by Sricharan R

[permalink] [raw]

Subject: RE: [PATCH V4 4/7] i2c: qup: Transfer each i2c_msg in i2c_msgs without a stop bit

Hi Ivan,

> -----Original Message-----
> From: Ivan T. Ivanov [mailto:[email protected]]
> Sent: Monday, July 20, 2015 4:53 PM
> To: Sricharan R
> Cc: [email protected]; [email protected];
> [email protected]; [email protected]; linux-
> [email protected]; [email protected]; [email protected];
> [email protected]
> Subject: Re: [PATCH V4 4/7] i2c: qup: Transfer each i2c_msg in i2c_msgs
> without a stop bit
>
>
> Hi,
>
> On Thu, 2015-07-09 at 08:55 +0530, Sricharan R wrote:
>
> <snip>
>
> > #define ONE_BYTE 0x1
> > +#define QUP_I2C_MX_CONFIG_DURING_RUN BIT(31)
> >
> > struct qup_i2c_block {
> > int count;
> > @@ -121,6 +122,7 @@ struct qup_i2c_block {
> > int rx_tag_len;
> > int data_len;
> > u8 tags[6];
> > + int config_run;
>
> This is not directly related to "block" control logic, right?
> Could it made part of qup_i2c_dev structure?
>
Yes, can move it there.

> > };
> >
> > struct qup_i2c_dev {
> > @@ -152,6 +154,10 @@ struct qup_i2c_dev {
> >
> > int (*qup_i2c_write_one)(struct qup_i2c_dev *qup,
> > struct i2c_msg *msg);
> > + /* Current i2c_msg in i2c_msgs */
> > + int cmsg;
> > + /* total num of i2c_msgs */
> > + int num;
>
> I think it will be simpler with just "bool is_last" evaluated in main xfer loop.

Hmm ok. Will change it that way. Set it one place and use it other places.
>
> <snip>
>
> >
> > @@ -374,6 +383,9 @@ static void qup_i2c_get_blk_data(struct
> qup_i2c_dev *qup,
> > /* There are 2 tag bytes that are read in to fifo for every block */
> > if (msg->flags & I2C_M_RD)
> > qup->blk.rx_tag_len = qup->blk.count * 2;
> > +
> > + if (qup->cmsg)
> > + qup->blk.config_run = QUP_I2C_MX_CONFIG_DURING_RUN;
>
> This could be moved to qup_i2c_xfer_v2() to avoid repeatedly setting it.
Ok, correct.

Regards,
Sricharan

2015-07-21 11:10:54

by Sricharan R

[permalink] [raw]

Subject: RE: [PATCH V4 5/7] i2c: qup: Add bam dma capabilities

Hi Ivan,

> -----Original Message-----
> From: linux-arm-kernel [mailto:linux-arm-kernel-
> [email protected]] On Behalf Of Ivan T. Ivanov
> Sent: Monday, July 20, 2015 8:17 PM
> To: Sricharan R
> Cc: [email protected]; [email protected];
> [email protected]; [email protected]; linux-
> [email protected]; [email protected]; [email protected];
> [email protected]
> Subject: Re: [PATCH V4 5/7] i2c: qup: Add bam dma capabilities
>
>
> Hi Sricharan,
>
> On Thu, 2015-07-09 at 08:55 +0530, Sricharan R wrote:
> > QUP cores can be attached to a BAM module, which acts as a dma engine
> > for the QUP core. When DMA with BAM is enabled, the BAM consumer
> pipe
> > transmitted data is written to the output FIFO and the BAM producer
> > pipe received data is read from the input FIFO.
> >
> > With BAM capabilities, qup-i2c core can transfer more than 256 bytes,
> > without a 'stop' which is not possible otherwise.
> >
> > Signed-off-by: Sricharan R <[email protected]>
> > ---
> > drivers/i2c/busses/i2c-qup.c | 431
> > +++++++++++++++++++++++++++++++++++++++++--
> > 1 file changed, 415 insertions(+), 16 deletions(-)
> >
> > diff --git a/drivers/i2c/busses/i2c-qup.c
> > b/drivers/i2c/busses/i2c-qup.c index c0757d9..810b021 100644
> > --- a/drivers/i2c/busses/i2c-qup.c
> > +++ b/drivers/i2c/busses/i2c-qup.c
> > @@ -24,6 +24,11 @@
> > #include <linux/of.h>
> > #include <linux/platform_device.h>
> > #include <linux/pm_runtime.h>
> > +#include <linux/dma-mapping.h>
> > +#include <linux/scatterlist.h>
> > +#include <linux/atomic.h>
> > +#include <linux/dmaengine.h>
> > +#include <linux/dmapool.h>
>
> Keep includes sorted alphabetically.
Ok.

