diff mbox series

[1/7] mmc: sdhci: add support for using external DMA devices

Message ID 20190111110851.6805-2-faiz_abbas@ti.com (mailing list archive)
State New, archived
Headers show
Series Port am335 and am437 devices to sdhci-omap | expand

Commit Message

Faiz Abbas Jan. 11, 2019, 11:08 a.m. UTC
From: Chunyan Zhang <zhang.chunyan@linaro.org>

Some standard SD host controllers can support both external dma
controllers as well as ADMA/SDMA in which the SD host controller
acts as DMA master. TI's omap controller is the case as an example.

Currently the generic SDHCI code supports ADMA/SDMA integrated in
the host controller but does not have any support for external DMA
controllers implemented using dmaengine, meaning that custom code is
needed for any systems that use an external DMA controller with SDHCI.

Fixes by Faiz Abbas <faiz_abbas@ti.com>:
1. Map scatterlists before dmaengine_prep_slave_sg()
2. Use dma_async() functions inside of the send_command() path and
synchronize once at the start of each request.

Signed-off-by: Chunyan Zhang <zhang.chunyan@linaro.org>
Signed-off-by: Faiz Abbas <faiz_abbas@ti.com>
---
 drivers/mmc/host/Kconfig |   3 +
 drivers/mmc/host/sdhci.c | 266 ++++++++++++++++++++++++++++++++++++++-
 drivers/mmc/host/sdhci.h |   8 ++
 3 files changed, 273 insertions(+), 4 deletions(-)

Comments

Adrian Hunter Jan. 14, 2019, 8:45 a.m. UTC | #1
On 11/01/19 1:08 PM, Faiz Abbas wrote:
> From: Chunyan Zhang <zhang.chunyan@linaro.org>
> 
> Some standard SD host controllers can support both external dma
> controllers as well as ADMA/SDMA in which the SD host controller
> acts as DMA master. TI's omap controller is the case as an example.
> 
> Currently the generic SDHCI code supports ADMA/SDMA integrated in
> the host controller but does not have any support for external DMA
> controllers implemented using dmaengine, meaning that custom code is
> needed for any systems that use an external DMA controller with SDHCI.
> 
> Fixes by Faiz Abbas <faiz_abbas@ti.com>:
> 1. Map scatterlists before dmaengine_prep_slave_sg()
> 2. Use dma_async() functions inside of the send_command() path and
> synchronize once at the start of each request.
> 
> Signed-off-by: Chunyan Zhang <zhang.chunyan@linaro.org>
> Signed-off-by: Faiz Abbas <faiz_abbas@ti.com>

Acked-by: Adrian Hunter <adrian.hunter@intel.com>

> ---
>  drivers/mmc/host/Kconfig |   3 +
>  drivers/mmc/host/sdhci.c | 266 ++++++++++++++++++++++++++++++++++++++-
>  drivers/mmc/host/sdhci.h |   8 ++
>  3 files changed, 273 insertions(+), 4 deletions(-)
> 
> diff --git a/drivers/mmc/host/Kconfig b/drivers/mmc/host/Kconfig
> index e26b8145efb3..333292e8ecdd 100644
> --- a/drivers/mmc/host/Kconfig
> +++ b/drivers/mmc/host/Kconfig
> @@ -999,3 +999,6 @@ config MMC_SDHCI_AM654
>  	  If you have a controller with this interface, say Y or M here.
>  
>  	  If unsure, say N.
> +
> +config MMC_SDHCI_EXTERNAL_DMA
> +        bool
> diff --git a/drivers/mmc/host/sdhci.c b/drivers/mmc/host/sdhci.c
> index a22e11a65658..4a9044c06e21 100644
> --- a/drivers/mmc/host/sdhci.c
> +++ b/drivers/mmc/host/sdhci.c
> @@ -14,6 +14,7 @@
>   */
>  
>  #include <linux/delay.h>
> +#include <linux/dmaengine.h>
>  #include <linux/ktime.h>
>  #include <linux/highmem.h>
>  #include <linux/io.h>
> @@ -1118,6 +1119,226 @@ static void sdhci_prepare_data(struct sdhci_host *host, struct mmc_command *cmd)
>  	}
>  }
>  
> +#if IS_ENABLED(CONFIG_MMC_SDHCI_EXTERNAL_DMA)
> +static int sdhci_external_dma_init(struct sdhci_host *host)
> +{
> +	int ret = 0;
> +	struct mmc_host *mmc = host->mmc;
> +
> +	host->tx_chan = dma_request_chan(mmc->parent, "tx");
> +	if (IS_ERR(host->tx_chan)) {
> +		ret = PTR_ERR(host->tx_chan);
> +		if (ret != -EPROBE_DEFER)
> +			pr_warn("Failed to request TX DMA channel.\n");
> +		host->tx_chan = NULL;
> +		return ret;
> +	}
> +
> +	host->rx_chan = dma_request_chan(mmc->parent, "rx");
> +	if (IS_ERR(host->rx_chan)) {
> +		if (host->tx_chan) {
> +			dma_release_channel(host->tx_chan);
> +			host->tx_chan = NULL;
> +		}
> +
> +		ret = PTR_ERR(host->rx_chan);
> +		if (ret != -EPROBE_DEFER)
> +			pr_warn("Failed to request RX DMA channel.\n");
> +		host->rx_chan = NULL;
> +	}
> +
> +	return ret;
> +}
> +
> +static inline struct dma_chan *
> +sdhci_external_dma_channel(struct sdhci_host *host, struct mmc_data *data)
> +{
> +	return data->flags & MMC_DATA_WRITE ? host->tx_chan : host->rx_chan;
> +}
> +
> +static int sdhci_external_dma_setup(struct sdhci_host *host,
> +				    struct mmc_command *cmd)
> +{
> +	int ret, i;
> +	struct dma_async_tx_descriptor *desc;
> +	struct mmc_data *data = cmd->data;
> +	struct dma_chan *chan;
> +	struct dma_slave_config cfg;
> +	dma_cookie_t cookie;
> +	int sg_cnt;
> +
> +	if (!host->mapbase)
> +		return -EINVAL;
> +
> +	cfg.src_addr = host->mapbase + SDHCI_BUFFER;
> +	cfg.dst_addr = host->mapbase + SDHCI_BUFFER;
> +	cfg.src_addr_width = DMA_SLAVE_BUSWIDTH_4_BYTES;
> +	cfg.dst_addr_width = DMA_SLAVE_BUSWIDTH_4_BYTES;
> +	cfg.src_maxburst = data->blksz / 4;
> +	cfg.dst_maxburst = data->blksz / 4;
> +
> +	/* Sanity check: all the SG entries must be aligned by block size. */
> +	for (i = 0; i < data->sg_len; i++) {
> +		if ((data->sg + i)->length % data->blksz)
> +			return -EINVAL;
> +	}
> +
> +	chan = sdhci_external_dma_channel(host, data);
> +
> +	ret = dmaengine_slave_config(chan, &cfg);
> +	if (ret)
> +		return ret;
> +
> +	sg_cnt = sdhci_pre_dma_transfer(host, data, COOKIE_MAPPED);
> +	if (sg_cnt <= 0)
> +		return -EINVAL;
> +
> +	desc = dmaengine_prep_slave_sg(chan, data->sg, data->sg_len,
> +				       mmc_get_dma_dir(data),
> +				       DMA_PREP_INTERRUPT | DMA_CTRL_ACK);
> +	if (!desc)
> +		return -EINVAL;
> +
> +	desc->callback = NULL;
> +	desc->callback_param = NULL;
> +
> +	cookie = dmaengine_submit(desc);
> +	if (cookie < 0)
> +		ret = cookie;
> +
> +	return ret;
> +}
> +
> +static void sdhci_external_dma_release(struct sdhci_host *host)
> +{
> +	if (host->tx_chan) {
> +		dma_release_channel(host->tx_chan);
> +		host->tx_chan = NULL;
> +	}
> +
> +	if (host->rx_chan) {
> +		dma_release_channel(host->rx_chan);
> +		host->rx_chan = NULL;
> +	}
> +
> +	sdhci_switch_external_dma(host, false);
> +}
> +
> +static int __sdhci_external_dma_prepare_data(struct sdhci_host *host,
> +					     struct mmc_command *cmd)
> +{
> +	struct mmc_data *data = cmd->data;
> +
> +	host->data_timeout = 0;
> +
> +	if (sdhci_data_line_cmd(cmd))
> +		sdhci_set_timeout(host, cmd);
> +
> +	WARN_ON(host->data);
> +
> +	/* Sanity checks */
> +	WARN_ON(data->blksz * data->blocks > 524288);
> +	WARN_ON(data->blksz > host->mmc->max_blk_size);
> +	WARN_ON(data->blocks > 65535);
> +
> +	host->flags |= SDHCI_REQ_USE_DMA;
> +	host->data = data;
> +	host->data_early = 0;
> +	host->data->bytes_xfered = 0;
> +
> +	sdhci_set_transfer_irqs(host);
> +
> +	/*
> +	 * For Version 4.10 onwards, if v4 mode is enabled, 32-bit Block Count
> +	 * can be supported, in that case 16-bit block count register must be 0.
> +	 */
> +	if (host->version >= SDHCI_SPEC_410 && host->v4_mode &&
> +	    (host->quirks2 & SDHCI_QUIRK2_USE_32BIT_BLK_CNT)) {
> +		if (sdhci_readw(host, SDHCI_BLOCK_COUNT))
> +			sdhci_writew(host, 0, SDHCI_BLOCK_COUNT);
> +		sdhci_writew(host, data->blocks, SDHCI_32BIT_BLK_CNT);
> +	} else {
> +		sdhci_writew(host, data->blocks, SDHCI_BLOCK_COUNT);
> +	}
> +
> +	return 0;
> +}
> +
> +static void sdhci_external_dma_prepare_data(struct sdhci_host *host,
> +					    struct mmc_command *cmd)
> +{
> +	struct mmc_data *data = cmd->data;
> +
> +	if (!data)
> +		return;
> +
> +	if (sdhci_external_dma_setup(host, cmd) ||
> +	    __sdhci_external_dma_prepare_data(host, cmd)) {
> +		sdhci_external_dma_release(host);
> +		pr_err("%s: Cannot use external DMA, switch to the DMA/PIO which standard SDHCI provides.\n",
> +		       mmc_hostname(host->mmc));
> +		sdhci_prepare_data(host, cmd);
> +	}
> +}
> +
> +static void sdhci_external_dma_pre_transfer(struct sdhci_host *host,
> +					    struct mmc_command *cmd)
> +{
> +	struct dma_chan *chan;
> +
> +	if (!cmd->data || cmd->opcode == MMC_SET_BLOCK_COUNT)
> +		return;
> +
> +	sdhci_writew(host, cmd->data->blksz, SDHCI_BLOCK_SIZE);
> +	chan = sdhci_external_dma_channel(host, cmd->data);
> +	if (chan)
> +		dma_async_issue_pending(chan);
> +}
> +
> +static int sdhci_external_dma_cleanup(struct sdhci_host *host,
> +				    struct mmc_data *data)
> +{
> +	struct dma_chan *chan = sdhci_external_dma_channel(host, data);
> +	int ret = 0;
> +
> +	if (chan)
> +		ret = dmaengine_terminate_async(chan);
> +
> +	return ret;
> +}
> +#else
> +static int sdhci_external_dma_init(struct sdhci_host *host)
> +{
> +	return -EOPNOTSUPP;
> +}
> +
> +static void sdhci_external_dma_release(struct sdhci_host *host)
> +{}
> +
> +static void sdhci_external_dma_prepare_data(struct sdhci_host *host,
> +					    struct mmc_command *cmd)
> +{
> +	/* If MMC_SDHCI_EXTERNAL_DMA not supported, PIO will be used */
> +	sdhci_prepare_data(host, cmd);
> +}
> +
> +static void sdhci_external_dma_pre_transfer(struct sdhci_host *host,
> +					    struct mmc_command *cmd)
> +{}
> +
> +static int sdhci_external_dma_cleanup(struct sdhci_host *host,
> +				    struct mmc_data *data)
> +{
> +	return 0;
> +}
> +#endif
> +
> +void sdhci_switch_external_dma(struct sdhci_host *host, bool en)
> +{
> +	host->use_external_dma = en;
> +}
> +EXPORT_SYMBOL_GPL(sdhci_switch_external_dma);
> +
>  static inline bool sdhci_auto_cmd12(struct sdhci_host *host,
>  				    struct mmc_request *mrq)
>  {
> @@ -1374,7 +1595,10 @@ void sdhci_send_command(struct sdhci_host *host, struct mmc_command *cmd)
>  		host->data_cmd = cmd;
>  	}
>  
> -	sdhci_prepare_data(host, cmd);
> +	if (host->use_external_dma)
> +		sdhci_external_dma_prepare_data(host, cmd);
> +	else
> +		sdhci_prepare_data(host, cmd);
>  
>  	sdhci_writel(host, cmd->arg, SDHCI_ARGUMENT);
>  
> @@ -1416,6 +1640,9 @@ void sdhci_send_command(struct sdhci_host *host, struct mmc_command *cmd)
>  		timeout += 10 * HZ;
>  	sdhci_mod_timer(host, cmd->mrq, timeout);
>  
> +	if (host->use_external_dma)
> +		sdhci_external_dma_pre_transfer(host, cmd);
> +
>  	sdhci_writew(host, SDHCI_MAKE_CMD(cmd->opcode, flags), SDHCI_COMMAND);
>  }
>  EXPORT_SYMBOL_GPL(sdhci_send_command);
> @@ -1781,6 +2008,11 @@ void sdhci_request(struct mmc_host *mmc, struct mmc_request *mrq)
>  
>  	sdhci_led_activate(host);
>  
> +	if (host->use_external_dma && mrq->data) {
> +		struct dma_chan *chan = sdhci_external_dma_channel(host,
> +								   mrq->data);
> +		dmaengine_synchronize(chan);
> +	}
>  	/*
>  	 * Ensure we don't send the STOP for non-SET_BLOCK_COUNTED
>  	 * requests if Auto-CMD12 is enabled.
> @@ -2658,6 +2890,8 @@ static bool sdhci_request_done(struct sdhci_host *host)
>  				dma_unmap_sg(mmc_dev(host->mmc), data->sg,
>  					     data->sg_len,
>  					     mmc_get_dma_dir(data));
> +				if (host->use_external_dma)
> +					sdhci_external_dma_cleanup(host, data);
>  			}
>  			data->host_cookie = COOKIE_UNMAPPED;
>  		}
> @@ -3692,12 +3926,15 @@ int sdhci_setup_host(struct sdhci_host *host)
>  		       mmc_hostname(mmc), host->version);
>  	}
>  
> -	if (host->quirks & SDHCI_QUIRK_FORCE_DMA)
> +	if (host->quirks & SDHCI_QUIRK_FORCE_DMA) {
>  		host->flags |= SDHCI_USE_SDMA;
> -	else if (!(host->caps & SDHCI_CAN_DO_SDMA))
> +	} else if (!(host->caps & SDHCI_CAN_DO_SDMA)) {
>  		DBG("Controller doesn't have SDMA capability\n");
> -	else
> +	} else if (host->use_external_dma) {
> +		/* Using dma-names to detect external dma capability */
> +	} else {
>  		host->flags |= SDHCI_USE_SDMA;
> +	}
>  
>  	if ((host->quirks & SDHCI_QUIRK_BROKEN_DMA) &&
>  		(host->flags & SDHCI_USE_SDMA)) {
> @@ -3785,6 +4022,19 @@ int sdhci_setup_host(struct sdhci_host *host)
>  		}
>  	}
>  
> +	if (host->use_external_dma) {
> +		ret = sdhci_external_dma_init(host);
> +		if (ret == -EPROBE_DEFER)
> +			goto unreg;
> +
> +		/*
> +		 * Fall back to use the DMA/PIO integrated in standard SDHCI
> +		 * instead of external DMA devices.
> +		 */
> +		if (ret)
> +			sdhci_switch_external_dma(host, false);
> +	}
> +
>  	/*
>  	 * If we use DMA, then it's up to the caller to set the DMA
>  	 * mask, but PIO does not need the hw shim so we set a new
> @@ -4201,6 +4451,10 @@ void sdhci_cleanup_host(struct sdhci_host *host)
>  		dma_free_coherent(mmc_dev(mmc), host->align_buffer_sz +
>  				  host->adma_table_sz, host->align_buffer,
>  				  host->align_addr);
> +
> +	if (host->use_external_dma)
> +		sdhci_external_dma_release(host);
> +
>  	host->adma_table = NULL;
>  	host->align_buffer = NULL;
>  }
> @@ -4247,6 +4501,7 @@ int __sdhci_add_host(struct sdhci_host *host)
>  
>  	pr_info("%s: SDHCI controller on %s [%s] using %s\n",
>  		mmc_hostname(mmc), host->hw_name, dev_name(mmc_dev(mmc)),
> +		host->use_external_dma ? "External DMA" :
>  		(host->flags & SDHCI_USE_ADMA) ?
>  		(host->flags & SDHCI_USE_64_BIT_DMA) ? "ADMA 64-bit" : "ADMA" :
>  		(host->flags & SDHCI_USE_SDMA) ? "DMA" : "PIO");
> @@ -4335,6 +4590,9 @@ void sdhci_remove_host(struct sdhci_host *host, int dead)
>  				  host->adma_table_sz, host->align_buffer,
>  				  host->align_addr);
>  
> +	if (host->use_external_dma)
> +		sdhci_external_dma_release(host);
> +
>  	host->adma_table = NULL;
>  	host->align_buffer = NULL;
>  }
> diff --git a/drivers/mmc/host/sdhci.h b/drivers/mmc/host/sdhci.h
> index 6cc9a3c2ac66..7a52823ebef4 100644
> --- a/drivers/mmc/host/sdhci.h
> +++ b/drivers/mmc/host/sdhci.h
> @@ -482,6 +482,7 @@ struct sdhci_host {
>  
>  	int irq;		/* Device IRQ */
>  	void __iomem *ioaddr;	/* Mapped address */
> +	phys_addr_t mapbase;	/* physical address base */
>  	char *bounce_buffer;	/* For packing SDMA reads/writes */
>  	dma_addr_t bounce_addr;
>  	unsigned int bounce_buffer_size;
> @@ -531,6 +532,7 @@ struct sdhci_host {
>  	bool pending_reset;	/* Cmd/data reset is pending */
>  	bool irq_wake_enabled;	/* IRQ wakeup is enabled */
>  	bool v4_mode;		/* Host Version 4 Enable */
> +	bool use_external_dma; /* Host selects to use external DMA */
>  
>  	struct mmc_request *mrqs_done[SDHCI_MAX_MRQS];	/* Requests done */
>  	struct mmc_command *cmd;	/* Current command */
> @@ -559,6 +561,11 @@ struct sdhci_host {
>  	struct timer_list timer;	/* Timer for timeouts */
>  	struct timer_list data_timer;	/* Timer for data timeouts */
>  
> +#if IS_ENABLED(CONFIG_MMC_SDHCI_EXTERNAL_DMA)
> +	struct dma_chan	*rx_chan;
> +	struct dma_chan	*tx_chan;
> +#endif
> +
>  	u32 caps;		/* CAPABILITY_0 */
>  	u32 caps1;		/* CAPABILITY_1 */
>  	bool read_caps;		/* Capability flags have been read */
> @@ -792,5 +799,6 @@ void sdhci_start_tuning(struct sdhci_host *host);
>  void sdhci_end_tuning(struct sdhci_host *host);
>  void sdhci_reset_tuning(struct sdhci_host *host);
>  void sdhci_send_tuning(struct sdhci_host *host, u32 opcode);
> +void sdhci_switch_external_dma(struct sdhci_host *host, bool en);
>  
>  #endif /* __SDHCI_HW_H */
>
Adrian Hunter Jan. 24, 2019, 11:40 a.m. UTC | #2
On 11/01/19 1:08 PM, Faiz Abbas wrote:
> From: Chunyan Zhang <zhang.chunyan@linaro.org>
> 
> Some standard SD host controllers can support both external dma
> controllers as well as ADMA/SDMA in which the SD host controller
> acts as DMA master. TI's omap controller is the case as an example.
> 
> Currently the generic SDHCI code supports ADMA/SDMA integrated in
> the host controller but does not have any support for external DMA
> controllers implemented using dmaengine, meaning that custom code is
> needed for any systems that use an external DMA controller with SDHCI.
> 
> Fixes by Faiz Abbas <faiz_abbas@ti.com>:
> 1. Map scatterlists before dmaengine_prep_slave_sg()
> 2. Use dma_async() functions inside of the send_command() path and
> synchronize once at the start of each request.

Sorry for the slow reply, but I do have some concerns.  Please see the comments.

