Re: [PATCH 06/14] qcom: mtd: nand: add bam dma descriptor handling
From: Archit Taneja
Date: Tue Jul 04 2017 - 02:11:12 EST
On 06/29/2017 12:45 PM, Abhishek Sahu wrote:
1. prepare_bam_async_desc is the function which will call
all the DMA APIâs. It will fetch the outstanding scatter gather
list for passed channel and will do the DMA descriptor formation.
The DMA flag is dependent upon the type of channel.
2. For ADM DMA, the descriptor is being formed for every DMA
request so its sgl count will be always 1 while in BAM DMA, the
clubbing of descriptor is being done to increase throughput.
3. ADM uses only one channel while in BAM, data descriptors
will be submitted to tx channel (for write) or rx channel
(for read) and all the registers read/write descriptors in
command channel.
Signed-off-by: Abhishek Sahu <absahu@xxxxxxxxxxxxxx>
---
drivers/mtd/nand/qcom_nandc.c | 119 ++++++++++++++++++++++++++++++++++++++++--
1 file changed, 114 insertions(+), 5 deletions(-)
diff --git a/drivers/mtd/nand/qcom_nandc.c b/drivers/mtd/nand/qcom_nandc.c
index f8d0bde..7042a65 100644
--- a/drivers/mtd/nand/qcom_nandc.c
+++ b/drivers/mtd/nand/qcom_nandc.c
@@ -206,14 +206,22 @@ struct bam_transaction {
* This data type corresponds to the nand dma descriptor
* @list - list for desc_info
* @dir - DMA transfer direction
- * @sgl - sgl which will be used for single sgl dma descriptor
+ * @sgl - sgl which will be used for single sgl dma descriptor. Only used by ADM
+ * @bam_sgl - sgl which will be used for dma descriptor. Only used by BAM
+ * @sgl_cnt - number of SGL in bam_sgl. Only used by BAM
* @dma_desc - low level dma engine descriptor
*/
struct desc_info {
struct list_head node;
enum dma_data_direction dir;
- struct scatterlist sgl;
+ union {
+ struct scatterlist sgl;
Can you make this adm_sgl instead for consistency? Also, please use only
two tabs instead of one here for indentation.
+ struct {
+ struct scatterlist *bam_sgl;
+ int sgl_cnt;
+ };
+ };
struct dma_async_tx_descriptor *dma_desc;
};
@@ -564,6 +572,68 @@ static void update_rw_regs(struct qcom_nand_host *host, int num_cw, bool read)
nandc_set_reg(nandc, NAND_EXEC_CMD, 1);
}
+/*
+ * Maps the scatter gather list for DMA transfer and forms the DMA descriptor
+ * for BAM. This descriptor will be added in the NAND DMA descriptor queue
+ * which will be submitted to DMA engine.
+ */
+static int prepare_bam_async_desc(struct qcom_nand_controller *nandc,
+ struct dma_chan *chan,
+ unsigned long flags)
From what I gathered in patch #10, this would be called by
prep_dma_desc_data_bam() and prep_dma_desc_command(). Can you rename these
two to something like prep_bam_dma_desc_data() and prep_bam_dma_desc_cmd()
+{
+ struct desc_info *desc;
+ struct scatterlist *sgl;
+ unsigned int sgl_cnt;
+ struct bam_transaction *bam_txn = nandc->bam_txn;
+ enum dma_transfer_direction dir_eng;
+ struct dma_async_tx_descriptor *dma_desc;
+
+ desc = kzalloc(sizeof(*desc), GFP_KERNEL);
+ if (!desc)
+ return -ENOMEM;
+
+ if (chan == nandc->cmd_chan) {
+ sgl = &bam_txn->cmd_sgl[bam_txn->cmd_sgl_start];
+ sgl_cnt = bam_txn->cmd_sgl_pos - bam_txn->cmd_sgl_start;
+ bam_txn->cmd_sgl_start = bam_txn->cmd_sgl_pos;
+ dir_eng = DMA_MEM_TO_DEV;
+ desc->dir = DMA_TO_DEVICE;
+ } else if (chan == nandc->tx_chan) {
+ sgl = &bam_txn->data_sg[bam_txn->tx_sgl_start];
+ sgl_cnt = bam_txn->tx_sgl_pos - bam_txn->tx_sgl_start;
+ bam_txn->tx_sgl_start = bam_txn->tx_sgl_pos;
+ dir_eng = DMA_MEM_TO_DEV;
+ desc->dir = DMA_TO_DEVICE;
+ } else {
+ sgl = &bam_txn->data_sg[bam_txn->rx_sgl_start];
+ sgl_cnt = bam_txn->rx_sgl_pos - bam_txn->rx_sgl_start;
+ bam_txn->rx_sgl_start = bam_txn->rx_sgl_pos;
+ desc->dir = DMA_FROM_DEVICE;
+ dir_eng = DMA_DEV_TO_MEM;
+ }
+
+ sg_mark_end(sgl + sgl_cnt - 1);
+ dma_map_sg(nandc->dev, sgl, sgl_cnt, desc->dir);
Is it safe to assume here that dma_map_sg won't return an error?
