| Message ID | 20180123132220.29886-1-linus.walleij@linaro.org |
|---|---|
| State | New |
| Headers | show |
| Series | drm/pl111: Use max memory bandwidth for resolution | expand |
Linus Walleij <linus.walleij@linaro.org> writes: > We were previously selecting 1024x768 and 32BPP as the default > set-up for the PL111 consumers. > > This does not work on elder systems: the device tree bindings > support a property "max-memory-bandwidth" in bytes/second that > states that if you exceed this the memory bus will saturate. > The result is flickering and unstable images. > > Parse the "max-memory-bandwidth" and respect it when > intializing the driver. On the RealView PB11MP we get a nice > 1024x768 console with RGB565 as default with this code. > > If the device tree does not specify the maximum memory > bandwidth we keep the old assumption that we can support > 1024x768, 32BPP. > > Signed-off-by: Linus Walleij <linus.walleij@linaro.org> > --- > drivers/gpu/drm/pl111/pl111_drm.h | 1 + > drivers/gpu/drm/pl111/pl111_drv.c | 99 +++++++++++++++++++++++++++++++++++++-- > 2 files changed, 97 insertions(+), 3 deletions(-) > > diff --git a/drivers/gpu/drm/pl111/pl111_drm.h b/drivers/gpu/drm/pl111/pl111_drm.h > index 440f53ebee8c..770ef5ce3645 100644 > --- a/drivers/gpu/drm/pl111/pl111_drm.h > +++ b/drivers/gpu/drm/pl111/pl111_drm.h > @@ -56,6 +56,7 @@ struct pl111_drm_dev_private { > struct drm_fbdev_cma *fbdev; > > void *regs; > + u32 memory_bw; > u32 ienb; > u32 ctrl; > /* The pixel clock (a reference to our clock divider off of CLCDCLK). */ > diff --git a/drivers/gpu/drm/pl111/pl111_drv.c b/drivers/gpu/drm/pl111/pl111_drv.c > index 101a9c7db6ff..1b47ee67097d 100644 > --- a/drivers/gpu/drm/pl111/pl111_drv.c > +++ b/drivers/gpu/drm/pl111/pl111_drv.c > @@ -58,6 +58,7 @@ > #include <linux/dma-buf.h> > #include <linux/module.h> > #include <linux/slab.h> > +#include <linux/of.h> > > #include <drm/drmP.h> > #include <drm/drm_atomic_helper.h> > @@ -80,21 +81,107 @@ static const struct drm_mode_config_funcs mode_config_funcs = { > .atomic_commit = drm_atomic_helper_commit, > }; > > +/** > + * pl111_choose_max_resolution() - choose max resolution > + * @dev: DRM device > + * @memory_bw: the graphics maximum memory bandwidth in bytes/s > + * @max_width: returns the maximum width > + * @max_height: returns the maximum height > + * @bpp: returns the maximum bips per pixed (32 or 16) > + * > + * This function attempts to cut down the maximum supported resolution > + * to something the system memory bus can handle. The PL111 and pixel > + * clocks may be able to support higher resolutions and color depths > + * but as we go up the memory bus get saturated and becomes the > + * bottleneck for the resolution. On several systems using e.g. > + * 1024x768 with 32 BPP results in instable flickering images. > + */ > +static void pl111_choose_max_resolution(struct drm_device *dev, > + u32 memory_bw, > + int *max_width, > + int *max_height, > + unsigned int *bpp) > +{ > + /* No limitations, this is the most aggressive */ > + if (!memory_bw) { > + *max_width = 1024; > + *max_height = 768; > + *bpp = 32; > + return; > + } > + > + /* > + * 1024x768 with RGBX8888 requires a memory bandwidth of > + * 65Mhz * 4 bytes = 260000000 bytes per second. > + */ > + if (memory_bw >= 260000000) { > + *max_width = 1024; > + *max_height = 768; > + *bpp = 32; > + return; > + } > + > + /* > + * 800x600 with RGB8888 requires a memory bandwidth of > + * 36Mhz * 4 bytes = 144000000 bytes per second. But we > + * assume the user prefer higher resolution over more > + * color depth, so we do not add this mode here. > + */ > + > + /* > + * 1024x768 with RGB565 requires a memory bandwidth of > + * 65Mhz * 2 bytes = 130000000 bytes per second. > + */ > + if (memory_bw >= 130000000) { > + *max_width = 1024; > + *max_height = 768; > + *bpp = 16; > + return; > + } > + > + /* > + * 800x600 with RGB565 requires a memory bandwidth of > + * 36Mhz * 2 bytes = 72000000 bytes per second. > + */ > + if (memory_bw >= 72000000) { > + *max_width = 800; > + *max_height = 600; > + *bpp = 16; > + return; > + } > + > + /* > + * 640x480 with RGB565 requires a memory bandwidth of > + * 25.175Mhz * 2 bytes = 50350000 bytes per second. > + */ > + if (memory_bw < 50350000) > + dev_err(dev->dev, "can't even do 640x480 VGA RGB565, proceed anyway\n"); > + > + *max_width = 640; > + *max_height = 480; > + *bpp = 16; > +} > + > static int pl111_modeset_init(struct drm_device *dev) > { > struct drm_mode_config *mode_config; > struct pl111_drm_dev_private *priv = dev->dev_private; > struct drm_panel *panel; > struct drm_bridge *bridge; > + unsigned int bpp; /* bits per pixel */ > int ret = 0; > > drm_mode_config_init(dev); > mode_config = &dev->mode_config; > mode_config->funcs = &mode_config_funcs; > mode_config->min_width = 1; > - mode_config->max_width = 1024; > mode_config->min_height = 1; > - mode_config->max_height = 768; > + > + pl111_choose_max_resolution(dev, priv->memory_bw, > + &mode_config->max_width, > + &mode_config->max_height, &bpp); > + dev_info(dev->dev, "cap resolution at %u x %u, %u BPP\n", > + mode_config->max_width, mode_config->max_height, bpp); I think this is the wrong place in the pipeline to be doing this, but I don't have a complete solution so I'm not necessarily saying no. Things I think we should do for bandwidth limits: A new pl111_mode_valid() rejects modes with width*height*2 > bandwidth (if we can't scan it out with our smallest format, don't advertise it). pl111_display_check() rejects modes with width*height*bpp > bandwidth (if we can't scan out this particular configuration, let them know we can't set the mode). Ideally given those two things, fbdev and X11 would notice that the preferred mode fails at 24bpp and fall back to 16bpp. I don't think either of those does so today, though. Interested in tackling any of these?