> 
> Signed-off-by: Chunyan Zhang <zhang.chunyan@linaro.org>
> Signed-off-by: Faiz Abbas <faiz_abbas@ti.com>
> ---
>  drivers/mmc/host/Kconfig |   3 +
>  drivers/mmc/host/sdhci.c | 266 ++++++++++++++++++++++++++++++++++++++-
>  drivers/mmc/host/sdhci.h |   8 ++
>  3 files changed, 273 insertions(+), 4 deletions(-)
> 
> diff --git a/drivers/mmc/host/Kconfig b/drivers/mmc/host/Kconfig
> index e26b8145efb3..333292e8ecdd 100644
> --- a/drivers/mmc/host/Kconfig
> +++ b/drivers/mmc/host/Kconfig
> @@ -999,3 +999,6 @@ config MMC_SDHCI_AM654
>  	  If you have a controller with this interface, say Y or M here.
>  
>  	  If unsure, say N.
> +
> +config MMC_SDHCI_EXTERNAL_DMA
> +        bool
> diff --git a/drivers/mmc/host/sdhci.c b/drivers/mmc/host/sdhci.c
> index a22e11a65658..4a9044c06e21 100644
> --- a/drivers/mmc/host/sdhci.c
> +++ b/drivers/mmc/host/sdhci.c
> @@ -14,6 +14,7 @@
>   */
>  
>  #include <linux/delay.h>
> +#include <linux/dmaengine.h>
>  #include <linux/ktime.h>
>  #include <linux/highmem.h>
>  #include <linux/io.h>
> @@ -1118,6 +1119,226 @@ static void sdhci_prepare_data(struct sdhci_host *host, struct mmc_command *cmd)
>  	}
>  }
>  
> +#if IS_ENABLED(CONFIG_MMC_SDHCI_EXTERNAL_DMA)
> +static int sdhci_external_dma_init(struct sdhci_host *host)
> +{
> +	int ret = 0;
> +	struct mmc_host *mmc = host->mmc;
> +
> +	host->tx_chan = dma_request_chan(mmc->parent, "tx");
> +	if (IS_ERR(host->tx_chan)) {
> +		ret = PTR_ERR(host->tx_chan);
> +		if (ret != -EPROBE_DEFER)
> +			pr_warn("Failed to request TX DMA channel.\n");
> +		host->tx_chan = NULL;
> +		return ret;
> +	}
> +
> +	host->rx_chan = dma_request_chan(mmc->parent, "rx");
> +	if (IS_ERR(host->rx_chan)) {
> +		if (host->tx_chan) {
> +			dma_release_channel(host->tx_chan);
> +			host->tx_chan = NULL;
> +		}
> +
> +		ret = PTR_ERR(host->rx_chan);
> +		if (ret != -EPROBE_DEFER)
> +			pr_warn("Failed to request RX DMA channel.\n");
> +		host->rx_chan = NULL;
> +	}
> +
> +	return ret;
> +}
> +
> +static inline struct dma_chan *
> +sdhci_external_dma_channel(struct sdhci_host *host, struct mmc_data *data)
> +{
> +	return data->flags & MMC_DATA_WRITE ? host->tx_chan : host->rx_chan;
> +}
> +
> +static int sdhci_external_dma_setup(struct sdhci_host *host,
> +				    struct mmc_command *cmd)
> +{
> +	int ret, i;
> +	struct dma_async_tx_descriptor *desc;
> +	struct mmc_data *data = cmd->data;
> +	struct dma_chan *chan;
> +	struct dma_slave_config cfg;
> +	dma_cookie_t cookie;
> +	int sg_cnt;
> +
> +	if (!host->mapbase)
> +		return -EINVAL;
> +
> +	cfg.src_addr = host->mapbase + SDHCI_BUFFER;
> +	cfg.dst_addr = host->mapbase + SDHCI_BUFFER;
> +	cfg.src_addr_width = DMA_SLAVE_BUSWIDTH_4_BYTES;
> +	cfg.dst_addr_width = DMA_SLAVE_BUSWIDTH_4_BYTES;
> +	cfg.src_maxburst = data->blksz / 4;
> +	cfg.dst_maxburst = data->blksz / 4;
> +
> +	/* Sanity check: all the SG entries must be aligned by block size. */
> +	for (i = 0; i < data->sg_len; i++) {
> +		if ((data->sg + i)->length % data->blksz)
> +			return -EINVAL;
> +	}
> +
> +	chan = sdhci_external_dma_channel(host, data);
> +
> +	ret = dmaengine_slave_config(chan, &cfg);
> +	if (ret)
> +		return ret;
> +
> +	sg_cnt = sdhci_pre_dma_transfer(host, data, COOKIE_MAPPED);
> +	if (sg_cnt <= 0)
> +		return -EINVAL;
> +
> +	desc = dmaengine_prep_slave_sg(chan, data->sg, data->sg_len,
> +				       mmc_get_dma_dir(data),
> +				       DMA_PREP_INTERRUPT | DMA_CTRL_ACK);
> +	if (!desc)
> +		return -EINVAL;
> +
> +	desc->callback = NULL;
> +	desc->callback_param = NULL;
> +
> +	cookie = dmaengine_submit(desc);
> +	if (cookie < 0)
> +		ret = cookie;
> +
> +	return ret;
> +}
> +
> +static void sdhci_external_dma_release(struct sdhci_host *host)
> +{
> +	if (host->tx_chan) {
> +		dma_release_channel(host->tx_chan);
> +		host->tx_chan = NULL;
> +	}
> +
> +	if (host->rx_chan) {
> +		dma_release_channel(host->rx_chan);
> +		host->rx_chan = NULL;
> +	}
> +
> +	sdhci_switch_external_dma(host, false);
> +}
> +
> +static int __sdhci_external_dma_prepare_data(struct sdhci_host *host,
> +					     struct mmc_command *cmd)
> +{
> +	struct mmc_data *data = cmd->data;
> +
> +	host->data_timeout = 0;
> +
> +	if (sdhci_data_line_cmd(cmd))
> +		sdhci_set_timeout(host, cmd);
> +
> +	WARN_ON(host->data);
> +
> +	/* Sanity checks */
> +	WARN_ON(data->blksz * data->blocks > 524288);
> +	WARN_ON(data->blksz > host->mmc->max_blk_size);
> +	WARN_ON(data->blocks > 65535);
> +
> +	host->flags |= SDHCI_REQ_USE_DMA;
> +	host->data = data;
> +	host->data_early = 0;
> +	host->data->bytes_xfered = 0;
> +
> +	sdhci_set_transfer_irqs(host);
> +
> +	/*
> +	 * For Version 4.10 onwards, if v4 mode is enabled, 32-bit Block Count
> +	 * can be supported, in that case 16-bit block count register must be 0.
> +	 */
> +	if (host->version >= SDHCI_SPEC_410 && host->v4_mode &&
> +	    (host->quirks2 & SDHCI_QUIRK2_USE_32BIT_BLK_CNT)) {
> +		if (sdhci_readw(host, SDHCI_BLOCK_COUNT))
> +			sdhci_writew(host, 0, SDHCI_BLOCK_COUNT);
> +		sdhci_writew(host, data->blocks, SDHCI_32BIT_BLK_CNT);
> +	} else {
> +		sdhci_writew(host, data->blocks, SDHCI_BLOCK_COUNT);
> +	}

It is probably worth factoring out the code that is shared with
sdhci_prepare_data() where possible.

> +
> +	return 0;
> +}
> +
> +static void sdhci_external_dma_prepare_data(struct sdhci_host *host,
> +					    struct mmc_command *cmd)
> +{
> +	struct mmc_data *data = cmd->data;
> +
> +	if (!data)
> +		return;

Even in the !data case, we still need to set up a timeout for commands with
busy waiting.  I suggest checking the !data case before calling
sdhci_external_dma_prepare_data()

> +
> +	if (sdhci_external_dma_setup(host, cmd) ||
> +	    __sdhci_external_dma_prepare_data(host, cmd)) {
> +		sdhci_external_dma_release(host);
> +		pr_err("%s: Cannot use external DMA, switch to the DMA/PIO which standard SDHCI provides.\n",
> +		       mmc_hostname(host->mmc));
> +		sdhci_prepare_data(host, cmd);
> +	}
> +}
> +
> +static void sdhci_external_dma_pre_transfer(struct sdhci_host *host,
> +					    struct mmc_command *cmd)
> +{
> +	struct dma_chan *chan;
> +
> +	if (!cmd->data || cmd->opcode == MMC_SET_BLOCK_COUNT)

MMC_SET_BLOCK_COUNT never has cmd->data and so does not need to be checked.

> +		return;
> +
> +	sdhci_writew(host, cmd->data->blksz, SDHCI_BLOCK_SIZE);

Block size is set in __sdhci_external_dma_prepare_data() so does it need to
be set here also.

> +	chan = sdhci_external_dma_channel(host, cmd->data);
> +	if (chan)
> +		dma_async_issue_pending(chan);
> +}
> +
> +static int sdhci_external_dma_cleanup(struct sdhci_host *host,
> +				    struct mmc_data *data)

Please align parameters with open parenthesis

> +{
> +	struct dma_chan *chan = sdhci_external_dma_channel(host, data);
> +	int ret = 0;
> +
> +	if (chan)
> +		ret = dmaengine_terminate_async(chan);
> +
> +	return ret;
> +}
> +#else
> +static int sdhci_external_dma_init(struct sdhci_host *host)
> +{
> +	return -EOPNOTSUPP;
> +}
> +
> +static void sdhci_external_dma_release(struct sdhci_host *host)
> +{}
> +
> +static void sdhci_external_dma_prepare_data(struct sdhci_host *host,
> +					    struct mmc_command *cmd)
> +{
> +	/* If MMC_SDHCI_EXTERNAL_DMA not supported, PIO will be used */
> +	sdhci_prepare_data(host, cmd);
> +}
> +
> +static void sdhci_external_dma_pre_transfer(struct sdhci_host *host,
> +					    struct mmc_command *cmd)
> +{}
> +
> +static int sdhci_external_dma_cleanup(struct sdhci_host *host,
> +				    struct mmc_data *data)

Please align parameters with open parenthesis

> +{
> +	return 0;
> +}
> +#endif
> +
> +void sdhci_switch_external_dma(struct sdhci_host *host, bool en)
> +{
> +	host->use_external_dma = en;
> +}
> +EXPORT_SYMBOL_GPL(sdhci_switch_external_dma);
> +
>  static inline bool sdhci_auto_cmd12(struct sdhci_host *host,
>  				    struct mmc_request *mrq)
>  {
> @@ -1374,7 +1595,10 @@ void sdhci_send_command(struct sdhci_host *host, struct mmc_command *cmd)
>  		host->data_cmd = cmd;
>  	}
>  
> -	sdhci_prepare_data(host, cmd);
> +	if (host->use_external_dma)

As mentioned above wrt sdhci_external_dma_prepare_data():

	if (host->use_external_dma && cmd->data)

> +		sdhci_external_dma_prepare_data(host, cmd);
> +	else
> +		sdhci_prepare_data(host, cmd);
>  
>  	sdhci_writel(host, cmd->arg, SDHCI_ARGUMENT);
>  
> @@ -1416,6 +1640,9 @@ void sdhci_send_command(struct sdhci_host *host, struct mmc_command *cmd)
>  		timeout += 10 * HZ;
>  	sdhci_mod_timer(host, cmd->mrq, timeout);
>  
> +	if (host->use_external_dma)
> +		sdhci_external_dma_pre_transfer(host, cmd);

Why is sdhci_external_dma_pre_transfer() needed here - couldn't it be done
in sdhci_external_dma_prepare_data()?

> +
>  	sdhci_writew(host, SDHCI_MAKE_CMD(cmd->opcode, flags), SDHCI_COMMAND);
>  }
>  EXPORT_SYMBOL_GPL(sdhci_send_command);
> @@ -1781,6 +2008,11 @@ void sdhci_request(struct mmc_host *mmc, struct mmc_request *mrq)
>  
>  	sdhci_led_activate(host);
>  
> +	if (host->use_external_dma && mrq->data) {
> +		struct dma_chan *chan = sdhci_external_dma_channel(host,
> +								   mrq->data);

sdhci_external_dma_channel is not declared if
!IS_ENABLED(CONFIG_MMC_SDHCI_EXTERNAL_DMA)

> +		dmaengine_synchronize(chan);

So this is to cover for using dmaengine_terminate_async()?

> +	}
>  	/*
>  	 * Ensure we don't send the STOP for non-SET_BLOCK_COUNTED
>  	 * requests if Auto-CMD12 is enabled.
> @@ -2658,6 +2890,8 @@ static bool sdhci_request_done(struct sdhci_host *host)
>  				dma_unmap_sg(mmc_dev(host->mmc), data->sg,
>  					     data->sg_len,
>  					     mmc_get_dma_dir(data));
> +				if (host->use_external_dma)
> +					sdhci_external_dma_cleanup(host, data);

Is sdhci_external_dma_cleanup() only needed in the error case?

The DMA must be stopped before the memory is unmapped and potentially freed.

Isn't the DMA cleanup also needed in the bounce buffer case?

Isn't the DMA cleanup also needed in the COOKIE_PRE_MAPPED case?

dmaengine_terminate_async() doesn't stop the DMA but
dmaengine_terminate_sync() is not atomic, which looks like a problem.

Perhaps you look at scheduling some work for the external dma error case
instead of calling __sdhci_finish_mrq()?  Then the work can do the
dmaengine_terminate_sync() and call __sdhci_finish_mrq().

>  			}
>  			data->host_cookie = COOKIE_UNMAPPED;
>  		}
> @@ -3692,12 +3926,15 @@ int sdhci_setup_host(struct sdhci_host *host)
>  		       mmc_hostname(mmc), host->version);
>  	}
>  
> -	if (host->quirks & SDHCI_QUIRK_FORCE_DMA)
> +	if (host->quirks & SDHCI_QUIRK_FORCE_DMA) {
>  		host->flags |= SDHCI_USE_SDMA;
> -	else if (!(host->caps & SDHCI_CAN_DO_SDMA))
> +	} else if (!(host->caps & SDHCI_CAN_DO_SDMA)) {
>  		DBG("Controller doesn't have SDMA capability\n");
> -	else
> +	} else if (host->use_external_dma) {
> +		/* Using dma-names to detect external dma capability */
> +	} else {
>  		host->flags |= SDHCI_USE_SDMA;
> +	}

These if-statements are about setting SDHCI_USE_SDMA but why is a change
needed for the host->use_external_dma case?

>  
>  	if ((host->quirks & SDHCI_QUIRK_BROKEN_DMA) &&
>  		(host->flags & SDHCI_USE_SDMA)) {
> @@ -3785,6 +4022,19 @@ int sdhci_setup_host(struct sdhci_host *host)
>  		}
>  	}
>  
> +	if (host->use_external_dma) {
> +		ret = sdhci_external_dma_init(host);
> +		if (ret == -EPROBE_DEFER)
> +			goto unreg;
> +
> +		/*
> +		 * Fall back to use the DMA/PIO integrated in standard SDHCI
> +		 * instead of external DMA devices.
> +		 */
> +		if (ret)
> +			sdhci_switch_external_dma(host, false);
> +	}
> +
>  	/*
>  	 * If we use DMA, then it's up to the caller to set the DMA
>  	 * mask, but PIO does not need the hw shim so we set a new
> @@ -4201,6 +4451,10 @@ void sdhci_cleanup_host(struct sdhci_host *host)
>  		dma_free_coherent(mmc_dev(mmc), host->align_buffer_sz +
>  				  host->adma_table_sz, host->align_buffer,
>  				  host->align_addr);
> +
> +	if (host->use_external_dma)
> +		sdhci_external_dma_release(host);
> +
>  	host->adma_table = NULL;
>  	host->align_buffer = NULL;
>  }
> @@ -4247,6 +4501,7 @@ int __sdhci_add_host(struct sdhci_host *host)
>  
>  	pr_info("%s: SDHCI controller on %s [%s] using %s\n",
>  		mmc_hostname(mmc), host->hw_name, dev_name(mmc_dev(mmc)),
> +		host->use_external_dma ? "External DMA" :
>  		(host->flags & SDHCI_USE_ADMA) ?
>  		(host->flags & SDHCI_USE_64_BIT_DMA) ? "ADMA 64-bit" : "ADMA" :
>  		(host->flags & SDHCI_USE_SDMA) ? "DMA" : "PIO");
> @@ -4335,6 +4590,9 @@ void sdhci_remove_host(struct sdhci_host *host, int dead)
>  				  host->adma_table_sz, host->align_buffer,
>  				  host->align_addr);
>  
> +	if (host->use_external_dma)
> +		sdhci_external_dma_release(host);
> +
>  	host->adma_table = NULL;
>  	host->align_buffer = NULL;
>  }
> diff --git a/drivers/mmc/host/sdhci.h b/drivers/mmc/host/sdhci.h
> index 6cc9a3c2ac66..7a52823ebef4 100644
> --- a/drivers/mmc/host/sdhci.h
> +++ b/drivers/mmc/host/sdhci.h
> @@ -482,6 +482,7 @@ struct sdhci_host {
>  
>  	int irq;		/* Device IRQ */
>  	void __iomem *ioaddr;	/* Mapped address */
> +	phys_addr_t mapbase;	/* physical address base */
>  	char *bounce_buffer;	/* For packing SDMA reads/writes */
>  	dma_addr_t bounce_addr;
>  	unsigned int bounce_buffer_size;
> @@ -531,6 +532,7 @@ struct sdhci_host {
>  	bool pending_reset;	/* Cmd/data reset is pending */
>  	bool irq_wake_enabled;	/* IRQ wakeup is enabled */
>  	bool v4_mode;		/* Host Version 4 Enable */
> +	bool use_external_dma; /* Host selects to use external DMA */

Please align /**/ with above i.e. use tab

>  
>  	struct mmc_request *mrqs_done[SDHCI_MAX_MRQS];	/* Requests done */
>  	struct mmc_command *cmd;	/* Current command */
> @@ -559,6 +561,11 @@ struct sdhci_host {
>  	struct timer_list timer;	/* Timer for timeouts */
>  	struct timer_list data_timer;	/* Timer for data timeouts */
>  
> +#if IS_ENABLED(CONFIG_MMC_SDHCI_EXTERNAL_DMA)
> +	struct dma_chan	*rx_chan;
> +	struct dma_chan	*tx_chan;
> +#endif
> +
>  	u32 caps;		/* CAPABILITY_0 */
>  	u32 caps1;		/* CAPABILITY_1 */
>  	bool read_caps;		/* Capability flags have been read */
> @@ -792,5 +799,6 @@ void sdhci_start_tuning(struct sdhci_host *host);
>  void sdhci_end_tuning(struct sdhci_host *host);
>  void sdhci_reset_tuning(struct sdhci_host *host);
>  void sdhci_send_tuning(struct sdhci_host *host, u32 opcode);
> +void sdhci_switch_external_dma(struct sdhci_host *host, bool en);
>  
>  #endif /* __SDHCI_HW_H */
>
Chunyan Zhang Jan. 28, 2019, 10:20 a.m. UTC | #3
On Thu, 24 Jan 2019 at 19:41, Adrian Hunter <adrian.hunter@intel.com> wrote:
>
> On 11/01/19 1:08 PM, Faiz Abbas wrote:
> > From: Chunyan Zhang <zhang.chunyan@linaro.org>
> >
> > Some standard SD host controllers can support both external dma
> > controllers as well as ADMA/SDMA in which the SD host controller
> > acts as DMA master. TI's omap controller is the case as an example.
> >
> > Currently the generic SDHCI code supports ADMA/SDMA integrated in
> > the host controller but does not have any support for external DMA
> > controllers implemented using dmaengine, meaning that custom code is
> > needed for any systems that use an external DMA controller with SDHCI.
> >
> > Fixes by Faiz Abbas <faiz_abbas@ti.com>:
> > 1. Map scatterlists before dmaengine_prep_slave_sg()
> > 2. Use dma_async() functions inside of the send_command() path and
> > synchronize once at the start of each request.
>
> Sorry for the slow reply, but I do have some concerns.  Please see the comments.
>
> >
> > Signed-off-by: Chunyan Zhang <zhang.chunyan@linaro.org>
> > Signed-off-by: Faiz Abbas <faiz_abbas@ti.com>
> > ---
> >  drivers/mmc/host/Kconfig |   3 +
> >  drivers/mmc/host/sdhci.c | 266 ++++++++++++++++++++++++++++++++++++++-
> >  drivers/mmc/host/sdhci.h |   8 ++
> >  3 files changed, 273 insertions(+), 4 deletions(-)
> >
> > diff --git a/drivers/mmc/host/Kconfig b/drivers/mmc/host/Kconfig
> > index e26b8145efb3..333292e8ecdd 100644
> > --- a/drivers/mmc/host/Kconfig
> > +++ b/drivers/mmc/host/Kconfig
> > @@ -999,3 +999,6 @@ config MMC_SDHCI_AM654
> >         If you have a controller with this interface, say Y or M here.
> >
> >         If unsure, say N.
> > +
> > +config MMC_SDHCI_EXTERNAL_DMA
> > +        bool
> > diff --git a/drivers/mmc/host/sdhci.c b/drivers/mmc/host/sdhci.c
> > index a22e11a65658..4a9044c06e21 100644
> > --- a/drivers/mmc/host/sdhci.c
> > +++ b/drivers/mmc/host/sdhci.c
> > @@ -14,6 +14,7 @@
> >   */
> >
> >  #include <linux/delay.h>
> > +#include <linux/dmaengine.h>
> >  #include <linux/ktime.h>
> >  #include <linux/highmem.h>
> >  #include <linux/io.h>
> > @@ -1118,6 +1119,226 @@ static void sdhci_prepare_data(struct sdhci_host *host, struct mmc_command *cmd)
> >       }
> >  }
> >
> > +#if IS_ENABLED(CONFIG_MMC_SDHCI_EXTERNAL_DMA)
> > +static int sdhci_external_dma_init(struct sdhci_host *host)
> > +{
> > +     int ret = 0;
> > +     struct mmc_host *mmc = host->mmc;
> > +
> > +     host->tx_chan = dma_request_chan(mmc->parent, "tx");
> > +     if (IS_ERR(host->tx_chan)) {
> > +             ret = PTR_ERR(host->tx_chan);
> > +             if (ret != -EPROBE_DEFER)
> > +                     pr_warn("Failed to request TX DMA channel.\n");
> > +             host->tx_chan = NULL;
> > +             return ret;
> > +     }
> > +
> > +     host->rx_chan = dma_request_chan(mmc->parent, "rx");
> > +     if (IS_ERR(host->rx_chan)) {
> > +             if (host->tx_chan) {
> > +                     dma_release_channel(host->tx_chan);
> > +                     host->tx_chan = NULL;
> > +             }
> > +
> > +             ret = PTR_ERR(host->rx_chan);
> > +             if (ret != -EPROBE_DEFER)
> > +                     pr_warn("Failed to request RX DMA channel.\n");
> > +             host->rx_chan = NULL;
> > +     }
> > +
> > +     return ret;
> > +}
> > +
> > +static inline struct dma_chan *
> > +sdhci_external_dma_channel(struct sdhci_host *host, struct mmc_data *data)
> > +{
> > +     return data->flags & MMC_DATA_WRITE ? host->tx_chan : host->rx_chan;
> > +}
> > +
> > +static int sdhci_external_dma_setup(struct sdhci_host *host,
> > +                                 struct mmc_command *cmd)
> > +{
> > +     int ret, i;
> > +     struct dma_async_tx_descriptor *desc;
> > +     struct mmc_data *data = cmd->data;
> > +     struct dma_chan *chan;
> > +     struct dma_slave_config cfg;
> > +     dma_cookie_t cookie;
> > +     int sg_cnt;
> > +
> > +     if (!host->mapbase)
> > +             return -EINVAL;
> > +
> > +     cfg.src_addr = host->mapbase + SDHCI_BUFFER;
> > +     cfg.dst_addr = host->mapbase + SDHCI_BUFFER;
> > +     cfg.src_addr_width = DMA_SLAVE_BUSWIDTH_4_BYTES;
> > +     cfg.dst_addr_width = DMA_SLAVE_BUSWIDTH_4_BYTES;
> > +     cfg.src_maxburst = data->blksz / 4;
> > +     cfg.dst_maxburst = data->blksz / 4;
> > +
> > +     /* Sanity check: all the SG entries must be aligned by block size. */
> > +     for (i = 0; i < data->sg_len; i++) {
> > +             if ((data->sg + i)->length % data->blksz)
> > +                     return -EINVAL;
> > +     }
> > +
> > +     chan = sdhci_external_dma_channel(host, data);
> > +
> > +     ret = dmaengine_slave_config(chan, &cfg);
> > +     if (ret)
> > +             return ret;
> > +
> > +     sg_cnt = sdhci_pre_dma_transfer(host, data, COOKIE_MAPPED);
> > +     if (sg_cnt <= 0)
> > +             return -EINVAL;
> > +
> > +     desc = dmaengine_prep_slave_sg(chan, data->sg, data->sg_len,
> > +                                    mmc_get_dma_dir(data),
> > +                                    DMA_PREP_INTERRUPT | DMA_CTRL_ACK);
> > +     if (!desc)
> > +             return -EINVAL;
> > +
> > +     desc->callback = NULL;
> > +     desc->callback_param = NULL;
> > +
> > +     cookie = dmaengine_submit(desc);
> > +     if (cookie < 0)
> > +             ret = cookie;
> > +
> > +     return ret;
> > +}
> > +
> > +static void sdhci_external_dma_release(struct sdhci_host *host)
> > +{
> > +     if (host->tx_chan) {
> > +             dma_release_channel(host->tx_chan);
> > +             host->tx_chan = NULL;
> > +     }
> > +
> > +     if (host->rx_chan) {
> > +             dma_release_channel(host->rx_chan);
> > +             host->rx_chan = NULL;
> > +     }
> > +
> > +     sdhci_switch_external_dma(host, false);
> > +}
> > +
> > +static int __sdhci_external_dma_prepare_data(struct sdhci_host *host,
> > +                                          struct mmc_command *cmd)
> > +{
> > +     struct mmc_data *data = cmd->data;
> > +
> > +     host->data_timeout = 0;
> > +
> > +     if (sdhci_data_line_cmd(cmd))
> > +             sdhci_set_timeout(host, cmd);
> > +
> > +     WARN_ON(host->data);
> > +
> > +     /* Sanity checks */
> > +     WARN_ON(data->blksz * data->blocks > 524288);
> > +     WARN_ON(data->blksz > host->mmc->max_blk_size);
> > +     WARN_ON(data->blocks > 65535);
> > +
> > +     host->flags |= SDHCI_REQ_USE_DMA;
> > +     host->data = data;
> > +     host->data_early = 0;
> > +     host->data->bytes_xfered = 0;
> > +
> > +     sdhci_set_transfer_irqs(host);
> > +
> > +     /*
> > +      * For Version 4.10 onwards, if v4 mode is enabled, 32-bit Block Count
> > +      * can be supported, in that case 16-bit block count register must be 0.
> > +      */
> > +     if (host->version >= SDHCI_SPEC_410 && host->v4_mode &&
> > +         (host->quirks2 & SDHCI_QUIRK2_USE_32BIT_BLK_CNT)) {
> > +             if (sdhci_readw(host, SDHCI_BLOCK_COUNT))
> > +                     sdhci_writew(host, 0, SDHCI_BLOCK_COUNT);
> > +             sdhci_writew(host, data->blocks, SDHCI_32BIT_BLK_CNT);
> > +     } else {
> > +             sdhci_writew(host, data->blocks, SDHCI_BLOCK_COUNT);
> > +     }
>
> It is probably worth factoring out the code that is shared with
> sdhci_prepare_data() where possible.
>
> > +
> > +     return 0;
> > +}
> > +
> > +static void sdhci_external_dma_prepare_data(struct sdhci_host *host,
> > +                                         struct mmc_command *cmd)
> > +{
> > +     struct mmc_data *data = cmd->data;
> > +
> > +     if (!data)
> > +             return;
>
> Even in the !data case, we still need to set up a timeout for commands with
> busy waiting.  I suggest checking the !data case before calling
> sdhci_external_dma_prepare_data()

Ok.

