On Tue, Sep 01, 2015 at 02:13:01PM +0300, Ville Syrjälä wrote:
> On Tue, Sep 01, 2015 at 11:14:43AM +0200, Daniel Vetter wrote:
> > On Wed, Aug 26, 2015 at 10:55:06PM +0300, ville.syrjala at linux.intel.com
> > wrote:
> > > From: Ville Syrjälä <ville.syrjala at linux.intel.com>
> > >
> > > Calculate the number of retries we should do for each i2c-over-aux
> > > message based on the time it takes to perform the i2c transfer vs. the
> > > aux transfer. We assume the shortest possible length for the aux
> > > transfer, and the longest possible (exluding clock stretching) for the
> > > i2c transfer.
> > >
> > > The DP spec has some examples on how to calculate this, but we don't
> > > calculate things quite the same way. The spec doesn't account for the
> > > retry interval (assumes immediate retry on defer), and doesn't assume
> > > the best/worst case behaviour as we do.
> > >
> > > Note that currently we assume 10 kHz speed for the i2c bus. Some real
> > > world devices (eg. some Apple DP->VGA dongle) fails with less than 16
> > > retries. and that would correspond to something close to 15 kHz (with
> > > our method of calculating things) But let's just go for 10 kHz to be
> > > on the safe side. Ideally we should query/set the i2c bus speed via
> > > DPCD but for now this should at leaast remove the regression from the
> > > 1->16 byte trasnfer size change. And of course if the sink completes
> > > the transfer quicker this shouldn't slow things down since we don't
> > > change the interval between retries.
> > >
> > > I did a few experiments with a DP->DVI dongle I have that allows you
> > > to change the i2c bus speed. Here are the results of me changing the
> > > actual bus speed and the assumed bus speed and seeing when we start
> > > to fail the operation:
> > >
> > > actual i2c khz assumed i2c khz max retries
> > > 1 1 ok -> 2 fail 211 ok -> 106 fail
> > > 5 8 ok -> 9 fail 27 ok -> 24 fail
> > > 10 17 ok -> 18 fail 13 ok -> 12 fail
> > > 100 210 ok -> 211 fail 2 ok -> 1 fail
> > >
> > > So based on that we have a fairly decent safety margin baked into
> > > the formula to calculate the max number of retries.
> > >
> > > Fixes a regression with some DP dongles from:
> > > commit 1d002fa720738bcd0bddb9178e9ea0773288e1dd
> > > Author: Simon Farnsworth <simon.farnsworth at onelan.co.uk>
> > > Date: Tue Feb 10 18:38:08 2015 +0000
> > >
> > > drm/dp: Use large transactions for I2C over AUX
> > >
> > > v2: Use best case for AUX and worst case for i2c (Simon Farnsworth)
> > > Add a define our AUX retry interval and account for it
> > >
> > > Cc: Simon Farnsworth <simon.farnsworth at onelan.com>
> > > Cc: moosotc at gmail.com
> > > Tested-by: moosotc at gmail.com
> > > Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=91451
> > > Signed-off-by: Ville Syrjälä <ville.syrjala at linux.intel.com>
> > > ---
> > > drivers/gpu/drm/drm_dp_helper.c | 81
> > > ++++++++++++++++++++++++++++++++++++++++-
> > > 1 file changed, 79 insertions(+), 2 deletions(-)
> > >
> > > diff --git a/drivers/gpu/drm/drm_dp_helper.c
> > > b/drivers/gpu/drm/drm_dp_helper.c
> > > index 7069e54..23b9fcc 100644
> > > --- a/drivers/gpu/drm/drm_dp_helper.c
> > > +++ b/drivers/gpu/drm/drm_dp_helper.c
> > > @@ -424,6 +424,78 @@ static u32 drm_dp_i2c_functionality(struct
> > > i2c_adapter *adapter)
> > > I2C_FUNC_10BIT_ADDR;
> > > }
> > >
> > > +#define AUX_PRECHARGE_LEN 10 /* 10 to 16 */
> > > +#define AUX_SYNC_LEN (16 + 4) /* preamble + AUX_SYNC_END */
> > > +#define AUX_STOP_LEN 4
> > > +#define AUX_CMD_LEN 4
> > > +#define AUX_ADDRESS_LEN 20
> > > +#define AUX_REPLY_PAD_LEN 4
> > > +#define AUX_LENGTH_LEN 8
> > > +
> > > +/*
> > > + * Calculate the length of the AUX request/reply. Gives the "best"
> > > + * case estimate, ie. successful while as short as possible.