>
> <snip>
>
> > +#define MX_TX_RX_LEN SZ_64K
> > +#define MX_BLOCKS (MX_TX_RX_LEN / QUP_READ_LIMIT)
> > +
> > +/* Max timeout in ms for 32k bytes */
> > +#define TOUT_MAX 300
> > +
> > struct qup_i2c_block {
> > int count;
> > int pos;
> > @@ -125,6 +143,23 @@ struct qup_i2c_block {
> > int config_run;
> > };
> >
> > +struct qup_i2c_tag {
> > + u8 *start;
> > + dma_addr_t addr;
> > +};
> > +
> > +struct qup_i2c_bam_rx {
> > + struct qup_i2c_tag scratch_tag;
> > + struct dma_chan *dma_rx;
> > + struct scatterlist *sg_rx;
> > +};
> > +
> > +struct qup_i2c_bam_tx {
> > + struct qup_i2c_tag footer_tag;
> > + struct dma_chan *dma_tx;
> > + struct scatterlist *sg_tx;
> > +};
> > +
>
> The only difference between above 2 structures is name of the fields.
> Please, just define one struct qup_i2c_bam and instantiate it twice.
Ok.

>
> > struct qup_i2c_dev {
> > struct device*dev;
> > void __iomem*base;
> > @@ -154,14 +189,20 @@ struct qup_i2c_dev {
> >
> > int (*qup_i2c_write_one)(struct qup_i2c_dev *qup,
> > struct i2c_msg *msg);
> > + int (*qup_i2c_read_one)(struct qup_i2c_dev *qup,
> > + struct i2c_msg *msg);
> > +
> > /* Current i2c_msg in i2c_msgs */
> > int cmsg;
> > /* total num of i2c_msgs */
> > int num;
> >
> > - int (*qup_i2c_read_one)(struct qup_i2c_dev *qup,
> > - struct i2c_msg *msg);
> > -
> > + /* dma parameters */
> > + bool is_dma;
> > + struct dma_pool *dpool;
> > + struct qup_i2c_tag start_tag;
> > + struct qup_i2c_bam_rx brx;
> > + struct qup_i2c_bam_tx btx;
> > struct completionxfer;
> > };
> >
> > @@ -238,6 +279,14 @@ static int qup_i2c_poll_state(struct qup_i2c_dev
> *qup, u32 req_state)
> > return qup_i2c_poll_state_mask(qup, req_state,
> > QUP_STATE_MASK); }
> >
> > +static void qup_i2c_flush(struct qup_i2c_dev *qup) {
> > + u32 val = readl(qup->base + QUP_STATE);
> > +
> > + val |= QUP_I2C_FLUSH;
> > + writel(val, qup->base + QUP_STATE); }
> > +
>
> Used in only one place.
So you mean no need to separate this as function ?

