| Message ID | 20220206185643.275811-1-marex@denx.de (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [RFC] drm: mxsfb: Implement LCDIF scanout CRC32 support | expand |
Hi Marek, On Sun, 2022-02-06 at 19:56 +0100, Marek Vasut wrote: > The LCDIF controller as present in i.MX6SX/i.MX8M Mini/Nano has a CRC_STAT > register, which contains CRC32 of the frame as it was clocked out of the > DPI interface of the LCDIF. This is likely meant as a functional safety > register. > > Unfortunatelly, there is zero documentation on how the CRC32 is calculated, > there is no documentation of the polynomial, the init value, nor on which > data is the checksum applied. > > By applying brute-force on 8 pixel / 2 line frame, which is the minimum > size LCDIF would work with, it turns out the polynomial is CRC32_POLY_LE > 0xedb88320 , init value is 0xffffffff , the input data are bitrev32() > of the entire frame and the resulting CRC has to be also bitrev32()ed. No idea how the HW calculates the CRC value. I didn't hear anyone internal tried this feature. > > Doing this calculation in software for each frame is unrealistic due to > the CPU demand, implement at least a sysfs attribute which permits testing > the current frame on demand. Why not using the existing debugfs CRC support implemented in drivers/gpu/drm/drm_debugfs_crc.c? > > Unfortunatelly, this functionality has another problem. On all of those SoCs, > it is possible to overload interconnect e.g. by concurrent USB and uSDHC > transfers, at which point the LCDIF LFIFO suffers an UNDERFLOW condition, > which results in the image being shifted to the right by exactly LFIFO size > pixels. On i.MX8M Mini, the LFIFO is 76x256 bits = 2432 Byte ~= 810 pixel > at 24bpp. In this case, the LCDIF does not assert UNDERFLOW_IRQ bit, the > frame CRC32 indicated in CRC_STAT register matches the CRC32 of the frame > in DRAM, the RECOVER_ON_UNDERFLOW bit has no effect, so if this mode of > failure occurs, the failure gets undetected and uncorrected. Hmmm, interesting, no UNDERFLOW_IRQ bit asserted when LCDIF suffers an UNDERFLOW condition? Are you sure LCDIF really underflows? If the shifted image is seen on a MIPI DSI display, could that be a MIPI DSI or DPHY issue, like wrong horizontal parameter(s)? > > Signed-off-by: Marek Vasut <marex@denx.de> > Cc: Alexander Stein <alexander.stein@ew.tq-group.com> > Cc: Laurent Pinchart <laurent.pinchart@ideasonboard.com> > Cc: Lucas Stach <l.stach@pengutronix.de> > Cc: Peng Fan <peng.fan@nxp.com> > Cc: Robby Cai <robby.cai@nxp.com> > Cc: Sam Ravnborg <sam@ravnborg.org> > Cc: Stefan Agner <stefan@agner.ch> > --- > drivers/gpu/drm/mxsfb/mxsfb_drv.c | 38 ++++++++++++++++++++++++++++++ > drivers/gpu/drm/mxsfb/mxsfb_drv.h | 3 +++ > drivers/gpu/drm/mxsfb/mxsfb_kms.c | 11 +++++---- > drivers/gpu/drm/mxsfb/mxsfb_regs.h | 1 + > 4 files changed, 49 insertions(+), 4 deletions(-) > > diff --git a/drivers/gpu/drm/mxsfb/mxsfb_drv.c b/drivers/gpu/drm/mxsfb/mxsfb_drv.c > index 4ff3c6195dd0c..6f296b398f28c 100644 > --- a/drivers/gpu/drm/mxsfb/mxsfb_drv.c > +++ b/drivers/gpu/drm/mxsfb/mxsfb_drv.c > @@ -9,6 +9,7 @@ > */ > > #include <linux/clk.h> > +#include <linux/crc32.h> > #include <linux/dma-mapping.h> > #include <linux/io.h> > #include <linux/module.h> > @@ -292,6 +293,37 @@ static void mxsfb_unload(struct drm_device *drm) > pm_runtime_disable(drm->dev); > } > > +static ssize_t mxsfb_frame_checksum_show(struct device *dev, > + struct device_attribute *attr, > + char *buf) > +{ > + struct drm_device *drm = dev_get_drvdata(dev); > + struct mxsfb_drm_private *mxsfb = drm->dev_private; > + u32 hwcrc = readl(mxsfb->base, LCDC_V4_CRC_STAT); Access register without relevant clock(s) enabled? LCDC_V4_CRC_STAT seems to hint that there should be some verion control logic for MXSFB_V3/4/6. Regards, Liu Ying > + u32 swcrc = 0xffffffff; > + int i; > + > + if (mxsfb->gem_vaddr) { > + for (i = 0; i < mxsfb->gem_size / 4; i++) { > + u32 data = bitrev32(((u32 *)mxsfb->gem_vaddr)[i]); > + swcrc = crc32(swcrc, &data, 4); > + } > + swcrc = bitrev32(swcrc); > + } > + > + return sysfs_emit(buf, "HW:%08x,SW:%08x,OK:%d\n", hwcrc, swcrc, hwcrc == swcrc); > +} > +static DEVICE_ATTR(frame_checksum, 0444, mxsfb_frame_checksum_show, NULL); > + > +static struct attribute *mxsfb_attributes[] = { > + &dev_attr_frame_checksum.attr, > + NULL, > +}; > + > +static const struct attribute_group mxsfb_attr_group = { > + .attrs = mxsfb_attributes, > +}; > + > DEFINE_DRM_GEM_CMA_FOPS(fops); > > static const struct drm_driver mxsfb_driver = { > @@ -335,10 +367,16 @@ static int mxsfb_probe(struct platform_device *pdev) > if (ret) > goto err_unload; > > + ret = devm_device_add_group(drm->dev, &mxsfb_attr_group); > + if (ret) > + goto err_attr; > + > drm_fbdev_generic_setup(drm, 32); > > return 0; > > +err_attr: > + drm_dev_unregister(drm); > err_unload: > mxsfb_unload(drm); > err_free: > diff --git a/drivers/gpu/drm/mxsfb/mxsfb_drv.h b/drivers/gpu/drm/mxsfb/mxsfb_drv.h > index ddb5b0417a82c..0a3e5dd1e8bab 100644 > --- a/drivers/gpu/drm/mxsfb/mxsfb_drv.h > +++ b/drivers/gpu/drm/mxsfb/mxsfb_drv.h > @@ -44,6 +44,9 @@ struct mxsfb_drm_private { > struct drm_encoder encoder; > struct drm_connector *connector; > struct drm_bridge *bridge; > + > + void *gem_vaddr; > + size_t gem_size; > }; > > static inline struct mxsfb_drm_private * > diff --git a/drivers/gpu/drm/mxsfb/mxsfb_kms.c b/drivers/gpu/drm/mxsfb/mxsfb_kms.c > index 03743a84c8e79..2a4edf5a2ac57 100644 > --- a/drivers/gpu/drm/mxsfb/mxsfb_kms.c > +++ b/drivers/gpu/drm/mxsfb/mxsfb_kms.c > @@ -196,7 +196,7 @@ static int mxsfb_reset_block(struct mxsfb_drm_private *mxsfb) > return clear_poll_bit(mxsfb->base + LCDC_CTRL, CTRL_CLKGATE); > } > > -static dma_addr_t mxsfb_get_fb_paddr(struct drm_plane *plane) > +static dma_addr_t mxsfb_get_fb_paddr(struct mxsfb_drm_private *mxsfb, struct drm_plane *plane) > { > struct drm_framebuffer *fb = plane->state->fb; > struct drm_gem_cma_object *gem; > @@ -208,6 +208,9 @@ static dma_addr_t mxsfb_get_fb_paddr(struct drm_plane *plane) > if (!gem) > return 0; > > + mxsfb->gem_vaddr = gem->vaddr; > + mxsfb->gem_size = gem->base.size; > + > return gem->paddr; > } > > @@ -370,7 +373,7 @@ static void mxsfb_crtc_atomic_enable(struct drm_crtc *crtc, > mxsfb_crtc_mode_set_nofb(mxsfb, bus_format); > > /* Write cur_buf as well to avoid an initial corrupt frame */ > - paddr = mxsfb_get_fb_paddr(crtc->primary); > + paddr = mxsfb_get_fb_paddr(mxsfb, crtc->primary); > if (paddr) { > writel(paddr, mxsfb->base + mxsfb->devdata->cur_buf); > writel(paddr, mxsfb->base + mxsfb->devdata->next_buf); > @@ -476,7 +479,7 @@ static void mxsfb_plane_primary_atomic_update(struct drm_plane *plane, > struct mxsfb_drm_private *mxsfb = to_mxsfb_drm_private(plane->dev); > dma_addr_t paddr; > > - paddr = mxsfb_get_fb_paddr(plane); > + paddr = mxsfb_get_fb_paddr(mxsfb, plane); > if (paddr) > writel(paddr, mxsfb->base + mxsfb->devdata->next_buf); > } > @@ -492,7 +495,7 @@ static void mxsfb_plane_overlay_atomic_update(struct drm_plane *plane, > dma_addr_t paddr; > u32 ctrl; > > - paddr = mxsfb_get_fb_paddr(plane); > + paddr = mxsfb_get_fb_paddr(mxsfb, plane); > if (!paddr) { > writel(0, mxsfb->base + LCDC_AS_CTRL); > return; > diff --git a/drivers/gpu/drm/mxsfb/mxsfb_regs.h b/drivers/gpu/drm/mxsfb/mxsfb_regs.h > index 694fea13e893e..cf813a1da1d78 100644 > --- a/drivers/gpu/drm/mxsfb/mxsfb_regs.h > +++ b/drivers/gpu/drm/mxsfb/mxsfb_regs.h > @@ -26,6 +26,7 @@ > #define LCDC_VDCTRL2 0x90 > #define LCDC_VDCTRL3 0xa0 > #define LCDC_VDCTRL4 0xb0 > +#define LCDC_V4_CRC_STAT 0x1a0 > #define LCDC_V4_DEBUG0 0x1d0 > #define LCDC_V3_DEBUG0 0x1f0 > #define LCDC_AS_CTRL 0x210
On 2/7/22 06:13, Liu Ying wrote: > Hi Marek, Hi, > On Sun, 2022-02-06 at 19:56 +0100, Marek Vasut wrote: >> The LCDIF controller as present in i.MX6SX/i.MX8M Mini/Nano has a CRC_STAT >> register, which contains CRC32 of the frame as it was clocked out of the >> DPI interface of the LCDIF. This is likely meant as a functional safety >> register. >> >> Unfortunatelly, there is zero documentation on how the CRC32 is calculated, >> there is no documentation of the polynomial, the init value, nor on which >> data is the checksum applied. >> >> By applying brute-force on 8 pixel / 2 line frame, which is the minimum >> size LCDIF would work with, it turns out the polynomial is CRC32_POLY_LE >> 0xedb88320 , init value is 0xffffffff , the input data are bitrev32() >> of the entire frame and the resulting CRC has to be also bitrev32()ed. > > No idea how the HW calculates the CRC value. > I didn't hear anyone internal tried this feature. It would be nice if the datasheet could be improved. There are many blank areas which are undocumented, this LCDIF CRC32 feature, i.MX8M Mini Arteris NOC at 0x32700000 , the ARM GPV NIC-301 at 0x32{0,1,2,3,4,5,6,8}00000 and their master/slave port mapping. The NOC and NICs were documented at least up to i.MX6QP and then that information disappeared from NXP datasheets. I think reconfiguring the NOC/NIC QoS would help mitigate this shift issue described below (*). Do you know if there is some additional