+
+ desc->sgl_cnt = sgl_cnt;
+ desc->bam_sgl = sgl;
+
+ dma_desc = dmaengine_prep_slave_sg(chan, sgl, sgl_cnt, dir_eng,
+ flags);
+
+ if (!dma_desc) {
+ dev_err(nandc->dev, "failure in prep desc\n");
+ kfree(desc);
+ return -EINVAL;
+ }
+
+ desc->dma_desc = dma_desc;
+
+ list_add_tail(&desc->node, &nandc->desc_list);
+
+ return 0;
+}
+
static int prep_dma_desc(struct qcom_nand_controller *nandc, bool read,
int reg_off, const void *vaddr, int size,
bool flow_control)
Could you rename this to prep_adm_dma_desc for consistency?
@@ -891,12 +961,44 @@ static int submit_descs(struct qcom_nand_controller *nandc)
{
struct desc_info *desc;
dma_cookie_t cookie = 0;
+ struct bam_transaction *bam_txn = nandc->bam_txn;
+ int r;
+
+ if (nandc->dma_bam_enabled) {
+ if (bam_txn->rx_sgl_pos - bam_txn->rx_sgl_start) {
Using '>' instead of '-' here should be more legible.
+ r = prepare_bam_async_desc(nandc, nandc->rx_chan, 0);
+ if (r)
+ return r;
+ }
+
+ if (bam_txn->tx_sgl_pos - bam_txn->tx_sgl_start) {
+ r = prepare_bam_async_desc(nandc, nandc->tx_chan,
+ DMA_PREP_INTERRUPT);
+ if (r)
+ return r;
+ }
+
+ if (bam_txn->cmd_sgl_pos - bam_txn->cmd_sgl_start) {
+ r = prepare_bam_async_desc(nandc, nandc->cmd_chan,
+ DMA_PREP_CMD);
+ if (r)
+ return r;
+ }
+ }
list_for_each_entry(desc, &nandc->desc_list, node)
cookie = dmaengine_submit(desc->dma_desc);
- if (dma_sync_wait(nandc->chan, cookie) != DMA_COMPLETE)
- return -ETIMEDOUT;
+ if (nandc->dma_bam_enabled) {
+ dma_async_issue_pending(nandc->tx_chan);
+ dma_async_issue_pending(nandc->rx_chan);
+
+ if (dma_sync_wait(nandc->cmd_chan, cookie) != DMA_COMPLETE)
+ return -ETIMEDOUT;
+ } else {
+ if (dma_sync_wait(nandc->chan, cookie) != DMA_COMPLETE)
+ return -ETIMEDOUT;
+ }
return 0;
}
@@ -907,7 +1009,14 @@ static void free_descs(struct qcom_nand_controller *nandc)
list_for_each_entry_safe(desc, n, &nandc->desc_list, node) {
list_del(&desc->node);
- dma_unmap_sg(nandc->dev, &desc->sgl, 1, desc->dir);
+
+ if (nandc->dma_bam_enabled)
+ dma_unmap_sg(nandc->dev, desc->bam_sgl,
+ desc->sgl_cnt, desc->dir);
+ else
+ dma_unmap_sg(nandc->dev, &desc->sgl, 1,
+ desc->dir);
+
kfree(desc);
}
}
--
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
a Linux Foundation Collaborative Project