>
> <snip>
>
> >
> > +static void qup_i2c_bam_cb(void *data) {
> > + struct qup_i2c_dev *qup = data;
> > +
> > + complete(&qup->xfer);
> > +}
> > +
> > +void qup_sg_set_buf(struct scatterlist *sg, void *buf, struct
qup_i2c_tag
> *tg,
> > + unsigned int buflen,
struct qup_i2c_dev *qup,
> > + int map, int dir) {
> > + sg_set_buf(sg, buf, buflen);
> > + dma_map_sg(qup->dev, sg, 1, dir);
> > +
> > + if (!map)
> > + sg_dma_address(sg) = tg->addr + ((u8 *)buf -
> > + tg->start);
>
> Changing DMA address that we just mapped?

The reason was tags that were allocated during probe
using dma_alloc could also be from highmem. dma_map_sg
assumes linear kernel addresses for virtual addresses to be mapped.
So the sg->addr set using sg_set_buf goes wrong for such addresses.
sg->len is fine for both cases. So setting the addresses explicitly
here.

>
> > +}
> > +
> > +static void qup_i2c_rel_dma(struct qup_i2c_dev *qup) {
> > + if (qup->btx.dma_tx)
> > + dma_release_channel(qup->btx.dma_tx);
> > + if (qup->brx.dma_rx)
> > + dma_release_channel(qup->brx.dma_rx);
> > + qup->btx.dma_tx = NULL;
> > + qup->brx.dma_rx = NULL;
> > +}
> > +
> > +static int qup_i2c_req_dma(struct qup_i2c_dev *qup) {
> > + if (!qup->btx.dma_tx) {
> > + qup->btx.dma_tx = dma_request_slave_channel(qup->dev,
> > +"tx");
>
> Please use dma_request_slave_channel_reason() and let deferred probe
> work.
Ok, right. Did this for something else previously. Missed it here. Will
change this.

>
> > + if (!qup->btx.dma_tx) {
> > + dev_err(qup->dev, "\n tx channel not
available");
> > + return -ENODEV;
> > + }
> > + }
> > +
> > + if (!qup->brx.dma_rx) {
> > + qup->brx.dma_rx = dma_request_slave_channel(qup->dev,
"rx");
> > + if (!qup->brx.dma_rx) {
> > + dev_err(qup->dev, "\n rx channel not
available");
> > + qup_i2c_rel_dma(qup);
> > + return -ENODEV;
> > + }
> > + }
> > + return 0;
> > +}
> > +
> > +static int bam_do_xfer(struct qup_i2c_dev *qup, struct i2c_msg *msg)
> > +{
>
> Please use consistent naming convention.
Ok.

