diff mbox series

[2/2] spi: spi-geni-qcom: Do not do DMA map/unmap inside driver, use framework instead

Message ID 1683911265-3395-3-git-send-email-quic_vnivarth@quicinc.com (mailing list archive)
State New, archived
Headers show
Series spi: spi-geni-qcom: Add new interfaces and utilise them to do map/unmap in framework for SE DMA | expand

Commit Message

Vijaya Krishna Nivarthi May 12, 2023, 5:07 p.m. UTC 
The spi geni driver in SE DMA mode, unlike GSI DMA, is not making use of
DMA mapping functionality available in the framework.
The driver does mapping internally which makes dma buffer fields available
in spi_transfer struct superfluous while requiring additional members in
spi_geni_master struct.

Conform to the design by having framework handle map/unmap and do only
DMA transfer in the driver; this also simplifies code a bit.

Fixes: e5f0dfa78ac7 ("spi: spi-geni-qcom: Add support for SE DMA mode")
Suggested-by: Douglas Anderson <dianders@chromium.org>
Signed-off-by: Vijaya Krishna Nivarthi <quic_vnivarth@quicinc.com>
---
 drivers/spi/spi-geni-qcom.c | 101 +++++++++++++++++++++++---------------------
 1 file changed, 52 insertions(+), 49 deletions(-)

Comments

Doug Anderson May 15, 2023, 4:13 p.m. UTC | #1 
Hi,

On Fri, May 12, 2023 at 10:07 AM Vijaya Krishna Nivarthi
<quic_vnivarth@quicinc.com> wrote:
>
> @@ -836,35 +858,24 @@ static int setup_se_xfer(struct spi_transfer *xfer,
>         geni_se_setup_m_cmd(se, m_cmd, FRAGMENTATION);
>
>         if (mas->cur_xfer_mode == GENI_SE_DMA) {
> +               dma_addr_t dma_ptr_sg;
> +               unsigned int dma_len_sg;
> +
>                 if (m_cmd & SPI_RX_ONLY) {
> -                       ret =  geni_se_rx_dma_prep(se, xfer->rx_buf,
> -                               xfer->len, &mas->rx_se_dma);
> -                       if (ret) {
> -                               dev_err(mas->dev, "Failed to setup Rx dma %d\n", ret);
> -                               mas->rx_se_dma = 0;
> -                               goto unlock_and_return;
> -                       }
> +                       dma_ptr_sg = sg_dma_address(xfer->rx_sg.sgl);
> +                       dma_len_sg = sg_dma_len(xfer->rx_sg.sgl);
> +                       geni_se_rx_init_dma(se, &dma_ptr_sg, dma_len_sg);

nit: probably don't need local variables if you change patch set #1
like I suggested and don't pass in a pointer for the iova.


One last question: should you call:

dma_set_max_seg_size(dev, INT_MAX)

...in your probe function? I don't think you have any limitations of
maximum segment size, right? Right now if you don't set anything it
looks as if it considers your max to be 64K. That would cause the SPI
framework to break things up into multiple chunks which would make you
fall back to FIFO mode, right?

Other than that this looks good to me.

-Doug
Vijaya Krishna Nivarthi May 17, 2023, 12:16 p.m. UTC | #2 
Hi,

Thank you very much for the review...


On 5/15/2023 9:43 PM, Doug Anderson wrote:
> Hi,
>
> On Fri, May 12, 2023 at 10:07 AM Vijaya Krishna Nivarthi
> <quic_vnivarth@quicinc.com> wrote:
>> @@ -836,35 +858,24 @@ static int setup_se_xfer(struct spi_transfer *xfer,
>>          geni_se_setup_m_cmd(se, m_cmd, FRAGMENTATION);
>>
>>          if (mas->cur_xfer_mode == GENI_SE_DMA) {
>> +               dma_addr_t dma_ptr_sg;
>> +               unsigned int dma_len_sg;
>> +
>>                  if (m_cmd & SPI_RX_ONLY) {
>> -                       ret =  geni_se_rx_dma_prep(se, xfer->rx_buf,
>> -                               xfer->len, &mas->rx_se_dma);
>> -                       if (ret) {
>> -                               dev_err(mas->dev, "Failed to setup Rx dma %d\n", ret);
>> -                               mas->rx_se_dma = 0;
>> -                               goto unlock_and_return;
>> -                       }
>> +                       dma_ptr_sg = sg_dma_address(xfer->rx_sg.sgl);
>> +                       dma_len_sg = sg_dma_len(xfer->rx_sg.sgl);
>> +                       geni_se_rx_init_dma(se, &dma_ptr_sg, dma_len_sg);
> nit: probably don't need local variables if you change patch set #1
> like I suggested and don't pass in a pointer for the iova.
>
Done.

