Message ID | 1505903026.7865.6.camel@pengutronix.de (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
On 20 September 2017 at 11:23, Philipp Zabel <p.zabel@pengutronix.de> wrote: > Hi, > > On Wed, 2017-09-20 at 10:14 +0100, Dave Stevenson wrote: >> Hi Mauro & Philipp >> >> On 19 September 2017 at 17:49, Mauro Carvalho Chehab >> <mchehab@s-opensource.com> wrote: >> > Em Tue, 19 Sep 2017 17:24:45 +0200 >> > Philipp Zabel <p.zabel@pengutronix.de> escreveu: >> > >> > > Hi Dave, >> > > >> > > On Tue, 2017-09-19 at 14:08 +0100, Dave Stevenson wrote: >> > > > The existing fixed value of 16 worked for UYVY 720P60 over >> > > > 2 lanes at 594MHz, or UYVY 1080P60 over 4 lanes. (RGB888 >> > > > 1080P60 needs 6 lanes at 594MHz). >> > > > It doesn't allow for lower resolutions to work as the FIFO >> > > > underflows. >> > > > >> > > > Using a value of 300 works for all resolutions down to VGA60, >> > > > and the increase in frame delay is <4usecs for 1080P60 UYVY >> > > > (2.55usecs for RGB888). >> > > > >> > > > Signed-off-by: Dave Stevenson <dave.stevenson@raspberrypi.org> >> > > >> > > Can we increase this to 320? This would also allow >> > > 720p60 at 594 Mbps / 4 lanes, according to the xls. >> >> Unless I've missed something then the driver would currently request >> only 2 lanes for 720p60 UYVY, and that works with the existing FIFO >> setting of 16. Likewise 720p60 RGB888 requests 3 lanes and also works >> on a FIFO setting of 16. >> How/why were you thinking we need to run all four lanes for 720p60 >> without other significant driver mods around lane config? > > The driver currently silently changes the number of active lanes > depending on required data rate, with no way to communicate it to the > receiver. It is communicated over the subdevice API - tc358743_g_mbus_config reports back the appropriate number of lanes to the receiver subdevice. A suitable v4l2_subdev_has_op(dev->sensor, video, g_mbus_config) call as you're starting streaming therefore gives you the correct information. That's what I've just done for the BCM283x Unicam driver[1], but admittedly I'm not using the media controller API which i.MX6 is. [1] http://www.spinics.net/lists/linux-media/msg121813.html, as part of the unicam_start_streaming function. > The i.MX6 MIPI CSI-2 receiver driver can't cope with that, as it always > activates all four lanes that are configured in the device tree. I can > work around that with the following patch: It can't cope running at less than 4 lanes, or it can't cope with a change? > ----------8<---------- > Subject: [PATCH] [media] tc358743: do not dynamically reduce number of lanes > > Dynamic lane number reduction does not work with receivers that > configure a fixed lane number according to the device tree settings. > To allow 720p60 at 594 Mbit/s on 4 lanes, increase the fifo_level > and tclk_trailcnt settings. > > Signed-off-by: Philipp Zabel <p.zabel@pengutronix.de> > --- > drivers/media/i2c/tc358743.c | 6 +++--- > 1 file changed, 3 insertions(+), 3 deletions(-) > > diff --git a/drivers/media/i2c/tc358743.c b/drivers/media/i2c/tc358743.c > index 64f504542a819..70a9435928cdb 100644 > --- a/drivers/media/i2c/tc358743.c > +++ b/drivers/media/i2c/tc358743.c > @@ -683,7 +683,7 @@ static void tc358743_set_csi(struct v4l2_subdev *sd) > { > struct tc358743_state *state = to_state(sd); > struct tc358743_platform_data *pdata = &state->pdata; > - unsigned lanes = tc358743_num_csi_lanes_needed(sd); > + unsigned lanes = state->bus.num_data_lanes; > > v4l2_dbg(3, debug, sd, "%s:\n", __func__); > > @@ -1906,7 +1906,7 @@ static int tc358743_probe_of(struct tc358743_state *state) > state->pdata.ddc5v_delay = DDC5V_DELAY_100_MS; > state->pdata.enable_hdcp = false; > /* A FIFO level of 16 should be enough for 2-lane 720p60 at 594 MHz. */ > - state->pdata.fifo_level = 16; > + state->pdata.fifo_level = 320; > /* > * The PLL input clock is obtained by dividing refclk by pll_prd. > * It must be between 6 MHz and 40 MHz, lower frequency is better. > @@ -1948,7 +1948,7 @@ static int tc358743_probe_of(struct tc358743_state *state) > state->pdata.lptxtimecnt = 0x003; > /* tclk-preparecnt: 3, tclk-zerocnt: 20 */ > state->pdata.tclk_headercnt = 0x1403; > - state->pdata.tclk_trailcnt = 0x00; > + state->pdata.tclk_trailcnt = 0x01; > /* ths-preparecnt: 3, ths-zerocnt: 1 */ > state->pdata.ths_headercnt = 0x0103; > state->pdata.twakeup = 0x4882; > -- > 2.11.0 > ---------->8---------- > > Just adding the same heuristic as tc358743_num_csi_lanes_needed in the > imx6-mipi-csi2 driver doesn't work, as the heuristic is specific to the > Toshiba chip. There are MIPI CSI-2 sensors that only support a fixed > number of lanes, for example. > > I'd need a way to communicate the number of active MIPI CSI-2 lanes > between transmitting and receiving subdevice driver. > >> Once I've got a v3 done on the Unicam driver I'll bash through the >> standard HDMI modes and check what value they need - I can see a big >> spreadsheet coming on. > > Oh dear. Unless the point you make below can be resolved, I think we > have no other choice. > >> I'll ignore interlaced modes as I can't see any support for it in the >> driver. Receiving the fields on different CSI-2 data types is >> something I know the Unicam hardware won't handle nicely, and I >> suspect it'd be an issue for many other platforms too. > > Yes, let's pretend interlacing doesn't exist as long as possible. > >> > Hmm... if this is dependent on the resolution and frame rate, >> > wouldn't >> > it be better to dynamically adjust it accordingly? >> >> It's setting up the FIFO matching the incoming HDMI data rate and >> outgoing CSI rate. That means it's dependent on the incoming pixel >> clock, blanking, colour format and resolution, and output CSI link >> frequency, number of lanes, and colour format. >> Whilst it could be set dynamically based on all those parameters, is >> there a significant enough gain in doing so? > > Ideally there would be no need to maintain a timing database with - > worst case - (number of link frequencies * number of HDMI modes) entries > in the driver ... > >> The value