Am Montag, dem 21.06.2021 um 18:45 +0200 schrieb Marek Vasut: > On 6/21/21 2:08 PM, Lucas Stach wrote: > > Am Montag, dem 21.06.2021 um 00:47 +0200 schrieb Marek Vasut: > > > In case the DRAM is under high load, the MXSFB FIFO might underflow > > > and that causes visible artifacts. This could be triggered on i.MX8MM > > > using e.g. "$ memtester 128M" on a device with 1920x1080 panel. The > > > first "Stuck Address" test of the memtester will completely corrupt > > > the image on the panel and leave the MXSFB FIFO in odd state. > > > > > > To avoid this underflow, increase number of outstanding requests to > > > DRAM from 2 to 16, which is the maximum. This mitigates the issue > > > and it can no longer be triggered. > > > > > I see the obvious benefit of this change, but I'm not sure if enabling > > this on older SoCs is without any drawbacks. The newer SoCs have a good > > transaction scheduling on the NOC (i.MX8M) or at least a somewhat okay > > one in the DRAM controller (i.MX6). I'm not so sure about the older > > SoCs, where I could imagine too many outstanding transactions to delay > > memory traffic for other masters on the SoC. > > > > You don't happen to have any experience with this on the older SoCs, do > > you? > > The only older SoC which would be affected by this, except for the ones > you already listed, is MX28. And the MX28 has rather decent DRAM > controller, so I wouldn't expect problems there either. > > You can look at it the other way around too however, if the DRAM gets > saturated, the LCDIF controller suffers from FIFO underflows and that > completely messes up the FIFO state, at which point the image on the > display is distorted, shifted, wrapped around, or any other such odd > effect. The underflow auto-recovery bit helps with it, but with this > patch in place you are unlikely to run into those effects at all. Yea, I just wanted to have this thought considered. If you think the probability of introducing regressions on MX28 is low, that's fine with me. I guess MX28 systems likely don't have a big enough screen attached to drown the memory controller in read requests. Reviewed-by: Lucas Stach <l.stach@xxxxxxxxxxxxxx>