> > > + */
> > > +static int drm_dp_aux_req_len(const struct drm_dp_aux_msg *msg)
> > > +{
> > > + int len = AUX_PRECHARGE_LEN + AUX_SYNC_LEN + AUX_STOP_LEN +
> > > + AUX_CMD_LEN + AUX_ADDRESS_LEN + AUX_LENGTH_LEN;
> > > +
> > > + if ((msg->request & DP_AUX_I2C_READ) == 0)
> > > + len += msg->size * 8;
> > > +
> > > + return len;
> > > +}
> > > +
> > > +static int drm_dp_aux_reply_len(const struct drm_dp_aux_msg *msg)
> > > +{
> > > + int len = AUX_PRECHARGE_LEN + AUX_SYNC_LEN + AUX_STOP_LEN +
> > > + AUX_CMD_LEN + AUX_REPLY_PAD_LEN;
> > > +
> > > + /*
> > > + * For read we expect what was asked. For writes there will
> > > + * be 0 or 1 data bytes. Assume 0 for the "best" case.
> > > + */
> > > + if (msg->request & DP_AUX_I2C_READ)
> > > + len += msg->size * 8;
> > > +
> > > + return len;
> > > +}
> > > +
> > > +#define I2C_START_LEN 1
> > > +#define I2C_STOP_LEN 1
> > > +#define I2C_ADDR_LEN 9 /* ADDRESS + R/W + ACK/NACK */
> > > +#define I2C_DATA_LEN 9 /* DATA + ACK/NACK */
> > > +
> > > +/*
> > > + * Calculate the length of the i2c transfer (in AUX clocks)
> > > + * assuming the i2c bus speed is as specified. Gives the the
> > > + * "worst" case estimate, ie. successful while as long as possible.
> > > + * Doesn't account the the "MOT" bit, and instead assumes each
> > > + * message includes a START, ADDRESS and STOP. Neither does it
> > > + * account for additional random variables such as clock stretching.
> > > + */
> > > +static int drm_dp_i2c_msg_len(const struct drm_dp_aux_msg *msg,
> > > + int i2c_speed_khz)
> > > +{
> > > + return (I2C_START_LEN + I2C_ADDR_LEN + msg->size * I2C_DATA_LEN +
> > > + I2C_STOP_LEN) * 1000 / i2c_speed_khz;
> >
> > This doesn't seem to compute the lenght, but the time a transfer takes, in
> > usec (if I haven't screwed up my numbers ...). Also DIV_ROUND_DOWN to make
> > the defensiveness clear?
>
> Well it started out as length, and then I figured I'll just shove the
> i2c->aux unit conversion in here. It's a length in time if you will ;)
>
> And it should rather be DIV_ROUND_UP() since we want the "maximum" here.
Hm right, mixed that up.
> > > +}
> > > +
> > > +/*
> > > + * Deterine how many retries should be attempted to successfully transfer
> >
> > Deteri_m_e.
> >
> > > + * the specified message, based on the estimated durations of the
> > > + * i2c and AUX transfers.
> > > + */
> > > +static int drm_dp_i2c_retry_count(const struct drm_dp_aux_msg *msg,
> > > + int i2c_speed_khz)
> > > +{
> > > + int aux_len = drm_dp_aux_req_len(msg) + drm_dp_aux_reply_len(msg);
> > > + int i2c_len = drm_dp_i2c_msg_len(msg, i2c_speed_khz);
> > > +
> > > + return DIV_ROUND_UP(i2c_len, aux_len + AUX_RETRY_INTERVAL);
> >
> > So assuming I get things right i2c_len is actually i2c_time_us, but
> > aux_len is still just a length (and since the bus runs at 2MHz it doesn't
> > seem to just add up correctly).
>
> AUX bitrate is 1MHz
Right I remembered only that we need to put in the clock divider for 2MHz
for intel dp aux transfers and got confused by that hw requirement.
Still a comment that the aux clock is 1MHz plus renaming aux_len to
aux_time_usecs would be great for clarity.
Thanks, Daniel
--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