of 300 works for all cases I've tried, and referencing back >> it is also the value that Hans said Cisco use via platform data on >> their hardware [1]. Generally I'm seeing that values of 0-130 are >> required, so 300 is giving a fair safety margin. > > It does not work for 720p60 on 4 lanes at 594 Mbit/s, as the spreadsheet > warns, and testing shows. If it doesn't work with 720p60, then I guess it has no hope with many other resolutions. It sounds like confirming whether g_mbus_config is a potential solution for i.MX6 (sorry I'm not familiar enough with that code to do my own quick search), but otherwise cranking it up to 320 is reasonable, and I'll see what other numbers fall out of the spreadsheet. >> Second question is does anyone have a suitable relationship with >> Toshiba to get permission to release details of these register >> calculations? The datasheet and value spreadsheet are marked as >> confidential, and probably under NDA in almost all cases. Whilst they >> can't object to drivers containing values to make them work, they >> might over releasing significant details. > > Unfortunately, I don't. I'm guessing Cisco may be in the best position to try pushing that. Certainly all our info is under NDA, and we don't really have a significant working relationship with them. Dave
On 09/20/17 13:00, Dave Stevenson wrote: > On 20 September 2017 at 11:23, Philipp Zabel <p.zabel@pengutronix.de> wrote: >> Hi, >> >> On Wed, 2017-09-20 at 10:14 +0100, Dave Stevenson wrote: >>> Hi Mauro & Philipp >>> >>> On 19 September 2017 at 17:49, Mauro Carvalho Chehab >>> <mchehab@s-opensource.com> wrote: >>>> Em Tue, 19 Sep 2017 17:24:45 +0200 >>>> Philipp Zabel <p.zabel@pengutronix.de> escreveu: >>>> >>>>> Hi Dave, >>>>> >>>>> On Tue, 2017-09-19 at 14:08 +0100, Dave Stevenson wrote: >>>>>> The existing fixed value of 16 worked for UYVY 720P60 over >>>>>> 2 lanes at 594MHz, or UYVY 1080P60 over 4 lanes. (RGB888 >>>>>> 1080P60 needs 6 lanes at 594MHz). >>>>>> It doesn't allow for lower resolutions to work as the FIFO >>>>>> underflows. >>>>>> >>>>>> Using a value of 300 works for all resolutions down to VGA60, >>>>>> and the increase in frame delay is <4usecs for 1080P60 UYVY >>>>>> (2.55usecs for RGB888). >>>>>> >>>>>> Signed-off-by: Dave Stevenson <dave.stevenson@raspberrypi.org> >>>>> >>>>> Can we increase this to 320? This would also allow >>>>> 720p60 at 594 Mbps / 4 lanes, according to the xls. >>> >>> Unless I've missed something then the driver would currently request >>> only 2 lanes for 720p60 UYVY, and that works with the existing FIFO >>> setting of 16. Likewise 720p60 RGB888 requests 3 lanes and also works >>> on a FIFO setting of 16. >>> How/why were you thinking we need to run all four lanes for 720p60 >>> without other significant driver mods around lane config? >> >> The driver currently silently changes the number of active lanes >> depending on required data rate, with no way to communicate it to the >> receiver. > > It is communicated over the subdevice API - tc358743_g_mbus_config > reports back the appropriate number of lanes to the receiver > subdevice. > A suitable v4l2_subdev_has_op(dev->sensor, video, g_mbus_config) call > as you're starting streaming therefore gives you the correct > information. That's what I've just done for the BCM283x Unicam > driver[1], but admittedly I'm not using the media controller API which > i.MX6 is. Shouldn't this information come from the device tree? The g_mbus_config op is close to being deprecated or even removed. There are currently only two obscure V4L2 bridge drivers that call it. It dates from pre-DT times I rather not see it used in a new bridge driver. The problem is that contains data that belongs to the DT (hardware capabilities). Things that can actually change dynamically should be communicated via another op. We don't have that, so that should be created. I've CC-ed Sakari, he is the specialist for such things. > > [1] http://www.spinics.net/lists/linux-media/msg121813.html, as part > of the unicam_start_streaming function. > >> The i.MX6 MIPI CSI-2 receiver driver can't cope with that, as it always >> activates all four lanes that are configured in the device tree. I can >> work around that with the following patch: > > It can't cope running at less than 4 lanes, or it can't cope with a change? > >> ----------8<---------- >> Subject: [PATCH] [media] tc358743: do not dynamically reduce number of lanes >> >> Dynamic lane number reduction does not work with receivers that >> configure a fixed lane number according to the device tree settings. >> To allow 720p60 at 594 Mbit/s on 4 lanes, increase the fifo_level >> and tclk_trailcnt settings. >> >> Signed-off-by: Philipp Zabel <p.zabel@pengutronix.de> >> --- >> drivers/media/i2c/tc358743.c | 6 +++--- >> 1 file changed, 3 insertions(+), 3 deletions(-) >> >> diff --git a/drivers/media/i2c/tc358743.c b/drivers/media/i2c/tc358743.c >> index 64f504542a819..70a9435928cdb 100644 >> --- a/drivers/media/i2c/tc358743.c >> +++ b/drivers/media/i2c/tc358743.c >> @@ -683,7 +683,7 @@ static void tc358743_set_csi(struct v4l2_subdev *sd) >> { >> struct tc358743_state *state = to_state(sd); >> struct tc358743_platform_data *pdata = &state->pdata; >> - unsigned lanes = tc358743_num_csi_lanes_needed(sd); >> + unsigned lanes = state->bus.num_data_lanes; >> >> v4l2_dbg(3, debug, sd, "%s:\n", __func__); >> >> @@ -1906,7 +1906,7 @@ static int tc358743_probe_of(struct tc358743_state *state) >> state->pdata.ddc5v_delay = DDC5V_DELAY_100_MS; >> state->pdata.enable_hdcp = false; >> /* A FIFO level of 16 should be enough for 2-lane 720p60 at 594 MHz. */ >> - state->pdata.fifo_level = 16; >> + state->pdata.fifo_level = 320; >> /* >> * The PLL input clock is obtained by dividing refclk by pll_prd. >> * It must be between 6 MHz and 40 MHz, lower frequency is better. >> @@ -1948,7 +1948,7 @@ static int tc358743_probe_of(struct tc358743_state *state) >> state->pdata.lptxtimecnt = 0x003; >> /* tclk-preparecnt: 3, tclk-zerocnt: 20 */ >> state->pdata.tclk_headercnt = 0x1403; >> - state->pdata.tclk_trailcnt = 0x00; >> + state->pdata.tclk_trailcnt = 0x01; >> /* ths-preparecnt: 3, ths-zerocnt: 1 */ >> state->pdata.ths_headercnt = 0x0103; >> state->pdata.twakeup = 0x4882; >> -- >> 2.11.0 >> ---------->8---------- >> >> Just adding the same heuristic as tc358743_num_csi_lanes_needed in the >> imx6-mipi-csi2 driver doesn't work, as the heuristic is specific to the >> Toshiba chip. There are MIPI CSI-2 sensors that only support a fixed >> number of lanes, for example. >> >> I'd need a way to communicate the number of active MIPI CSI-2 lanes >> between transmitting and receiving subdevice driver. >> >>> Once I've got a v3 done on the Unicam driver I'll bash through the >>> standard HDMI modes and check what value they need - I can see a big >>> spreadsheet coming on. >> >> Oh dear. Unless the point you make below can be resolved, I think we >> have no other choice. >> >>> I'll ignore interlaced modes as I can't see any support for it in the >>> driver. Receiving the fields on different CSI-2 data types is >>> something I know the Unicam hardware won't handle nicely, and I >>> suspect it'd be an issue for many other platforms too. >> >> Yes, let's pretend interlacing doesn't exist as long as possible. >> >>>> Hmm... if this is dependent on the resolution and frame rate, >>>> wouldn't >>>> it be better to dynamically adjust it accordingly? >>> >>> It's setting up the FIFO matching the incoming HDMI data rate and >>> outgoing CSI rate. That means it's dependent on the incoming pixel >>> clock, blanking, colour format and resolution, and output CSI link >>> frequency, number of lanes, and colour format. >>> Whilst it could be set dynamically based on all those parameters, is >>> there a significant enough gain in doing so? >> >> Ideally there would be no need to maintain a timing database with - >> worst case - (number of link frequencies * number of HDMI modes) entries >> in the driver ... >> >>> The value of 300 works for all cases I've tried, and referencing back >>> it is also the value that Hans said Cisco use via platform data on >>> their hardware [1]. Generally I'm seeing that values of 0-130 are >>> required, so 300 is giving a fair safety margin. >> >> It does not work for 720p60 on 4 lanes at 594 Mbit/s, as the spreadsheet >> warns, and testing shows. > > If it doesn't work with 720p60, then I guess it has no hope with many > other resolutions. > It sounds like confirming whether g_mbus_config is a potential > solution for i.MX6 (sorry I'm not familiar enough with that code to do > my own quick search), but otherwise cranking it up to 320 is > reasonable, and I'll see what other numbers fall out of the > spreadsheet. > >>> Second question is does anyone have a suitable relationship with >>> Toshiba to get permission to release details of these register >>> calculations? The datasheet and value spreadsheet are marked as >>> confidential, and probably under NDA in almost all cases. Whilst they >>> can't object to drivers containing values to make them work, they >>> might over releasing significant details. >> >> Unfortunately, I don't. > > I'm guessing Cisco may be in the best position to try pushing that. > Certainly all our info is under NDA, and we don't really have a > significant working relationship with them. I think we tried at the time, but no luck. Regards, Hans > > Dave >
On 20 September 2017 at 12:24, Hans Verkuil <hansverk@cisco.com> wrote: > On 09/20/17 13:00, Dave Stevenson wrote: >> On 20 September 2017 at 11:23, Philipp Zabel <p.zabel@pengutronix.de> wrote: >>> Hi, >>> >>> On Wed, 2017-09-20 at 10:14 +0100, Dave Stevenson wrote: >>>> Hi Mauro & Philipp >>>> >>>> On 19 September 2017 at 17:49, Mauro Carvalho Chehab >>>> <mchehab@s-opensource.com> wrote: >>>>> Em Tue, 19 Sep 2017 17:24:45 +0200 >>>>> Philipp Zabel <p.zabel@pengutronix.de> escreveu: >>>>> >>>>>> Hi Dave, >>>>>> >>>>>> On Tue, 2017-09-19 at 14:08 +0100, Dave Stevenson wrote: >>>>>>> The existing fixed value of 16 worked for UYVY 720P60 over >>>>>>> 2 lanes at 594MHz, or UYVY 1080P60 over 4 lanes. (RGB888 >>>>>>> 1080P60 needs 6 lanes at 594MHz). >>>>>>> It doesn't allow for lower resolutions to work as the FIFO >>>>>>> underflows. >>>>>>> >>>>>>> Using a value of 300 works for all resolutions down to VGA60, >>>>>>> and the increase in frame delay is <4usecs for 1080P60 UYVY >>>>>>> (2.55usecs for RGB888). >>>>>>> >>>>>>> Signed-off-by: Dave Stevenson <dave.stevenson@raspberrypi.org> >>>>>> >>>>>> Can we increase this to 320? This would also allow >>>>>> 720p60 at 594 Mbps / 4 lanes, according to the xls. >>>> >>>> Unless I've missed something then the driver would currently request >>>> only 2 lanes for 720p60 UYVY, and that works with the existing FIFO >>>> setting of 16. Likewise 720p60 RGB888 requests 3 lanes and also works >>>> on a FIFO setting of 16. >>>> How/why were you thinking we need to run all four lanes for 720p60 >>>> without other significant driver mods around lane config? >>> >>> The driver currently silently changes the number of active lanes >>> depending on required data rate, with no way to communicate it to the >>> receiver. >> >> It is communicated over the subdevice API - tc358743_g_mbus_config >> reports back the appropriate number of lanes to the receiver >> subdevice. >> A suitable v4l2_subdev_has_op(dev->sensor, video, g_mbus_config) call >> as you're starting streaming therefore gives you the correct >> information. That's what I've just done for the BCM283x Unicam >> driver[1], but admittedly I'm not using the media controller API which >> i.MX6 is. > > Shouldn't this information come from the device tree? The g_mbus_config > op is close to being deprecated or even removed. There are currently only > two obscure V4L2 bridge drivers that call it. It dates from pre-DT times > I rather not see it used in a new bridge driver. > > The problem is that contains data that belongs to the DT (hardware > capabilities). Things that can actually change dynamically should be > communicated via another op. We don't have that, so that should be created. > > I've CC-ed Sakari, he is the specialist for such things. You've reminded me that I asked that same