NOC/NIC documentation for i.MX8M Mini available ? >> Doing this calculation in software for each frame is unrealistic due to >> the CPU demand, implement at least a sysfs attribute which permits testing >> the current frame on demand. > > Why not using the existing debugfs CRC support implemented > in drivers/gpu/drm/drm_debugfs_crc.c? I wasn't aware of that, thanks. >> Unfortunatelly, this functionality has another problem. On all of those SoCs, >> it is possible to overload interconnect e.g. by concurrent USB and uSDHC >> transfers, at which point the LCDIF LFIFO suffers an UNDERFLOW condition, >> which results in the image being shifted to the right by exactly LFIFO size >> pixels. On i.MX8M Mini, the LFIFO is 76x256 bits = 2432 Byte ~= 810 pixel >> at 24bpp. In this case, the LCDIF does not assert UNDERFLOW_IRQ bit, the >> frame CRC32 indicated in CRC_STAT register matches the CRC32 of the frame >> in DRAM, the RECOVER_ON_UNDERFLOW bit has no effect, so if this mode of >> failure occurs, the failure gets undetected and uncorrected. > > Hmmm, interesting, no UNDERFLOW_IRQ bit asserted when LCDIF suffers an > UNDERFLOW condition? Yes > Are you sure LCDIF really underflows? Mostly sure. This problem occurs also on i.MX6SX which has no DSIM. The failure is triggered by many short writes into DRAM to different addresses (I was successful at triggering it by using i.MX8M Mini with ASIX 88772 USB ethernet adapter, running iperf3 on the device, iperf3 -c ... -t 0 -R -P 16 on the PC). This effectively makes the CI HDRC behave as a DRAM thrashing AXI master, since it triggers a lot of short USB qTD READs from DRAM and a lot of short ethernet packet WRITEs to DRAM. And that either clogs the DRAM itself, or the NOC or DISPLAY/HSIO NIC-301, and prevents LCDIF from getting data long enough for this underflow condition to happen, LFIFO to underflow, and this shift to appear. And the shift does not disappear automatically itself, it just stays there until the LCDIF is reinitialized. And it apparently also happens on iMXRT, where a suggestion was made to tweak the QoS settings of the interconnect (which cannot be tested on i.MX8M Mini, because neither of that documentation is available, see above (*)): https://community.nxp.com/t5/i-MX-RT/iMXRT1052-LCD-Screen-shifted/td-p/1069978 > If the shifted image is seen on a MIPI DSI display, could that be a > MIPI DSI or DPHY issue, like wrong horizontal parameter(s)? No, it happens also on i.MX6SX without DSIM, so this is LCDIF problem.
On Mon, 2022-02-07 at 09:14 +0100, Marek Vasut wrote: > On 2/7/22 06:13, Liu Ying wrote: > > Hi Marek, > > Hi, > > > On Sun, 2022-02-06 at 19:56 +0100, Marek Vasut wrote: > > > The LCDIF controller as present in i.MX6SX/i.MX8M Mini/Nano has a > > > CRC_STAT > > > register, which contains CRC32 of the frame as it was clocked out > > > of the > > > DPI interface of the LCDIF. This is likely meant as a functional > > > safety > > > register. > > > > > > Unfortunatelly, there is zero documentation on how the CRC32 is > > > calculated, > > > there is no documentation of the polynomial, the init value, nor > > > on which > > > data is the checksum applied. > > > > > > By applying brute-force on 8 pixel / 2 line frame, which is the > > > minimum > > > size LCDIF would work with, it turns out the polynomial is > > > CRC32_POLY_LE > > > 0xedb88320 , init value is 0xffffffff , the input data are > > > bitrev32() > > > of the entire frame and the resulting CRC has to be also > > > bitrev32()ed. > > > > No idea how the HW calculates the CRC value. > > I didn't hear anyone internal tried this feature. > > It would be nice if the datasheet could be improved. Agreed. > > There are many blank areas which are undocumented, this LCDIF CRC32 > feature, i.MX8M Mini Arteris NOC at 0x32700000 , the ARM GPV NIC-301 > at > 0x32{0,1,2,3,4,5,6,8}00000 and their master/slave port mapping. The > NOC > and NICs were documented at least up to i.MX6QP and then that > information disappeared from NXP datasheets. I think reconfiguring > the > NOC/NIC QoS would help mitigate this shift issue described below (*). I also think the QoS would help if it is configureable. > > Do you know if there is some additional NOC/NIC documentation for > i.MX8M > Mini available ? No. > > > > Doing this calculation in software for each frame is unrealistic > > > due to > > > the CPU demand, implement at least a sysfs attribute which > > > permits testing > > > the current frame on demand. > > > > Why not using the existing debugfs CRC support implemented > > in drivers/gpu/drm/drm_debugfs_crc.c? > > I wasn't aware of that, thanks. No problem. > > > > Unfortunatelly, this functionality has another problem. On all of > > > those SoCs, > > > it is possible to overload interconnect e.g. by concurrent USB > > > and uSDHC > > > transfers, at which point the LCDIF LFIFO suffers an UNDERFLOW > > > condition, > > > which results in the image being shifted to the right by exactly > > > LFIFO size > > > pixels. On i.MX8M Mini, the LFIFO is 76x256 bits = 2432 Byte ~= > > > 810 pixel > > > at 24bpp. In this case, the LCDIF does not assert UNDERFLOW_IRQ > > > bit, the > > > frame CRC32 indicated in CRC_STAT register matches the CRC32 of > > > the frame > > > in DRAM, the RECOVER_ON_UNDERFLOW bit has no effect, so if this > > > mode of > > > failure occurs, the failure gets undetected and uncorrected. > > > > Hmmm, interesting, no UNDERFLOW_IRQ bit asserted when LCDIF suffers > > an > > UNDERFLOW condition? > > Yes Did you ever see UNDERFLOW_IRQ bit asserted in any case? Liu Ying
On 2/7/22 10:18, Liu Ying wrote: Hi, >>> On Sun, 2022-02-06 at 19:56 +0100, Marek Vasut wrote: >>>> The LCDIF controller as present in i.MX6SX/i.MX8M Mini/Nano has a >>>> CRC_STAT >>>> register, which contains CRC32 of the frame as it was clocked out >>>> of the >>>> DPI interface of the LCDIF. This is likely meant as a functional >>>> safety >>>> register. >>>> >>>> Unfortunatelly, there is zero documentation on how the CRC32 is >>>> calculated, >>>> there is no documentation of the polynomial, the init value, nor >>>> on which >>>> data is the checksum applied. >>>> >>>> By applying brute-force on 8 pixel / 2 line frame, which is the >>>> minimum >>>> size LCDIF would work with, it turns out the polynomial is >>>> CRC32_POLY_LE >>>> 0xedb88320 , init value is 0xffffffff , the input data are >>>> bitrev32() >>>> of the entire frame and the resulting CRC has to be also >>>> bitrev32()ed. >>> >>> No idea how the HW calculates the CRC value. >>> I didn't hear anyone internal tried this feature. >> >> It would be nice if the datasheet could be improved. > > Agreed. > >> >> There are many blank areas which are undocumented, this LCDIF CRC32 >> feature, i.MX8M Mini Arteris NOC at 0x32700000 , the ARM GPV NIC-301 >> at >> 0x32{0,1,2,3,4,5,6,8}00000 and their master/slave port mapping. The >> NOC >> and NICs were documented at least up to i.MX6QP and then that >> information disappeared from NXP datasheets. I think reconfiguring >> the >> NOC/NIC QoS would help mitigate this shift issue described below (*). > > I also think the QoS would help if it is configureable. It is programmable, it's just the port mapping which is undocumented. >> Do you know if there is some additional NOC/NIC documentation for >> i.MX8M >> Mini available ? > > No. Can you ask someone internally in NXP maybe ? >>>> Doing this calculation in software for each frame is unrealistic >>>> due to >>>> the CPU demand, implement at least a sysfs attribute which >>>> permits testing >>>> the current frame on demand. >>> >>> Why not using the existing debugfs CRC support implemented >>> in drivers/gpu/drm/drm_debugfs_crc.c? >> >> I wasn't aware of that, thanks. > > No problem. > >> >>>> Unfortunatelly, this functionality has another problem. On all of >>>> those SoCs, >>>> it is possible to overload interconnect e.g. by concurrent USB >>>> and uSDHC >>>> transfers, at which point the LCDIF LFIFO suffers an UNDERFLOW >>>> condition, >>>> which results in the image being shifted to the right by exactly >>>> LFIFO size >>>> pixels. On i.MX8M Mini, the LFIFO is 76x256 bits = 2432 Byte ~= >>>> 810 pixel >>>> at 24bpp. In this case, the LCDIF does not assert UNDERFLOW_IRQ >>>> bit, the >>>> frame CRC32 indicated in CRC_STAT register matches the CRC32 of >>>> the frame >>>> in DRAM, the RECOVER_ON_UNDERFLOW bit has no effect, so if this >>>> mode of >>>> failure occurs, the failure gets undetected and uncorrected. >>> >>> Hmmm, interesting, no UNDERFLOW_IRQ bit asserted when LCDIF suffers >>> an >>> UNDERFLOW condition? >> >> Yes > > Did you ever see UNDERFLOW_IRQ bit asserted in any case? I didn't see the UNDERFLOW_IRQ bit asserted during my tests, either with this IRQ enabled (UNDERFLOW_IRQ_EN=1) or with the IRQ disabled (UNDERFLOW_IRQ_EN=0) by reading the CTRL1 register in interrupt handler when CUR_FRAME_DONE_IRQ triggered the IRQ handler. I did see a few auto-recoveries of the panel back into non-shifted image, that happened once in some 100-200 tests. Mostly the LCDIF does not recover automatically.