>
> > +
> > +     if (sdhci_external_dma_setup(host, cmd) ||
> > +         __sdhci_external_dma_prepare_data(host, cmd)) {
> > +             sdhci_external_dma_release(host);
> > +             pr_err("%s: Cannot use external DMA, switch to the DMA/PIO which standard SDHCI provides.\n",
> > +                    mmc_hostname(host->mmc));
> > +             sdhci_prepare_data(host, cmd);
> > +     }
> > +}
> > +
> > +static void sdhci_external_dma_pre_transfer(struct sdhci_host *host,
> > +                                         struct mmc_command *cmd)
> > +{
> > +     struct dma_chan *chan;
> > +
> > +     if (!cmd->data || cmd->opcode == MMC_SET_BLOCK_COUNT)
>
> MMC_SET_BLOCK_COUNT never has cmd->data and so does not need to be checked.

Ok.

>
> > +             return;
> > +
> > +     sdhci_writew(host, cmd->data->blksz, SDHCI_BLOCK_SIZE);
>
> Block size is set in __sdhci_external_dma_prepare_data() so does it need to
> be set here also.

Ok.

>
> > +     chan = sdhci_external_dma_channel(host, cmd->data);
> > +     if (chan)
> > +             dma_async_issue_pending(chan);
> > +}
> > +
> > +static int sdhci_external_dma_cleanup(struct sdhci_host *host,
> > +                                 struct mmc_data *data)
>
> Please align parameters with open parenthesis
>
> > +{
> > +     struct dma_chan *chan = sdhci_external_dma_channel(host, data);
> > +     int ret = 0;
> > +
> > +     if (chan)
> > +             ret = dmaengine_terminate_async(chan);
> > +
> > +     return ret;
> > +}
> > +#else
> > +static int sdhci_external_dma_init(struct sdhci_host *host)
> > +{
> > +     return -EOPNOTSUPP;
> > +}
> > +
> > +static void sdhci_external_dma_release(struct sdhci_host *host)
> > +{}
> > +
> > +static void sdhci_external_dma_prepare_data(struct sdhci_host *host,
> > +                                         struct mmc_command *cmd)
> > +{
> > +     /* If MMC_SDHCI_EXTERNAL_DMA not supported, PIO will be used */
> > +     sdhci_prepare_data(host, cmd);
> > +}
> > +
> > +static void sdhci_external_dma_pre_transfer(struct sdhci_host *host,
> > +                                         struct mmc_command *cmd)
> > +{}
> > +
> > +static int sdhci_external_dma_cleanup(struct sdhci_host *host,
> > +                                 struct mmc_data *data)
>
> Please align parameters with open parenthesis
>
> > +{
> > +     return 0;
> > +}
> > +#endif
> > +
> > +void sdhci_switch_external_dma(struct sdhci_host *host, bool en)
> > +{
> > +     host->use_external_dma = en;
> > +}
> > +EXPORT_SYMBOL_GPL(sdhci_switch_external_dma);
> > +
> >  static inline bool sdhci_auto_cmd12(struct sdhci_host *host,
> >                                   struct mmc_request *mrq)
> >  {
> > @@ -1374,7 +1595,10 @@ void sdhci_send_command(struct sdhci_host *host, struct mmc_command *cmd)
> >               host->data_cmd = cmd;
> >       }
> >
> > -     sdhci_prepare_data(host, cmd);
> > +     if (host->use_external_dma)
>
> As mentioned above wrt sdhci_external_dma_prepare_data():
>
>         if (host->use_external_dma && cmd->data)
>

Maybe we should move checking the !data case out of both
_prepare_data, and add it to sdhci_send_command() before calling
_prepare_data(), that's saying we can do set up a timeout in
sdhci_send_command().

> > +             sdhci_external_dma_prepare_data(host, cmd);
> > +     else
> > +             sdhci_prepare_data(host, cmd);
> >
> >       sdhci_writel(host, cmd->arg, SDHCI_ARGUMENT);
> >
> > @@ -1416,6 +1640,9 @@ void sdhci_send_command(struct sdhci_host *host, struct mmc_command *cmd)
> >               timeout += 10 * HZ;
> >       sdhci_mod_timer(host, cmd->mrq, timeout);
> >
> > +     if (host->use_external_dma)
> > +             sdhci_external_dma_pre_transfer(host, cmd);
>
> Why is sdhci_external_dma_pre_transfer() needed here - couldn't it be done
> in sdhci_external_dma_prepare_data()?
>

I'm not sure dma_async_issue_pending() can be done so early in
sdhci_external_dma_prepare_data().

> > +
> >       sdhci_writew(host, SDHCI_MAKE_CMD(cmd->opcode, flags), SDHCI_COMMAND);
> >  }
> >  EXPORT_SYMBOL_GPL(sdhci_send_command);
> > @@ -1781,6 +2008,11 @@ void sdhci_request(struct mmc_host *mmc, struct mmc_request *mrq)
> >
> >       sdhci_led_activate(host);
> >
> > +     if (host->use_external_dma && mrq->data) {
> > +             struct dma_chan *chan = sdhci_external_dma_channel(host,
> > +                                                                mrq->data);
>
> sdhci_external_dma_channel is not declared if
> !IS_ENABLED(CONFIG_MMC_SDHCI_EXTERNAL_DMA)

Ok, I guess this can be moved to sdhci_external_dma_prepare_data().

>
> > +             dmaengine_synchronize(chan);
>
> So this is to cover for using dmaengine_terminate_async()?

Ok.

>
> > +     }
> >       /*
> >        * Ensure we don't send the STOP for non-SET_BLOCK_COUNTED
> >        * requests if Auto-CMD12 is enabled.
> > @@ -2658,6 +2890,8 @@ static bool sdhci_request_done(struct sdhci_host *host)
> >                               dma_unmap_sg(mmc_dev(host->mmc), data->sg,
> >                                            data->sg_len,
> >                                            mmc_get_dma_dir(data));
> > +                             if (host->use_external_dma)
> > +                                     sdhci_external_dma_cleanup(host, data);
>
> Is sdhci_external_dma_cleanup() only needed in the error case?
>
> The DMA must be stopped before the memory is unmapped and potentially freed.
>
> Isn't the DMA cleanup also needed in the bounce buffer case?
>
> Isn't the DMA cleanup also needed in the COOKIE_PRE_MAPPED case?
>
> dmaengine_terminate_async() doesn't stop the DMA but
> dmaengine_terminate_sync() is not atomic, which looks like a problem.
>
> Perhaps you look at scheduling some work for the external dma error case
> instead of calling __sdhci_finish_mrq()?  Then the work can do the
> dmaengine_terminate_sync() and call __sdhci_finish_mrq().

Ok, I will look at these issues.

>
> >                       }
> >                       data->host_cookie = COOKIE_UNMAPPED;
> >               }
> > @@ -3692,12 +3926,15 @@ int sdhci_setup_host(struct sdhci_host *host)
> >                      mmc_hostname(mmc), host->version);
> >       }
> >
> > -     if (host->quirks & SDHCI_QUIRK_FORCE_DMA)
> > +     if (host->quirks & SDHCI_QUIRK_FORCE_DMA) {
> >               host->flags |= SDHCI_USE_SDMA;
> > -     else if (!(host->caps & SDHCI_CAN_DO_SDMA))
> > +     } else if (!(host->caps & SDHCI_CAN_DO_SDMA)) {
> >               DBG("Controller doesn't have SDMA capability\n");
> > -     else
> > +     } else if (host->use_external_dma) {
> > +             /* Using dma-names to detect external dma capability */
> > +     } else {
> >               host->flags |= SDHCI_USE_SDMA;
> > +     }
>
> These if-statements are about setting SDHCI_USE_SDMA but why is a change
> needed for the host->use_external_dma case?

Yes, this is not needed, otherwise the controller cannot switch back
to SDMA if it is supported in the controller.

>
> >
> >       if ((host->quirks & SDHCI_QUIRK_BROKEN_DMA) &&
> >               (host->flags & SDHCI_USE_SDMA)) {
> > @@ -3785,6 +4022,19 @@ int sdhci_setup_host(struct sdhci_host *host)
> >               }
> >       }
> >
> > +     if (host->use_external_dma) {
> > +             ret = sdhci_external_dma_init(host);
> > +             if (ret == -EPROBE_DEFER)
> > +                     goto unreg;
> > +
> > +             /*
> > +              * Fall back to use the DMA/PIO integrated in standard SDHCI
> > +              * instead of external DMA devices.
> > +              */
> > +             if (ret)
> > +                     sdhci_switch_external_dma(host, false);
> > +     }
> > +
> >       /*
> >        * If we use DMA, then it's up to the caller to set the DMA
> >        * mask, but PIO does not need the hw shim so we set a new
> > @@ -4201,6 +4451,10 @@ void sdhci_cleanup_host(struct sdhci_host *host)
> >               dma_free_coherent(mmc_dev(mmc), host->align_buffer_sz +
> >                                 host->adma_table_sz, host->align_buffer,
> >                                 host->align_addr);
> > +
> > +     if (host->use_external_dma)
> > +             sdhci_external_dma_release(host);
> > +
> >       host->adma_table = NULL;
> >       host->align_buffer = NULL;
> >  }
> > @@ -4247,6 +4501,7 @@ int __sdhci_add_host(struct sdhci_host *host)
> >
> >       pr_info("%s: SDHCI controller on %s [%s] using %s\n",
> >               mmc_hostname(mmc), host->hw_name, dev_name(mmc_dev(mmc)),
> > +             host->use_external_dma ? "External DMA" :
> >               (host->flags & SDHCI_USE_ADMA) ?
> >               (host->flags & SDHCI_USE_64_BIT_DMA) ? "ADMA 64-bit" : "ADMA" :
> >               (host->flags & SDHCI_USE_SDMA) ? "DMA" : "PIO");
> > @@ -4335,6 +4590,9 @@ void sdhci_remove_host(struct sdhci_host *host, int dead)
> >                                 host->adma_table_sz, host->align_buffer,
> >                                 host->align_addr);
> >
> > +     if (host->use_external_dma)
> > +             sdhci_external_dma_release(host);
> > +
> >       host->adma_table = NULL;
> >       host->align_buffer = NULL;
> >  }
> > diff --git a/drivers/mmc/host/sdhci.h b/drivers/mmc/host/sdhci.h
> > index 6cc9a3c2ac66..7a52823ebef4 100644
> > --- a/drivers/mmc/host/sdhci.h
> > +++ b/drivers/mmc/host/sdhci.h
> > @@ -482,6 +482,7 @@ struct sdhci_host {
> >
> >       int irq;                /* Device IRQ */
> >       void __iomem *ioaddr;   /* Mapped address */
> > +     phys_addr_t mapbase;    /* physical address base */
> >       char *bounce_buffer;    /* For packing SDMA reads/writes */
> >       dma_addr_t bounce_addr;
> >       unsigned int bounce_buffer_size;
> > @@ -531,6 +532,7 @@ struct sdhci_host {
> >       bool pending_reset;     /* Cmd/data reset is pending */
> >       bool irq_wake_enabled;  /* IRQ wakeup is enabled */
> >       bool v4_mode;           /* Host Version 4 Enable */
> > +     bool use_external_dma; /* Host selects to use external DMA */
>
> Please align /**/ with above i.e. use tab
>
> >
> >       struct mmc_request *mrqs_done[SDHCI_MAX_MRQS];  /* Requests done */
> >       struct mmc_command *cmd;        /* Current command */
> > @@ -559,6 +561,11 @@ struct sdhci_host {
> >       struct timer_list timer;        /* Timer for timeouts */
> >       struct timer_list data_timer;   /* Timer for data timeouts */
> >
> > +#if IS_ENABLED(CONFIG_MMC_SDHCI_EXTERNAL_DMA)
> > +     struct dma_chan *rx_chan;
> > +     struct dma_chan *tx_chan;
> > +#endif
> > +
> >       u32 caps;               /* CAPABILITY_0 */
> >       u32 caps1;              /* CAPABILITY_1 */
> >       bool read_caps;         /* Capability flags have been read */
> > @@ -792,5 +799,6 @@ void sdhci_start_tuning(struct sdhci_host *host);
> >  void sdhci_end_tuning(struct sdhci_host *host);
> >  void sdhci_reset_tuning(struct sdhci_host *host);
> >  void sdhci_send_tuning(struct sdhci_host *host, u32 opcode);
> > +void sdhci_switch_external_dma(struct sdhci_host *host, bool en);
> >
> >  #endif /* __SDHCI_HW_H */
> >
>
Faiz Abbas Jan. 28, 2019, 10:43 a.m. UTC | #4
Hi,

On 28/01/19 3:50 PM, Chunyan Zhang wrote:
> On Thu, 24 Jan 2019 at 19:41, Adrian Hunter <adrian.hunter@intel.com> wrote:
>>
>> On 11/01/19 1:08 PM, Faiz Abbas wrote:
>>> From: Chunyan Zhang <zhang.chunyan@linaro.org>
>>>
>>> Some standard SD host controllers can support both external dma
>>> controllers as well as ADMA/SDMA in which the SD host controller
>>> acts as DMA master. TI's omap controller is the case as an example.
>>>
>>> Currently the generic SDHCI code supports ADMA/SDMA integrated in
>>> the host controller but does not have any support for external DMA
>>> controllers implemented using dmaengine, meaning that custom code is
>>> needed for any systems that use an external DMA controller with SDHCI.
>>>
>>> Fixes by Faiz Abbas <faiz_abbas@ti.com>:
>>> 1. Map scatterlists before dmaengine_prep_slave_sg()
>>> 2. Use dma_async() functions inside of the send_command() path and
>>> synchronize once at the start of each request.
>>
>> Sorry for the slow reply, but I do have some concerns.  Please see the comments.
>>
>>>
>>> Signed-off-by: Chunyan Zhang <zhang.chunyan@linaro.org>
>>> Signed-off-by: Faiz Abbas <faiz_abbas@ti.com>
>>> ---
>>>  drivers/mmc/host/Kconfig |   3 +
>>>  drivers/mmc/host/sdhci.c | 266 ++++++++++++++++++++++++++++++++++++++-
>>>  drivers/mmc/host/sdhci.h |   8 ++
>>>  3 files changed, 273 insertions(+), 4 deletions(-)
>>>
>>> diff --git a/drivers/mmc/host/Kconfig b/drivers/mmc/host/Kconfig
>>> index e26b8145efb3..333292e8ecdd 100644
>>> --- a/drivers/mmc/host/Kconfig
>>> +++ b/drivers/mmc/host/Kconfig
>>> @@ -999,3 +999,6 @@ config MMC_SDHCI_AM654
>>>         If you have a controller with this interface, say Y or M here.
>>>
>>>         If unsure, say N.
>>> +
>>> +config MMC_SDHCI_EXTERNAL_DMA
>>> +        bool
>>> diff --git a/drivers/mmc/host/sdhci.c b/drivers/mmc/host/sdhci.c
>>> index a22e11a65658..4a9044c06e21 100644
>>> --- a/drivers/mmc/host/sdhci.c
>>> +++ b/drivers/mmc/host/sdhci.c
>>> @@ -14,6 +14,7 @@
>>>   */
>>>
>>>  #include <linux/delay.h>
>>> +#include <linux/dmaengine.h>
>>>  #include <linux/ktime.h>
>>>  #include <linux/highmem.h>
>>>  #include <linux/io.h>
>>> @@ -1118,6 +1119,226 @@ static void sdhci_prepare_data(struct sdhci_host *host, struct mmc_command *cmd)
>>>       }
>>>  }
>>>
>>> +#if IS_ENABLED(CONFIG_MMC_SDHCI_EXTERNAL_DMA)
>>> +static int sdhci_external_dma_init(struct sdhci_host *host)
>>> +{
>>> +     int ret = 0;
>>> +     struct mmc_host *mmc = host->mmc;
>>> +
>>> +     host->tx_chan = dma_request_chan(mmc->parent, "tx");
>>> +     if (IS_ERR(host->tx_chan)) {
>>> +             ret = PTR_ERR(host->tx_chan);
>>> +             if (ret != -EPROBE_DEFER)
>>> +                     pr_warn("Failed to request TX DMA channel.\n");
>>> +             host->tx_chan = NULL;
>>> +             return ret;
>>> +     }
>>> +
>>> +     host->rx_chan = dma_request_chan(mmc->parent, "rx");
>>> +     if (IS_ERR(host->rx_chan)) {
>>> +             if (host->tx_chan) {
>>> +                     dma_release_channel(host->tx_chan);
>>> +                     host->tx_chan = NULL;
>>> +             }
>>> +
>>> +             ret = PTR_ERR(host->rx_chan);
>>> +             if (ret != -EPROBE_DEFER)
>>> +                     pr_warn("Failed to request RX DMA channel.\n");
>>> +             host->rx_chan = NULL;
>>> +     }
>>> +
>>> +     return ret;
>>> +}
>>> +
>>> +static inline struct dma_chan *
>>> +sdhci_external_dma_channel(struct sdhci_host *host, struct mmc_data *data)
>>> +{
>>> +     return data->flags & MMC_DATA_WRITE ? host->tx_chan : host->rx_chan;
>>> +}
>>> +
>>> +static int sdhci_external_dma_setup(struct sdhci_host *host,
>>> +                                 struct mmc_command *cmd)
>>> +{
>>> +     int ret, i;
>>> +     struct dma_async_tx_descriptor *desc;
>>> +     struct mmc_data *data = cmd->data;
>>> +     struct dma_chan *chan;
>>> +     struct dma_slave_config cfg;
>>> +     dma_cookie_t cookie;
>>> +     int sg_cnt;
>>> +
>>> +     if (!host->mapbase)
>>> +             return -EINVAL;
>>> +
>>> +     cfg.src_addr = host->mapbase + SDHCI_BUFFER;
>>> +     cfg.dst_addr = host->mapbase + SDHCI_BUFFER;
>>> +     cfg.src_addr_width = DMA_SLAVE_BUSWIDTH_4_BYTES;
>>> +     cfg.dst_addr_width = DMA_SLAVE_BUSWIDTH_4_BYTES;
>>> +     cfg.src_maxburst = data->blksz / 4;
>>> +     cfg.dst_maxburst = data->blksz / 4;
>>> +
>>> +     /* Sanity check: all the SG entries must be aligned by block size. */
>>> +     for (i = 0; i < data->sg_len; i++) {
>>> +             if ((data->sg + i)->length % data->blksz)
>>> +                     return -EINVAL;
>>> +     }
>>> +
>>> +     chan = sdhci_external_dma_channel(host, data);
>>> +
>>> +     ret = dmaengine_slave_config(chan, &cfg);
>>> +     if (ret)
>>> +             return ret;
>>> +
>>> +     sg_cnt = sdhci_pre_dma_transfer(host, data, COOKIE_MAPPED);
>>> +     if (sg_cnt <= 0)
>>> +             return -EINVAL;
>>> +
>>> +     desc = dmaengine_prep_slave_sg(chan, data->sg, data->sg_len,
>>> +                                    mmc_get_dma_dir(data),
>>> +                                    DMA_PREP_INTERRUPT | DMA_CTRL_ACK);
>>> +     if (!desc)
>>> +             return -EINVAL;
>>> +
>>> +     desc->callback = NULL;
>>> +     desc->callback_param = NULL;
>>> +
>>> +     cookie = dmaengine_submit(desc);
>>> +     if (cookie < 0)
>>> +             ret = cookie;
>>> +
>>> +     return ret;
>>> +}
>>> +
>>> +static void sdhci_external_dma_release(struct sdhci_host *host)
>>> +{
>>> +     if (host->tx_chan) {
>>> +             dma_release_channel(host->tx_chan);
>>> +             host->tx_chan = NULL;
>>> +     }
>>> +
>>> +     if (host->rx_chan) {
>>> +             dma_release_channel(host->rx_chan);
>>> +             host->rx_chan = NULL;
>>> +     }
>>> +
>>> +     sdhci_switch_external_dma(host, false);
>>> +}
>>> +
>>> +static int __sdhci_external_dma_prepare_data(struct sdhci_host *host,
>>> +                                          struct mmc_command *cmd)
>>> +{
>>> +     struct mmc_data *data = cmd->data;
>>> +
>>> +     host->data_timeout = 0;
>>> +
>>> +     if (sdhci_data_line_cmd(cmd))
>>> +             sdhci_set_timeout(host, cmd);
>>> +
>>> +     WARN_ON(host->data);
>>> +
>>> +     /* Sanity checks */
>>> +     WARN_ON(data->blksz * data->blocks > 524288);
>>> +     WARN_ON(data->blksz > host->mmc->max_blk_size);
>>> +     WARN_ON(data->blocks > 65535);
>>> +
>>> +     host->flags |= SDHCI_REQ_USE_DMA;
>>> +     host->data = data;
>>> +     host->data_early = 0;
>>> +     host->data->bytes_xfered = 0;
>>> +
>>> +     sdhci_set_transfer_irqs(host);
>>> +
>>> +     /*
>>> +      * For Version 4.10 onwards, if v4 mode is enabled, 32-bit Block Count
>>> +      * can be supported, in that case 16-bit block count register must be 0.
>>> +      */
>>> +     if (host->version >= SDHCI_SPEC_410 && host->v4_mode &&
>>> +         (host->quirks2 & SDHCI_QUIRK2_USE_32BIT_BLK_CNT)) {
>>> +             if (sdhci_readw(host, SDHCI_BLOCK_COUNT))
>>> +                     sdhci_writew(host, 0, SDHCI_BLOCK_COUNT);
>>> +             sdhci_writew(host, data->blocks, SDHCI_32BIT_BLK_CNT);
>>> +     } else {
>>> +             sdhci_writew(host, data->blocks, SDHCI_BLOCK_COUNT);
>>> +     }
>>
>> It is probably worth factoring out the code that is shared with
>> sdhci_prepare_data() where possible.
>>
>>> +
>>> +     return 0;
>>> +}
>>> +
>>> +static void sdhci_external_dma_prepare_data(struct sdhci_host *host,
>>> +                                         struct mmc_command *cmd)
>>> +{
>>> +     struct mmc_data *data = cmd->data;
>>> +
>>> +     if (!data)
>>> +             return;
>>
>> Even in the !data case, we still need to set up a timeout for commands with
>> busy waiting.  I suggest checking the !data case before calling
>> sdhci_external_dma_prepare_data()
> 
> Ok.
> 
>>
>>> +
>>> +     if (sdhci_external_dma_setup(host, cmd) ||
>>> +         __sdhci_external_dma_prepare_data(host, cmd)) {
>>> +             sdhci_external_dma_release(host);
>>> +             pr_err("%s: Cannot use external DMA, switch to the DMA/PIO which standard SDHCI provides.\n",
>>> +                    mmc_hostname(host->mmc));
>>> +             sdhci_prepare_data(host, cmd);
>>> +     }
>>> +}
>>> +
>>> +static void sdhci_external_dma_pre_transfer(struct sdhci_host *host,
>>> +                                         struct mmc_command *cmd)
>>> +{
>>> +     struct dma_chan *chan;
>>> +
>>> +     if (!cmd->data || cmd->opcode == MMC_SET_BLOCK_COUNT)
>>
>> MMC_SET_BLOCK_COUNT never has cmd->data and so does not need to be checked.
> 
> Ok.
> 
>>
>>> +             return;
>>> +
>>> +     sdhci_writew(host, cmd->data->blksz, SDHCI_BLOCK_SIZE);
>>
>> Block size is set in __sdhci_external_dma_prepare_data() so does it need to
>> be set here also.
> 
> Ok.