> One last question: should you call:
>
> dma_set_max_seg_size(dev, INT_MAX)
>
> ...in your probe function? I don't think you have any limitations of
> maximum segment size, right? Right now if you don't set anything it
> looks as if it considers your max to be 64K. That would cause the SPI
> framework to break things up into multiple chunks which would make you
> fall back to FIFO mode, right?


Actually we would need to call:

dma_set_max_seg_size(dev->parent, INT_MAX)

Please note that in probe()

spi->dma_map_dev = dev->parent;

and in __spi_map_msg()

tx_dev = ctlr->dma_map_dev;

ret = spi_map_buf(ctlr, tx_dev, &xfer->tx_sg,...


Since the dev->parent is QUP containing other SEs and its max_seg_size 
seems to be getting set from elsewhere than code (perhaps kernel 
scripts) it seemed safer not to modify that.

So I made below change and uploaded v2...

spi->max_dma_len = 0xffff0; /* 24 bits for tx/rx dma length */

which actually doesnt help much because spi_map_buf() picks the lower of 
these 2 but added it anyway

desc_len = min_t(size_t, max_seg_size, ctlr->max_dma_len);


Any case as next step we will look into scatter list support to DMA; and 
practically we may not have transfers over 64KB, so its ok for now?

Thank you,

Vijay/


>
> Other than that this looks good to me.
>
> -Doug
Doug Anderson May 17, 2023, 2:23 p.m. UTC | #3 
Hi,

On Wed, May 17, 2023 at 5:17 AM Vijaya Krishna Nivarthi
<quic_vnivarth@quicinc.com> wrote:
>
> > One last question: should you call:
> >
> > dma_set_max_seg_size(dev, INT_MAX)
> >
> > ...in your probe function? I don't think you have any limitations of
> > maximum segment size, right? Right now if you don't set anything it
> > looks as if it considers your max to be 64K. That would cause the SPI
> > framework to break things up into multiple chunks which would make you
> > fall back to FIFO mode, right?
>
>
> Actually we would need to call:
>
> dma_set_max_seg_size(dev->parent, INT_MAX)
>
> Please note that in probe()
>
> spi->dma_map_dev = dev->parent;
>
> and in __spi_map_msg()
>
> tx_dev = ctlr->dma_map_dev;
>
> ret = spi_map_buf(ctlr, tx_dev, &xfer->tx_sg,...
>
>
> Since the dev->parent is QUP containing other SEs and its max_seg_size
> seems to be getting set from elsewhere than code (perhaps kernel
> scripts) it seemed safer not to modify that.
>
> So I made below change and uploaded v2...
>
> spi->max_dma_len = 0xffff0; /* 24 bits for tx/rx dma length */
>
> which actually doesnt help much because spi_map_buf() picks the lower of
> these 2 but added it anyway
>
> desc_len = min_t(size_t, max_seg_size, ctlr->max_dma_len);

OK, fair enough.


> Any case as next step we will look into scatter list support to DMA; and
> practically we may not have transfers over 64KB, so its ok for now?

I think it's OK for now, especially since you do properly fall back to
FIFO in cases that aren't handled. Thanks!

-Doug
diff mbox series

Patch

diff --git a/drivers/spi/spi-geni-qcom.c b/drivers/spi/spi-geni-qcom.c
index e423efc..bfe2a1a 100644
--- a/drivers/spi/spi-geni-qcom.c
+++ b/drivers/spi/spi-geni-qcom.c
@@ -97,8 +97,6 @@  struct spi_geni_master {
 	struct dma_chan *tx;
 	struct dma_chan *rx;
 	int cur_xfer_mode;
-	dma_addr_t tx_se_dma;
-	dma_addr_t rx_se_dma;
 };
 