On Thu, Jan 25, 2018 at 4:46 AM, Eric Anholt <eric@anholt.net> wrote: >> + pl111_choose_max_resolution(dev, priv->memory_bw, >> + &mode_config->max_width, >> + &mode_config->max_height, &bpp); >> + dev_info(dev->dev, "cap resolution at %u x %u, %u BPP\n", >> + mode_config->max_width, mode_config->max_height, bpp); > > I think this is the wrong place in the pipeline to be doing this, but I > don't have a complete solution so I'm not necessarily saying no. So currently the driver does this: mode_config->max_width = 1024; mode_config->max_height = 768; And that is because it cannot really handle anything. I guess ideally the DRM driver should set these to -1 or something so that any widths and heights negotiated will work. > Things I think we should do for bandwidth limits: > > A new pl111_mode_valid() rejects modes with width*height*2 > bandwidth > (if we can't scan it out with our smallest format, don't advertise it). > > pl111_display_check() rejects modes with width*height*bpp > bandwidth > (if we can't scan out this particular configuration, let them know we > can't set the mode). > > Ideally given those two things, fbdev and X11 would notice that the > preferred mode fails at 24bpp and fall back to 16bpp. I don't think > either of those does so today, though. > > Interested in tackling any of these? I tried the pl111_display_check() version. It just made the driver fail to initialize anything, at least when using the dumb VGA bridge. There are .mode_valid() callbacks on the bridges we use (panel and dumb VGA) but neither uses it at the moment, hm. I could just assign my own .mode_valid() callback to the bridge, but it seems a bit fragile. But it's worth a hack, I'll try it. I sent a sent of lesser controversial patches in the meantime, and rebased this on top of those so we can deal with the memory BW issue separately. Yours, Linus Walleij
On Fri, Jan 26, 2018 at 2:27 PM, Linus Walleij <linus.walleij@linaro.org> wrote: > On Thu, Jan 25, 2018 at 4:46 AM, Eric Anholt <eric@anholt.net> wrote: > >>> + pl111_choose_max_resolution(dev, priv->memory_bw, >>> + &mode_config->max_width, >>> + &mode_config->max_height, &bpp); >>> + dev_info(dev->dev, "cap resolution at %u x %u, %u BPP\n", >>> + mode_config->max_width, mode_config->max_height, bpp); >> >> I think this is the wrong place in the pipeline to be doing this, but I >> don't have a complete solution so I'm not necessarily saying no. > > So currently the driver does this: > > mode_config->max_width = 1024; > mode_config->max_height = 768; > > And that is because it cannot really handle anything. I guess ideally > the DRM driver should set these to -1 or something so that any widths > and heights negotiated will work. > >> Things I think we should do for bandwidth limits: >> >> A new pl111_mode_valid() rejects modes with width*height*2 > bandwidth >> (if we can't scan it out with our smallest format, don't advertise it). >> >> pl111_display_check() rejects modes with width*height*bpp > bandwidth >> (if we can't scan out this particular configuration, let them know we >> can't set the mode). >> >> Ideally given those two things, fbdev and X11 would notice that the >> preferred mode fails at 24bpp and fall back to 16bpp. I don't think >> either of those does so today, though. >> >> Interested in tackling any of these? > > I tried the pl111_display_check() version. It just made the driver > fail to initialize anything, at least when using the dumb VGA > bridge. I guess this is because it gets called from drm_simple_display_pipe_funcs at which point the driver framework has already decided to go with this format. And that is backed by crtc. We would need to extend this with a new function such as .crtc_valid() that can check both mode (for resolution) and format (for BPP). But then I start to wonde how "simple" drm_simple_display_pipe is becoming :/ I can't figure out if the crtc is even the right place to address this... > There are .mode_valid() callbacks on the bridges we use > (panel and dumb VGA) but neither uses it at the moment, hm. > I could just assign my own .mode_valid() callback to the bridge, > but it seems a bit fragile. But it's worth a hack, I'll try it. It turns out that this passes only an struct drm_display_mode which does not concern itself with display engine details like BPP. So the bridges just put limitations on modes, not on BPP, which makes a lot of sense, it corresponds to what the hardware does. It's evident when I think about it... The check needs to be done in the drm_simple_display_pipe_funcs or setting that up as per above. I just don't really see exactly where? Yours, Linus Walleij