On Mon, 2022-02-07 at 11:43 +0100, Marek Vasut wrote: > On 2/7/22 10:18, Liu Ying wrote: > > Hi, > > > > > On Sun, 2022-02-06 at 19:56 +0100, Marek Vasut wrote: > > > > > The LCDIF controller as present in i.MX6SX/i.MX8M Mini/Nano > > > > > has a > > > > > CRC_STAT > > > > > register, which contains CRC32 of the frame as it was clocked > > > > > out > > > > > of the > > > > > DPI interface of the LCDIF. This is likely meant as a > > > > > functional > > > > > safety > > > > > register. > > > > > > > > > > Unfortunatelly, there is zero documentation on how the CRC32 > > > > > is > > > > > calculated, > > > > > there is no documentation of the polynomial, the init value, > > > > > nor > > > > > on which > > > > > data is the checksum applied. > > > > > > > > > > By applying brute-force on 8 pixel / 2 line frame, which is > > > > > the > > > > > minimum > > > > > size LCDIF would work with, it turns out the polynomial is > > > > > CRC32_POLY_LE > > > > > 0xedb88320 , init value is 0xffffffff , the input data are > > > > > bitrev32() > > > > > of the entire frame and the resulting CRC has to be also > > > > > bitrev32()ed. > > > > > > > > No idea how the HW calculates the CRC value. > > > > I didn't hear anyone internal tried this feature. > > > > > > It would be nice if the datasheet could be improved. > > > > Agreed. > > > > > > > > There are many blank areas which are undocumented, this LCDIF > > > CRC32 > > > feature, i.MX8M Mini Arteris NOC at 0x32700000 , the ARM GPV NIC- > > > 301 > > > at > > > 0x32{0,1,2,3,4,5,6,8}00000 and their master/slave port mapping. > > > The > > > NOC > > > and NICs were documented at least up to i.MX6QP and then that > > > information disappeared from NXP datasheets. I think > > > reconfiguring > > > the > > > NOC/NIC QoS would help mitigate this shift issue described below > > > (*). > > > > I also think the QoS would help if it is configureable. > > It is programmable, it's just the port mapping which is undocumented. > > > > Do you know if there is some additional NOC/NIC documentation for > > > i.MX8M > > > Mini available ? > > > > No. > > Can you ask someone internally in NXP maybe ? Maybe, you may try community.nxp.com, like the i.MXRT case. Liu Ying
Hello Liu Ying, On Tue, Feb 08, 2022 at 10:41:59AM +0800, Liu Ying wrote: > On Mon, 2022-02-07 at 11:43 +0100, Marek Vasut wrote: > > On 2/7/22 10:18, Liu Ying wrote: > > > > > On Sun, 2022-02-06 at 19:56 +0100, Marek Vasut wrote: > > > > > > The LCDIF controller as present in i.MX6SX/i.MX8M Mini/Nano has a CRC_STAT > > > > > > register, which contains CRC32 of the frame as it was clocked out of the > > > > > > DPI interface of the LCDIF. This is likely meant as a functional safety > > > > > > register. > > > > > > > > > > > > Unfortunatelly, there is zero documentation on how the CRC32 is calculated, > > > > > > there is no documentation of the polynomial, the init value, nor on which > > > > > > data is the checksum applied. > > > > > > > > > > > > By applying brute-force on 8 pixel / 2 line frame, which is the minimum > > > > > > size LCDIF would work with, it turns out the polynomial is CRC32_POLY_LE > > > > > > 0xedb88320 , init value is 0xffffffff , the input data are bitrev32() > > > > > > of the entire frame and the resulting CRC has to be also bitrev32()ed. > > > > > > > > > > No idea how the HW calculates the CRC value. > > > > > I didn't hear anyone internal tried this feature. > > > > > > > > It would be nice if the datasheet could be improved. > > > > > > Agreed. > > > > > > > There are many blank areas which are undocumented, this LCDIF CRC32 > > > > feature, i.MX8M Mini Arteris NOC at 0x32700000 , the ARM GPV NIC-301 at > > > > 0x32{0,1,2,3,4,5,6,8}00000 and their master/slave port mapping. The NOC > > > > and NICs were documented at least up to i.MX6QP and then that > > > > information disappeared from NXP datasheets. I think reconfiguring the > > > > NOC/NIC QoS would help mitigate this shift issue described below > > > > (*). > > > > > > I also think the QoS would help if it is configureable. > > > > It is programmable, it's just the port mapping which is undocumented. > > > > > > Do you know if there is some additional NOC/NIC documentation for i.MX8M > > > > Mini available ? > > > > > > No. > > > > Can you ask someone internally in NXP maybe ? > > Maybe, you may try community.nxp.com, like the i.MXRT case. Overall we seem to have had little luck with community.nxp.com. I wonder if it would be possible for key community members to get some more direct access to support when working on upstream drivers. I'm pretty sure nobody will try to abuse it :-)
On 2/8/22 03:41, Liu Ying wrote: Hello everyone, >>>> There are many blank areas which are undocumented, this LCDIF >>>> CRC32 >>>> feature, i.MX8M Mini Arteris NOC at 0x32700000 , the ARM GPV NIC- >>>> 301 >>>> at >>>> 0x32{0,1,2,3,4,5,6,8}00000 and their master/slave port mapping. >>>> The >>>> NOC >>>> and NICs were documented at least up to i.MX6QP and then that >>>> information disappeared from NXP datasheets. I think >>>> reconfiguring >>>> the >>>> NOC/NIC QoS would help mitigate this shift issue described below >>>> (*). >>> >>> I also think the QoS would help if it is configureable. >> >> It is programmable, it's just the port mapping which is undocumented. >> >>>> Do you know if there is some additional NOC/NIC documentation for >>>> i.MX8M >>>> Mini available ? >>> >>> No. >> >> Can you ask someone internally in NXP maybe ? > > Maybe, you may try community.nxp.com, like the i.MXRT case. The community.nxp.com is unhelpful, the i.MXRT case it a good example -- the solution to the problem has been found by the person who asked the question on their own, and elsewhere too. But note that the i.MXRT interconnect documentation is available in the i.MXRT datasheet, which made that possible in the first place. On i.MX, all that information has been removed from the datasheet in i.MX7 and i.MX8M, so I cannot even help myself, even if I wanted to. This is very bad.