question earlier in the year, and Sakari had replied - http://www.spinics.net/lists/linux-media/msg115550.html Is it specifically device tree related? Just because the lanes are physically there doesn't necessarily mean they have to be used. A quick test with the spreadsheet appears to say that 1080p24 UYVY over 4 lanes at 594Mbps needs a FIFO setting >=480 (the max is 511). I would anticipate that to be one of the worst situations as we're dealing with a FIFO underflow herewhen there is a significantly faster CSI rate than HDMI. It can't be supported with a 972Mbps link frequency over 4 lanes (needs >=667), and 2 lanes needs a FIFO setting >=374. I'll see what numbers fall out of the new spreadsheet for all standard modes. If there are some modes that can't be supported over 4 lanes then there is an absolute requirement for communicating the number of lanes to use. Seeing as Cisco have kit shipping with this chip and driver, can I ask how they are managing the choice over number of lanes in use? As for this patch it sounds like we need to crank the FIFO setting up to the maximum of 511, and potentially a second patch that removes g_mbus_config and only reads DT if that is the desired behaviour. >> >> [1] http://www.spinics.net/lists/linux-media/msg121813.html, as part >> of the unicam_start_streaming function. >> >>> The i.MX6 MIPI CSI-2 receiver driver can't cope with that, as it always >>> activates all four lanes that are configured in the device tree. I can >>> work around that with the following patch: >> >> It can't cope running at less than 4 lanes, or it can't cope with a change? >> >>> ----------8<---------- >>> Subject: [PATCH] [media] tc358743: do not dynamically reduce number of lanes >>> >>> Dynamic lane number reduction does not work with receivers that >>> configure a fixed lane number according to the device tree settings. >>> To allow 720p60 at 594 Mbit/s on 4 lanes, increase the fifo_level >>> and tclk_trailcnt settings. >>> >>> Signed-off-by: Philipp Zabel <p.zabel@pengutronix.de> >>> --- >>> drivers/media/i2c/tc358743.c | 6 +++--- >>> 1 file changed, 3 insertions(+), 3 deletions(-) >>> >>> diff --git a/drivers/media/i2c/tc358743.c b/drivers/media/i2c/tc358743.c >>> index 64f504542a819..70a9435928cdb 100644 >>> --- a/drivers/media/i2c/tc358743.c >>> +++ b/drivers/media/i2c/tc358743.c >>> @@ -683,7 +683,7 @@ static void tc358743_set_csi(struct v4l2_subdev *sd) >>> { >>> struct tc358743_state *state = to_state(sd); >>> struct tc358743_platform_data *pdata = &state->pdata; >>> - unsigned lanes = tc358743_num_csi_lanes_needed(sd); >>> + unsigned lanes = state->bus.num_data_lanes; >>> >>> v4l2_dbg(3, debug, sd, "%s:\n", __func__); >>> >>> @@ -1906,7 +1906,7 @@ static int tc358743_probe_of(struct tc358743_state *state) >>> state->pdata.ddc5v_delay = DDC5V_DELAY_100_MS; >>> state->pdata.enable_hdcp = false; >>> /* A FIFO level of 16 should be enough for 2-lane 720p60 at 594 MHz. */ >>> - state->pdata.fifo_level = 16; >>> + state->pdata.fifo_level = 320; >>> /* >>> * The PLL input clock is obtained by dividing refclk by pll_prd. >>> * It must be between 6 MHz and 40 MHz, lower frequency is better. >>> @@ -1948,7 +1948,7 @@ static int tc358743_probe_of(struct tc358743_state *state) >>> state->pdata.lptxtimecnt = 0x003; >>> /* tclk-preparecnt: 3, tclk-zerocnt: 20 */ >>> state->pdata.tclk_headercnt = 0x1403; >>> - state->pdata.tclk_trailcnt = 0x00; >>> + state->pdata.tclk_trailcnt = 0x01; >>> /* ths-preparecnt: 3, ths-zerocnt: 1 */ >>> state->pdata.ths_headercnt = 0x0103; >>> state->pdata.twakeup = 0x4882; >>> -- >>> 2.11.0 >>> ---------->8---------- >>> >>> Just adding the same heuristic as tc358743_num_csi_lanes_needed in the >>> imx6-mipi-csi2 driver doesn't work, as the heuristic is specific to the >>> Toshiba chip. There are MIPI CSI-2 sensors that only support a fixed >>> number of lanes, for example. >>> >>> I'd need a way to communicate the number of active MIPI CSI-2 lanes >>> between transmitting and receiving subdevice driver. >>> >>>> Once I've got a v3 done on the Unicam driver I'll bash through the >>>> standard HDMI modes and check what value they need - I can see a big >>>> spreadsheet coming on. >>> >>> Oh dear. Unless the point you make below can be resolved, I think we >>> have no other choice. >>> >>>> I'll ignore interlaced modes as I can't see any support for it in the >>>> driver. Receiving the fields on different CSI-2 data types is >>>> something I know the Unicam hardware won't handle nicely, and I >>>> suspect it'd be an issue for many other platforms too. >>> >>> Yes, let's pretend interlacing doesn't exist as long as possible. >>> >>>>> Hmm... if this is dependent on the resolution and frame rate, >>>>> wouldn't >>>>> it be better to dynamically adjust it accordingly? >>>> >>>> It's setting up the FIFO matching the incoming HDMI data rate and >>>> outgoing CSI rate. That means it's dependent on the incoming pixel >>>> clock, blanking, colour format and resolution, and output CSI link >>>> frequency, number of lanes, and colour format. >>>> Whilst it could be set dynamically based on all those parameters, is >>>> there a significant enough gain in doing so? >>> >>> Ideally there would be no need to maintain a timing database with - >>> worst case - (number of link frequencies * number of HDMI modes) entries >>> in the driver ... >>> >>>> The value of 300 works for all cases I've tried, and referencing back >>>> it is also the value that Hans said Cisco use via platform data on >>>> their hardware [1]. Generally I'm seeing that values of 0-130 are >>>> required, so 300 is giving a fair safety margin. >>> >>> It does not work for 720p60 on 4 lanes at 594 Mbit/s, as the spreadsheet >>> warns, and testing shows. >> >> If it doesn't work with 720p60, then I guess it has no hope with many >> other resolutions. >> It sounds like confirming whether g_mbus_config is a potential >> solution for i.MX6 (sorry I'm not familiar enough with that code to do >> my own quick search), but otherwise cranking it up to 320 is >> reasonable, and I'll see what other numbers fall out of the >> spreadsheet. >> >>>> Second question is does anyone have a suitable relationship with >>>> Toshiba to get permission to release details of these register >>>> calculations? The datasheet and value spreadsheet are marked as >>>> confidential, and probably under NDA in almost all cases. Whilst they >>>> can't object to drivers containing values to make them work, they >>>> might over releasing significant details. >>> >>> Unfortunately, I don't. >> >> I'm guessing Cisco may be in the best position to try pushing that. >> Certainly all our info is under NDA, and we don't really have a >> significant working relationship with them. > > I think we tried at the time, but no luck. Thanks. So unless someone else has a relationship with Toshiba then it has to be some fixed value - no real option for calculating it dynamically. (That also means my intent to support arbitrary link frequencies won't fly as that would require publishing details of how to compute lineinintcnt, lptxtimecnt, etc. Oh well.) Dave
On Wed, 2017-09-20 at 13:24 +0200, Hans Verkuil wrote: > On 09/20/17 13:00, Dave Stevenson wrote: [...] > > It is communicated over the subdevice API - tc358743_g_mbus_config > > reports back the appropriate number of lanes to the receiver > > subdevice. > > A suitable v4l2_subdev_has_op(dev->sensor, video, g_mbus_config) call > > as you're starting streaming therefore gives you the correct > > information. That's what I've just done for the BCM283x Unicam > > driver[1], but admittedly I'm not using the media controller API which > > i.MX6 is. > > Shouldn't this information come from the device tree? The g_mbus_config > op is close to being deprecated or even removed. There are currently only > two obscure V4L2 bridge drivers that call it. It dates from pre-DT times > I rather not see it used in a new bridge driver. > > The problem is that contains data that belongs to the DT (hardware > capabilities). Things that can actually change dynamically should be > communicated via another op. We don't have that, so that should be > created. > > I've CC-ed Sakari, he is the specialist for such things. The total number of MIPI CSI-2 lanes (as well as their order) and the list of allowed link frequencies are static and come from the device tree. But the possible combinations of link frequency and number of active lanes out of those for a given resolution and format can vary depending on both transmitter and receiver capabilities. For example, if the DT specifies 4 lanes and both 148.5 MHz and 297 MHz link frequencies, the Toshiba chip could send 720p60 YUYV via 4 lanes at 148.5 MHz, via 2 lanes at 297 MHz, or even via 4 lanes at 297 MHz, with the longer FIFO delay. > > > > [1] http://www.spinics.net/lists/linux-media/msg121813.html, as part > > of the unicam_start_streaming function. > > > > > The i.MX6 MIPI CSI-2 receiver driver can't cope with that, as it always > > > activates all four lanes that are configured in the device tree. I can > > > work around that with the following patch: > > > > It can't cope running at less than 4 lanes, or it can't cope with a > > change? The hardware can cope with both, although I don't know if there are receivers that can not. In my case this is just about not knowing how many lanes to activate (see below) and which link frequency to choose (currently fixed to the first). [...] > > > [...] 300 is giving a fair safety margin. > > > > > > It does not work for 720p60 on 4 lanes at 594 Mbit/s, as the spreadsheet > > > warns, and testing shows. > > > > If it doesn't work with 720p60, then I guess it has no hope with many > > other resolutions. That would have to be checked on a case by case basis, unfortunately. I have a usecase that only supports 1080p50/60 and 720p50/60. > > It sounds like confirming whether g_mbus_config is a potential > > solution for i.MX6 (sorry I'm not familiar enough with that code to do > > my own quick search), but otherwise cranking it up to 320 is > > reasonable, and I'll see what other numbers fall out of the > > spreadsheet. It seems we need a replacement for g_mbus_config that only returns the dynamic parameters. regards Philipp
On 09/20/17 14:23, Dave Stevenson wrote: > On 20 September 2017 at 12:24, Hans Verkuil <hansverk@cisco.com> wrote: >> On 09/20/17 13:00, Dave Stevenson wrote: >>> On 20 September 2017 at 11:23, Philipp Zabel <p.zabel@pengutronix.de> wrote: >>>> Hi, >>>> >>>> On Wed, 2017-09-20 at 10:14 +0100, Dave Stevenson wrote: >>>>> Hi Mauro & Philipp >>>>> >>>>> On 19 September 2017 at 17:49, Mauro Carvalho Chehab >>>>> <mchehab@s-opensource.com> wrote: >>>>>> Em Tue, 19 Sep 2017 17:24:45 +0200 >>>>>> Philipp Zabel <p.zabel@pengutronix.de> escreveu: >>>>>> >>>>>>> Hi Dave, >>>>>>> >>>>>>> On Tue, 2017-09-19 at 14:08 +0100, Dave Stevenson wrote: >>>>>>>> The existing fixed value of 16 worked for UYVY 720P60 over >>>>>>>> 2 lanes at 594MHz, or UYVY 1080P60 over 4 lanes. (RGB888 >>>>>>>> 1080P60 needs 6 lanes at 594MHz). >>>>>>>> It doesn't allow for lower resolutions to work as the FIFO >>>>>>>> underflows. >>>>>>>> >>>>>>>> Using a value of 300 works for all resolutions down to VGA60, >>>>>>>> and the increase in frame delay is <4usecs for 1080P60 UYVY >>>>>>>> (2.55usecs for RGB888). >>>>>>>> >>>>>>>> Signed-off-by: Dave Stevenson <dave.stevenson@raspberrypi.org> >>>>>>> >>>>>>> Can we increase this to 320? This would also allow >>>>>>> 720p60 at 594 Mbps / 4 lanes, according to the xls. >>>>> >>>>> Unless I've missed something then the driver would currently request >>>>> only 2 lanes for 720p60 UYVY, and that works with the existing FIFO >>>>> setting of 16. Likewise 720p60 RGB888 requests 3 lanes and also works >>>>> on a FIFO setting of 16. >>>>> How/why were you thinking we need to run all four lanes for 720p60 >>>>> without other significant driver mods around lane config? >>>> >>>> The driver currently silently changes the number of active lanes >>>> depending on required data rate, with no way to communicate it to the >>>> receiver. >>> >>> It is communicated over the subdevice API - tc358743_g_mbus_config >>> reports back the appropriate number of lanes