>
> > + struct dma_async_tx_descriptor *txd, *rxd = NULL;
> > + int ret = 0;
> > + dma_cookie_t cookie_rx, cookie_tx;
> > + u32 rx_nents = 0, tx_nents = 0, len, blocks, rem;
> > + u32 i, tlen, tx_len, tx_buf = 0, rx_buf = 0, off = 0;
> > + u8 *tags;
> > +
> > + while (qup->cmsg < qup->num) {
> > + blocks = (msg->len + QUP_READ_LIMIT) / QUP_READ_LIMIT;
> > + rem = msg->len % QUP_READ_LIMIT;
> > + tx_len = 0, len = 0, i = 0;
> > +
> > + qup_i2c_get_blk_data(qup, msg);
> > +
> > + if (msg->flags & I2C_M_RD) {
> > + rx_nents += (blocks * 2) + 1;
> > + tx_nents += 1;
> > +
> > + while (qup->blk.pos < blocks) {
> > + /* length set to '0' implies 256 bytes
*/
> > + tlen = (i == (blocks - 1)) ? rem : 0;
> > + tags = &qup->start_tag.start[off + len];
> > + len += qup_i2c_get_tags(tags, qup,
> > + msg, 1);
> > +
> > + /* scratch buf to read the start and len
tags */
> > +
qup_sg_set_buf(&qup->brx.sg_rx[rx_buf++],
> > +
&qup-
> >brx.scratch_tag.start[0],
> > +
&qup-
> >brx.scratch_tag,
> > +
> > + 2, qup, 0, 0);
> > +
> > +
qup_sg_set_buf(&qup->brx.sg_rx[rx_buf++],
> > +
&msg-
> >buf[QUP_READ_LIMIT * i],
> > +
NULL, tlen, qup,
> > +
1,
> DMA_FROM_DEVICE);
> > + i++;
> > + qup->blk.pos = i;
> > + }
> > + qup_sg_set_buf(&qup->btx.sg_tx[tx_buf++],
> > +
&qup-
> >start_tag.start[off],
> > +
&qup->start_tag, len, qup,
> 0, 0);
> > + off += len;
> > + /* scratch buf to read the BAM EOT and FLUSH
tags */
> > + qup_sg_set_buf(&qup->brx.sg_rx[rx_buf++],
> > +
&qup-
> >brx.scratch_tag.start[0],
> > +
&qup->brx.scratch_tag, 2,
> > +
qup, 0, 0);
> > + } else {
> > + tx_nents += (blocks * 2);
> > +
> > + while (qup->blk.pos < blocks) {
> > + tlen = (i == (blocks - 1)) ? rem : 0;
> > + tags = &qup->start_tag.start[off +
tx_len];
> > + len = qup_i2c_get_tags(tags, qup, msg,
> > + 1);
> > +
> > +
qup_sg_set_buf(&qup->btx.sg_tx[tx_buf++],
> > +
tags,
> > +
&qup->start_tag, len,
> > +
> > + qup, 0, 0);
> > +
> > + tx_len += len;
> > +
qup_sg_set_buf(&qup->btx.sg_tx[tx_buf++],
> > +
&msg-
> >buf[QUP_READ_LIMIT * i],
> > +
NULL, tlen, qup, 1,
> > +
DMA_TO_DEVICE);
> > + i++;
> > + qup->blk.pos = i;
> > + }
> > + off += tx_len;
> > +
> > + if (qup->cmsg == (qup->num - 1)) {
> > + qup->btx.footer_tag.start[0] =
> > +
QUP_BAM_FLUSH_STOP;
> > + qup->btx.footer_tag.start[1] =
> > +
QUP_BAM_FLUSH_STOP;
> > +
qup_sg_set_buf(&qup->btx.sg_tx[tx_buf++],
> > +
&qup-
> >btx.footer_tag.start[0],
> > +
&qup-
> >btx.footer_tag, 2,
> > +
qup, 0, 0);
> > + tx_nents += 1;
> > + }
> > + }
> > + qup->cmsg++;
> > + msg++;
> > + }
> > +
> > + txd = dmaengine_prep_slave_sg(qup->btx.dma_tx, qup->btx.sg_tx,
> tx_nents,
> > +
DMA_MEM_TO_DEV,
> > +
DMA_PREP_INTERRUPT |
> DMA_PREP_FENCE);
> > + if (!txd) {
> > + dev_err(qup->dev, "failed to get tx desc\n");
> > + ret = -EINVAL;
> > + goto desc_err;
> > + }
> > +
> > + if (!rx_nents) {
> > + txd->callback = qup_i2c_bam_cb;
> > + txd->callback_param = qup;
> > + }
> > +
> > + cookie_tx = dmaengine_submit(txd);
>
> Check this cookie for error? Same bellow.
Ok.

>
> > + dma_async_issue_pending(qup->btx.dma_tx);
>
> Why TX messages are started first?
For rx to start, tx pipe of tags have to start first.
So tx always have to happen and rx happens only when
there are rx_nents as below.