Hello Laurent, On Tue, 2022-02-08 at 05:03 +0200, Laurent Pinchart wrote: > Hello Liu Ying, > > On Tue, Feb 08, 2022 at 10:41:59AM +0800, Liu Ying wrote: > > On Mon, 2022-02-07 at 11:43 +0100, Marek Vasut wrote: > > > On 2/7/22 10:18, Liu Ying wrote: > > > > > > On Sun, 2022-02-06 at 19:56 +0100, Marek Vasut wrote: > > > > > > > The LCDIF controller as present in i.MX6SX/i.MX8M > > > > > > > Mini/Nano has a CRC_STAT > > > > > > > register, which contains CRC32 of the frame as it was > > > > > > > clocked out of the > > > > > > > DPI interface of the LCDIF. This is likely meant as a > > > > > > > functional safety > > > > > > > register. > > > > > > > > > > > > > > Unfortunatelly, there is zero documentation on how the > > > > > > > CRC32 is calculated, > > > > > > > there is no documentation of the polynomial, the init > > > > > > > value, nor on which > > > > > > > data is the checksum applied. > > > > > > > > > > > > > > By applying brute-force on 8 pixel / 2 line frame, which > > > > > > > is the minimum > > > > > > > size LCDIF would work with, it turns out the polynomial > > > > > > > is CRC32_POLY_LE > > > > > > > 0xedb88320 , init value is 0xffffffff , the input data > > > > > > > are bitrev32() > > > > > > > of the entire frame and the resulting CRC has to be also > > > > > > > bitrev32()ed. > > > > > > > > > > > > No idea how the HW calculates the CRC value. > > > > > > I didn't hear anyone internal tried this feature. > > > > > > > > > > It would be nice if the datasheet could be improved. > > > > > > > > Agreed. > > > > > > > > > There are many blank areas which are undocumented, this LCDIF > > > > > CRC32 > > > > > feature, i.MX8M Mini Arteris NOC at 0x32700000 , the ARM GPV > > > > > NIC-301 at > > > > > 0x32{0,1,2,3,4,5,6,8}00000 and their master/slave port > > > > > mapping. The NOC > > > > > and NICs were documented at least up to i.MX6QP and then that > > > > > information disappeared from NXP datasheets. I think > > > > > reconfiguring the > > > > > NOC/NIC QoS would help mitigate this shift issue described > > > > > below > > > > > (*). > > > > > > > > I also think the QoS would help if it is configureable. > > > > > > It is programmable, it's just the port mapping which is > > > undocumented. > > > > > > > > Do you know if there is some additional NOC/NIC documentation > > > > > for i.MX8M > > > > > Mini available ? > > > > > > > > No. > > > > > > Can you ask someone internally in NXP maybe ? > > > > Maybe, you may try community.nxp.com, like the i.MXRT case. > > Overall we seem to have had little luck with community.nxp.com. I > wonder > if it would be possible for key community members to get some more > direct access to support when working on upstream drivers. I'm pretty > sure nobody will try to abuse it :-) I'm not sure if it is possible. It's not a bad idea in my personal opinion. Liu Ying
On Tue, 2022-02-08 at 11:02 +0100, Marek Vasut wrote: > On 2/8/22 03:41, Liu Ying wrote: > > Hello everyone, > > > > > > There are many blank areas which are undocumented, this LCDIF > > > > > CRC32 > > > > > feature, i.MX8M Mini Arteris NOC at 0x32700000 , the ARM GPV > > > > > NIC- > > > > > 301 > > > > > at > > > > > 0x32{0,1,2,3,4,5,6,8}00000 and their master/slave port > > > > > mapping. > > > > > The > > > > > NOC > > > > > and NICs were documented at least up to i.MX6QP and then that > > > > > information disappeared from NXP datasheets. I think > > > > > reconfiguring > > > > > the > > > > > NOC/NIC QoS would help mitigate this shift issue described > > > > > below > > > > > (*). > > > > > > > > I also think the QoS would help if it is configureable. > > > > > > It is programmable, it's just the port mapping which is > > > undocumented. > > > > > > > > Do you know if there is some additional NOC/NIC documentation > > > > > for > > > > > i.MX8M > > > > > Mini available ? > > > > > > > > No. > > > > > > Can you ask someone internally in NXP maybe ? > > > > Maybe, you may try community.nxp.com, like the i.MXRT case. > > The community.nxp.com is unhelpful, the i.MXRT case it a good example > -- > the solution to the problem has been found by the person who asked > the > question on their own, and elsewhere too. AFAIK, there are questions answered by internal support team and RnD team at that community. I personally take it as a resource to use. > > But note that the i.MXRT interconnect documentation is available in > the > i.MXRT datasheet, which made that possible in the first place. On > i.MX, > all that information has been removed from the datasheet in i.MX7 > and > i.MX8M, so I cannot even help myself, even if I wanted to. This is > very bad. I'm not familiar with the documention in that area, so I personally will not be helpful at the documention topic. The main purpose I jumped in this thread is to review the patch and share the idea to use the existing drm debugfs crc support instead of creating a sysfs attribute. Liu Ying