to the receiver >>> subdevice. >>> A suitable v4l2_subdev_has_op(dev->sensor, video, g_mbus_config) call >>> as you're starting streaming therefore gives you the correct >>> information. That's what I've just done for the BCM283x Unicam >>> driver[1], but admittedly I'm not using the media controller API which >>> i.MX6 is. >> >> Shouldn't this information come from the device tree? The g_mbus_config >> op is close to being deprecated or even removed. There are currently only >> two obscure V4L2 bridge drivers that call it. It dates from pre-DT times >> I rather not see it used in a new bridge driver. >> >> The problem is that contains data that belongs to the DT (hardware >> capabilities). Things that can actually change dynamically should be >> communicated via another op. We don't have that, so that should be created. >> >> I've CC-ed Sakari, he is the specialist for such things. > > You've reminded me that I asked that same question earlier in the > year, and Sakari had replied - > http://www.spinics.net/lists/linux-media/msg115550.html > > Is it specifically device tree related? Just because the lanes are > physically there doesn't necessarily mean they have to be used. The DT should tell how many lanes are connected. g/s_mbus_config was really doing the job that the device tree does today, but it does so badly. My recommendation is that you don't use it at all. Instead look at the DT for the number of lanes. *If* it becomes clear that you need to communicate the actual number of lanes in use, then we need to make a new op or whatever. > > A quick test with the spreadsheet appears to say that 1080p24 UYVY > over 4 lanes at 594Mbps needs a FIFO setting >=480 (the max is 511). I > would anticipate that to be one of the worst situations as we're > dealing with a FIFO underflow herewhen there is a significantly faster > CSI rate than HDMI. > It can't be supported with a 972Mbps link frequency over 4 lanes > (needs >=667), and 2 lanes needs a FIFO setting >=374. > > I'll see what numbers fall out of the new spreadsheet for all standard > modes. If there are some modes that can't be supported over 4 lanes > then there is an absolute requirement for communicating the number of > lanes to use. > > Seeing as Cisco have kit shipping with this chip and driver, can I ask > how they are managing the choice over number of lanes in use? It's using g_mbus_config since it runs on an old pre-DT kernel. I personally would be perfectly happy with a simple new op to just communicate the number of lanes in use, but there may be more things that should be passed on to the bridge driver. Sakari knows this better than I do. But g_mbus_config shouldn't be used here. No way you could have known that, we really need to clarify that in v4l2-subdev.h. Hmm, it DOES say that for s_mbus_config, but not for g_mbus_config. Regards, Hans
Hi Hans and others, On Wed, Sep 20, 2017 at 01:24:02PM +0200, Hans Verkuil wrote: > On 09/20/17 13:00, Dave Stevenson wrote: > > On 20 September 2017 at 11:23, Philipp Zabel <p.zabel@pengutronix.de> wrote: > >> Hi, > >> > >> On Wed, 2017-09-20 at 10:14 +0100, Dave Stevenson wrote: > >>> Hi Mauro & Philipp > >>> > >>> On 19 September 2017 at 17:49, Mauro Carvalho Chehab > >>> <mchehab@s-opensource.com> wrote: > >>>> Em Tue, 19 Sep 2017 17:24:45 +0200 > >>>> Philipp Zabel <p.zabel@pengutronix.de> escreveu: > >>>> > >>>>> Hi Dave, > >>>>> > >>>>> On Tue, 2017-09-19 at 14:08 +0100, Dave Stevenson wrote: > >>>>>> The existing fixed value of 16 worked for UYVY 720P60 over > >>>>>> 2 lanes at 594MHz, or UYVY 1080P60 over 4 lanes. (RGB888 > >>>>>> 1080P60 needs 6 lanes at 594MHz). > >>>>>> It doesn't allow for lower resolutions to work as the FIFO > >>>>>> underflows. > >>>>>> > >>>>>> Using a value of 300 works for all resolutions down to VGA60, > >>>>>> and the increase in frame delay is <4usecs for 1080P60 UYVY > >>>>>> (2.55usecs for RGB888). > >>>>>> > >>>>>> Signed-off-by: Dave Stevenson <dave.stevenson@raspberrypi.org> > >>>>> > >>>>> Can we increase this to 320? This would also allow > >>>>> 720p60 at 594 Mbps / 4 lanes, according to the xls. > >>> > >>> Unless I've missed something then the driver would currently request > >>> only 2 lanes for 720p60 UYVY, and that works with the existing FIFO > >>> setting of 16. Likewise 720p60 RGB888 requests 3 lanes and also works > >>> on a FIFO setting of 16. > >>> How/why were you thinking we need to run all four lanes for 720p60 > >>> without other significant driver mods around lane config? > >> > >> The driver currently silently changes the number of active lanes > >> depending on required data rate, with no way to communicate it to the > >> receiver. > > > > It is communicated over the subdevice API - tc358743_g_mbus_config > > reports back the appropriate number of lanes to the receiver > > subdevice. > > A suitable v4l2_subdev_has_op(dev->sensor, video, g_mbus_config) call > > as you're starting streaming therefore gives you the correct > > information. That's what I've just done for the BCM283x Unicam > > driver[1], but admittedly I'm not using the media controller API which > > i.MX6 is. > > Shouldn't this information come from the device tree? The g_mbus_config > op is close to being deprecated or even removed. There are currently only > two obscure V4L2 bridge drivers that call it. It dates from pre-DT times > I rather not see it used in a new bridge driver. > > The problem is that contains data that belongs to the DT (hardware > capabilities). Things that can actually change dynamically should be > communicated via another op. We don't have that, so that should be created. The DT tells how many lanes are connected in hardware, but up to now that's also been the number of lanes actually used. The g_mbus_config() is there, and I'd like to replace that with the more generic frame descriptors, with CSI-2 virtual channel and data type support. They're however not quite ready yet nor I've recently worked on them. I think using g_mbus_config() for the purpose right now is entirely acceptable, we can rework that later on when adding support for frame descriptors.