That is BLOCK_COUNT. This is BLOCK_SIZE.

> 
>>
>>> +     chan = sdhci_external_dma_channel(host, cmd->data);
>>> +     if (chan)
>>> +             dma_async_issue_pending(chan);
>>> +}
>>> +
>>> +static int sdhci_external_dma_cleanup(struct sdhci_host *host,
>>> +                                 struct mmc_data *data)
>>
>> Please align parameters with open parenthesis
>>
>>> +{
>>> +     struct dma_chan *chan = sdhci_external_dma_channel(host, data);
>>> +     int ret = 0;
>>> +
>>> +     if (chan)
>>> +             ret = dmaengine_terminate_async(chan);
>>> +
>>> +     return ret;
>>> +}
>>> +#else
>>> +static int sdhci_external_dma_init(struct sdhci_host *host)
>>> +{
>>> +     return -EOPNOTSUPP;
>>> +}
>>> +
>>> +static void sdhci_external_dma_release(struct sdhci_host *host)
>>> +{}
>>> +
>>> +static void sdhci_external_dma_prepare_data(struct sdhci_host *host,
>>> +                                         struct mmc_command *cmd)
>>> +{
>>> +     /* If MMC_SDHCI_EXTERNAL_DMA not supported, PIO will be used */
>>> +     sdhci_prepare_data(host, cmd);
>>> +}
>>> +
>>> +static void sdhci_external_dma_pre_transfer(struct sdhci_host *host,
>>> +                                         struct mmc_command *cmd)
>>> +{}
>>> +
>>> +static int sdhci_external_dma_cleanup(struct sdhci_host *host,
>>> +                                 struct mmc_data *data)
>>
>> Please align parameters with open parenthesis
>>
>>> +{
>>> +     return 0;
>>> +}
>>> +#endif
>>> +
>>> +void sdhci_switch_external_dma(struct sdhci_host *host, bool en)
>>> +{
>>> +     host->use_external_dma = en;
>>> +}
>>> +EXPORT_SYMBOL_GPL(sdhci_switch_external_dma);
>>> +
>>>  static inline bool sdhci_auto_cmd12(struct sdhci_host *host,
>>>                                   struct mmc_request *mrq)
>>>  {
>>> @@ -1374,7 +1595,10 @@ void sdhci_send_command(struct sdhci_host *host, struct mmc_command *cmd)
>>>               host->data_cmd = cmd;
>>>       }
>>>
>>> -     sdhci_prepare_data(host, cmd);
>>> +     if (host->use_external_dma)
>>
>> As mentioned above wrt sdhci_external_dma_prepare_data():
>>
>>         if (host->use_external_dma && cmd->data)
>>
> 
> Maybe we should move checking the !data case out of both
> _prepare_data, and add it to sdhci_send_command() before calling
> _prepare_data(), that's saying we can do set up a timeout in
> sdhci_send_command().

Lets just factor out everything in common as Adrian said. The !data,
BLOCK_COUNT writes and busy waiting.

> 
>>> +             sdhci_external_dma_prepare_data(host, cmd);
>>> +     else
>>> +             sdhci_prepare_data(host, cmd);
>>>
>>>       sdhci_writel(host, cmd->arg, SDHCI_ARGUMENT);
>>>
>>> @@ -1416,6 +1640,9 @@ void sdhci_send_command(struct sdhci_host *host, struct mmc_command *cmd)
>>>               timeout += 10 * HZ;
>>>       sdhci_mod_timer(host, cmd->mrq, timeout);
>>>
>>> +     if (host->use_external_dma)
>>> +             sdhci_external_dma_pre_transfer(host, cmd);
>>
>> Why is sdhci_external_dma_pre_transfer() needed here - couldn't it be done
>> in sdhci_external_dma_prepare_data()?
>>
> 
> I'm not sure dma_async_issue_pending() can be done so early in
> sdhci_external_dma_prepare_data().
> 
>>> +
>>>       sdhci_writew(host, SDHCI_MAKE_CMD(cmd->opcode, flags), SDHCI_COMMAND);
>>>  }
>>>  EXPORT_SYMBOL_GPL(sdhci_send_command);
>>> @@ -1781,6 +2008,11 @@ void sdhci_request(struct mmc_host *mmc, struct mmc_request *mrq)
>>>
>>>       sdhci_led_activate(host);
>>>
>>> +     if (host->use_external_dma && mrq->data) {
>>> +             struct dma_chan *chan = sdhci_external_dma_channel(host,
>>> +                                                                mrq->data);
>>
>> sdhci_external_dma_channel is not declared if
>> !IS_ENABLED(CONFIG_MMC_SDHCI_EXTERNAL_DMA)
> 
> Ok, I guess this can be moved to sdhci_external_dma_prepare_data().

No. We have to synchronize from a guaranteed non-atomic context. Maybe
add another empty sdhci_external_dma_channel definition in the #else part?

> 
>>
>>> +             dmaengine_synchronize(chan);
>>
>> So this is to cover for using dmaengine_terminate_async()?

Yes.

> 
> Ok.
> 
>>
>>> +     }
>>>       /*
>>>        * Ensure we don't send the STOP for non-SET_BLOCK_COUNTED
>>>        * requests if Auto-CMD12 is enabled.
>>> @@ -2658,6 +2890,8 @@ static bool sdhci_request_done(struct sdhci_host *host)
>>>                               dma_unmap_sg(mmc_dev(host->mmc), data->sg,
>>>                                            data->sg_len,
>>>                                            mmc_get_dma_dir(data));
>>> +                             if (host->use_external_dma)
>>> +                                     sdhci_external_dma_cleanup(host, data);
>>
>> Is sdhci_external_dma_cleanup() only needed in the error case?
>>
>> The DMA must be stopped before the memory is unmapped and potentially freed.
>>
>> Isn't the DMA cleanup also needed in the bounce buffer case?
>>
>> Isn't the DMA cleanup also needed in the COOKIE_PRE_MAPPED case?
>>
>> dmaengine_terminate_async() doesn't stop the DMA but
>> dmaengine_terminate_sync() is not atomic, which looks like a problem.
>>
>> Perhaps you look at scheduling some work for the external dma error case
>> instead of calling __sdhci_finish_mrq()?  Then the work can do the
>> dmaengine_terminate_sync() and call __sdhci_finish_mrq().
> 
> Ok, I will look at these issues.
> 
>>
>>>                       }
>>>                       data->host_cookie = COOKIE_UNMAPPED;
>>>               }
>>> @@ -3692,12 +3926,15 @@ int sdhci_setup_host(struct sdhci_host *host)
>>>                      mmc_hostname(mmc), host->version);
>>>       }
>>>
>>> -     if (host->quirks & SDHCI_QUIRK_FORCE_DMA)
>>> +     if (host->quirks & SDHCI_QUIRK_FORCE_DMA) {
>>>               host->flags |= SDHCI_USE_SDMA;
>>> -     else if (!(host->caps & SDHCI_CAN_DO_SDMA))
>>> +     } else if (!(host->caps & SDHCI_CAN_DO_SDMA)) {
>>>               DBG("Controller doesn't have SDMA capability\n");
>>> -     else
>>> +     } else if (host->use_external_dma) {
>>> +             /* Using dma-names to detect external dma capability */
>>> +     } else {
>>>               host->flags |= SDHCI_USE_SDMA;
>>> +     }
>>
>> These if-statements are about setting SDHCI_USE_SDMA but why is a change
>> needed for the host->use_external_dma case?
> 
> Yes, this is not needed, otherwise the controller cannot switch back
> to SDMA if it is supported in the controller.

The else case means that the host always uses SDMA. Needed to make sure
this doesn't happen in the external_dma case. Its not SDHCI_CAN_DO_SDMA,
its SDHCI_USE_SDMA.

> 
>>
>>>
>>>       if ((host->quirks & SDHCI_QUIRK_BROKEN_DMA) &&
>>>               (host->flags & SDHCI_USE_SDMA)) {
>>> @@ -3785,6 +4022,19 @@ int sdhci_setup_host(struct sdhci_host *host)
>>>               }
>>>       }
>>>
>>> +     if (host->use_external_dma) {
>>> +             ret = sdhci_external_dma_init(host);
>>> +             if (ret == -EPROBE_DEFER)
>>> +                     goto unreg;
>>> +
>>> +             /*
>>> +              * Fall back to use the DMA/PIO integrated in standard SDHCI
>>> +              * instead of external DMA devices.
>>> +              */
>>> +             if (ret)
>>> +                     sdhci_switch_external_dma(host, false);
>>> +     }
>>> +
>>>       /*
>>>        * If we use DMA, then it's up to the caller to set the DMA
>>>        * mask, but PIO does not need the hw shim so we set a new
>>> @@ -4201,6 +4451,10 @@ void sdhci_cleanup_host(struct sdhci_host *host)
>>>               dma_free_coherent(mmc_dev(mmc), host->align_buffer_sz +
>>>                                 host->adma_table_sz, host->align_buffer,
>>>                                 host->align_addr);
>>> +
>>> +     if (host->use_external_dma)
>>> +             sdhci_external_dma_release(host);
>>> +
>>>       host->adma_table = NULL;
>>>       host->align_buffer = NULL;
>>>  }
>>> @@ -4247,6 +4501,7 @@ int __sdhci_add_host(struct sdhci_host *host)
>>>
>>>       pr_info("%s: SDHCI controller on %s [%s] using %s\n",
>>>               mmc_hostname(mmc), host->hw_name, dev_name(mmc_dev(mmc)),
>>> +             host->use_external_dma ? "External DMA" :
>>>               (host->flags & SDHCI_USE_ADMA) ?
>>>               (host->flags & SDHCI_USE_64_BIT_DMA) ? "ADMA 64-bit" : "ADMA" :
>>>               (host->flags & SDHCI_USE_SDMA) ? "DMA" : "PIO");
>>> @@ -4335,6 +4590,9 @@ void sdhci_remove_host(struct sdhci_host *host, int dead)
>>>                                 host->adma_table_sz, host->align_buffer,
>>>                                 host->align_addr);
>>>
>>> +     if (host->use_external_dma)
>>> +             sdhci_external_dma_release(host);
>>> +
>>>       host->adma_table = NULL;
>>>       host->align_buffer = NULL;
>>>  }
>>> diff --git a/drivers/mmc/host/sdhci.h b/drivers/mmc/host/sdhci.h
>>> index 6cc9a3c2ac66..7a52823ebef4 100644
>>> --- a/drivers/mmc/host/sdhci.h
>>> +++ b/drivers/mmc/host/sdhci.h
>>> @@ -482,6 +482,7 @@ struct sdhci_host {
>>>
>>>       int irq;                /* Device IRQ */
>>>       void __iomem *ioaddr;   /* Mapped address */
>>> +     phys_addr_t mapbase;    /* physical address base */
>>>       char *bounce_buffer;    /* For packing SDMA reads/writes */
>>>       dma_addr_t bounce_addr;
>>>       unsigned int bounce_buffer_size;
>>> @@ -531,6 +532,7 @@ struct sdhci_host {
>>>       bool pending_reset;     /* Cmd/data reset is pending */
>>>       bool irq_wake_enabled;  /* IRQ wakeup is enabled */
>>>       bool v4_mode;           /* Host Version 4 Enable */
>>> +     bool use_external_dma; /* Host selects to use external DMA */
>>
>> Please align /**/ with above i.e. use tab

Will fix.

Thanks,
Faiz
Chunyan Zhang Jan. 28, 2019, 11:46 a.m. UTC | #5
On Mon, 28 Jan 2019 at 18:40, Faiz Abbas <faiz_abbas@ti.com> wrote:
>
> Hi,
>
> On 28/01/19 3:50 PM, Chunyan Zhang wrote:
> > On Thu, 24 Jan 2019 at 19:41, Adrian Hunter <adrian.hunter@intel.com> wrote:
> >>
> >> On 11/01/19 1:08 PM, Faiz Abbas wrote:
> >>> From: Chunyan Zhang <zhang.chunyan@linaro.org>
> >>>
> >>> Some standard SD host controllers can support both external dma
> >>> controllers as well as ADMA/SDMA in which the SD host controller
> >>> acts as DMA master. TI's omap controller is the case as an example.
> >>>
> >>> Currently the generic SDHCI code supports ADMA/SDMA integrated in
> >>> the host controller but does not have any support for external DMA
> >>> controllers implemented using dmaengine, meaning that custom code is
> >>> needed for any systems that use an external DMA controller with SDHCI.
> >>>
> >>> Fixes by Faiz Abbas <faiz_abbas@ti.com>:
> >>> 1. Map scatterlists before dmaengine_prep_slave_sg()
> >>> 2. Use dma_async() functions inside of the send_command() path and
> >>> synchronize once at the start of each request.
> >>
> >> Sorry for the slow reply, but I do have some concerns.  Please see the comments.
> >>
> >>>
> >>> Signed-off-by: Chunyan Zhang <zhang.chunyan@linaro.org>
> >>> Signed-off-by: Faiz Abbas <faiz_abbas@ti.com>
> >>> ---
> >>>  drivers/mmc/host/Kconfig |   3 +
> >>>  drivers/mmc/host/sdhci.c | 266 ++++++++++++++++++++++++++++++++++++++-
> >>>  drivers/mmc/host/sdhci.h |   8 ++
> >>>  3 files changed, 273 insertions(+), 4 deletions(-)
> >>>
> >>> diff --git a/drivers/mmc/host/Kconfig b/drivers/mmc/host/Kconfig
> >>> index e26b8145efb3..333292e8ecdd 100644
> >>> --- a/drivers/mmc/host/Kconfig
> >>> +++ b/drivers/mmc/host/Kconfig
> >>> @@ -999,3 +999,6 @@ config MMC_SDHCI_AM654
> >>>         If you have a controller with this interface, say Y or M here.
> >>>
> >>>         If unsure, say N.
> >>> +
> >>> +config MMC_SDHCI_EXTERNAL_DMA
> >>> +        bool
> >>> diff --git a/drivers/mmc/host/sdhci.c b/drivers/mmc/host/sdhci.c
> >>> index a22e11a65658..4a9044c06e21 100644
> >>> --- a/drivers/mmc/host/sdhci.c
> >>> +++ b/drivers/mmc/host/sdhci.c
> >>> @@ -14,6 +14,7 @@
> >>>   */
> >>>
> >>>  #include <linux/delay.h>
> >>> +#include <linux/dmaengine.h>
> >>>  #include <linux/ktime.h>
> >>>  #include <linux/highmem.h>
> >>>  #include <linux/io.h>
> >>> @@ -1118,6 +1119,226 @@ static void sdhci_prepare_data(struct sdhci_host *host, struct mmc_command *cmd)
> >>>       }
> >>>  }
> >>>
> >>> +#if IS_ENABLED(CONFIG_MMC_SDHCI_EXTERNAL_DMA)
> >>> +static int sdhci_external_dma_init(struct sdhci_host *host)
> >>> +{
> >>> +     int ret = 0;
> >>> +     struct mmc_host *mmc = host->mmc;
> >>> +
> >>> +     host->tx_chan = dma_request_chan(mmc->parent, "tx");
> >>> +     if (IS_ERR(host->tx_chan)) {
> >>> +             ret = PTR_ERR(host->tx_chan);
> >>> +             if (ret != -EPROBE_DEFER)
> >>> +                     pr_warn("Failed to request TX DMA channel.\n");
> >>> +             host->tx_chan = NULL;
> >>> +             return ret;
> >>> +     }
> >>> +
> >>> +     host->rx_chan = dma_request_chan(mmc->parent, "rx");
> >>> +     if (IS_ERR(host->rx_chan)) {
> >>> +             if (host->tx_chan) {
> >>> +                     dma_release_channel(host->tx_chan);
> >>> +                     host->tx_chan = NULL;
> >>> +             }
> >>> +
> >>> +             ret = PTR_ERR(host->rx_chan);
> >>> +             if (ret != -EPROBE_DEFER)
> >>> +                     pr_warn("Failed to request RX DMA channel.\n");
> >>> +             host->rx_chan = NULL;
> >>> +     }
> >>> +
> >>> +     return ret;
> >>> +}
> >>> +
> >>> +static inline struct dma_chan *
> >>> +sdhci_external_dma_channel(struct sdhci_host *host, struct mmc_data *data)
> >>> +{
> >>> +     return data->flags & MMC_DATA_WRITE ? host->tx_chan : host->rx_chan;
> >>> +}
> >>> +
> >>> +static int sdhci_external_dma_setup(struct sdhci_host *host,
> >>> +                                 struct mmc_command *cmd)
> >>> +{
> >>> +     int ret, i;
> >>> +     struct dma_async_tx_descriptor *desc;
> >>> +     struct mmc_data *data = cmd->data;
> >>> +     struct dma_chan *chan;
> >>> +     struct dma_slave_config cfg;
> >>> +     dma_cookie_t cookie;
> >>> +     int sg_cnt;
> >>> +
> >>> +     if (!host->mapbase)
> >>> +             return -EINVAL;
> >>> +
> >>> +     cfg.src_addr = host->mapbase + SDHCI_BUFFER;
> >>> +     cfg.dst_addr = host->mapbase + SDHCI_BUFFER;
> >>> +     cfg.src_addr_width = DMA_SLAVE_BUSWIDTH_4_BYTES;
> >>> +     cfg.dst_addr_width = DMA_SLAVE_BUSWIDTH_4_BYTES;
> >>> +     cfg.src_maxburst = data->blksz / 4;
> >>> +     cfg.dst_maxburst = data->blksz / 4;
> >>> +
> >>> +     /* Sanity check: all the SG entries must be aligned by block size. */
> >>> +     for (i = 0; i < data->sg_len; i++) {
> >>> +             if ((data->sg + i)->length % data->blksz)
> >>> +                     return -EINVAL;
> >>> +     }
> >>> +
> >>> +     chan = sdhci_external_dma_channel(host, data);
> >>> +
> >>> +     ret = dmaengine_slave_config(chan, &cfg);
> >>> +     if (ret)
> >>> +             return ret;
> >>> +
> >>> +     sg_cnt = sdhci_pre_dma_transfer(host, data, COOKIE_MAPPED);
> >>> +     if (sg_cnt <= 0)
> >>> +             return -EINVAL;
> >>> +
> >>> +     desc = dmaengine_prep_slave_sg(chan, data->sg, data->sg_len,
> >>> +                                    mmc_get_dma_dir(data),
> >>> +                                    DMA_PREP_INTERRUPT | DMA_CTRL_ACK);
> >>> +     if (!desc)
> >>> +             return -EINVAL;
> >>> +
> >>> +     desc->callback = NULL;
> >>> +     desc->callback_param = NULL;
> >>> +
> >>> +     cookie = dmaengine_submit(desc);
> >>> +     if (cookie < 0)
> >>> +             ret = cookie;
> >>> +
> >>> +     return ret;
> >>> +}
> >>> +
> >>> +static void sdhci_external_dma_release(struct sdhci_host *host)
> >>> +{
> >>> +     if (host->tx_chan) {
> >>> +             dma_release_channel(host->tx_chan);
> >>> +             host->tx_chan = NULL;
> >>> +     }
> >>> +
> >>> +     if (host->rx_chan) {
> >>> +             dma_release_channel(host->rx_chan);
> >>> +             host->rx_chan = NULL;
> >>> +     }
> >>> +
> >>> +     sdhci_switch_external_dma(host, false);
> >>> +}
> >>> +
> >>> +static int __sdhci_external_dma_prepare_data(struct sdhci_host *host,
> >>> +                                          struct mmc_command *cmd)
> >>> +{
> >>> +     struct mmc_data *data = cmd->data;
> >>> +
> >>> +     host->data_timeout = 0;
> >>> +
> >>> +     if (sdhci_data_line_cmd(cmd))
> >>> +             sdhci_set_timeout(host, cmd);
> >>> +
> >>> +     WARN_ON(host->data);
> >>> +
> >>> +     /* Sanity checks */
> >>> +     WARN_ON(data->blksz * data->blocks > 524288);
> >>> +     WARN_ON(data->blksz > host->mmc->max_blk_size);
> >>> +     WARN_ON(data->blocks > 65535);
> >>> +
> >>> +     host->flags |= SDHCI_REQ_USE_DMA;
> >>> +     host->data = data;
> >>> +     host->data_early = 0;
> >>> +     host->data->bytes_xfered = 0;
> >>> +
> >>> +     sdhci_set_transfer_irqs(host);
> >>> +
> >>> +     /*
> >>> +      * For Version 4.10 onwards, if v4 mode is enabled, 32-bit Block Count
> >>> +      * can be supported, in that case 16-bit block count register must be 0.
> >>> +      */
> >>> +     if (host->version >= SDHCI_SPEC_410 && host->v4_mode &&
> >>> +         (host->quirks2 & SDHCI_QUIRK2_USE_32BIT_BLK_CNT)) {
> >>> +             if (sdhci_readw(host, SDHCI_BLOCK_COUNT))
> >>> +                     sdhci_writew(host, 0, SDHCI_BLOCK_COUNT);
> >>> +             sdhci_writew(host, data->blocks, SDHCI_32BIT_BLK_CNT);
> >>> +     } else {
> >>> +             sdhci_writew(host, data->blocks, SDHCI_BLOCK_COUNT);
> >>> +     }
> >>
> >> It is probably worth factoring out the code that is shared with
> >> sdhci_prepare_data() where possible.
> >>
> >>> +
> >>> +     return 0;
> >>> +}
> >>> +
> >>> +static void sdhci_external_dma_prepare_data(struct sdhci_host *host,
> >>> +                                         struct mmc_command *cmd)
> >>> +{
> >>> +     struct mmc_data *data = cmd->data;
> >>> +
> >>> +     if (!data)
> >>> +             return;
> >>
> >> Even in the !data case, we still need to set up a timeout for commands with
> >> busy waiting.  I suggest checking the !data case before calling
> >> sdhci_external_dma_prepare_data()
> >
> > Ok.
> >
> >>
> >>> +
> >>> +     if (sdhci_external_dma_setup(host, cmd) ||
> >>> +         __sdhci_external_dma_prepare_data(host, cmd)) {
> >>> +             sdhci_external_dma_release(host);
> >>> +             pr_err("%s: Cannot use external DMA, switch to the DMA/PIO which standard SDHCI provides.\n",
> >>> +                    mmc_hostname(host->mmc));
> >>> +             sdhci_prepare_data(host, cmd);
> >>> +     }
> >>> +}
> >>> +
> >>> +static void sdhci_external_dma_pre_transfer(struct sdhci_host *host,
> >>> +                                         struct mmc_command *cmd)
> >>> +{
> >>> +     struct dma_chan *chan;
> >>> +
> >>> +     if (!cmd->data || cmd->opcode == MMC_SET_BLOCK_COUNT)
> >>
> >> MMC_SET_BLOCK_COUNT never has cmd->data and so does not need to be checked.
> >
> > Ok.
> >
> >>
> >>> +             return;
> >>> +
> >>> +     sdhci_writew(host, cmd->data->blksz, SDHCI_BLOCK_SIZE);
> >>
> >> Block size is set in __sdhci_external_dma_prepare_data() so does it need to
> >> be set here also.
> >
> > Ok.
>
> That is BLOCK_COUNT. This is BLOCK_SIZE.
>
> >
> >>
> >>> +     chan = sdhci_external_dma_channel(host, cmd->data);
> >>> +     if (chan)
> >>> +             dma_async_issue_pending(chan);
> >>> +}
> >>> +
> >>> +static int sdhci_external_dma_cleanup(struct sdhci_host *host,
> >>> +                                 struct mmc_data *data)
> >>
> >> Please align parameters with open parenthesis
> >>
> >>> +{
> >>> +     struct dma_chan *chan = sdhci_external_dma_channel(host, data);
> >>> +     int ret = 0;
> >>> +
> >>> +     if (chan)
> >>> +             ret = dmaengine_terminate_async(chan);
> >>> +
> >>> +     return ret;
> >>> +}
> >>> +#else
> >>> +static int sdhci_external_dma_init(struct sdhci_host *host)
> >>> +{
> >>> +     return -EOPNOTSUPP;
> >>> +}
> >>> +
> >>> +static void sdhci_external_dma_release(struct sdhci_host *host)
> >>> +{}
> >>> +
> >>> +static void sdhci_external_dma_prepare_data(struct sdhci_host *host,
> >>> +                                         struct mmc_command *cmd)
> >>> +{
> >>> +     /* If MMC_SDHCI_EXTERNAL_DMA not supported, PIO will be used */
> >>> +     sdhci_prepare_data(host, cmd);
> >>> +}
> >>> +
> >>> +static void sdhci_external_dma_pre_transfer(struct sdhci_host *host,
> >>> +                                         struct mmc_command *cmd)
> >>> +{}
> >>> +
> >>> +static int sdhci_external_dma_cleanup(struct sdhci_host *host,
> >>> +                                 struct mmc_data *data)
> >>
> >> Please align parameters with open parenthesis
> >>
> >>> +{
> >>> +     return 0;
> >>> +}
> >>> +#endif
> >>> +
> >>> +void sdhci_switch_external_dma(struct sdhci_host *host, bool en)
> >>> +{
> >>> +     host->use_external_dma = en;
> >>> +}
> >>> +EXPORT_SYMBOL_GPL(sdhci_switch_external_dma);
> >>> +
> >>>  static inline bool sdhci_auto_cmd12(struct sdhci_host *host,
> >>>                                   struct mmc_request *mrq)
> >>>  {
> >>> @@ -1374,7 +1595,10 @@ void sdhci_send_command(struct sdhci_host *host, struct mmc_command *cmd)
> >>>               host->data_cmd = cmd;
> >>>       }
> >>>
> >>> -     sdhci_prepare_data(host, cmd);
> >>> +     if (host->use_external_dma)
> >>
> >> As mentioned above wrt sdhci_external_dma_prepare_data():
> >>
> >>         if (host->use_external_dma && cmd->data)
> >>
> >
> > Maybe we should move checking the !data case out of both
> > _prepare_data, and add it to sdhci_send_command() before calling
> > _prepare_data(), that's saying we can do set up a timeout in
> > sdhci_send_command().
>
> Lets just factor out everything in common as Adrian said. The !data,
> BLOCK_COUNT writes and busy waiting.

