Hi Marek, Liu, On 22-07-26, Liu Ying wrote: > On Tue, 2022-07-26 at 16:19 +0200, Marek Vasut wrote: > > On 7/26/22 11:43, Marco Felsch wrote: > > > FIFO underruns are seen if a AXI bus master with a higher priority > > > do a > > > lot of memory access. Increase the burst size to 256B to avoid such > > > underruns and to improve the memory access efficiency. > > > > Sigh, this again ...
I know.. we also tried the PANIC mode but this somehow didn't worked as documented. So this was the only way to reduce the underruns without adapting the interconnect prio for the hdmi-lcdif. > > > diff --git a/drivers/gpu/drm/mxsfb/lcdif_kms.c > > > b/drivers/gpu/drm/mxsfb/lcdif_kms.c > > > index 1bec1279c8b5..1f22ea5896d5 100644 > > > --- a/drivers/gpu/drm/mxsfb/lcdif_kms.c > > > +++ b/drivers/gpu/drm/mxsfb/lcdif_kms.c > > > @@ -143,8 +143,20 @@ static void lcdif_set_mode(struct > > > lcdif_drm_private *lcdif, u32 bus_flags) > > > CTRLDESCL0_1_WIDTH(m->crtc_hdisplay), > > > lcdif->base + LCDC_V8_CTRLDESCL0_1); > > > > > > - writel(CTRLDESCL0_3_PITCH(lcdif->crtc.primary->state->fb- > > > >pitches[0]), > > > - lcdif->base + LCDC_V8_CTRLDESCL0_3); > > > + /* > > > + * Undocumented P_SIZE and T_SIZE bit fields but according the > > > + * downstream kernel they control the AXI burst size. As of now > > > there > > I'm not sure if it is AXI burst size or any other burst size, though it > seems to be AXI burst size. > > Cc'ing Jian who mentioned 'burst size' and changed it from 128B to 256B > in the downstream kernel. Thanks. > > > + * are two known values: > > > + * 1 - 128Byte > > > + * 2 - 256Byte > > > + * > > > + * Downstream has set the burst size to 256Byte to improve the > > > memory > > > + * efficiency so set it here too. This also reduces the FIFO > > > underrun > > > + * possibility. > > > + */ > > > + ctrl = CTRLDESCL0_3_P_SIZE(2) | CTRLDESCL0_3_T_SIZE(2) | > > > + CTRLDESCL0_3_PITCH(lcdif->crtc.primary->state->fb- > > > >pitches[0]); > > > + writel(ctrl, lcdif->base + LCDC_V8_CTRLDESCL0_3); > > Nit: I would write the register directly, instead of caching the value > in ctrl. IMHO it's more readable that way. Regards, Marco > > > } > > > > Sometimes I wonder whether this might be some successor of MXSFB > > LCDIF_CTRL2n OUTSTANDING_REQS and BURST_LEN_B fields. > > No idea... > > Liu Ying > > > > > +CC Liu, who seems to have a lot of knowledge about this IP. > > > > Reviewed-by: Marek Vasut <ma...@denx.de> > >