Linus Walleij <linus.walleij@linaro.org> writes: > On Thu, Jan 25, 2018 at 4:46 AM, Eric Anholt <eric@anholt.net> wrote: > >>> + pl111_choose_max_resolution(dev, priv->memory_bw, >>> + &mode_config->max_width, >>> + &mode_config->max_height, &bpp); >>> + dev_info(dev->dev, "cap resolution at %u x %u, %u BPP\n", >>> + mode_config->max_width, mode_config->max_height, bpp); >> >> I think this is the wrong place in the pipeline to be doing this, but I >> don't have a complete solution so I'm not necessarily saying no. > > So currently the driver does this: > > mode_config->max_width = 1024; > mode_config->max_height = 768; > > And that is because it cannot really handle anything. I guess ideally > the DRM driver should set these to -1 or something so that any widths > and heights negotiated will work. The PPL field only gets you up to 1024 width. Looks like LPP lets you get up to 1024 height, though, even if the top level of the PL111 docs say 1024x768 is the max. >> Things I think we should do for bandwidth limits: >> >> A new pl111_mode_valid() rejects modes with width*height*2 > bandwidth >> (if we can't scan it out with our smallest format, don't advertise it). >> >> pl111_display_check() rejects modes with width*height*bpp > bandwidth >> (if we can't scan out this particular configuration, let them know we >> can't set the mode). >> >> Ideally given those two things, fbdev and X11 would notice that the >> preferred mode fails at 24bpp and fall back to 16bpp. I don't think >> either of those does so today, though. >> >> Interested in tackling any of these? > > I tried the pl111_display_check() version. It just made the driver > fail to initialize anything, at least when using the dumb VGA > bridge. > > There are .mode_valid() callbacks on the bridges we use > (panel and dumb VGA) but neither uses it at the moment, hm. > I could just assign my own .mode_valid() callback to the bridge, > but it seems a bit fragile. But it's worth a hack, I'll try it. > > I sent a sent of lesser controversial patches in the meantime, > and rebased this on top of those so we can deal with the > memory BW issue separately. The memory-bandwidth field is in the pl111 binding, so I do think we should handle it there.
diff --git a/drivers/gpu/drm/pl111/pl111_drm.h b/drivers/gpu/drm/pl111/pl111_drm.h index 440f53ebee8c..770ef5ce3645 100644 --- a/drivers/gpu/drm/pl111/pl111_drm.h +++ b/drivers/gpu/drm/pl111/pl111_drm.h @@ -56,6 +56,7 @@ struct pl111_drm_dev_private { struct drm_fbdev_cma *fbdev; void *regs; + u32 memory_bw; u32 ienb; u32 ctrl; /* The pixel clock (a reference to our clock divider off of CLCDCLK). */ diff --git a/drivers/gpu/drm/pl111/pl111_drv.c b/drivers/gpu/drm/pl111/pl111_drv.c index 101a9c7db6ff..1b47ee67097d 100644 --- a/drivers/gpu/drm/pl111/pl111_drv.c +++ b/drivers/gpu/drm/pl111/pl111_drv.c @@ -58,6 +58,7 @@ #include <linux/dma-buf.h> #include <linux/module.h> #include <linux/slab.h> +#include <linux/of.h> #include <drm/drmP.h> #include <drm/drm_atomic_helper.h> @@ -80,21 +81,107 @@ static const struct drm_mode_config_funcs mode_config_funcs = { .atomic_commit = drm_atomic_helper_commit, }; +/** + * pl111_choose_max_resolution() - choose max resolution + * @dev: DRM device + * @memory_bw: the graphics maximum memory bandwidth in bytes/s + * @max_width: returns the maximum width + * @max_height: returns the maximum height + * @bpp: returns the maximum bips per pixed (32 or 16) + * + * This function attempts to cut down the maximum supported resolution + * to something the system