Ok, and also BLOCK_SIZE which can be set in _prepare_data().

>
> >
> >>> +             sdhci_external_dma_prepare_data(host, cmd);
> >>> +     else
> >>> +             sdhci_prepare_data(host, cmd);
> >>>
> >>>       sdhci_writel(host, cmd->arg, SDHCI_ARGUMENT);
> >>>
> >>> @@ -1416,6 +1640,9 @@ void sdhci_send_command(struct sdhci_host *host, struct mmc_command *cmd)
> >>>               timeout += 10 * HZ;
> >>>       sdhci_mod_timer(host, cmd->mrq, timeout);
> >>>
> >>> +     if (host->use_external_dma)
> >>> +             sdhci_external_dma_pre_transfer(host, cmd);
> >>
> >> Why is sdhci_external_dma_pre_transfer() needed here - couldn't it be done
> >> in sdhci_external_dma_prepare_data()?
> >>
> >
> > I'm not sure dma_async_issue_pending() can be done so early in
> > sdhci_external_dma_prepare_data().
> >
> >>> +
> >>>       sdhci_writew(host, SDHCI_MAKE_CMD(cmd->opcode, flags), SDHCI_COMMAND);
> >>>  }
> >>>  EXPORT_SYMBOL_GPL(sdhci_send_command);
> >>> @@ -1781,6 +2008,11 @@ void sdhci_request(struct mmc_host *mmc, struct mmc_request *mrq)
> >>>
> >>>       sdhci_led_activate(host);
> >>>
> >>> +     if (host->use_external_dma && mrq->data) {
> >>> +             struct dma_chan *chan = sdhci_external_dma_channel(host,
> >>> +                                                                mrq->data);
> >>
> >> sdhci_external_dma_channel is not declared if
> >> !IS_ENABLED(CONFIG_MMC_SDHCI_EXTERNAL_DMA)
> >
> > Ok, I guess this can be moved to sdhci_external_dma_prepare_data().
>
> No. We have to synchronize from a guaranteed non-atomic context. Maybe
> add another empty sdhci_external_dma_channel definition in the #else part?
>
> >
> >>
> >>> +             dmaengine_synchronize(chan);
> >>
> >> So this is to cover for using dmaengine_terminate_async()?
>
> Yes.
>
> >
> > Ok.
> >
> >>
> >>> +     }
> >>>       /*
> >>>        * Ensure we don't send the STOP for non-SET_BLOCK_COUNTED
> >>>        * requests if Auto-CMD12 is enabled.
> >>> @@ -2658,6 +2890,8 @@ static bool sdhci_request_done(struct sdhci_host *host)
> >>>                               dma_unmap_sg(mmc_dev(host->mmc), data->sg,
> >>>                                            data->sg_len,
> >>>                                            mmc_get_dma_dir(data));
> >>> +                             if (host->use_external_dma)
> >>> +                                     sdhci_external_dma_cleanup(host, data);
> >>
> >> Is sdhci_external_dma_cleanup() only needed in the error case?
> >>
> >> The DMA must be stopped before the memory is unmapped and potentially freed.
> >>
> >> Isn't the DMA cleanup also needed in the bounce buffer case?
> >>
> >> Isn't the DMA cleanup also needed in the COOKIE_PRE_MAPPED case?
> >>
> >> dmaengine_terminate_async() doesn't stop the DMA but
> >> dmaengine_terminate_sync() is not atomic, which looks like a problem.
> >>
> >> Perhaps you look at scheduling some work for the external dma error case
> >> instead of calling __sdhci_finish_mrq()?  Then the work can do the
> >> dmaengine_terminate_sync() and call __sdhci_finish_mrq().
> >
> > Ok, I will look at these issues.
> >
> >>
> >>>                       }
> >>>                       data->host_cookie = COOKIE_UNMAPPED;
> >>>               }
> >>> @@ -3692,12 +3926,15 @@ int sdhci_setup_host(struct sdhci_host *host)
> >>>                      mmc_hostname(mmc), host->version);
> >>>       }
> >>>
> >>> -     if (host->quirks & SDHCI_QUIRK_FORCE_DMA)
> >>> +     if (host->quirks & SDHCI_QUIRK_FORCE_DMA) {
> >>>               host->flags |= SDHCI_USE_SDMA;
> >>> -     else if (!(host->caps & SDHCI_CAN_DO_SDMA))
> >>> +     } else if (!(host->caps & SDHCI_CAN_DO_SDMA)) {
> >>>               DBG("Controller doesn't have SDMA capability\n");
> >>> -     else
> >>> +     } else if (host->use_external_dma) {
> >>> +             /* Using dma-names to detect external dma capability */
> >>> +     } else {
> >>>               host->flags |= SDHCI_USE_SDMA;
> >>> +     }
> >>
> >> These if-statements are about setting SDHCI_USE_SDMA but why is a change
> >> needed for the host->use_external_dma case?
> >
> > Yes, this is not needed, otherwise the controller cannot switch back
> > to SDMA if it is supported in the controller.
>
> The else case means that the host always uses SDMA. Needed to make sure
> this doesn't happen in the external_dma case. Its not SDHCI_CAN_DO_SDMA,
> its SDHCI_USE_SDMA.
>
> >
> >>
> >>>
> >>>       if ((host->quirks & SDHCI_QUIRK_BROKEN_DMA) &&
> >>>               (host->flags & SDHCI_USE_SDMA)) {
> >>> @@ -3785,6 +4022,19 @@ int sdhci_setup_host(struct sdhci_host *host)
> >>>               }
> >>>       }
> >>>
> >>> +     if (host->use_external_dma) {
> >>> +             ret = sdhci_external_dma_init(host);
> >>> +             if (ret == -EPROBE_DEFER)
> >>> +                     goto unreg;
> >>> +
> >>> +             /*
> >>> +              * Fall back to use the DMA/PIO integrated in standard SDHCI
> >>> +              * instead of external DMA devices.
> >>> +              */
> >>> +             if (ret)
> >>> +                     sdhci_switch_external_dma(host, false);
> >>> +     }
> >>> +
> >>>       /*
> >>>        * If we use DMA, then it's up to the caller to set the DMA
> >>>        * mask, but PIO does not need the hw shim so we set a new
> >>> @@ -4201,6 +4451,10 @@ void sdhci_cleanup_host(struct sdhci_host *host)
> >>>               dma_free_coherent(mmc_dev(mmc), host->align_buffer_sz +
> >>>                                 host->adma_table_sz, host->align_buffer,
> >>>                                 host->align_addr);
> >>> +
> >>> +     if (host->use_external_dma)
> >>> +             sdhci_external_dma_release(host);
> >>> +
> >>>       host->adma_table = NULL;
> >>>       host->align_buffer = NULL;
> >>>  }
> >>> @@ -4247,6 +4501,7 @@ int __sdhci_add_host(struct sdhci_host *host)
> >>>
> >>>       pr_info("%s: SDHCI controller on %s [%s] using %s\n",
> >>>               mmc_hostname(mmc), host->hw_name, dev_name(mmc_dev(mmc)),
> >>> +             host->use_external_dma ? "External DMA" :
> >>>               (host->flags & SDHCI_USE_ADMA) ?
> >>>               (host->flags & SDHCI_USE_64_BIT_DMA) ? "ADMA 64-bit" : "ADMA" :
> >>>               (host->flags & SDHCI_USE_SDMA) ? "DMA" : "PIO");
> >>> @@ -4335,6 +4590,9 @@ void sdhci_remove_host(struct sdhci_host *host, int dead)
> >>>                                 host->adma_table_sz, host->align_buffer,
> >>>                                 host->align_addr);
> >>>
> >>> +     if (host->use_external_dma)
> >>> +             sdhci_external_dma_release(host);
> >>> +
> >>>       host->adma_table = NULL;
> >>>       host->align_buffer = NULL;
> >>>  }
> >>> diff --git a/drivers/mmc/host/sdhci.h b/drivers/mmc/host/sdhci.h
> >>> index 6cc9a3c2ac66..7a52823ebef4 100644
> >>> --- a/drivers/mmc/host/sdhci.h
> >>> +++ b/drivers/mmc/host/sdhci.h
> >>> @@ -482,6 +482,7 @@ struct sdhci_host {
> >>>
> >>>       int irq;                /* Device IRQ */
> >>>       void __iomem *ioaddr;   /* Mapped address */
> >>> +     phys_addr_t mapbase;    /* physical address base */
> >>>       char *bounce_buffer;    /* For packing SDMA reads/writes */
> >>>       dma_addr_t bounce_addr;
> >>>       unsigned int bounce_buffer_size;
> >>> @@ -531,6 +532,7 @@ struct sdhci_host {
> >>>       bool pending_reset;     /* Cmd/data reset is pending */
> >>>       bool irq_wake_enabled;  /* IRQ wakeup is enabled */
> >>>       bool v4_mode;           /* Host Version 4 Enable */
> >>> +     bool use_external_dma; /* Host selects to use external DMA */
> >>
> >> Please align /**/ with above i.e. use tab
>
> Will fix.

Do you need me to address these comments and then send you a patch or
you'd like to do instead?

