcpy) will never call dmaengine_slave_config before > > > > prep, > > > > > > and that should have been a hint to you that you should not expect > > > that > > > > > > > so jobs in dmaengine_slave_config need to be moved into somewhere > > > > before device_prep_dma_memcpy. Besides, dmatest never setup chan > > > > ->private as other common case like uart/audio/spi will always > > > > ->setup > > > > chan->private. Here check it to judge if it's dmatest case and do > > > > jobs in slave_config. > > > > > > and you should not do anything for dmatest. Supporting it means > > > memcpy implementation is not correct :) > > Okay, I will any word about dmatest here since memcpy assume no > > calling slave_config. > > > > > > > > > > > Signed-off-by: Robin Gong <yibin.gong@nxp.com> > > > > --- > > > > drivers/dma/imx-sdma.c | 37 ++++++++++++++++++++++++++++--------- > > > > 1 file changed, 28 insertions(+), 9 deletions(-) > > > > > > > > diff --git a/drivers/dma/imx-sdma.c b/drivers/dma/imx-sdma.c index > > > > ed2267d..48f3749 100644 > > > > --- a/drivers/dma/imx-sdma.c > > > > +++ b/drivers/dma/imx-sdma.c > > > > @@ -1222,10 +1222,36 @@ static int > > > > sdma_alloc_chan_resources(struct dma_chan *chan) { > > > > struct sdma_channel *sdmac = to_sdma_chan(chan); > > > > struct imx_dma_data *data = chan->private; > > > > + struct imx_dma_data default_data; > > > > int prio, ret; > > > > > > > > - if (!data) > > > > - return -EINVAL; > > > > + ret = clk_enable(sdmac->sdma->clk_ipg); > > > > + if (ret) > > > > + return ret; > > > > + ret = clk_enable(sdmac->sdma->clk_ahb); > > > > + if (ret) > > > > + goto disable_clk_ipg; > > > > + /* > > > > + * dmatest(memcpy) will never call dmaengine_slave_config before > prep, > > > > + * so jobs in dmaengine_slave_config need to be moved into > somewhere > > > > + * before device_prep_dma_memcpy. Besides, dmatest never setup > chan > > > > + * ->private as other common cases like uart/audio/spi will setup > > > > + * chan->private always. Here check it to judge if it's dmatest case > > > > + * and do jobs in slave_config. > > > > + */ > > > > + if (!data) { > > > > + dev_warn(sdmac->sdma->dev, "dmatest is running?\n"); > > > > > > why is that a warning! > > Current SDMA driver assume filter function to set chan->private with > > specific data (struct imx_dma_data dma_data)like below > (sound/soc/fsl/fsl_asrc_dma.c): > > static bool filter(struct dma_chan *chan, void *param) { > > if (!imx_dma_is_general_purpose(chan)) > > return false; > > chan->private = param; > > return true; > > } > > > > But in memcpy case, at lease dmatest case, no chan->private set in its filter > function. > > So here take dmatest a special case and do some prepare jobs for > > memcpy. But if the Upper device driver call dma_request_channel() with > > their specific filter without 'chan->private' setting in the future. > > The warning message is a useful hint to them to Add 'chan->private' in filter > function. Or doc it somewhere? > > Instead of doing heuristics to guess whether we are doing memcpy you could > instead make memcpy the default when slave_config is not called, i.e. drop the > if (!data) check completely. Yes, for memcpy case, that's a good way, but how to warning the future case Without setup 'chan->private'... > > > > > > > > + sdmac->word_size = > sdmac->sdma->dma_device.copy_align; > > > > + default_data.priority = 2; > > > > + default_data.peripheral_type = IMX_DMATYPE_MEMORY; > > > > + default_data.dma_request = 0; > > > > + default_data.dma_request2 = 0; > > > > + data = &default_data; > > > > + > > > > + sdma_config_ownership(sdmac, false, true, false); > > > > + sdma_get_pc(sdmac, IMX_DMATYPE_MEMORY); > > > > + sdma_load_context(sdmac); > > > > + } > > > > > > this needs to be default for memcpy > > The problem