>
> > +
> > + if (rx_nents) {
> > + rxd = dmaengine_prep_slave_sg(qup->brx.dma_rx, qup-
> >brx.sg_rx,
> > +
rx_nents,
> DMA_DEV_TO_MEM,
> > +
DMA_PREP_INTERRUPT);
> > + if (!rxd) {
> > + dev_err(qup->dev, "failed to get rx desc\n");
> > + ret = -EINVAL;
> > +
> > + /* abort TX descriptors */
> > + dmaengine_terminate_all(qup->btx.dma_tx);
> > + goto desc_err;
> > + }
> > +
> > + rxd->callback = qup_i2c_bam_cb;
> > + rxd->callback_param = qup;
> > + cookie_rx = dmaengine_submit(rxd);
> > + dma_async_issue_pending(qup->brx.dma_rx);
> > + }
> > +
> > + if (!wait_for_completion_timeout(&qup->xfer, TOUT_MAX * HZ)) {
> > + dev_err(qup->dev, "normal trans timed out\n");
> > + ret = -ETIMEDOUT;
> > + }
>
> There could be multiple messages for RX and TX queued for transfer, and
> they could be mixed. Using just one completion didn't look right.

Even though there are multiple message mixed, all of them are packed as
one tx and rx descriptor finally. Two things here,

1) when only tx pipe is used, we set the txd->callback which gets
fired after
writing the last desc which has a 'EOT' bit set.

2) For rx, we set the callback only for rx. Because
for any RX transfer tx is used for writing tags. So when doing a

read command, qup signals a completion on tx and rx pipe when it
sees a
'BAM_INPUT_EOT' for rx. So rx callback being called for read
implies
corresponding tx is also done. I have seen this while doing
reads,
tx and rx bam interrupts fire together. So even when mixed TX
and RX are there
getting a callback for the last RX data implies a completion of
all.

Having said above things, one thing missing is if tx is last
message
then BAM_INPUT_EOT tag should be sent there as well. I did not
have
a case where tx was the last. But will have to just add that tag
in above
for tx. But not sure if there are clients to test all
combinations.
>
> > +
> > + if (ret || qup->bus_err || qup->qup_err) {
> > + if (qup->bus_err & QUP_I2C_NACK_FLAG) {
> > + msg--;
> > + dev_err(qup->dev, "NACK from %x\n", msg->addr);
> > + ret = -EIO;
> > +
> > + if (qup_i2c_change_state(qup, QUP_RUN_STATE)) {
> > + dev_err(qup->dev, "change to run state
timed out");
> > + return ret;
> > + }
> > +
> > + if (rx_nents)
> > + writel(QUP_BAM_INPUT_EOT,
> > +
> > + qup->base + QUP_OUT_FIFO_BASE);
> > +
> > + writel(QUP_BAM_FLUSH_STOP,
> > +
> > + qup->base + QUP_OUT_FIFO_BASE);
> > +
> > + qup_i2c_flush(qup);
> > +
> > + /* wait for remaining interrupts to occur */
> > + if (!wait_for_completion_timeout(&qup->xfer,
HZ))
> > + dev_err(qup->dev, "flush timed
> > + out\n");
> > +
> > + qup_i2c_rel_dma(qup);
> > + }
> > + }
> > +
> > + dma_unmap_sg(qup->dev, qup->btx.sg_tx, tx_nents,
> > + DMA_TO_DEVICE);
> > +
> > + if (rx_nents)
> > + dma_unmap_sg(qup->dev, qup->brx.sg_rx, rx_nents,
> > +
> > +DMA_FROM_DEVICE);
> > +desc_err:
> > + return ret;
> > +}
> > +
> > +static int qup_bam_xfer(struct i2c_adapter *adap, struct i2c_msg
> > +*msg) {
>
> Please use consistent naming convention.
Ok.

>
> > + struct qup_i2c_dev *qup = i2c_get_adapdata(adap);
> > + int ret = 0;
> > +
> > + enable_irq(qup->irq);
> > + if (qup_i2c_req_dma(qup))
>
> Why?
Have to request it again because, when the device 'Nacks out'
on the previous transfers, then the bam channels used for that transfer
have to be reset and init to be used again. Release and requesting them
does this.