 static int get_spi_clk_cfg(unsigned int speed_hz,
@@ -174,7 +172,7 @@  static void handle_se_timeout(struct spi_master *spi,
 unmap_if_dma:
 	if (mas->cur_xfer_mode == GENI_SE_DMA) {
 		if (xfer) {
-			if (xfer->tx_buf && mas->tx_se_dma) {
+			if (xfer->tx_buf) {
 				spin_lock_irq(&mas->lock);
 				reinit_completion(&mas->tx_reset_done);
 				writel(1, se->base + SE_DMA_TX_FSM_RST);
@@ -182,9 +180,8 @@  static void handle_se_timeout(struct spi_master *spi,
 				time_left = wait_for_completion_timeout(&mas->tx_reset_done, HZ);
 				if (!time_left)
 					dev_err(mas->dev, "DMA TX RESET failed\n");
-				geni_se_tx_dma_unprep(se, mas->tx_se_dma, xfer->len);
 			}
-			if (xfer->rx_buf && mas->rx_se_dma) {
+			if (xfer->rx_buf) {
 				spin_lock_irq(&mas->lock);
 				reinit_completion(&mas->rx_reset_done);
 				writel(1, se->base + SE_DMA_RX_FSM_RST);
@@ -192,7 +189,6 @@  static void handle_se_timeout(struct spi_master *spi,
 				time_left = wait_for_completion_timeout(&mas->rx_reset_done, HZ);
 				if (!time_left)
 					dev_err(mas->dev, "DMA RX RESET failed\n");
-				geni_se_rx_dma_unprep(se, mas->rx_se_dma, xfer->len);
 			}
 		} else {
 			/*
@@ -523,17 +519,36 @@  static int setup_gsi_xfer(struct spi_transfer *xfer, struct spi_geni_master *mas
 	return 1;
 }
 
+static u32 get_xfer_len_in_words(struct spi_transfer *xfer,
+				struct spi_geni_master *mas)
+{
+	u32 len;
+
+	if (!(mas->cur_bits_per_word % MIN_WORD_LEN))
+		len = xfer->len * BITS_PER_BYTE / mas->cur_bits_per_word;
+	else
+		len = xfer->len / (mas->cur_bits_per_word / BITS_PER_BYTE + 1);
+	len &= TRANS_LEN_MSK;
+
+	return len;
+}
+
 static bool geni_can_dma(struct spi_controller *ctlr,
 			 struct spi_device *slv, struct spi_transfer *xfer)
 {
 	struct spi_geni_master *mas = spi_master_get_devdata(slv->master);
+	u32 len, fifo_size;
 
-	/*
-	 * Return true if transfer needs to be mapped prior to
-	 * calling transfer_one which is the case only for GPI_DMA.
-	 * For SE_DMA mode, map/unmap is done in geni_se_*x_dma_prep.
-	 */
-	return mas->cur_xfer_mode == GENI_GPI_DMA;
+	if (mas->cur_xfer_mode == GENI_GPI_DMA)
+		return true;
+
+	len = get_xfer_len_in_words(xfer, mas);
+	fifo_size = mas->tx_fifo_depth * mas->fifo_width_bits / mas->cur_bits_per_word;
+
+	if (len > fifo_size)
+		return true;
+	else
+		return false;
 }
 
 static int spi_geni_prepare_message(struct spi_master *spi,
@@ -772,7 +787,7 @@  static int setup_se_xfer(struct spi_transfer *xfer,
 				u16 mode, struct spi_master *spi)
 {
 	u32 m_cmd = 0;
-	u32 len, fifo_size;
+	u32 len;
 	struct geni_se *se = &mas->se;
 	int ret;
 
@@ -804,11 +819,7 @@  static int setup_se_xfer(struct spi_transfer *xfer,
 	mas->tx_rem_bytes = 0;
 	mas->rx_rem_bytes = 0;
 
-	if (!(mas->cur_bits_per_word % MIN_WORD_LEN))
-		len = xfer->len * BITS_PER_BYTE / mas->cur_bits_per_word;
-	else
-		len = xfer->len / (mas->cur_bits_per_word / BITS_PER_BYTE + 1);
-	len &= TRANS_LEN_MSK;
+	len = get_xfer_len_in_words(xfer, mas);
 
 	mas->cur_xfer = xfer;
 	if (xfer->tx_buf) {
@@ -823,9 +834,20 @@  static int setup_se_xfer(struct spi_transfer *xfer,
 		mas->rx_rem_bytes = xfer->len;
 	}
 