seems to be that we do not know whether we are doing memcpy > or not. Normally we get the information how a channel is to be configured in > dma_device->device_config, but this function is not called in the memcpy case. > > An alternative might also be to do the setup in > dma_device->device_prep_dma_memcpy. Yes, I've think about it before, but such prepare steps only needed in once time... > > Sascha > > -- > Pengutronix e.K. | > | > Industrial Linux Solutions | > https://emea01.safelinks.protection.outlook.com/?url=http%3A%2F%2Fwww. > pengutronix.de%2F&data=02%7C01%7Cyibin.gong%40nxp.com%7C3fcf03 > db12f441398fbb08d5e6fb142b%7C686ea1d3bc2b4c6fa92cd99c5c301635%7C0 > %7C0%7C636668888416328960&sdata=E1DT1BW4b5Q1VWgkMNZqA28 > oK%2FVVQviC8qF2%2BqG0Feo%3D&reserved=0 | > Peiner Str. 6-8, 31137 Hildesheim, Germany | Phone: +49-5121-206917-0 | > Amtsgericht Hildesheim, HRA 2686 | Fax: > +49-5121-206917-5555 |
On 11-07-18, 08:53, s.hauer@pengutronix.de wrote: > On Wed, Jul 11, 2018 at 06:37:02AM +0000, Robin Gong wrote: > > > > > -----Original Message----- > > > From: Vinod [mailto:vkoul@kernel.org] > > > Sent: 2018年7月10日 23:33 > > > To: Robin Gong <yibin.gong@nxp.com> > > > Cc: dan.j.williams@intel.com; shawnguo@kernel.org; > > > s.hauer@pengutronix.de; Fabio Estevam <fabio.estevam@nxp.com>; > > > linux@armlinux.org.uk; linux-arm-kernel@lists.infradead.org; > > > kernel@pengutronix.de; dmaengine@vger.kernel.org; > > > linux-kernel@vger.kernel.org; dl-linux-imx <linux-imx@nxp.com> > > > Subject: Re: [PATCH v1 3/4] dmaengine: imx-sdma: support dmatest > > > > > > On 11-07-18, 00:23, Robin Gong wrote: > > > > dmatest(memcpy) will never call dmaengine_slave_config before prep, > > > > > > and that should have been a hint to you that you should not expect that > > > > > > > so jobs in dmaengine_slave_config need to be moved into somewhere > > > > before device_prep_dma_memcpy. Besides, dmatest never setup chan > > > > ->private as other common case like uart/audio/spi will always setup > > > > chan->private. Here check it to judge if it's dmatest case and do > > > > jobs in slave_config. > > > > > > and you should not do anything for dmatest. Supporting it means memcpy > > > implementation is not correct :) > > Okay, I will any word about dmatest here since memcpy assume no calling > > slave_config. > > > > > > > > > > > Signed-off-by: Robin Gong <yibin.gong@nxp.com> > > > > --- > > > > drivers/dma/imx-sdma.c | 37 ++++++++++++++++++++++++++++--------- > > > > 1 file changed, 28 insertions(+), 9 deletions(-) > > > > > > > > diff --git a/drivers/dma/imx-sdma.c b/drivers/dma/imx-sdma.c index > > > > ed2267d..48f3749 100644 > > > > --- a/drivers/dma/imx-sdma.c > > > > +++ b/drivers/dma/imx-sdma.c > > > > @@ -1222,10 +1222,36 @@ static int sdma_alloc_chan_resources(struct > > > > dma_chan *chan) { > > > > struct sdma_channel *sdmac = to_sdma_chan(chan); > > > > struct imx_dma_data *data = chan->private; > > > > + struct imx_dma_data default_data; > > > > int prio, ret; > > > > > > > > - if (!data) > > > > - return -EINVAL; > > > > + ret = clk_enable(sdmac->sdma->clk_ipg); > > > > + if (ret) > > > > + return ret; > > > > + ret = clk_enable(sdmac->sdma->clk_ahb); > > > > + if (ret) > > > > + goto disable_clk_ipg; > > > > + /* > > > > + * dmatest(memcpy) will never call dmaengine_slave_config before prep, > > > > + * so jobs in dmaengine_slave_config need to be moved into somewhere > > > > + * before device_prep_dma_memcpy. Besides, dmatest never setup chan > > > > + * ->private as other common cases like uart/audio/spi will setup > > > > + * chan->private always. Here check it to judge if it's dmatest case > > > > + * and do jobs in slave_config. > > > > + */ > > > > + if (!data) { > > > > + dev_warn(sdmac->sdma->dev, "dmatest is running?