Thanks,
Chunyan

>
> Thanks,
> Faiz
Faiz Abbas Jan. 29, 2019, 11:53 a.m. UTC | #6
Chunyan,

On 28/01/19 5:16 PM, Chunyan Zhang wrote:
> On Mon, 28 Jan 2019 at 18:40, Faiz Abbas <faiz_abbas@ti.com> wrote:
>>
>> Hi,
>>
>> On 28/01/19 3:50 PM, Chunyan Zhang wrote:
>>> On Thu, 24 Jan 2019 at 19:41, Adrian Hunter <adrian.hunter@intel.com> wrote:
>>>>
>>>> On 11/01/19 1:08 PM, Faiz Abbas wrote:
>>>>> From: Chunyan Zhang <zhang.chunyan@linaro.org>
>>>>>
>>>>> Some standard SD host controllers can support both external dma
>>>>> controllers as well as ADMA/SDMA in which the SD host controller
>>>>> acts as DMA master. TI's omap controller is the case as an example.
>>>>>
>>>>> Currently the generic SDHCI code supports ADMA/SDMA integrated in
>>>>> the host controller but does not have any support for external DMA
>>>>> controllers implemented using dmaengine, meaning that custom code is
>>>>> needed for any systems that use an external DMA controller with SDHCI.
>>>>>
>>>>> Fixes by Faiz Abbas <faiz_abbas@ti.com>:
>>>>> 1. Map scatterlists before dmaengine_prep_slave_sg()
>>>>> 2. Use dma_async() functions inside of the send_command() path and
>>>>> synchronize once at the start of each request.
>>>>
>>>> Sorry for the slow reply, but I do have some concerns.  Please see the comments.
>>>>
>>>>>
>>>>> Signed-off-by: Chunyan Zhang <zhang.chunyan@linaro.org>
>>>>> Signed-off-by: Faiz Abbas <faiz_abbas@ti.com>
>>>>> ---
>>>>>  drivers/mmc/host/Kconfig |   3 +
>>>>>  drivers/mmc/host/sdhci.c | 266 ++++++++++++++++++++++++++++++++++++++-
>>>>>  drivers/mmc/host/sdhci.h |   8 ++
>>>>>  3 files changed, 273 insertions(+), 4 deletions(-)
>>>>>
>>>>> diff --git a/drivers/mmc/host/Kconfig b/drivers/mmc/host/Kconfig
>>>>> index e26b8145efb3..333292e8ecdd 100644
>>>>> --- a/drivers/mmc/host/Kconfig
>>>>> +++ b/drivers/mmc/host/Kconfig
>>>>> @@ -999,3 +999,6 @@ config MMC_SDHCI_AM654
>>>>>         If you have a controller with this interface, say Y or M here.
>>>>>
>>>>>         If unsure, say N.
>>>>> +
>>>>> +config MMC_SDHCI_EXTERNAL_DMA
>>>>> +        bool
>>>>> diff --git a/drivers/mmc/host/sdhci.c b/drivers/mmc/host/sdhci.c
>>>>> index a22e11a65658..4a9044c06e21 100644
>>>>> --- a/drivers/mmc/host/sdhci.c
>>>>> +++ b/drivers/mmc/host/sdhci.c
>>>>> @@ -14,6 +14,7 @@
>>>>>   */
>>>>>
>>>>>  #include <linux/delay.h>
>>>>> +#include <linux/dmaengine.h>
>>>>>  #include <linux/ktime.h>
>>>>>  #include <linux/highmem.h>
>>>>>  #include <linux/io.h>
>>>>> @@ -1118,6 +1119,226 @@ static void sdhci_prepare_data(struct sdhci_host *host, struct mmc_command *cmd)
>>>>>       }
>>>>>  }
>>>>>
>>>>> +#if IS_ENABLED(CONFIG_MMC_SDHCI_EXTERNAL_DMA)
>>>>> +static int sdhci_external_dma_init(struct sdhci_host *host)
>>>>> +{
>>>>> +     int ret = 0;
>>>>> +     struct mmc_host *mmc = host->mmc;
>>>>> +
>>>>> +     host->tx_chan = dma_request_chan(mmc->parent, "tx");
>>>>> +     if (IS_ERR(host->tx_chan)) {
>>>>> +             ret = PTR_ERR(host->tx_chan);
>>>>> +             if (ret != -EPROBE_DEFER)
>>>>> +                     pr_warn("Failed to request TX DMA channel.\n");
>>>>> +             host->tx_chan = NULL;
>>>>> +             return ret;
>>>>> +     }
>>>>> +
>>>>> +     host->rx_chan = dma_request_chan(mmc->parent, "rx");
>>>>> +     if (IS_ERR(host->rx_chan)) {
>>>>> +             if (host->tx_chan) {
>>>>> +                     dma_release_channel(host->tx_chan);
>>>>> +                     host->tx_chan = NULL;
>>>>> +             }
>>>>> +
>>>>> +             ret = PTR_ERR(host->rx_chan);
>>>>> +             if (ret != -EPROBE_DEFER)
>>>>> +                     pr_warn("Failed to request RX DMA channel.\n");
>>>>> +             host->rx_chan = NULL;
>>>>> +     }
>>>>> +
>>>>> +     return ret;
>>>>> +}
>>>>> +
>>>>> +static inline struct dma_chan *
>>>>> +sdhci_external_dma_channel(struct sdhci_host *host, struct mmc_data *data)
>>>>> +{
>>>>> +     return data->flags & MMC_DATA_WRITE ? host->tx_chan : host->rx_chan;
>>>>> +}
>>>>> +
>>>>> +static int sdhci_external_dma_setup(struct sdhci_host *host,
>>>>> +                                 struct mmc_command *cmd)
>>>>> +{
>>>>> +     int ret, i;
>>>>> +     struct dma_async_tx_descriptor *desc;
>>>>> +     struct mmc_data *data = cmd->data;
>>>>> +     struct dma_chan *chan;
>>>>> +     struct dma_slave_config cfg;
>>>>> +     dma_cookie_t cookie;
>>>>> +     int sg_cnt;
>>>>> +
>>>>> +     if (!host->mapbase)
>>>>> +             return -EINVAL;
>>>>> +
>>>>> +     cfg.src_addr = host->mapbase + SDHCI_BUFFER;
>>>>> +     cfg.dst_addr = host->mapbase + SDHCI_BUFFER;
>>>>> +     cfg.src_addr_width = DMA_SLAVE_BUSWIDTH_4_BYTES;
>>>>> +     cfg.dst_addr_width = DMA_SLAVE_BUSWIDTH_4_BYTES;
>>>>> +     cfg.src_maxburst = data->blksz / 4;
>>>>> +     cfg.dst_maxburst = data->blksz / 4;
>>>>> +
>>>>> +     /* Sanity check: all the SG entries must be aligned by block size. */
>>>>> +     for (i = 0; i < data->sg_len; i++) {
>>>>> +             if ((data->sg + i)->length % data->blksz)
>>>>> +                     return -EINVAL;
>>>>> +     }
>>>>> +
>>>>> +     chan = sdhci_external_dma_channel(host, data);
>>>>> +
>>>>> +     ret = dmaengine_slave_config(chan, &cfg);
>>>>> +     if (ret)
>>>>> +             return ret;
>>>>> +
>>>>> +     sg_cnt = sdhci_pre_dma_transfer(host, data, COOKIE_MAPPED);
>>>>> +     if (sg_cnt <= 0)
>>>>> +             return -EINVAL;
>>>>> +
>>>>> +     desc = dmaengine_prep_slave_sg(chan, data->sg, data->sg_len,
>>>>> +                                    mmc_get_dma_dir(data),
>>>>> +                                    DMA_PREP_INTERRUPT | DMA_CTRL_ACK);
>>>>> +     if (!desc)
>>>>> +             return -EINVAL;
>>>>> +
>>>>> +     desc->callback = NULL;
>>>>> +     desc->callback_param = NULL;
>>>>> +
>>>>> +     cookie = dmaengine_submit(desc);
>>>>> +     if (cookie < 0)
>>>>> +             ret = cookie;
>>>>> +
>>>>> +     return ret;
>>>>> +}
>>>>> +
>>>>> +static void sdhci_external_dma_release(struct sdhci_host *host)
>>>>> +{
>>>>> +     if (host->tx_chan) {
>>>>> +             dma_release_channel(host->tx_chan);
>>>>> +             host->tx_chan = NULL;
>>>>> +     }
>>>>> +
>>>>> +     if (host->rx_chan) {
>>>>> +             dma_release_channel(host->rx_chan);
>>>>> +             host->rx_chan = NULL;
>>>>> +     }
>>>>> +
>>>>> +     sdhci_switch_external_dma(host, false);
>>>>> +}
>>>>> +
>>>>> +static int __sdhci_external_dma_prepare_data(struct sdhci_host *host,
>>>>> +                                          struct mmc_command *cmd)
>>>>> +{
>>>>> +     struct mmc_data *data = cmd->data;
>>>>> +
>>>>> +     host->data_timeout = 0;
>>>>> +
>>>>> +     if (sdhci_data_line_cmd(cmd))
>>>>> +             sdhci_set_timeout(host, cmd);
>>>>> +
>>>>> +     WARN_ON(host->data);
>>>>> +
>>>>> +     /* Sanity checks */
>>>>> +     WARN_ON(data->blksz * data->blocks > 524288);
>>>>> +     WARN_ON(data->blksz > host->mmc->max_blk_size);
>>>>> +     WARN_ON(data->blocks > 65535);
>>>>> +
>>>>> +     host->flags |= SDHCI_REQ_USE_DMA;
>>>>> +     host->data = data;
>>>>> +     host->data_early = 0;
>>>>> +     host->data->bytes_xfered = 0;
>>>>> +
>>>>> +     sdhci_set_transfer_irqs(host);
>>>>> +
>>>>> +     /*
>>>>> +      * For Version 4.10 onwards, if v4 mode is enabled, 32-bit Block Count
>>>>> +      * can be supported, in that case 16-bit block count register must be 0.
>>>>> +      */
>>>>> +     if (host->version >= SDHCI_SPEC_410 && host->v4_mode &&
>>>>> +         (host->quirks2 & SDHCI_QUIRK2_USE_32BIT_BLK_CNT)) {
>>>>> +             if (sdhci_readw(host, SDHCI_BLOCK_COUNT))
>>>>> +                     sdhci_writew(host, 0, SDHCI_BLOCK_COUNT);
>>>>> +             sdhci_writew(host, data->blocks, SDHCI_32BIT_BLK_CNT);
>>>>> +     } else {
>>>>> +             sdhci_writew(host, data->blocks, SDHCI_BLOCK_COUNT);
>>>>> +     }
>>>>
>>>> It is probably worth factoring out the code that is shared with
>>>> sdhci_prepare_data() where possible.
>>>>
>>>>> +
>>>>> +     return 0;
>>>>> +}
>>>>> +
>>>>> +static void sdhci_external_dma_prepare_data(struct sdhci_host *host,
>>>>> +                                         struct mmc_command *cmd)
>>>>> +{
>>>>> +     struct mmc_data *data = cmd->data;
>>>>> +
>>>>> +     if (!data)
>>>>> +             return;
>>>>
>>>> Even in the !data case, we still need to set up a timeout for commands with
>>>> busy waiting.  I suggest checking the !data case before calling
>>>> sdhci_external_dma_prepare_data()
>>>
>>> Ok.
>>>
>>>>
>>>>> +
>>>>> +     if (sdhci_external_dma_setup(host, cmd) ||
>>>>> +         __sdhci_external_dma_prepare_data(host, cmd)) {
>>>>> +             sdhci_external_dma_release(host);
>>>>> +             pr_err("%s: Cannot use external DMA, switch to the DMA/PIO which standard SDHCI provides.\n",
>>>>> +                    mmc_hostname(host->mmc));
>>>>> +             sdhci_prepare_data(host, cmd);
>>>>> +     }
>>>>> +}
>>>>> +
>>>>> +static void sdhci_external_dma_pre_transfer(struct sdhci_host *host,
>>>>> +                                         struct mmc_command *cmd)
>>>>> +{
>>>>> +     struct dma_chan *chan;
>>>>> +
>>>>> +     if (!cmd->data || cmd->opcode == MMC_SET_BLOCK_COUNT)
>>>>
>>>> MMC_SET_BLOCK_COUNT never has cmd->data and so does not need to be checked.
>>>
>>> Ok.
>>>
>>>>
>>>>> +             return;
>>>>> +
>>>>> +     sdhci_writew(host, cmd->data->blksz, SDHCI_BLOCK_SIZE);
>>>>
>>>> Block size is set in __sdhci_external_dma_prepare_data() so does it need to
>>>> be set here also.
>>>
>>> Ok.
>>
>> That is BLOCK_COUNT. This is BLOCK_SIZE.
>>
>>>
>>>>
>>>>> +     chan = sdhci_external_dma_channel(host, cmd->data);
>>>>> +     if (chan)
>>>>> +             dma_async_issue_pending(chan);
>>>>> +}
>>>>> +
>>>>> +static int sdhci_external_dma_cleanup(struct sdhci_host *host,
>>>>> +                                 struct mmc_data *data)
>>>>
>>>> Please align parameters with open parenthesis
>>>>
>>>>> +{
>>>>> +     struct dma_chan *chan = sdhci_external_dma_channel(host, data);
>>>>> +     int ret = 0;
>>>>> +
>>>>> +     if (chan)
>>>>> +             ret = dmaengine_terminate_async(chan);
>>>>> +
>>>>> +     return ret;
>>>>> +}
>>>>> +#else
>>>>> +static int sdhci_external_dma_init(struct sdhci_host *host)
>>>>> +{
>>>>> +     return -EOPNOTSUPP;
>>>>> +}
>>>>> +
>>>>> +static void sdhci_external_dma_release(struct sdhci_host *host)
>>>>> +{}
>>>>> +
>>>>> +static void sdhci_external_dma_prepare_data(struct sdhci_host *host,
>>>>> +                                         struct mmc_command *cmd)
>>>>> +{
>>>>> +     /* If MMC_SDHCI_EXTERNAL_DMA not supported, PIO will be used */
>>>>> +     sdhci_prepare_data(host, cmd);
>>>>> +}
>>>>> +
>>>>> +static void sdhci_external_dma_pre_transfer(struct sdhci_host *host,
>>>>> +                                         struct mmc_command *cmd)
>>>>> +{}
>>>>> +
>>>>> +static int sdhci_external_dma_cleanup(struct sdhci_host *host,
>>>>> +                                 struct mmc_data *data)
>>>>
>>>> Please align parameters with open parenthesis
>>>>
>>>>> +{
>>>>> +     return 0;
>>>>> +}
>>>>> +#endif
>>>>> +
>>>>> +void sdhci_switch_external_dma(struct sdhci_host *host, bool en)
>>>>> +{
>>>>> +     host->use_external_dma = en;
>>>>> +}
>>>>> +EXPORT_SYMBOL_GPL(sdhci_switch_external_dma);
>>>>> +
>>>>>  static inline bool sdhci_auto_cmd12(struct sdhci_host *host,
>>>>>                                   struct mmc_request *mrq)
>>>>>  {
>>>>> @@ -1374,7 +1595,10 @@ void sdhci_send_command(struct sdhci_host *host, struct mmc_command *cmd)
>>>>>               host->data_cmd = cmd;
>>>>>       }
>>>>>
>>>>> -     sdhci_prepare_data(host, cmd);
>>>>> +     if (host->use_external_dma)
>>>>
>>>> As mentioned above wrt sdhci_external_dma_prepare_data():
>>>>
>>>>         if (host->use_external_dma && cmd->data)
>>>>
>>>
>>> Maybe we should move checking the !data case out of both
>>> _prepare_data, and add it to sdhci_send_command() before calling
>>> _prepare_data(), that's saying we can do set up a timeout in
>>> sdhci_send_command().
>>
>> Lets just factor out everything in common as Adrian said. The !data,
>> BLOCK_COUNT writes and busy waiting.
> 
> Ok, and also BLOCK_SIZE which can be set in _prepare_data().
> 
>>
>>>
>>>>> +             sdhci_external_dma_prepare_data(host, cmd);
>>>>> +     else
>>>>> +             sdhci_prepare_data(host, cmd);
>>>>>
>>>>>       sdhci_writel(host, cmd->arg, SDHCI_ARGUMENT);
>>>>>
>>>>> @@ -1416,6 +1640,9 @@ void sdhci_send_command(struct sdhci_host *host, struct mmc_command *cmd)
>>>>>               timeout += 10 * HZ;
>>>>>       sdhci_mod_timer(host, cmd->mrq, timeout);
>>>>>
>>>>> +     if (host->use_external_dma)
>>>>> +             sdhci_external_dma_pre_transfer(host, cmd);
>>>>
>>>> Why is sdhci_external_dma_pre_transfer() needed here - couldn't it be done
>>>> in sdhci_external_dma_prepare_data()?
>>>>
>>>
>>> I'm not sure dma_async_issue_pending() can be done so early in
>>> sdhci_external_dma_prepare_data().
>>>
>>>>> +
>>>>>       sdhci_writew(host, SDHCI_MAKE_CMD(cmd->opcode, flags), SDHCI_COMMAND);
>>>>>  }
>>>>>  EXPORT_SYMBOL_GPL(sdhci_send_command);
>>>>> @@ -1781,6 +2008,11 @@ void sdhci_request(struct mmc_host *mmc, struct mmc_request *mrq)
>>>>>
>>>>>       sdhci_led_activate(host);
>>>>>
>>>>> +     if (host->use_external_dma && mrq->data) {
>>>>> +             struct dma_chan *chan = sdhci_external_dma_channel(host,
>>>>> +                                                                mrq->data);
>>>>
>>>> sdhci_external_dma_channel is not declared if
>>>> !IS_ENABLED(CONFIG_MMC_SDHCI_EXTERNAL_DMA)
>>>
>>> Ok, I guess this can be moved to sdhci_external_dma_prepare_data().
>>
>> No. We have to synchronize from a guaranteed non-atomic context. Maybe
>> add another empty sdhci_external_dma_channel definition in the #else part?
>>
>>>
>>>>
>>>>> +             dmaengine_synchronize(chan);
>>>>
>>>> So this is to cover for using dmaengine_terminate_async()?
>>
>> Yes.
>>
>>>
>>> Ok.
>>>
>>>>
>>>>> +     }
>>>>>       /*
>>>>>        * Ensure we don't send the STOP for non-SET_BLOCK_COUNTED
>>>>>        * requests if Auto-CMD12 is enabled.
>>>>> @@ -2658,6 +2890,8 @@ static bool sdhci_request_done(struct sdhci_host *host)
>>>>>                               dma_unmap_sg(mmc_dev(host->mmc), data->sg,
>>>>>                                            data->sg_len,
>>>>>                                            mmc_get_dma_dir(data));
>>>>> +                             if (host->use_external_dma)
>>>>> +                                     sdhci_external_dma_cleanup(host, data);
>>>>
>>>> Is sdhci_external_dma_cleanup() only needed in the error case?
>>>>
>>>> The DMA must be stopped before the memory is unmapped and potentially freed.
>>>>
>>>> Isn't the DMA cleanup also needed in the bounce buffer case?
>>>>
>>>> Isn't the DMA cleanup also needed in the COOKIE_PRE_MAPPED case?
>>>>
>>>> dmaengine_terminate_async() doesn't stop the DMA but
>>>> dmaengine_terminate_sync() is not atomic, which looks like a problem.
>>>>
>>>> Perhaps you look at scheduling some work for the external dma error case
>>>> instead of calling __sdhci_finish_mrq()?  Then the work can do the
>>>> dmaengine_terminate_sync() and call __sdhci_finish_mrq().
>>>
>>> Ok, I will look at these issues.
>>>
>>>>
>>>>>                       }
>>>>>                       data->host_cookie = COOKIE_UNMAPPED;
>>>>>               }
>>>>> @@ -3692,12 +3926,15 @@ int sdhci_setup_host(struct sdhci_host *host)
>>>>>                      mmc_hostname(mmc), host->version);
>>>>>       }
>>>>>
>>>>> -     if (host->quirks & SDHCI_QUIRK_FORCE_DMA)
>>>>> +     if (host->quirks & SDHCI_QUIRK_FORCE_DMA) {
>>>>>               host->flags |= SDHCI_USE_SDMA;
>>>>> -     else if (!(host->caps & SDHCI_CAN_DO_SDMA))
>>>>> +     } else if (!(host->caps & SDHCI_CAN_DO_SDMA)) {
>>>>>               DBG("Controller doesn't have SDMA capability\n");
>>>>> -     else
>>>>> +     } else if (host->use_external_dma) {
>>>>> +             /* Using dma-names to detect external dma capability */
>>>>> +     } else {
>>>>>               host->flags |= SDHCI_USE_SDMA;
>>>>> +     }
>>>>
>>>> These if-statements are about setting SDHCI_USE_SDMA but why is a change
>>>> needed for the host->use_external_dma case?
>>>
>>> Yes, this is not needed, otherwise the controller cannot switch back
>>> to SDMA if it is supported in the controller.
>>
>> The else case means that the host always uses SDMA. Needed to make sure
>> this doesn't happen in the external_dma case. Its not SDHCI_CAN_DO_SDMA,
>> its SDHCI_USE_SDMA.
>>
>>>
>>>>
>>>>>
>>>>>       if ((host->quirks & SDHCI_QUIRK_BROKEN_DMA) &&
>>>>>               (host->flags & SDHCI_USE_SDMA)) {
>>>>> @@ -3785,6 +4022,19 @@ int sdhci_setup_host(struct sdhci_host *host)
>>>>>               }
>>>>>       }
>>>>>
>>>>> +     if (host->use_external_dma) {
>>>>> +             ret = sdhci_external_dma_init(host);
>>>>> +             if (ret == -EPROBE_DEFER)
>>>>> +                     goto unreg;
>>>>> +
>>>>> +             /*
>>>>> +              * Fall back to use the DMA/PIO integrated in standard SDHCI
>>>>> +              * instead of external DMA devices.
>>>>> +              */
>>>>> +             if (ret)
>>>>> +                     sdhci_switch_external_dma(host, false);
>>>>> +     }
>>>>> +
>>>>>       /*
>>>>>        * If we use DMA, then it's up to the caller to set the DMA
>>>>>        * mask, but PIO does not need the hw shim so we set a new
>>>>> @@ -4201,6 +4451,10 @@ void sdhci_cleanup_host(struct sdhci_host *host)
>>>>>               dma_free_coherent(mmc_dev(mmc), host->align_buffer_sz +
>>>>>                                 host->adma_table_sz, host->align_buffer,
>>>>>                                 host->align_addr);
>>>>> +
>>>>> +     if (host->use_external_dma)
>>>>> +             sdhci_external_dma_release(host);
>>>>> +
>>>>>       host->adma_table = NULL;
>>>>>       host->align_buffer = NULL;
>>>>>  }
>>>>> @@ -4247,6 +4501,7 @@ int __sdhci_add_host(struct sdhci_host *host)
>>>>>
>>>>>       pr_info("%s: SDHCI controller on %s [%s] using %s\n",
>>>>>               mmc_hostname(mmc), host->hw_name, dev_name(mmc_dev(mmc)),
>>>>> +             host->use_external_dma ? "External DMA" :
>>>>>               (host->flags & SDHCI_USE_ADMA) ?
>>>>>               (host->flags & SDHCI_USE_64_BIT_DMA) ? "ADMA 64-bit" : "ADMA" :
>>>>>               (host->flags & SDHCI_USE_SDMA) ? "DMA" : "PIO");
>>>>> @@ -4335,6 +4590,9 @@ void sdhci_remove_host(struct sdhci_host *host, int dead)
>>>>>                                 host->adma_table_sz, host->align_buffer,
>>>>>                                 host->align_addr);
>>>>>
>>>>> +     if (host->use_external_dma)
>>>>> +             sdhci_external_dma_release(host);
>>>>> +
>>>>>       host->adma_table = NULL;
>>>>>       host->align_buffer = NULL;
>>>>>  }
>>>>> diff --git a/drivers/mmc/host/sdhci.h b/drivers/mmc/host/sdhci.h
>>>>> index 6cc9a3c2ac66..7a52823ebef4 100644
>>>>> --- a/drivers/mmc/host/sdhci.h
>>>>> +++ b/drivers/mmc/host/sdhci.h
>>>>> @@ -482,6 +482,7 @@ struct sdhci_host {
>>>>>
>>>>>       int irq;                /* Device IRQ */
>>>>>       void __iomem *ioaddr;   /* Mapped address */
>>>>> +     phys_addr_t mapbase;    /* physical address base */
>>>>>       char *bounce_buffer;    /* For packing SDMA reads/writes */
>>>>>       dma_addr_t bounce_addr;
>>>>>       unsigned int bounce_buffer_size;
>>>>> @@ -531,6 +532,7 @@ struct sdhci_host {
>>>>>       bool pending_reset;     /* Cmd/data reset is pending */
>>>>>       bool irq_wake_enabled;  /* IRQ wakeup is enabled */
>>>>>       bool v4_mode;           /* Host Version 4 Enable */
>>>>> +     bool use_external_dma; /* Host selects to use external DMA */
>>>>
>>>> Please align /**/ with above i.e. use tab
>>
>> Will fix.
> 
> Do you need me to address these comments and then send you a patch or
> you'd like to do instead?
> 

This would be much appreciated. Would you be able to work on it this
week? I would like to send a v2 by Friday.