memory bus can handle. The PL111 and pixel + * clocks may be able to support higher resolutions and color depths + * but as we go up the memory bus get saturated and becomes the + * bottleneck for the resolution. On several systems using e.g. + * 1024x768 with 32 BPP results in instable flickering images. + */ +static void pl111_choose_max_resolution(struct drm_device *dev, + u32 memory_bw, + int *max_width, + int *max_height, + unsigned int *bpp) +{ + /* No limitations, this is the most aggressive */ + if (!memory_bw) { + *max_width = 1024; + *max_height = 768; + *bpp = 32; + return; + } + + /* + * 1024x768 with RGBX8888 requires a memory bandwidth of + * 65Mhz * 4 bytes = 260000000 bytes per second. + */ + if (memory_bw >= 260000000) { + *max_width = 1024; + *max_height = 768; + *bpp = 32; + return; + } + + /* + * 800x600 with RGB8888 requires a memory bandwidth of + * 36Mhz * 4 bytes = 144000000 bytes per second. But we + * assume the user prefer higher resolution over more + * color depth, so we do not add this mode here. + */ + + /* + * 1024x768 with RGB565 requires a memory bandwidth of + * 65Mhz * 2 bytes = 130000000 bytes per second. + */ + if (memory_bw >= 130000000) { + *max_width = 1024; + *max_height = 768; + *bpp = 16; + return; + } + + /* + * 800x600 with RGB565 requires a memory bandwidth of + * 36Mhz * 2 bytes = 72000000 bytes per second. + */ + if (memory_bw >= 72000000) { + *max_width = 800; + *max_height = 600; + *bpp = 16; + return; + } + + /* + * 640x480 with RGB565 requires a memory bandwidth of + * 25.175Mhz * 2 bytes = 50350000 bytes per second. + */ + if (memory_bw < 50350000) + dev_err(dev->dev, "can't even do 640x480 VGA RGB565, proceed anyway\n"); + + *max_width = 640; + *max_height = 480; + *bpp = 16; +} + static int pl111_modeset_init(struct drm_device *dev) { struct drm_mode_config *mode_config; struct pl111_drm_dev_private *priv = dev->dev_private; struct drm_panel *panel; struct drm_bridge *bridge; + unsigned int bpp; /* bits per pixel */ int ret = 0; drm_mode_config_init(dev); mode_config = &dev->mode_config; mode_config->funcs = &mode_config_funcs; mode_config->min_width = 1; - mode_config->max_width = 1024; mode_config->min_height = 1; - mode_config->max_height = 768; + + pl111_choose_max_resolution(dev, priv->memory_bw, + &mode_config->max_width, + &mode_config->max_height, &bpp); + dev_info(dev->dev, "cap resolution at %u x %u, %u BPP\n", + mode_config->max_width, mode_config->max_height, bpp); ret = drm_of_find_panel_or_bridge(dev->dev->of_node, 0, 0, &panel, &bridge); @@ -139,7 +226,7 @@ static int pl111_modeset_init(struct drm_device *dev) drm_mode_config_reset(dev); - priv->fbdev = drm_fbdev_cma_init(dev, 32, + priv->fbdev = drm_fbdev_cma_init(dev, bpp, dev->mode_config.num_connector); drm_kms_helper_poll_init(dev); @@ -216,6 +303,12 @@ static int pl111_amba_probe(struct amba_device *amba_dev, drm->dev_private = priv; priv->variant = variant; + if (of_property_read_u32(dev->of_node, "max-memory-bandwidth", + &priv->memory_bw)) { + dev_info(dev, "no max memory bandwidth specified, assume unlimited\n"); + priv->memory_bw = 0; + } + /* * The PL110 and PL111 variants have two registers * swapped: interrupt enable and control. For this reason
We were previously selecting 1024x768 and 32BPP as the default set-up for the PL111 consumers. This does not work on elder systems: the device tree bindings support a property "max-memory-bandwidth" in bytes/second that states that if you exceed this the memory bus will saturate. The result is flickering and unstable images. Parse the "max-memory-bandwidth" and respect it when intializing the driver. On the RealView PB11MP we get a nice 1024x768 console with RGB565 as default with this code. If the device tree does not specify the maximum memory bandwidth we keep the old assumption that we can support 1024x768, 32BPP. Signed-off-by: Linus Walleij <linus.walleij@linaro.org> --- drivers/gpu/drm/pl111/pl111_drm.h | 1 + drivers/gpu/drm/pl111/pl111_drv.c | 99 +++++++++++++++++++++++++++++++++++++-- 2 files changed, 97 insertions(+), 3 deletions(-)