\n"); > > > > > > why is that a warning! > > Current SDMA driver assume filter function to set chan->private with specific data > > (struct imx_dma_data dma_data)like below (sound/soc/fsl/fsl_asrc_dma.c): > > static bool filter(struct dma_chan *chan, void *param) > > { > > if (!imx_dma_is_general_purpose(chan)) > > return false; > > chan->private = param; > > return true; > > } > > > > But in memcpy case, at lease dmatest case, no chan->private set in its filter function. > > So here take dmatest a special case and do some prepare jobs for memcpy. But if the > > Upper device driver call dma_request_channel() with their specific filter without > > 'chan->private' setting in the future. The warning message is a useful hint to them to > > Add 'chan->private' in filter function. Or doc it somewhere? > > Instead of doing heuristics to guess whether we are doing memcpy you > could instead make memcpy the default when slave_config is not called, > i.e. drop the if (!data) check completely. > > > > > > > > + sdmac->word_size = sdmac->sdma->dma_device.copy_align; > > > > + default_data.priority = 2; > > > > + default_data.peripheral_type = IMX_DMATYPE_MEMORY; > > > > + default_data.dma_request = 0; > > > > + default_data.dma_request2 = 0; > > > > + data = &default_data; > > > > + > > > > + sdma_config_ownership(sdmac, false, true, false); > > > > + sdma_get_pc(sdmac, IMX_DMATYPE_MEMORY); > > > > + sdma_load_context(sdmac); > > > > + } > > > > > > this needs to be default for memcpy > > The problem seems to be that we do not know whether we are doing memcpy > or not. Normally we get the information how a channel is to be > configured in dma_device->device_config, but this function is not called > in the memcpy case. Not really true, device_config only provides parameters to be configured for next slave transaction > An alternative might also be to do the setup in dma_device->device_prep_dma_memcpy. Precisely, see how other drivers do this Let's roll back a bit and foresee why is this required. In case of memcpy, you need to tell DMA to do transfer from src to dstn and size. Additional parameters like buswidth etc should be derived for maximum throughput (after all we are dma, people want it to be done fastest) Now for slave, you are interfacing with a peripheral and don't know how that is setup. So you need to match the parameters, otherwise you get overflow or underflow and hence need for device_config Please do not derive additional notions from these, please do not assume anything else, unless provided in documentation :) In doubt, just ask! HTH
> -----Original Message----- > From: Vinod [mailto:vkoul@kernel.org] > Sent: 2018年7月11日 15:19 > To: s.hauer@pengutronix.de > Cc: Robin Gong <yibin.gong@nxp.com>; dan.j.williams@intel.com; > shawnguo@kernel.org; Fabio Estevam <fabio.estevam@nxp.com>; > linux@armlinux.org.uk; linux-arm-kernel@lists.infradead.org; > kernel@pengutronix.de; dmaengine@vger.kernel.org; > linux-kernel@vger.kernel.org; dl-linux-imx <linux-imx@nxp.com> > Subject: Re: [PATCH v1 3/4] dmaengine: imx-sdma: support dmatest > > On 11-07-18, 08:53, s.hauer@pengutronix.de wrote: > > On Wed, Jul 11, 2018 at 06:37:02AM +0000, Robin Gong wrote: > > > > > > > -----Original Message----- > > > > From: Vinod [mailto:vkoul@kernel.org] > > > > Sent: 2018年7月10日 23:33 > > > > To: Robin Gong <yibin.gong@nxp.com> > > > > Cc: dan.j.williams@intel.com; shawnguo@kernel.org; > > > > s.hauer@pengutronix.de; Fabio Estevam <fabio.estevam@nxp.com>; > > > > linux@armlinux.org.uk; linux-arm-kernel@lists.infradead.org; > > > > kernel@pengutronix.de; dmaengine@vger.kernel.org; > > > > linux-kernel@vger.kernel.org; dl-linux-imx <linux-imx@nxp.com> > > > > Subject: Re: [PATCH v1 3/4] dmaengine: imx-sdma: support dmatest > > > > > > > > On 11-07-18, 00:23, Robin Gong wrote: > > > > > dmatest(memcpy) will never call dmaengine_slave_config before > > > > > prep, > > > > > > > > and that should have been a hint to you that you should not expect > > > > that > > > > > > > > > so jobs in dmaengine_slave_config need to be moved into > > > > > somewhere before device_prep_dma_memcpy. Besides, dmatest never > > > > > setup chan > > > > > ->private as other common case like uart/audio/spi will always > > > > > ->setup > > > > > chan->private. Here check it to judge if it's dmatest case and > > > > > chan->do > > > > > jobs in slave_config. > > > > > > > > and you should not do anything for dmatest. Supporting it means > > > > memcpy implementation is not correct :) > > > Okay, I will any word about dmatest here since memcpy assume no > > > calling slave_config. > > > > > > > > > > > > > > Signed-off-by: Robin Gong <yibin.gong@nxp.com> > > > > > --- > > > > > drivers/dma/imx-sdma.c | 37 > > > > > ++++++++++++++++++++++++++++--------- > > > > > 1 file changed, 28 insertions(+), 9 deletions(-) > > > > > > > > > > diff --git a/drivers/dma/imx-sdma.c b/drivers/dma/imx-sdma.c > > > > > index > > > > > ed2267d..48f3749 100644 > > > > > --- a/drivers/dma/imx-sdma.c > > > > > +++ b/drivers/dma/imx-sdma.c > > > > > @@ -1222,10 +1222,36 @@ static int > > > > > sdma_alloc_chan_resources(struct dma_chan *chan) { > > > > > struct sdma_channel *sdmac = to_sdma_chan(chan); > > > > > struct imx_dma_data *data = chan->private; > > > > > + struct imx_dma_data default_data; > > > > > int prio, ret; > > > > > > > > > > - if (!data) > > > > > - return -EINVAL; > > > > > + ret = clk_enable(sdmac->sdma->clk_ipg); > > > > > + if (ret) > > > > > + return ret; > > > > > + ret = clk_enable(sdmac->sdma->clk_ahb); > > > > > + if (ret) > > > > > + goto disable_clk_ipg; > > > > > + /* > > > > > + * dmatest(memcpy) will never call dmaengine_slave_config before > prep, > > > > > + * so jobs in dmaengine_slave_config need to be moved into > somewhere > > > > > + * before device_prep_dma_memcpy. Besides, dmatest never setup > chan > > > > > + * ->private as other common cases like uart/audio/spi will setup > > > > > + * chan->private always. Here check it to judge if it's dmatest case > > > > > + * and do jobs in slave_config. > > > > > + */ > > > > > + if (!data) { > > > > > + dev_warn(sdmac->sdma->dev, "dmatest is running?\n"); > > > > > > > > why is that a warning! > > > Current SDMA driver assume filter function to set chan->private with > > > specific data (struct imx_dma_data dma_data)like below > (sound/soc/fsl/fsl_asrc_dma.c): > > > static bool filter(struct dma_chan *chan, void *param) { > > > if (!imx_dma_is_general_purpose(chan)) > > > return false; > > > chan->private = param; > > > return true; > > > } > > > > > > But in memcpy case, at lease dmatest case, no chan->private set in its filter > function. > > > So here take dmatest a special case and do some prepare jobs for > > > memcpy. But if the Upper device driver call dma_request_channel() > > > with their specific filter without 'chan->private' setting in the > > > future. The warning message is a useful hint to them to Add 'chan->private' > in filter function. Or doc it somewhere? > > > > Instead of doing heuristics to guess whether we are doing memcpy you > > could instead make memcpy the default when slave_config is not called, > > i.e. drop the if (!data) check completely. > > > > > > > > > > > + sdmac->word_size = > sdmac->sdma->dma_device.copy_align; > > > > > + default_data.priority = 2; > > > > > + default_data.peripheral_type = IMX_DMATYPE_MEMORY; > > > > > + default_data.dma_request = 0; > > > > > + default_data.dma_request2 = 0; > > > > > + data = &default_data; > > > > > + > > > > > + sdma_config_ownership(sdmac, false, true, false); > > > > > + sdma_get_pc(sdmac, IMX_DMATYPE_MEMORY); > > > > > + sdma_load_context(sdmac); > > > > > + } > > > > > > > > this