Thanks,
Faiz
Chunyan Zhang Jan. 30, 2019, 5:01 a.m. UTC | #7
Hi Faiz,

On Tue, 29 Jan 2019 at 19:50, Faiz Abbas <faiz_abbas@ti.com> wrote:
>
> Chunyan,
>
> On 28/01/19 5:16 PM, Chunyan Zhang wrote:
> > On Mon, 28 Jan 2019 at 18:40, Faiz Abbas <faiz_abbas@ti.com> wrote:
> >>
> >> Hi,
> >>
> >> On 28/01/19 3:50 PM, Chunyan Zhang wrote:
> >>> On Thu, 24 Jan 2019 at 19:41, Adrian Hunter <adrian.hunter@intel.com> wrote:
> >>>>
> >>>> On 11/01/19 1:08 PM, Faiz Abbas wrote:
> >>>>> From: Chunyan Zhang <zhang.chunyan@linaro.org>
> >>>>>
> >>>>> Some standard SD host controllers can support both external dma
> >>>>> controllers as well as ADMA/SDMA in which the SD host controller
> >>>>> acts as DMA master. TI's omap controller is the case as an example.
> >>>>>
> >>>>> Currently the generic SDHCI code supports ADMA/SDMA integrated in
> >>>>> the host controller but does not have any support for external DMA
> >>>>> controllers implemented using dmaengine, meaning that custom code is
> >>>>> needed for any systems that use an external DMA controller with SDHCI.
> >>>>>
> >>>>> Fixes by Faiz Abbas <faiz_abbas@ti.com>:
> >>>>> 1. Map scatterlists before dmaengine_prep_slave_sg()
> >>>>> 2. Use dma_async() functions inside of the send_command() path and
> >>>>> synchronize once at the start of each request.
> >>>>
> >>>> Sorry for the slow reply, but I do have some concerns.  Please see the comments.
> >>>>
> >>>>>
> >>>>> Signed-off-by: Chunyan Zhang <zhang.chunyan@linaro.org>
> >>>>> Signed-off-by: Faiz Abbas <faiz_abbas@ti.com>
> >>>>> ---
> >>>>>  drivers/mmc/host/Kconfig |   3 +
> >>>>>  drivers/mmc/host/sdhci.c | 266 ++++++++++++++++++++++++++++++++++++++-
> >>>>>  drivers/mmc/host/sdhci.h |   8 ++
> >>>>>  3 files changed, 273 insertions(+), 4 deletions(-)
> >>>>>
> >>>>> diff --git a/drivers/mmc/host/Kconfig b/drivers/mmc/host/Kconfig
> >>>>> index e26b8145efb3..333292e8ecdd 100644
> >>>>> --- a/drivers/mmc/host/Kconfig
> >>>>> +++ b/drivers/mmc/host/Kconfig
> >>>>> @@ -999,3 +999,6 @@ config MMC_SDHCI_AM654
> >>>>>         If you have a controller with this interface, say Y or M here.
> >>>>>
> >>>>>         If unsure, say N.
> >>>>> +
> >>>>> +config MMC_SDHCI_EXTERNAL_DMA
> >>>>> +        bool
> >>>>> diff --git a/drivers/mmc/host/sdhci.c b/drivers/mmc/host/sdhci.c
> >>>>> index a22e11a65658..4a9044c06e21 100644
> >>>>> --- a/drivers/mmc/host/sdhci.c
> >>>>> +++ b/drivers/mmc/host/sdhci.c
> >>>>> @@ -14,6 +14,7 @@
> >>>>>   */
> >>>>>
> >>>>>  #include <linux/delay.h>
> >>>>> +#include <linux/dmaengine.h>
> >>>>>  #include <linux/ktime.h>
> >>>>>  #include <linux/highmem.h>
> >>>>>  #include <linux/io.h>
> >>>>> @@ -1118,6 +1119,226 @@ static void sdhci_prepare_data(struct sdhci_host *host, struct mmc_command *cmd)
> >>>>>       }
> >>>>>  }
> >>>>>
> >>>>> +#if IS_ENABLED(CONFIG_MMC_SDHCI_EXTERNAL_DMA)
> >>>>> +static int sdhci_external_dma_init(struct sdhci_host *host)
> >>>>> +{
> >>>>> +     int ret = 0;
> >>>>> +     struct mmc_host *mmc = host->mmc;
> >>>>> +
> >>>>> +     host->tx_chan = dma_request_chan(mmc->parent, "tx");
> >>>>> +     if (IS_ERR(host->tx_chan)) {
> >>>>> +             ret = PTR_ERR(host->tx_chan);
> >>>>> +             if (ret != -EPROBE_DEFER)
> >>>>> +                     pr_warn("Failed to request TX DMA channel.\n");
> >>>>> +             host->tx_chan = NULL;
> >>>>> +             return ret;
> >>>>> +     }
> >>>>> +
> >>>>> +     host->rx_chan = dma_request_chan(mmc->parent, "rx");
> >>>>> +     if (IS_ERR(host->rx_chan)) {
> >>>>> +             if (host->tx_chan) {
> >>>>> +                     dma_release_channel(host->tx_chan);
> >>>>> +                     host->tx_chan = NULL;
> >>>>> +             }
> >>>>> +
> >>>>> +             ret = PTR_ERR(host->rx_chan);
> >>>>> +             if (ret != -EPROBE_DEFER)
> >>>>> +                     pr_warn("Failed to request RX DMA channel.\n");
> >>>>> +             host->rx_chan = NULL;
> >>>>> +     }
> >>>>> +
> >>>>> +     return ret;
> >>>>> +}
> >>>>> +
> >>>>> +static inline struct dma_chan *
> >>>>> +sdhci_external_dma_channel(struct sdhci_host *host, struct mmc_data *data)
> >>>>> +{
> >>>>> +     return data->flags & MMC_DATA_WRITE ? host->tx_chan : host->rx_chan;
> >>>>> +}
> >>>>> +
> >>>>> +static int sdhci_external_dma_setup(struct sdhci_host *host,
> >>>>> +                                 struct mmc_command *cmd)
> >>>>> +{
> >>>>> +     int ret, i;
> >>>>> +     struct dma_async_tx_descriptor *desc;
> >>>>> +     struct mmc_data *data = cmd->data;
> >>>>> +     struct dma_chan *chan;
> >>>>> +     struct dma_slave_config cfg;
> >>>>> +     dma_cookie_t cookie;
> >>>>> +     int sg_cnt;
> >>>>> +
> >>>>> +     if (!host->mapbase)
> >>>>> +             return -EINVAL;
> >>>>> +
> >>>>> +     cfg.src_addr = host->mapbase + SDHCI_BUFFER;
> >>>>> +     cfg.dst_addr = host->mapbase + SDHCI_BUFFER;
> >>>>> +     cfg.src_addr_width = DMA_SLAVE_BUSWIDTH_4_BYTES;
> >>>>> +     cfg.dst_addr_width = DMA_SLAVE_BUSWIDTH_4_BYTES;
> >>>>> +     cfg.src_maxburst = data->blksz / 4;
> >>>>> +     cfg.dst_maxburst = data->blksz / 4;
> >>>>> +
> >>>>> +     /* Sanity check: all the SG entries must be aligned by block size. */
> >>>>> +     for (i = 0; i < data->sg_len; i++) {
> >>>>> +             if ((data->sg + i)->length % data->blksz)
> >>>>> +                     return -EINVAL;
> >>>>> +     }
> >>>>> +
> >>>>> +     chan = sdhci_external_dma_channel(host, data);
> >>>>> +
> >>>>> +     ret = dmaengine_slave_config(chan, &cfg);
> >>>>> +     if (ret)
> >>>>> +             return ret;
> >>>>> +
> >>>>> +     sg_cnt = sdhci_pre_dma_transfer(host, data, COOKIE_MAPPED);
> >>>>> +     if (sg_cnt <= 0)
> >>>>> +             return -EINVAL;
> >>>>> +
> >>>>> +     desc = dmaengine_prep_slave_sg(chan, data->sg, data->sg_len,
> >>>>> +                                    mmc_get_dma_dir(data),
> >>>>> +                                    DMA_PREP_INTERRUPT | DMA_CTRL_ACK);
> >>>>> +     if (!desc)
> >>>>> +             return -EINVAL;
> >>>>> +
> >>>>> +     desc->callback = NULL;
> >>>>> +     desc->callback_param = NULL;
> >>>>> +
> >>>>> +     cookie = dmaengine_submit(desc);
> >>>>> +     if (cookie < 0)
> >>>>> +             ret = cookie;
> >>>>> +
> >>>>> +     return ret;
> >>>>> +}
> >>>>> +
> >>>>> +static void sdhci_external_dma_release(struct sdhci_host *host)
> >>>>> +{
> >>>>> +     if (host->tx_chan) {
> >>>>> +             dma_release_channel(host->tx_chan);
> >>>>> +             host->tx_chan = NULL;
> >>>>> +     }
> >>>>> +
> >>>>> +     if (host->rx_chan) {
> >>>>> +             dma_release_channel(host->rx_chan);
> >>>>> +             host->rx_chan = NULL;
> >>>>> +     }
> >>>>> +
> >>>>> +     sdhci_switch_external_dma(host, false);
> >>>>> +}
> >>>>> +
> >>>>> +static int __sdhci_external_dma_prepare_data(struct sdhci_host *host,
> >>>>> +                                          struct mmc_command *cmd)
> >>>>> +{
> >>>>> +     struct mmc_data *data = cmd->data;
> >>>>> +
> >>>>> +     host->data_timeout = 0;
> >>>>> +
> >>>>> +     if (sdhci_data_line_cmd(cmd))
> >>>>> +             sdhci_set_timeout(host, cmd);
> >>>>> +
> >>>>> +     WARN_ON(host->data);
> >>>>> +
> >>>>> +     /* Sanity checks */
> >>>>> +     WARN_ON(data->blksz * data->blocks > 524288);
> >>>>> +     WARN_ON(data->blksz > host->mmc->max_blk_size);
> >>>>> +     WARN_ON(data->blocks > 65535);
> >>>>> +
> >>>>> +     host->flags |= SDHCI_REQ_USE_DMA;
> >>>>> +     host->data = data;
> >>>>> +     host->data_early = 0;
> >>>>> +     host->data->bytes_xfered = 0;
> >>>>> +
> >>>>> +     sdhci_set_transfer_irqs(host);
> >>>>> +
> >>>>> +     /*
> >>>>> +      * For Version 4.10 onwards, if v4 mode is enabled, 32-bit Block Count
> >>>>> +      * can be supported, in that case 16-bit block count register must be 0.
> >>>>> +      */
> >>>>> +     if (host->version >= SDHCI_SPEC_410 && host->v4_mode &&
> >>>>> +         (host->quirks2 & SDHCI_QUIRK2_USE_32BIT_BLK_CNT)) {
> >>>>> +             if (sdhci_readw(host, SDHCI_BLOCK_COUNT))
> >>>>> +                     sdhci_writew(host, 0, SDHCI_BLOCK_COUNT);
> >>>>> +             sdhci_writew(host, data->blocks, SDHCI_32BIT_BLK_CNT);
> >>>>> +     } else {
> >>>>> +             sdhci_writew(host, data->blocks, SDHCI_BLOCK_COUNT);
> >>>>> +     }
> >>>>
> >>>> It is probably worth factoring out the code that is shared with
> >>>> sdhci_prepare_data() where possible.
> >>>>
> >>>>> +
> >>>>> +     return 0;
> >>>>> +}
> >>>>> +
> >>>>> +static void sdhci_external_dma_prepare_data(struct sdhci_host *host,
> >>>>> +                                         struct mmc_command *cmd)
> >>>>> +{
> >>>>> +     struct mmc_data *data = cmd->data;
> >>>>> +
> >>>>> +     if (!data)
> >>>>> +             return;
> >>>>
> >>>> Even in the !data case, we still need to set up a timeout for commands with
> >>>> busy waiting.  I suggest checking the !data case before calling
> >>>> sdhci_external_dma_prepare_data()
> >>>
> >>> Ok.
> >>>
> >>>>
> >>>>> +
> >>>>> +     if (sdhci_external_dma_setup(host, cmd) ||
> >>>>> +         __sdhci_external_dma_prepare_data(host, cmd)) {
> >>>>> +             sdhci_external_dma_release(host);
> >>>>> +             pr_err("%s: Cannot use external DMA, switch to the DMA/PIO which standard SDHCI provides.\n",
> >>>>> +                    mmc_hostname(host->mmc));
> >>>>> +             sdhci_prepare_data(host, cmd);
> >>>>> +     }
> >>>>> +}
> >>>>> +
> >>>>> +static void sdhci_external_dma_pre_transfer(struct sdhci_host *host,
> >>>>> +                                         struct mmc_command *cmd)
> >>>>> +{
> >>>>> +     struct dma_chan *chan;
> >>>>> +
> >>>>> +     if (!cmd->data || cmd->opcode == MMC_SET_BLOCK_COUNT)
> >>>>
> >>>> MMC_SET_BLOCK_COUNT never has cmd->data and so does not need to be checked.
> >>>
> >>> Ok.
> >>>
> >>>>
> >>>>> +             return;
> >>>>> +
> >>>>> +     sdhci_writew(host, cmd->data->blksz, SDHCI_BLOCK_SIZE);
> >>>>
> >>>> Block size is set in __sdhci_external_dma_prepare_data() so does it need to
> >>>> be set here also.
> >>>
> >>> Ok.
> >>
> >> That is BLOCK_COUNT. This is BLOCK_SIZE.
> >>
> >>>
> >>>>
> >>>>> +     chan = sdhci_external_dma_channel(host, cmd->data);
> >>>>> +     if (chan)
> >>>>> +             dma_async_issue_pending(chan);
> >>>>> +}
> >>>>> +
> >>>>> +static int sdhci_external_dma_cleanup(struct sdhci_host *host,
> >>>>> +                                 struct mmc_data *data)
> >>>>
> >>>> Please align parameters with open parenthesis
> >>>>
> >>>>> +{
> >>>>> +     struct dma_chan *chan = sdhci_external_dma_channel(host, data);
> >>>>> +     int ret = 0;
> >>>>> +
> >>>>> +     if (chan)
> >>>>> +             ret = dmaengine_terminate_async(chan);
> >>>>> +
> >>>>> +     return ret;
> >>>>> +}
> >>>>> +#else
> >>>>> +static int sdhci_external_dma_init(struct sdhci_host *host)
> >>>>> +{
> >>>>> +     return -EOPNOTSUPP;
> >>>>> +}
> >>>>> +
> >>>>> +static void sdhci_external_dma_release(struct sdhci_host *host)
> >>>>> +{}
> >>>>> +
> >>>>> +static void sdhci_external_dma_prepare_data(struct sdhci_host *host,
> >>>>> +                                         struct mmc_command *cmd)
> >>>>> +{
> >>>>> +     /* If MMC_SDHCI_EXTERNAL_DMA not supported, PIO will be used */
> >>>>> +     sdhci_prepare_data(host, cmd);
> >>>>> +}
> >>>>> +
> >>>>> +static void sdhci_external_dma_pre_transfer(struct sdhci_host *host,
> >>>>> +                                         struct mmc_command *cmd)
> >>>>> +{}
> >>>>> +
> >>>>> +static int sdhci_external_dma_cleanup(struct sdhci_host *host,
> >>>>> +                                 struct mmc_data *data)
> >>>>
> >>>> Please align parameters with open parenthesis
> >>>>
> >>>>> +{
> >>>>> +     return 0;
> >>>>> +}
> >>>>> +#endif
> >>>>> +
> >>>>> +void sdhci_switch_external_dma(struct sdhci_host *host, bool en)
> >>>>> +{
> >>>>> +     host->use_external_dma = en;
> >>>>> +}
> >>>>> +EXPORT_SYMBOL_GPL(sdhci_switch_external_dma);
> >>>>> +
> >>>>>  static inline bool sdhci_auto_cmd12(struct sdhci_host *host,
> >>>>>                                   struct mmc_request *mrq)
> >>>>>  {
> >>>>> @@ -1374,7 +1595,10 @@ void sdhci_send_command(struct sdhci_host *host, struct mmc_command *cmd)
> >>>>>               host->data_cmd = cmd;
> >>>>>       }
> >>>>>
> >>>>> -     sdhci_prepare_data(host, cmd);
> >>>>> +     if (host->use_external_dma)
> >>>>
> >>>> As mentioned above wrt sdhci_external_dma_prepare_data():
> >>>>
> >>>>         if (host->use_external_dma && cmd->data)
> >>>>
> >>>
> >>> Maybe we should move checking the !data case out of both
> >>> _prepare_data, and add it to sdhci_send_command() before calling
> >>> _prepare_data(), that's saying we can do set up a timeout in
> >>> sdhci_send_command().
> >>
> >> Lets just factor out everything in common as Adrian said. The !data,
> >> BLOCK_COUNT writes and busy waiting.
> >
> > Ok, and also BLOCK_SIZE which can be set in _prepare_data().
> >
> >>
> >>>
> >>>>> +             sdhci_external_dma_prepare_data(host, cmd);
> >>>>> +     else
> >>>>> +             sdhci_prepare_data(host, cmd);
> >>>>>
> >>>>>       sdhci_writel(host, cmd->arg, SDHCI_ARGUMENT);
> >>>>>
> >>>>> @@ -1416,6 +1640,9 @@ void sdhci_send_command(struct sdhci_host *host, struct mmc_command *cmd)
> >>>>>               timeout += 10 * HZ;
> >>>>>       sdhci_mod_timer(host, cmd->mrq, timeout);
> >>>>>
> >>>>> +     if (host->use_external_dma)
> >>>>> +             sdhci_external_dma_pre_transfer(host, cmd);
> >>>>
> >>>> Why is sdhci_external_dma_pre_transfer() needed here - couldn't it be done
> >>>> in sdhci_external_dma_prepare_data()?
> >>>>
> >>>
> >>> I'm not sure dma_async_issue_pending() can be done so early in
> >>> sdhci_external_dma_prepare_data().
> >>>
> >>>>> +
> >>>>>       sdhci_writew(host, SDHCI_MAKE_CMD(cmd->opcode, flags), SDHCI_COMMAND);
> >>>>>  }
> >>>>>  EXPORT_SYMBOL_GPL(sdhci_send_command);
> >>>>> @@ -1781,6 +2008,11 @@ void sdhci_request(struct mmc_host *mmc, struct mmc_request *mrq)
> >>>>>
> >>>>>       sdhci_led_activate(host);
> >>>>>
> >>>>> +     if (host->use_external_dma && mrq->data) {
> >>>>> +             struct dma_chan *chan = sdhci_external_dma_channel(host,
> >>>>> +                                                                mrq->data);
> >>>>
> >>>> sdhci_external_dma_channel is not declared if
> >>>> !IS_ENABLED(CONFIG_MMC_SDHCI_EXTERNAL_DMA)
> >>>
> >>> Ok, I guess this can be moved to sdhci_external_dma_prepare_data().
> >>
> >> No. We have to synchronize from a guaranteed non-atomic context. Maybe
> >> add another empty sdhci_external_dma_channel definition in the #else part?
> >>
> >>>
> >>>>
> >>>>> +             dmaengine_synchronize(chan);
> >>>>
> >>>> So this is to cover for using dmaengine_terminate_async()?
> >>
> >> Yes.
> >>
> >>>
> >>> Ok.
> >>>
> >>>>
> >>>>> +     }
> >>>>>       /*
> >>>>>        * Ensure we don't send the STOP for non-SET_BLOCK_COUNTED
> >>>>>        * requests if Auto-CMD12 is enabled.
> >>>>> @@ -2658,6 +2890,8 @@ static bool sdhci_request_done(struct sdhci_host *host)
> >>>>>                               dma_unmap_sg(mmc_dev(host->mmc), data->sg,
> >>>>>                                            data->sg_len,
> >>>>>                                            mmc_get_dma_dir(data));
> >>>>> +                             if (host->use_external_dma)
> >>>>> +                                     sdhci_external_dma_cleanup(host, data);
> >>>>
> >>>> Is sdhci_external_dma_cleanup() only needed in the error case?
> >>>>
> >>>> The DMA must be stopped before the memory is unmapped and potentially freed.
> >>>>
> >>>> Isn't the DMA cleanup also needed in the bounce buffer case?
> >>>>
> >>>> Isn't the DMA cleanup also needed in the COOKIE_PRE_MAPPED case?
> >>>>
> >>>> dmaengine_terminate_async() doesn't stop the DMA but
> >>>> dmaengine_terminate_sync() is not atomic, which looks like a problem.
> >>>>
> >>>> Perhaps you look at scheduling some work for the external dma error case
> >>>> instead of calling __sdhci_finish_mrq()?  Then the work can do the
> >>>> dmaengine_terminate_sync() and call __sdhci_finish_mrq().
> >>>
> >>> Ok, I will look at these issues.
> >>>
> >>>>
> >>>>>                       }
> >>>>>                       data->host_cookie = COOKIE_UNMAPPED;
> >>>>>               }
> >>>>> @@ -3692,12 +3926,15 @@ int sdhci_setup_host(struct sdhci_host *host)
> >>>>>                      mmc_hostname(mmc), host->version);
> >>>>>       }
> >>>>>
> >>>>> -     if (host->quirks & SDHCI_QUIRK_FORCE_DMA)
> >>>>> +     if (host->quirks & SDHCI_QUIRK_FORCE_DMA) {
> >>>>>               host->flags |= SDHCI_USE_SDMA;
> >>>>> -     else if (!(host->caps & SDHCI_CAN_DO_SDMA))
> >>>>> +     } else if (!(host->caps & SDHCI_CAN_DO_SDMA)) {
> >>>>>               DBG("Controller doesn't have SDMA capability\n");
> >>>>> -     else
> >>>>> +     } else if (host->use_external_dma) {
> >>>>> +             /* Using dma-names to detect external dma capability */
> >>>>> +     } else {
> >>>>>               host->flags |= SDHCI_USE_SDMA;
> >>>>> +     }
> >>>>
> >>>> These if-statements are about setting SDHCI_USE_SDMA but why is a change
> >>>> needed for the host->use_external_dma case?
> >>>
> >>> Yes, this is not needed, otherwise the controller cannot switch back
> >>> to SDMA if it is supported in the controller.
> >>
> >> The else case means that the host always uses SDMA. Needed to make sure
> >> this doesn't happen in the external_dma case. Its not SDHCI_CAN_DO_SDMA,
> >> its SDHCI_USE_SDMA.
> >>
> >>>
> >>>>
> >>>>>
> >>>>>       if ((host->quirks & SDHCI_QUIRK_BROKEN_DMA) &&
> >>>>>               (host->flags & SDHCI_USE_SDMA)) {
> >>>>> @@ -3785,6 +4022,19 @@ int sdhci_setup_host(struct sdhci_host *host)
> >>>>>               }
> >>>>>       }
> >>>>>
> >>>>> +     if (host->use_external_dma) {
> >>>>> +             ret = sdhci_external_dma_init(host);
> >>>>> +             if (ret == -EPROBE_DEFER)
> >>>>> +                     goto unreg;
> >>>>> +
> >>>>> +             /*
> >>>>> +              * Fall back to use the DMA/PIO integrated in standard SDHCI
> >>>>> +              * instead of external DMA devices.
> >>>>> +              */
> >>>>> +             if (ret)
> >>>>> +                     sdhci_switch_external_dma(host, false);
> >>>>> +     }
> >>>>> +
> >>>>>       /*
> >>>>>        * If we use DMA, then it's up to the caller to set the DMA
> >>>>>        * mask, but PIO does not need the hw shim so we set a new
> >>>>> @@ -4201,6 +4451,10 @@ void sdhci_cleanup_host(struct sdhci_host *host)
> >>>>>               dma_free_coherent(mmc_dev(mmc), host->align_buffer_sz +
> >>>>>                                 host->adma_table_sz, host->align_buffer,
> >>>>>                                 host->align_addr);
> >>>>> +
> >>>>> +     if (host->use_external_dma)
> >>>>> +             sdhci_external_dma_release(host);
> >>>>> +
> >>>>>       host->adma_table = NULL;
> >>>>>       host->align_buffer = NULL;
> >>>>>  }
> >>>>> @@ -4247,6 +4501,7 @@ int __sdhci_add_host(struct sdhci_host *host)
> >>>>>
> >>>>>       pr_info("%s: SDHCI controller on %s [%s] using %s\n",
> >>>>>               mmc_hostname(mmc), host->hw_name, dev_name(mmc_dev(mmc)),
> >>>>> +             host->use_external_dma ? "External DMA" :
> >>>>>               (host->flags & SDHCI_USE_ADMA) ?
> >>>>>               (host->flags & SDHCI_USE_64_BIT_DMA) ? "ADMA 64-bit" : "ADMA" :
> >>>>>               (host->flags & SDHCI_USE_SDMA) ? "DMA" : "PIO");
> >>>>> @@ -4335,6 +4590,9 @@ void sdhci_remove_host(struct sdhci_host *host, int dead)
> >>>>>                                 host->adma_table_sz, host->align_buffer,
> >>>>>                                 host->align_addr);
> >>>>>
> >>>>> +     if (host->use_external_dma)
> >>>>> +             sdhci_external_dma_release(host);
> >>>>> +
> >>>>>       host->adma_table = NULL;
> >>>>>       host->align_buffer = NULL;
> >>>>>  }
> >>>>> diff --git a/drivers/mmc/host/sdhci.h b/drivers/mmc/host/sdhci.h
> >>>>> index 6cc9a3c2ac66..7a52823ebef4 100644
> >>>>> --- a/drivers/mmc/host/sdhci.h
> >>>>> +++ b/drivers/mmc/host/sdhci.h
> >>>>> @@ -482,6 +482,7 @@ struct sdhci_host {
> >>>>>
> >>>>>       int irq;                /* Device IRQ */
> >>>>>       void __iomem *ioaddr;   /* Mapped address */
> >>>>> +     phys_addr_t mapbase;    /* physical address base */
> >>>>>       char *bounce_buffer;    /* For packing SDMA reads/writes */
> >>>>>       dma_addr_t bounce_addr;
> >>>>>       unsigned int bounce_buffer_size;
> >>>>> @@ -531,6 +532,7 @@ struct sdhci_host {
> >>>>>       bool pending_reset;     /* Cmd/data reset is pending */
> >>>>>       bool irq_wake_enabled;  /* IRQ wakeup is enabled */
> >>>>>       bool v4_mode;           /* Host Version 4 Enable */
> >>>>> +     bool use_external_dma; /* Host selects to use external DMA */
> >>>>
> >>>> Please align /**/ with above i.e. use tab
> >>
> >> Will fix.
> >
> > Do you need me to address these comments and then send you a patch or
> > you'd like to do instead?
> >
>
> This would be much appreciated. Would you be able to work on it this
> week? I would like to send a v2 by Friday.

I will try. I just have worries about handling the external dma error case.
Anyway I'll send you a patch by tomorrow, since I will leave 10 days
from this Friday.