On 2/10/22 06:22, Liu Ying wrote: Hi, [...] >>>>>> There are many blank areas which are undocumented, this LCDIF >>>>>> CRC32 >>>>>> feature, i.MX8M Mini Arteris NOC at 0x32700000 , the ARM GPV >>>>>> NIC- >>>>>> 301 >>>>>> at >>>>>> 0x32{0,1,2,3,4,5,6,8}00000 and their master/slave port >>>>>> mapping. >>>>>> The >>>>>> NOC >>>>>> and NICs were documented at least up to i.MX6QP and then that >>>>>> information disappeared from NXP datasheets. I think >>>>>> reconfiguring >>>>>> the >>>>>> NOC/NIC QoS would help mitigate this shift issue described >>>>>> below >>>>>> (*). >>>>> >>>>> I also think the QoS would help if it is configureable. >>>> >>>> It is programmable, it's just the port mapping which is >>>> undocumented. >>>> >>>>>> Do you know if there is some additional NOC/NIC documentation >>>>>> for >>>>>> i.MX8M >>>>>> Mini available ? >>>>> >>>>> No. >>>> >>>> Can you ask someone internally in NXP maybe ? >>> >>> Maybe, you may try community.nxp.com, like the i.MXRT case. >> >> The community.nxp.com is unhelpful, the i.MXRT case it a good example >> -- >> the solution to the problem has been found by the person who asked >> the >> question on their own, and elsewhere too. > > AFAIK, there are questions answered by internal support team and RnD > team at that community. I personally take it as a resource to use. Sure, here is a list of links to similar problem triggered by various people using the NXP BSP, neither of them in answered: https://community.nxp.com/t5/i-MX-Processors/Image-shift-right-for-LVDS/td-p/969581 https://community.nxp.com/t5/i-MX-Processors/iMX8M-display-shifted-after-playing-decoded-video-with-gstreamer/td-p/928269 https://community.nxp.com/t5/i-MX-Processors/Display-Wrap-Around-Issue/td-p/1084052 https://community.nxp.com/t5/i-MX-Processors/Display-Vertically-shifted-IMX8mq-evk-board-in-dual-display-use/m-p/965726 https://community.nxp.com/t5/i-MX-RT/iMXRT1052-LCD-Screen-shifted/td-p/1069978 >> But note that the i.MXRT interconnect documentation is available in >> the >> i.MXRT datasheet, which made that possible in the first place. On >> i.MX, >> all that information has been removed from the datasheet in i.MX7 >> and >> i.MX8M, so I cannot even help myself, even if I wanted to. This is >> very bad. > > I'm not familiar with the documention in that area, so I personally > will not be helpful at the documention topic. The main purpose I > jumped in this thread is to review the patch and share the idea to use > the existing drm debugfs crc support instead of creating a sysfs > attribute. Can you maybe ask someone inside NXP about this problem ?
diff --git a/drivers/gpu/drm/mxsfb/mxsfb_drv.c b/drivers/gpu/drm/mxsfb/mxsfb_drv.c index 4ff3c6195dd0c..6f296b398f28c 100644 --- a/drivers/gpu/drm/mxsfb/mxsfb_drv.c +++ b/drivers/gpu/drm/mxsfb/mxsfb_drv.c @@ -9,6 +9,7 @@ */ #include <linux/clk.h> +#include <linux/crc32.h> #include <linux/dma-mapping.h> #include <linux/io.h> #include <linux/module.h> @@ -292,6 +293,37 @@ static void mxsfb_unload(struct drm_device *drm) pm_runtime_disable(drm->dev); } +static ssize_t mxsfb_frame_checksum_show(struct device *dev, + struct device_attribute *attr, + char *buf) +{ + struct drm_device *drm = dev_get_drvdata(dev); + struct mxsfb_drm_private *mxsfb = drm->dev_private; + u32 hwcrc = readl(mxsfb->base, LCDC_V4_CRC_STAT); + u32 swcrc = 0xffffffff; + int i; + + if (mxsfb->gem_vaddr) { + for (i = 0; i < mxsfb->gem_size / 4; i++) { + u32 data = bitrev32(((u32 *)mxsfb->gem_vaddr)[i]); + swcrc = crc32(swcrc, &data, 4); + } + swcrc = bitrev32(swcrc); + } + + return sysfs_emit(buf, "HW:%08x,SW:%08x,OK:%d\n", hwcrc, swcrc, hwcrc == swcrc); +} +static DEVICE_ATTR(frame_checksum, 0444, mxsfb_frame_checksum_show, NULL); + +static struct attribute *mxsfb_attributes[] = { + &dev_attr_frame_checksum.attr, + NULL, +}; + +static const struct attribute_group mxsfb_attr_group = { + .attrs = mxsfb_attributes, +}; + DEFINE_DRM_GEM_CMA_FOPS(fops); static const struct drm_driver mxsfb_driver = { @@ -335,10 +367,16 @@ static int mxsfb_probe(struct platform_device *pdev) if (ret) goto err_unload; + ret = devm_device_add_group(drm->dev, &mxsfb_attr_group); + if (ret) + goto err_attr; + drm_fbdev_generic_setup(drm, 32); return 0; +err_attr: + drm_dev_unregister(drm); err_unload: mxsfb_unload(drm); err_free: diff --git a/drivers/gpu/drm/mxsfb/mxsfb_drv.h b/drivers/gpu/drm/mxsfb/mxsfb_drv.h