needs to be default for memcpy > > > > The problem seems to be that we do not know whether we are doing > > memcpy or not. Normally we get the information how a channel is to be > > configured in dma_device->device_config, but this function is not > > called in the memcpy case. > > Not really true, device_config only provides parameters to be configured for > next slave transaction > > > An alternative might also be to do the setup in > dma_device->device_prep_dma_memcpy. > > Precisely, see how other drivers do this > > Let's roll back a bit and foresee why is this required. > > In case of memcpy, you need to tell DMA to do transfer from src to dstn and > size. Additional parameters like buswidth etc should be derived for maximum > throughput (after all we are dma, people want it to be done > fastest) > > Now for slave, you are interfacing with a peripheral and don't know how that is > setup. So you need to match the parameters, otherwise you get overflow or > underflow and hence need for device_config > > Please do not derive additional notions from these, please do not assume > anything else, unless provided in documentation :) I will move such prepare jobs from slave_config to device_prep_dma_memcpy Instead of device_alloc_chan_resources as I did in v1, thus we have no 'chan->private' issue, just like drivers/dma/stm32-mdma.c. The only limitation is those prepare jobs (some register setting) will be done every time memcpy instead of only one time in slave_config or v1 case. Is that ok? > > In doubt, just ask! > > HTH > -- > ~Vinod
On 11-07-18, 08:16, Robin Gong wrote: > > > The problem seems to be that we do not know whether we are doing > > > memcpy or not. Normally we get the information how a channel is to be > > > configured in dma_device->device_config, but this function is not > > > called in the memcpy case. > > > > Not really true, device_config only provides parameters to be configured for > > next slave transaction > > > > > An alternative might also be to do the setup in > > dma_device->device_prep_dma_memcpy. > > > > Precisely, see how other drivers do this > > > > Let's roll back a bit and foresee why is this required. > > > > In case of memcpy, you need to tell DMA to do transfer from src to dstn and > > size. Additional parameters like buswidth etc should be derived for maximum > > throughput (after all we are dma, people want it to be done > > fastest) > > > > Now for slave, you are interfacing with a peripheral and don't know how that is > > setup. So you need to match the parameters, otherwise you get overflow or > > underflow and hence need for device_config > > > > Please do not derive additional notions from these, please do not assume > > anything else, unless provided in documentation :) > I will move such prepare jobs from slave_config to device_prep_dma_memcpy > Instead of device_alloc_chan_resources as I did in v1, thus we have no 'chan->private' > issue, just like drivers/dma/stm32-mdma.c. The only limitation is those prepare jobs > (some register setting) will be done every time memcpy instead of only one time in slave_config > or v1 case. Is that ok? sounds fine to me
This patchset is to add memcpy interface for imx-sdma, besides,to support dmatest and enable config related, so that easily test dma without any other device support such as uart/audio/spi... Robin Gong (4): dmaengine: imx-sdma: add memcpy interface dmaengine: imx-sdma: add check_bd_buswidth() to kill the dulicated code dmaengine: imx-sdma: support dmatest ARM: configs: imx_v6_v7_defconfig: add DMATEST support arch/arm/configs/imx_v6_v7_defconfig | 3 +- drivers/dma/imx-sdma.c | 172 ++++++++++++++++++++++++++++------- 2 files changed, 139 insertions(+), 36 deletions(-)