-	/* Select transfer mode based on transfer length */
-	fifo_size = mas->tx_fifo_depth * mas->fifo_width_bits / mas->cur_bits_per_word;
-	mas->cur_xfer_mode = (len <= fifo_size) ? GENI_SE_FIFO : GENI_SE_DMA;
+	/*
+	 * Select DMA mode if sgt are present; and with only 1 entry
+	 * This is not a serious limitation because the xfer buffers are
+	 * expected to fit into 1 entry almost always, and if any
+	 * doesn't for any reason, we fall back to FIFO mode anyway
+	 */
+	if (!xfer->tx_sg.nents && !xfer->rx_sg.nents)
+		mas->cur_xfer_mode = GENI_SE_FIFO;
+	else if (xfer->tx_sg.nents > 1 || xfer->rx_sg.nents > 1) {
+		dev_warn_once(mas->dev, "Doing FIFO, cannot handle tx_nents-%d, rx_nents-%d\n",
+			xfer->tx_sg.nents, xfer->rx_sg.nents);
+		mas->cur_xfer_mode = GENI_SE_FIFO;
+	} else
+		mas->cur_xfer_mode = GENI_SE_DMA;
 	geni_se_select_mode(se, mas->cur_xfer_mode);
 
 	/*
@@ -836,35 +858,24 @@  static int setup_se_xfer(struct spi_transfer *xfer,
 	geni_se_setup_m_cmd(se, m_cmd, FRAGMENTATION);
 
 	if (mas->cur_xfer_mode == GENI_SE_DMA) {
+		dma_addr_t dma_ptr_sg;
+		unsigned int dma_len_sg;
+
 		if (m_cmd & SPI_RX_ONLY) {
-			ret =  geni_se_rx_dma_prep(se, xfer->rx_buf,
-				xfer->len, &mas->rx_se_dma);
-			if (ret) {
-				dev_err(mas->dev, "Failed to setup Rx dma %d\n", ret);
-				mas->rx_se_dma = 0;
-				goto unlock_and_return;
-			}
+			dma_ptr_sg = sg_dma_address(xfer->rx_sg.sgl);
+			dma_len_sg = sg_dma_len(xfer->rx_sg.sgl);
+			geni_se_rx_init_dma(se, &dma_ptr_sg, dma_len_sg);
 		}
 		if (m_cmd & SPI_TX_ONLY) {
-			ret =  geni_se_tx_dma_prep(se, (void *)xfer->tx_buf,
-				xfer->len, &mas->tx_se_dma);
-			if (ret) {
-				dev_err(mas->dev, "Failed to setup Tx dma %d\n", ret);
-				mas->tx_se_dma = 0;
-				if (m_cmd & SPI_RX_ONLY) {
-					/* Unmap rx buffer if duplex transfer */
-					geni_se_rx_dma_unprep(se, mas->rx_se_dma, xfer->len);
-					mas->rx_se_dma = 0;
-				}
-				goto unlock_and_return;
-			}
+			dma_ptr_sg = sg_dma_address(xfer->tx_sg.sgl);
+			dma_len_sg = sg_dma_len(xfer->tx_sg.sgl);
+			geni_se_tx_init_dma(se, &dma_ptr_sg, dma_len_sg);
 		}
 	} else if (m_cmd & SPI_TX_ONLY) {
 		if (geni_spi_handle_tx(mas))
 			writel(mas->tx_wm, se->base + SE_GENI_TX_WATERMARK_REG);
 	}
 
-unlock_and_return:
 	spin_unlock_irq(&mas->lock);
 	return ret;
 }
@@ -965,14 +976,6 @@  static irqreturn_t geni_spi_isr(int irq, void *data)
 		if (dma_rx_status & RX_RESET_DONE)
 			complete(&mas->rx_reset_done);
 		if (!mas->tx_rem_bytes && !mas->rx_rem_bytes && xfer) {
-			if (xfer->tx_buf && mas->tx_se_dma) {
-				geni_se_tx_dma_unprep(se, mas->tx_se_dma, xfer->len);
-				mas->tx_se_dma = 0;
-			}
-			if (xfer->rx_buf && mas->rx_se_dma) {
-				geni_se_rx_dma_unprep(se, mas->rx_se_dma, xfer->len);
-				mas->rx_se_dma = 0;
-			}
 			spi_finalize_current_transfer(spi);
 			mas->cur_xfer = NULL;
 		}