index ddb5b0417a82c..0a3e5dd1e8bab 100644 --- a/drivers/gpu/drm/mxsfb/mxsfb_drv.h +++ b/drivers/gpu/drm/mxsfb/mxsfb_drv.h @@ -44,6 +44,9 @@ struct mxsfb_drm_private { struct drm_encoder encoder; struct drm_connector *connector; struct drm_bridge *bridge; + + void *gem_vaddr; + size_t gem_size; }; static inline struct mxsfb_drm_private * diff --git a/drivers/gpu/drm/mxsfb/mxsfb_kms.c b/drivers/gpu/drm/mxsfb/mxsfb_kms.c index 03743a84c8e79..2a4edf5a2ac57 100644 --- a/drivers/gpu/drm/mxsfb/mxsfb_kms.c +++ b/drivers/gpu/drm/mxsfb/mxsfb_kms.c @@ -196,7 +196,7 @@ static int mxsfb_reset_block(struct mxsfb_drm_private *mxsfb) return clear_poll_bit(mxsfb->base + LCDC_CTRL, CTRL_CLKGATE); } -static dma_addr_t mxsfb_get_fb_paddr(struct drm_plane *plane) +static dma_addr_t mxsfb_get_fb_paddr(struct mxsfb_drm_private *mxsfb, struct drm_plane *plane) { struct drm_framebuffer *fb = plane->state->fb; struct drm_gem_cma_object *gem; @@ -208,6 +208,9 @@ static dma_addr_t mxsfb_get_fb_paddr(struct drm_plane *plane) if (!gem) return 0; + mxsfb->gem_vaddr = gem->vaddr; + mxsfb->gem_size = gem->base.size; + return gem->paddr; } @@ -370,7 +373,7 @@ static void mxsfb_crtc_atomic_enable(struct drm_crtc *crtc, mxsfb_crtc_mode_set_nofb(mxsfb, bus_format); /* Write cur_buf as well to avoid an initial corrupt frame */ - paddr = mxsfb_get_fb_paddr(crtc->primary); + paddr = mxsfb_get_fb_paddr(mxsfb, crtc->primary); if (paddr) { writel(paddr, mxsfb->base + mxsfb->devdata->cur_buf); writel(paddr, mxsfb->base + mxsfb->devdata->next_buf); @@ -476,7 +479,7 @@ static void mxsfb_plane_primary_atomic_update(struct drm_plane *plane, struct mxsfb_drm_private *mxsfb = to_mxsfb_drm_private(plane->dev); dma_addr_t paddr; - paddr = mxsfb_get_fb_paddr(plane); + paddr = mxsfb_get_fb_paddr(mxsfb, plane); if (paddr) writel(paddr, mxsfb->base + mxsfb->devdata->next_buf); } @@ -492,7 +495,7 @@ static void mxsfb_plane_overlay_atomic_update(struct drm_plane *plane, dma_addr_t paddr; u32 ctrl; - paddr = mxsfb_get_fb_paddr(plane); + paddr = mxsfb_get_fb_paddr(mxsfb, plane); if (!paddr) { writel(0, mxsfb->base + LCDC_AS_CTRL); return; diff --git a/drivers/gpu/drm/mxsfb/mxsfb_regs.h b/drivers/gpu/drm/mxsfb/mxsfb_regs.h index 694fea13e893e..cf813a1da1d78 100644 --- a/drivers/gpu/drm/mxsfb/mxsfb_regs.h +++ b/drivers/gpu/drm/mxsfb/mxsfb_regs.h @@ -26,6 +26,7 @@ #define LCDC_VDCTRL2 0x90 #define LCDC_VDCTRL3 0xa0 #define LCDC_VDCTRL4 0xb0 +#define LCDC_V4_CRC_STAT 0x1a0 #define LCDC_V4_DEBUG0 0x1d0 #define LCDC_V3_DEBUG0 0x1f0 #define LCDC_AS_CTRL 0x210
The LCDIF controller as present in i.MX6SX/i.MX8M Mini/Nano has a CRC_STAT register, which contains CRC32 of the frame as it was clocked out of the DPI interface of the LCDIF. This is likely meant as a functional safety register. Unfortunatelly, there is zero documentation on how the CRC32 is calculated, there is no documentation of the polynomial, the init value, nor on which data is the checksum applied. By applying brute-force on 8 pixel / 2 line frame, which is the minimum size LCDIF would work with, it turns out the polynomial is CRC32_POLY_LE 0xedb88320 , init value is 0xffffffff , the input data are bitrev32() of the entire frame and the resulting CRC has to be also bitrev32()ed. Doing this calculation in software for each frame is unrealistic due to the CPU demand, implement at least a sysfs attribute which permits testing the current frame on demand. Unfortunatelly, this functionality has another problem. On all of those SoCs, it is possible to overload interconnect e.g. by concurrent USB and uSDHC transfers, at which point the LCDIF LFIFO suffers an UNDERFLOW condition, which results in the image being shifted to the right by exactly LFIFO size pixels. On i.MX8M Mini, the LFIFO is 76x256 bits = 2432 Byte ~= 810 pixel at 24bpp. In this case, the LCDIF does not assert UNDERFLOW_IRQ bit, the frame CRC32 indicated in CRC_STAT register matches the CRC32 of the frame in DRAM, the RECOVER_ON_UNDERFLOW bit has no effect, so if this mode of failure occurs, the failure gets undetected and uncorrected. Signed-off-by: Marek Vasut <marex@denx.de> Cc: Alexander Stein <alexander.stein@ew.tq-group.com> Cc: Laurent Pinchart <laurent.pinchart@ideasonboard.com> Cc: Lucas Stach <l.stach@pengutronix.de> Cc: Peng Fan <peng.fan@nxp.com> Cc: Robby Cai <robby.cai@nxp.com> Cc: Sam Ravnborg <sam@ravnborg.org> Cc: Stefan Agner <stefan@agner.ch> --- drivers/gpu/drm/mxsfb/mxsfb_drv.c | 38 ++++++++++++++++++++++++++++++ drivers/gpu/drm/mxsfb/mxsfb_drv.h | 3 +++ drivers/gpu/drm/mxsfb/mxsfb_kms.c | 11 +++++---- drivers/gpu/drm/mxsfb/mxsfb_regs.h | 1 + 4 files changed, 49 insertions(+), 4 deletions(-)