On 09/20/17 14:50, Sakari Ailus wrote: > Hi Hans and others, > > On Wed, Sep 20, 2017 at 01:24:02PM +0200, Hans Verkuil wrote: >> On 09/20/17 13:00, Dave Stevenson wrote: >>> On 20 September 2017 at 11:23, Philipp Zabel <p.zabel@pengutronix.de> wrote: >>>> Hi, >>>> >>>> On Wed, 2017-09-20 at 10:14 +0100, Dave Stevenson wrote: >>>>> Hi Mauro & Philipp >>>>> >>>>> On 19 September 2017 at 17:49, Mauro Carvalho Chehab >>>>> <mchehab@s-opensource.com> wrote: >>>>>> Em Tue, 19 Sep 2017 17:24:45 +0200 >>>>>> Philipp Zabel <p.zabel@pengutronix.de> escreveu: >>>>>> >>>>>>> Hi Dave, >>>>>>> >>>>>>> On Tue, 2017-09-19 at 14:08 +0100, Dave Stevenson wrote: >>>>>>>> The existing fixed value of 16 worked for UYVY 720P60 over >>>>>>>> 2 lanes at 594MHz, or UYVY 1080P60 over 4 lanes. (RGB888 >>>>>>>> 1080P60 needs 6 lanes at 594MHz). >>>>>>>> It doesn't allow for lower resolutions to work as the FIFO >>>>>>>> underflows. >>>>>>>> >>>>>>>> Using a value of 300 works for all resolutions down to VGA60, >>>>>>>> and the increase in frame delay is <4usecs for 1080P60 UYVY >>>>>>>> (2.55usecs for RGB888). >>>>>>>> >>>>>>>> Signed-off-by: Dave Stevenson <dave.stevenson@raspberrypi.org> >>>>>>> >>>>>>> Can we increase this to 320? This would also allow >>>>>>> 720p60 at 594 Mbps / 4 lanes, according to the xls. >>>>> >>>>> Unless I've missed something then the driver would currently request >>>>> only 2 lanes for 720p60 UYVY, and that works with the existing FIFO >>>>> setting of 16. Likewise 720p60 RGB888 requests 3 lanes and also works >>>>> on a FIFO setting of 16. >>>>> How/why were you thinking we need to run all four lanes for 720p60 >>>>> without other significant driver mods around lane config? >>>> >>>> The driver currently silently changes the number of active lanes >>>> depending on required data rate, with no way to communicate it to the >>>> receiver. >>> >>> It is communicated over the subdevice API - tc358743_g_mbus_config >>> reports back the appropriate number of lanes to the receiver >>> subdevice. >>> A suitable v4l2_subdev_has_op(dev->sensor, video, g_mbus_config) call >>> as you're starting streaming therefore gives you the correct >>> information. That's what I've just done for the BCM283x Unicam >>> driver[1], but admittedly I'm not using the media controller API which >>> i.MX6 is. >> >> Shouldn't this information come from the device tree? The g_mbus_config >> op is close to being deprecated or even removed. There are currently only >> two obscure V4L2 bridge drivers that call it. It dates from pre-DT times >> I rather not see it used in a new bridge driver. >> >> The problem is that contains data that belongs to the DT (hardware >> capabilities). Things that can actually change dynamically should be >> communicated via another op. We don't have that, so that should be created. > > The DT tells how many lanes are connected in hardware, but up to now that's > also been the number of lanes actually used. > > The g_mbus_config() is there, and I'd like to replace that with the more > generic frame descriptors, with CSI-2 virtual channel and data type > support. They're however not quite ready yet nor I've recently worked on > them. > > I think using g_mbus_config() for the purpose right now is entirely > acceptable, we can rework that later on when adding support for frame > descriptors. > I don't like it :-) Currently g_mbus_config returns (and I quote from v4l2-mediabus.h): "How many lanes the client can use". I.e. the capabilities of the HW. If we are going to use this to communicate how many lines are currently in use, then I would propose that we add a lane mask, i.e. something like this: /* Number of lanes in use, 0 == use all available lanes (default) */ #define V4L2_MBUS_CSI2_LANE_MASK (3 << 10) And add comments along the lines that this is a temporary fix. I would feel a lot happier (or a lot less unhappy) if we'd do it this way. Rather than re-interpreting bits that are not quite what they should be. I'd also add a comment that all other flags must be 0 if the device tree is used. This to avoid mixing the two. Regards, Hans
Hi Hans, On Wed, Sep 20, 2017 at 03:12:03PM +0200, Hans Verkuil wrote: > On 09/20/17 14:50, Sakari Ailus wrote: > > Hi Hans and others, > > > > On Wed, Sep 20, 2017 at 01:24:02PM +0200, Hans Verkuil wrote: > >> On 09/20/17 13:00, Dave Stevenson wrote: > >>> On 20 September 2017 at 11:23, Philipp Zabel <p.zabel@pengutronix.de> wrote: > >>>> Hi, > >>>> > >>>> On Wed, 2017-09-20 at 10:14 +0100, Dave Stevenson wrote: > >>>>> Hi Mauro & Philipp > >>>>> > >>>>> On 19 September 2017 at 17:49, Mauro Carvalho Chehab > >>>>> <mchehab@s-opensource.com> wrote: > >>>>>> Em Tue, 19 Sep 2017 17:24:45 +0200 > >>>>>> Philipp Zabel <p.zabel@pengutronix.de> escreveu: > >>>>>> > >>>>>>> Hi Dave, > >>>>>>> > >>>>>>> On Tue, 2017-09-19 at 14:08 +0100, Dave Stevenson wrote: > >>>>>>>> The existing fixed value of 16 worked for UYVY 720P60 over > >>>>>>>> 2 lanes at 594MHz, or UYVY 1080P60 over 4 lanes. (RGB888 > >>>>>>>> 1080P60 needs 6 lanes at 594MHz). > >>>>>>>> It doesn't allow for lower resolutions to work as the FIFO > >>>>>>>> underflows. > >>>>>>>> > >>>>>>>> Using a value of 300 works for all resolutions down to VGA60, > >>>>>>>> and the increase in frame delay is <4usecs for 1080P60 UYVY > >>>>>>>> (2.55usecs for RGB888). > >>>>>>>> > >>>>>>>> Signed-off-by: Dave Stevenson <dave.stevenson@raspberrypi.org> > >>>>>>> > >>>>>>> Can we increase this to 320? This would also allow > >>>>>>> 720p60 at 594 Mbps / 4 lanes, according to the xls. > >>>>> > >>>>> Unless I've missed something then the driver would currently request > >>>>> only 2 lanes for 720p60 UYVY, and that works with the existing FIFO > >>>>> setting of 16. Likewise 720p60 RGB888 requests 3 lanes and also works > >>>>> on a FIFO setting of 16. > >>>>> How/why were you thinking we need to run all four lanes for 720p60 > >>>>> without other significant driver mods around lane config? > >>>> > >>>> The driver currently silently changes the number of active lanes > >>>> depending on required data rate, with no way to communicate it to the > >>>> receiver. > >>> > >>> It is communicated over the subdevice API - tc358743_g_mbus_config > >>> reports back the appropriate number of lanes to the receiver > >>> subdevice. > >>> A suitable v4l2_subdev_has_op(dev->sensor, video, g_mbus_config) call > >>> as you're starting streaming therefore gives you the correct > >>> information. That's what I've just done for the BCM283x Unicam > >>> driver[1], but admittedly I'm not using the media controller API which > >>> i.MX6 is. > >> > >> Shouldn't this information come from the device tree? The g_mbus_config > >> op is close to being deprecated or even removed. There are currently only > >> two obscure V4L2 bridge drivers that call it. It dates from pre-DT times > >> I rather not see it used in a new bridge driver. > >> > >> The problem is that contains data that belongs to the DT (hardware > >> capabilities). Things that can actually change dynamically should be > >> communicated via another op. We don't have that, so that should be created. > > > > The DT tells how many lanes are connected in hardware, but up to now that's > > also been the number of lanes actually used. > > > > The g_mbus_config() is there, and I'd like to replace that with the more > > generic frame descriptors, with CSI-2 virtual channel and data type > > support. They're however not quite ready yet nor I've recently worked on > > them. > > > > I think using g_mbus_config() for the purpose right now is entirely > > acceptable, we can rework that later on when adding support for frame > > descriptors. > > > > I don't like it :-) > > Currently g_mbus_config returns (and I quote from v4l2-mediabus.h): "How > many lanes the client can use". I.e. the capabilities of the HW. > > If we are going to use this to communicate how many lines are currently > in use, then I would propose that we add a lane mask, i.e. something like > this: > > /* Number of lanes in use, 0 == use all available lanes (default) */ > #define V4L2_MBUS_CSI2_LANE_MASK (3 << 10) > > And add comments along the lines that this is a temporary fix. > > I would feel a lot happier (or a lot less unhappy) if we'd do it this way. > Rather than re-interpreting bits that are not quite what they should be. > > I'd also add a comment that all other flags must be 0 if the device tree is > used. This to avoid mixing the two. That would work for me as well. There are very few users of the g_mbus_config API and I bet the current users could get the configuration from DT as well: they're platform drivers used on ARM. With the possible exception of SoC camera drives.
Hi Hans, On Wed, 2017-09-20 at 15:12 +0200, Hans Verkuil wrote: [...] > I don't like it :-) > > Currently g_mbus_config returns (and I quote from v4l2-mediabus.h): "How > many lanes the client can use". I.e. the capabilities of the HW. > > If we are going to use this to communicate how many lines are currently > in use, then I would propose that we add a lane mask, i.e. something like > this: > > /* Number of lanes in use, 0 == use all available lanes (default) */ > #define V4L2_MBUS_CSI2_LANE_MASK (3 << 10) > > And add comments along the lines that this is a temporary fix. > > I would feel a lot happier (or a lot less unhappy) if we'd do it this way. > Rather than re-interpreting bits that are not quite what they should be. > > I'd also add a comment that all other flags must be 0 if the device tree is > used. This to avoid mixing the two. I would like to try this. Currently the driver sets the V4L2_MBUS_CSI2_[1-4]_LANE bits according to csi_lanes_in_use, which is wrong as you say. After moving the csi_lanes_in_use info into a new V4L2_MBUS_CSI2_LANE_MASK bitfield, the V4L2_MBUS_CSI2_[1-4]_LANE bits could be either set to zero or to the really connected lanes as configured in the device tree (csi->bus.num_data_lanes) in the DT case. What would the bits be set to in the pdata case, though? Should a lane count setting be added to tc358743_platform_data with, defaulting to all bits set? regards Philipp
diff --git a/drivers/media/i2c/tc358743.c b/drivers/media/i2c/tc358743.c index 64f504542a819..70a9435928cdb 100644 --- a/drivers/media/i2c/tc358743.c +++ b/drivers/media/i2c/tc358743.c @@ -683,7 +683,7 @@ static void tc358743_set_csi(struct v4l2_subdev *sd) { struct tc358743_state *state = to_state(sd); struct tc358743_platform_data *pdata = &state->pdata; - unsigned lanes = tc358743_num_csi_lanes_needed(sd); + unsigned lanes = state->bus.num_data_lanes; v4l2_dbg(3, debug, sd, "%s:\n", __func__); @@ -1906,7 +1906,7 @@ static int tc358743_probe_of(struct tc358743_state *state) state->pdata.ddc5v_delay = DDC5V_DELAY_100_MS; state->pdata.enable_hdcp = false; /* A FIFO level of 16 should be enough for 2-lane 720p60 at 594 MHz. */ - state->pdata.fifo_level = 16; + state->pdata.fifo_level = 320; /* * The PLL input clock is obtained by dividing refclk by pll_prd. * It must be between 6 MHz and 40 MHz, lower frequency is better. @@ -1948,7 +1948,7 @@ static int tc358743_probe_of(struct tc358743_state *state) state->pdata.lptxtimecnt = 0x003; /* tclk-preparecnt: 3, tclk-zerocnt: 20 */ state->pdata.tclk_headercnt = 0x1403; - state->pdata.tclk_trailcnt = 0x00; + state->pdata.tclk_trailcnt = 0x01; /* ths-preparecnt: 3, ths-zerocnt: 1 */ state->pdata.ths_headercnt = 0x0103; state->pdata.twakeup = 0x4882;