>
> >
> > +
> > static int qup_i2c_xfer(struct i2c_adapter *adap,
> > struct i2c_msg msgs[],
> > int num)
> > @@ -836,7 +1143,7 @@ static int qup_i2c_xfer_v2(struct i2c_adapter
> *adap,
> > int num) {
> > struct qup_i2c_dev *qup = i2c_get_adapdata(adap);
> > - int ret, idx;
> > + int ret, idx, len, use_dma = 0;
> >
> > qup->num = num;
> > qup->cmsg = 0;
> > @@ -854,12 +1161,27 @@ static int qup_i2c_xfer_v2(struct i2c_adapter
> *adap,
> > writel(I2C_MINI_CORE | I2C_N_VAL_V2, qup->base + QUP_CONFIG);
> > writel(QUP_V2_TAGS_EN, qup->base + QUP_I2C_MASTER_GEN);
> >
> > - for (idx = 0; idx < num; idx++) {
> > - if (msgs[idx].len == 0) {
> > - ret = -EINVAL;
> > - goto out;
> > + if ((qup->is_dma)) {
> > + /* All i2c_msgs should be transferred using either dma
or cpu */
> > + for (idx = 0; idx < num; idx++) {
> > + if (msgs[idx].len == 0) {
> > + ret = -EINVAL;
> > + goto out;
> > + }
> > +
> > + len = (msgs[idx].len > qup->out_fifo_sz) ||
> > +
> > + (msgs[idx].len > qup->in_fifo_sz);
> > +
> > + if ((!is_vmalloc_addr(msgs[idx].buf)) && len)
> > + {
>
> What is wrong with vmalloc addresses?
When not physically contiguous, then we will have to again split it in
to smaller sg buffers and then DMA them. This check was done to avoid
that
extra level and use cpu transfers to keep it simple.

>
> > + use_dma = 1;
> > + } else {
> > + use_dma = 0;
> > + break;
> > + }
> > }
> > + }
> >
> > + for (idx = 0; idx < num; idx++) {
> > if (qup_i2c_poll_state_i2c_master(qup)) {
> > ret = -EIO;
> > goto out;
> > @@ -867,10 +1189,17 @@ static int qup_i2c_xfer_v2(struct i2c_adapter
> > *adap,
> >
> > reinit_completion(&qup->xfer);
> >
> > - if (msgs[idx].flags & I2C_M_RD)
> > - ret = qup_i2c_read(qup, &msgs[idx]);
> > - else
> > - ret = qup_i2c_write(qup, &msgs[idx]);
> > + len = msgs[idx].len;
>
> Unused?
Oh right. Will remove this above assignment.
>
> > +
> > + if (use_dma) {
> > + ret = qup_bam_xfer(adap, &msgs[idx]);
> > + idx = num;
>
> Hacky.
Hmm, because looping is not required for BAM.
Will change this to look correct.
>
> > + } else {
> > + if (msgs[idx].flags & I2C_M_RD)
> > + ret = qup_i2c_read(qup, &msgs[idx]);
> > + else
> > + ret = qup_i2c_write(qup, &msgs[idx]);
> > + }
> >
> > if (ret)
> > break;
> > @@ -943,6 +1272,7 @@ static int qup_i2c_probe(struct platform_device
> *pdev)
> > int ret, fs_div, hs_div;
> > int src_clk_freq;
> > u32 clk_freq = 100000;
> > + int blocks;
> >
> > qup = devm_kzalloc(&pdev->dev, sizeof(*qup), GFP_KERNEL);
> > if (!qup)
> > @@ -963,8 +1293,60 @@ static int qup_i2c_probe(struct platform_device
> *pdev)
> > qup->qup_i2c_write_one = qup_i2c_write_one_v2;
> > qup->qup_i2c_read_one = qup_i2c_read_one_v2;
> > qup->use_v2_tags = 1;
> > +
> > + if (qup_i2c_req_dma(qup))
> > + goto nodma;
>
> Just return what request DMA functions return?
You mean return when DMA functions returns EPROBE_DEFER and
continue if it errors out.

Regards,
Sricharan