Thanks,
Chunyan
Faiz Abbas Feb. 11, 2019, 12:23 p.m. UTC | #8
Hi Adrian,

On 24/01/19 5:10 PM, Adrian Hunter wrote:
> On 11/01/19 1:08 PM, Faiz Abbas wrote:
>> From: Chunyan Zhang <zhang.chunyan@linaro.org>
>>
>> Some standard SD host controllers can support both external dma
>> controllers as well as ADMA/SDMA in which the SD host controller
>> acts as DMA master. TI's omap controller is the case as an example.
>>
>> Currently the generic SDHCI code supports ADMA/SDMA integrated in
>> the host controller but does not have any support for external DMA
>> controllers implemented using dmaengine, meaning that custom code is
>> needed for any systems that use an external DMA controller with SDHCI.
>>
>> Fixes by Faiz Abbas <faiz_abbas@ti.com>:
>> 1. Map scatterlists before dmaengine_prep_slave_sg()
>> 2. Use dma_async() functions inside of the send_command() path and
>> synchronize once at the start of each request.
> 
> Sorry for the slow reply, but I do have some concerns.  Please see the comments.
>[snip]>>  	/*
>>  	 * Ensure we don't send the STOP for non-SET_BLOCK_COUNTED
>>  	 * requests if Auto-CMD12 is enabled.
>> @@ -2658,6 +2890,8 @@ static bool sdhci_request_done(struct sdhci_host *host)
>>  				dma_unmap_sg(mmc_dev(host->mmc), data->sg,
>>  					     data->sg_len,
>>  					     mmc_get_dma_dir(data));
>> +				if (host->use_external_dma)
>> +					sdhci_external_dma_cleanup(host, data);
> 
> Is sdhci_external_dma_cleanup() only needed in the error case?
> 
> The DMA must be stopped before the memory is unmapped and potentially freed.
> 
> Isn't the DMA cleanup also needed in the bounce buffer case?
> 
> Isn't the DMA cleanup also needed in the COOKIE_PRE_MAPPED case?
> 
> dmaengine_terminate_async() doesn't stop the DMA but
> dmaengine_terminate_sync() is not atomic, which looks like a problem.
> 
> Perhaps you look at scheduling some work for the external dma error case
> instead of calling __sdhci_finish_mrq()?  Then the work can do the
> dmaengine_terminate_sync() and call __sdhci_finish_mrq().
> 

Why don't we get rid of the finish_tasklet and use the already existing
threaded_irq for everything? I tested the following patch out and it
seems to work well. This enables us to just call
dmaengine_terminate_sync() in sdhci_request_done().

diff --git a/drivers/mmc/host/sdhci.c b/drivers/mmc/host/sdhci.c
index a22e11a65658..beff2ac2dee5 100644
--- a/drivers/mmc/host/sdhci.c
+++ b/drivers/mmc/host/sdhci.c
@@ -1232,7 +1232,7 @@ static void __sdhci_finish_mrq(struct sdhci_host
*host, struct mmc_request *mrq)

 	WARN_ON(i >= SDHCI_MAX_MRQS);

-	tasklet_schedule(&host->finish_tasklet);
+	host->threaded_irq_needed = true;
 }

 static void sdhci_finish_mrq(struct sdhci_host *host, struct
mmc_request *mrq)
@@ -2705,14 +2705,6 @@ static bool sdhci_request_done(struct sdhci_host
*host)
 	return false;
 }

-static void sdhci_tasklet_finish(unsigned long param)
-{
-	struct sdhci_host *host = (struct sdhci_host *)param;
-
-	while (!sdhci_request_done(host))
-		;
-}
-
 static void sdhci_timeout_timer(struct timer_list *t)
 {
 	struct sdhci_host *host;
@@ -3000,6 +2992,8 @@ static irqreturn_t sdhci_irq(int irq, void *dev_id)
 	u32 intmask, mask, unexpected = 0;
 	int max_loops = 16;

+	host->threaded_irq_needed = false;
+
 	spin_lock(&host->lock);

 	if (host->runtime_suspended && !sdhci_sdio_irq_enabled(host)) {
@@ -3077,6 +3071,9 @@ static irqreturn_t sdhci_irq(int irq, void *dev_id)
 			result = IRQ_WAKE_THREAD;
 		}

+		if (host->threaded_irq_needed)
+			result = IRQ_WAKE_THREAD;
+
 		intmask &= ~(SDHCI_INT_CARD_INSERT | SDHCI_INT_CARD_REMOVE |
 			     SDHCI_INT_CMD_MASK | SDHCI_INT_DATA_MASK |
 			     SDHCI_INT_ERROR | SDHCI_INT_BUS_POWER |
@@ -3131,6 +3128,10 @@ static irqreturn_t sdhci_thread_irq(int irq, void
*dev_id)
 		spin_unlock_irqrestore(&host->lock, flags);
 	}

+	do {
+		isr = sdhci_request_done(host);
+	} while(isr);
+
 	return isr ? IRQ_HANDLED : IRQ_NONE;
 }

@@ -4211,12 +4212,6 @@ int __sdhci_add_host(struct sdhci_host *host)
 	struct mmc_host *mmc = host->mmc;
 	int ret;

-	/*
-	 * Init tasklets.
-	 */
-	tasklet_init(&host->finish_tasklet,
-		sdhci_tasklet_finish, (unsigned long)host);
-
 	timer_setup(&host->timer, sdhci_timeout_timer, 0);
 	timer_setup(&host->data_timer, sdhci_timeout_data_timer, 0);

@@ -4229,7 +4224,7 @@ int __sdhci_add_host(struct sdhci_host *host)
 	if (ret) {
 		pr_err("%s: Failed to request IRQ %d: %d\n",
 		       mmc_hostname(mmc), host->irq, ret);
-		goto untasklet;
+		return ret;
 	}

 	ret = sdhci_led_register(host);
@@ -4262,8 +4257,6 @@ int __sdhci_add_host(struct sdhci_host *host)
 	sdhci_writel(host, 0, SDHCI_INT_ENABLE);
 	sdhci_writel(host, 0, SDHCI_SIGNAL_ENABLE);
 	free_irq(host->irq, host);
-untasklet:
-	tasklet_kill(&host->finish_tasklet);

 	return ret;
 }
@@ -4325,8 +4318,6 @@ void sdhci_remove_host(struct sdhci_host *host,
int dead)
 	del_timer_sync(&host->timer);
 	del_timer_sync(&host->data_timer);

-	tasklet_kill(&host->finish_tasklet);
-
 	if (!IS_ERR(mmc->supply.vqmmc))
 		regulator_disable(mmc->supply.vqmmc);

diff --git a/drivers/mmc/host/sdhci.h b/drivers/mmc/host/sdhci.h
index 6cc9a3c2ac66..abf4f77650b5 100644
--- a/drivers/mmc/host/sdhci.h
+++ b/drivers/mmc/host/sdhci.h
@@ -554,7 +554,7 @@ struct sdhci_host {

 	unsigned int desc_sz;	/* ADMA descriptor size */

-	struct tasklet_struct finish_tasklet;	/* Tasklet structures */
+	bool threaded_irq_needed;

 	struct timer_list timer;	/* Timer for timeouts */
 	struct timer_list data_timer;	/* Timer for data timeouts */

Thanks,
Faiz
diff mbox series

Patch

diff --git a/drivers/mmc/host/Kconfig b/drivers/mmc/host/Kconfig
index e26b8145efb3..333292e8ecdd 100644
--- a/drivers/mmc/host/Kconfig
+++ b/drivers/mmc/host/Kconfig
@@ -999,3 +999,6 @@  config MMC_SDHCI_AM654
 	  If you have a controller with this interface, say Y or M here.
 
 	  If unsure, say N.
+
+config MMC_SDHCI_EXTERNAL_DMA
+        bool
diff --git a/drivers/mmc/host/sdhci.c b/drivers/mmc/host/sdhci.c
index a22e11a65658..4a9044c06e21 100644
--- a/drivers/mmc/host/sdhci.c
+++ b/drivers/mmc/host/sdhci.c
@@ -14,6 +14,7 @@ 
  */
 
 #include <linux/delay.h>
+#include <linux/dmaengine.h>
 #include <linux/ktime.h>
 #include <linux/highmem.h>
 #include <linux/io.h>
@@ -1118,6 +1119,226 @@  static void sdhci_prepare_data(struct sdhci_host *host, struct mmc_command *cmd)
 	}
 }
 
+#if IS_ENABLED(CONFIG_MMC_SDHCI_EXTERNAL_DMA)
+static int sdhci_external_dma_init(struct sdhci_host *host)
+{
+	int ret = 0;
+	struct mmc_host *mmc = host->mmc;
+
+	host->tx_chan = dma_request_chan(mmc->parent, "tx");
+	if (IS_ERR(host->tx_chan)) {
+		ret = PTR_ERR(host->tx_chan);
+		if (ret != -EPROBE_DEFER)
+			pr_warn("Failed to request TX DMA channel.\n");
+		host->tx_chan = NULL;
+		return ret;
+	}
+
+	host->rx_chan = dma_request_chan(mmc->parent, "rx");
+	if (IS_ERR(host->rx_chan)) {
+		if (host->tx_chan) {
+			dma_release_channel(host->tx_chan);
+			host->tx_chan = NULL;
+		}
+
+		ret = PTR_ERR(host->rx_chan);
+		if (ret != -EPROBE_DEFER)
+			pr_warn("Failed to request RX DMA channel.\n");
+		host->rx_chan = NULL;
+	}
+
+	return ret;
+}
+
+static inline struct dma_chan *
+sdhci_external_dma_channel(struct sdhci_host *host, struct mmc_data *data)
+{
+	return data->flags & MMC_DATA_WRITE ? host->tx_chan : host->rx_chan;
+}
+
+static int sdhci_external_dma_setup(struct sdhci_host *host,
+				    struct mmc_command *cmd)
+{
+	int ret, i;
+	struct dma_async_tx_descriptor *desc;
+	struct mmc_data *data = cmd->data;
+	struct dma_chan *chan;
+	struct dma_slave_config cfg;
+	dma_cookie_t cookie;
+	int sg_cnt;
+
+	if (!host->mapbase)
+		return -EINVAL;
+
+	cfg.src_addr = host->mapbase + SDHCI_BUFFER;
+	cfg.dst_addr = host->mapbase + SDHCI_BUFFER;
+	cfg.src_addr_width = DMA_SLAVE_BUSWIDTH_4_BYTES;
+	cfg.dst_addr_width = DMA_SLAVE_BUSWIDTH_4_BYTES;
+	cfg.src_maxburst = data->blksz / 4;
+	cfg.dst_maxburst = data->blksz / 4;
+
+	/* Sanity check: all the SG entries must be aligned by block size. */
+	for (i = 0; i < data->sg_len; i++) {
+		if ((data->sg + i)->length % data->blksz)
+			return -EINVAL;
+	}
+
+	chan = sdhci_external_dma_channel(host, data);
+
+	ret = dmaengine_slave_config(chan, &cfg);
+	if (ret)
+		return ret;
+
+	sg_cnt = sdhci_pre_dma_transfer(host, data, COOKIE_MAPPED);
+	if (sg_cnt <= 0)
+		return -EINVAL;
+
+	desc = dmaengine_prep_slave_sg(chan, data->sg, data->sg_len,
+				       mmc_get_dma_dir(data),
+				       DMA_PREP_INTERRUPT | DMA_CTRL_ACK);
+	if (!desc)
+		return -EINVAL;
+
+	desc->callback = NULL;
+	desc->callback_param = NULL;
+
+	cookie = dmaengine_submit(desc);
+	if (cookie < 0)
+		ret = cookie;
+
+	return ret;
+}
+
+static void sdhci_external_dma_release(struct sdhci_host *host)
+{
+	if (host->tx_chan) {
+		dma_release_channel(host->tx_chan);
+		host->tx_chan = NULL;
+	}
+
+	if (host->rx_chan) {
+		dma_release_channel(host->rx_chan);
+		host->rx_chan = NULL;
+	}
+
+	sdhci_switch_external_dma(host, false);
+}
+
+static int __sdhci_external_dma_prepare_data(struct sdhci_host *host,
+					     struct mmc_command *cmd)
+{
+	struct mmc_data *data = cmd->data;
+
+	host->data_timeout = 0;
+
+	if (sdhci_data_line_cmd(cmd))
+		sdhci_set_timeout(host, cmd);
+
+	WARN_ON(host->data);
+
+	/* Sanity checks */
+	WARN_ON(data->blksz * data->blocks > 524288);
+	WARN_ON(data->blksz > host->mmc->max_blk_size);
+	WARN_ON(data->blocks > 65535);
+
+	host->flags |= SDHCI_REQ_USE_DMA;
+	host->data = data;
+	host->data_early = 0;
+	host->data->bytes_xfered = 0;
+
+	sdhci_set_transfer_irqs(host);
+
+	/*
+	 * For Version 4.10 onwards, if v4 mode is enabled, 32-bit Block Count
+	 * can be supported, in that case 16-bit block count register must be 0.
+	 */
+	if (host->version >= SDHCI_SPEC_410 && host->v4_mode &&
+	    (host->quirks2 & SDHCI_QUIRK2_USE_32BIT_BLK_CNT)) {
+		if (sdhci_readw(host, SDHCI_BLOCK_COUNT))
+			sdhci_writew(host, 0, SDHCI_BLOCK_COUNT);
+		sdhci_writew(host, data->blocks, SDHCI_32BIT_BLK_CNT);
+	} else {
+		sdhci_writew(host, data->blocks, SDHCI_BLOCK_COUNT);
+	}
+
+	return 0;
+}
+
+static void sdhci_external_dma_prepare_data(struct sdhci_host *host,
+					    struct mmc_command *cmd)
+{
+	struct mmc_data *data = cmd->data;
+
+	if (!data)
+		return;
+
+	if (sdhci_external_dma_setup(host, cmd) ||
+	    __sdhci_external_dma_prepare_data(host, cmd)) {
+		sdhci_external_dma_release(host);
+		pr_err("%s: Cannot use external DMA, switch to the DMA/PIO which standard SDHCI provides.\n",
+		       mmc_hostname(host->mmc));
+		sdhci_prepare_data(host, cmd);
+	}
+}
+
+static void sdhci_external_dma_pre_transfer(struct sdhci_host *host,
+					    struct mmc_command *cmd)
+{
+	struct dma_chan *chan;
+
+	if (!cmd->data || cmd->opcode == MMC_SET_BLOCK_COUNT)
+		return;
+
+	sdhci_writew(host, cmd->data->blksz, SDHCI_BLOCK_SIZE);
+	chan = sdhci_external_dma_channel(host, cmd->data);
+	if (chan)
+		dma_async_issue_pending(chan);
+}
+
+static int sdhci_external_dma_cleanup(struct sdhci_host *host,
+				    struct mmc_data *data)
+{
+	struct dma_chan *chan = sdhci_external_dma_channel(host, data);
+	int ret = 0;
+
+	if (chan)
+		ret = dmaengine_terminate_async(chan);
+
+	return ret;
+}
+#else
+static int sdhci_external_dma_init(struct sdhci_host *host)
+{
+	return -EOPNOTSUPP;
+}
+
+static void sdhci_external_dma_release(struct sdhci_host *host)
+{}
+
+static void sdhci_external_dma_prepare_data(struct sdhci_host *host,
+					    struct mmc_command *cmd)
+{
+	/* If MMC_SDHCI_EXTERNAL_DMA not supported, PIO will be used */
+	sdhci_prepare_data(host, cmd);
+}
+
+static void sdhci_external_dma_pre_transfer(struct sdhci_host *host,
+					    struct mmc_command *cmd)
+{}
+
+static int sdhci_external_dma_cleanup(struct sdhci_host *host,
+				    struct mmc_data *data)
+{
+	return 0;
+}
+#endif
+
+void sdhci_switch_external_dma(struct sdhci_host *host, bool en)
+{
+	host->use_external_dma = en;
+}
+EXPORT_SYMBOL_GPL(sdhci_switch_external_dma);
+
 static inline bool sdhci_auto_cmd12(struct sdhci_host *host,
 				    struct mmc_request *mrq)
 {
@@ -1374,7 +1595,10 @@  void sdhci_send_command(struct sdhci_host *host, struct mmc_command *cmd)
 		host->data_cmd = cmd;
 	}
 
-	sdhci_prepare_data(host, cmd);
+	if (host->use_external_dma)
+		sdhci_external_dma_prepare_data(host, cmd);
+	else
+		sdhci_prepare_data(host, cmd);
 
 	sdhci_writel(host, cmd->arg, SDHCI_ARGUMENT);
 
@@ -1416,6 +1640,9 @@  void sdhci_send_command(struct sdhci_host *host, struct mmc_command *cmd)
 		timeout += 10 * HZ;
 	sdhci_mod_timer(host, cmd->mrq, timeout);
 
+	if (host->use_external_dma)
+		sdhci_external_dma_pre_transfer(host, cmd);
+
 	sdhci_writew(host, SDHCI_MAKE_CMD(cmd->opcode, flags), SDHCI_COMMAND);
 }
 EXPORT_SYMBOL_GPL(sdhci_send_command);
@@ -1781,6 +2008,11 @@  void sdhci_request(struct mmc_host *mmc, struct mmc_request *mrq)
 
 	sdhci_led_activate(host);
 
+	if (host->use_external_dma && mrq->data) {
+		struct dma_chan *chan = sdhci_external_dma_channel(host,
+								   mrq->data);
+		dmaengine_synchronize(chan);
+	}
 	/*
 	 * Ensure we don't send the STOP for non-SET_BLOCK_COUNTED
 	 * requests if Auto-CMD12 is enabled.
@@ -2658,6 +2890,8 @@  static bool sdhci_request_done(struct sdhci_host *host)
 				dma_unmap_sg(mmc_dev(host->mmc), data->sg,
 					     data->sg_len,
 					     mmc_get_dma_dir(data));
+				if (host->use_external_dma)
+					sdhci_external_dma_cleanup(host, data);
 			}
 			data->host_cookie = COOKIE_UNMAPPED;
 		}
@@ -3692,12 +3926,15 @@  int sdhci_setup_host(struct sdhci_host *host)
 		       mmc_hostname(mmc), host->version);
 	}
 
-	if (host->quirks & SDHCI_QUIRK_FORCE_DMA)
+	if (host->quirks & SDHCI_QUIRK_FORCE_DMA) {
 		host->flags |= SDHCI_USE_SDMA;
-	else if (!(host->caps & SDHCI_CAN_DO_SDMA))
+	} else if (!(host->caps & SDHCI_CAN_DO_SDMA)) {
 		DBG("Controller doesn't have SDMA capability\n");
-	else
+	} else if (host->use_external_dma) {
+		/* Using dma-names to detect external dma capability */
+	} else {
 		host->flags |= SDHCI_USE_SDMA;
+	}
 
 	if ((host->quirks & SDHCI_QUIRK_BROKEN_DMA) &&
 		(host->flags & SDHCI_USE_SDMA)) {
@@ -3785,6 +4022,19 @@  int sdhci_setup_host(struct sdhci_host *host)
 		}
 	}
 
+	if (host->use_external_dma) {
+		ret = sdhci_external_dma_init(host);
+		if (ret == -EPROBE_DEFER)
+			goto unreg;
+
+		/*
+		 * Fall back to use the DMA/PIO integrated in standard SDHCI
+		 * instead of external DMA devices.
+		 */
+		if (ret)
+			sdhci_switch_external_dma(host, false);
+	}
+
 	/*
 	 * If we use DMA, then it's up to the caller to set the DMA
 	 * mask, but PIO does not need the hw shim so we set a new
@@ -4201,6 +4451,10 @@  void sdhci_cleanup_host(struct sdhci_host *host)
 		dma_free_coherent(mmc_dev(mmc), host->align_buffer_sz +
 				  host->adma_table_sz, host->align_buffer,
 				  host->align_addr);
+
+	if (host->use_external_dma)
+		sdhci_external_dma_release(host);
+
 	host->adma_table = NULL;
 	host->align_buffer = NULL;
 }
@@ -4247,6 +4501,7 @@  int __sdhci_add_host(struct sdhci_host *host)
 
 	pr_info("%s: SDHCI controller on %s [%s] using %s\n",
 		mmc_hostname(mmc), host->hw_name, dev_name(mmc_dev(mmc)),
+		host->use_external_dma ? "External DMA" :
 		(host->flags & SDHCI_USE_ADMA) ?
 		(host->flags & SDHCI_USE_64_BIT_DMA) ? "ADMA 64-bit" : "ADMA" :
 		(host->flags & SDHCI_USE_SDMA) ? "DMA" : "PIO");
@@ -4335,6 +4590,9 @@  void sdhci_remove_host(struct sdhci_host *host, int dead)
 				  host->adma_table_sz, host->align_buffer,
 				  host->align_addr);
 
+	if (host->use_external_dma)
+		sdhci_external_dma_release(host);
+
 	host->adma_table = NULL;
 	host->align_buffer = NULL;
 }
diff --git a/drivers/mmc/host/sdhci.h b/drivers/mmc/host/sdhci.h
index 6cc9a3c2ac66..7a52823ebef4 100644
--- a/drivers/mmc/host/sdhci.h
+++ b/drivers/mmc/host/sdhci.h
@@ -482,6 +482,7 @@  struct sdhci_host {
 
 	int irq;		/* Device IRQ */
 	void __iomem *ioaddr;	/* Mapped address */
+	phys_addr_t mapbase;	/* physical address base */
 	char *bounce_buffer;	/* For packing SDMA reads/writes */
 	dma_addr_t bounce_addr;
 	unsigned int bounce_buffer_size;
@@ -531,6 +532,7 @@  struct sdhci_host {
 	bool pending_reset;	/* Cmd/data reset is pending */
 	bool irq_wake_enabled;	/* IRQ wakeup is enabled */
 	bool v4_mode;		/* Host Version 4 Enable */
+	bool use_external_dma; /* Host selects to use external DMA */
 
 	struct mmc_request *mrqs_done[SDHCI_MAX_MRQS];	/* Requests done */
 	struct mmc_command *cmd;	/* Current command */
@@ -559,6 +561,11 @@  struct sdhci_host {
 	struct timer_list timer;	/* Timer for timeouts */
 	struct timer_list data_timer;	/* Timer for data timeouts */
 
+#if IS_ENABLED(CONFIG_MMC_SDHCI_EXTERNAL_DMA)
+	struct dma_chan	*rx_chan;
+	struct dma_chan	*tx_chan;
+#endif
+
 	u32 caps;		/* CAPABILITY_0 */
 	u32 caps1;		/* CAPABILITY_1 */
 	bool read_caps;		/* Capability flags have been read */
@@ -792,5 +799,6 @@  void sdhci_start_tuning(struct sdhci_host *host);
 void sdhci_end_tuning(struct sdhci_host *host);
 void sdhci_reset_tuning(struct sdhci_host *host);
 void sdhci_send_tuning(struct sdhci_host *host, u32 opcode);
+void sdhci_switch_external_dma(struct sdhci_host *host, bool en);
 
 #endif /* __SDHCI_HW_H */