On Thu, Apr 23, 2020 at 11:13 AM Peter Geis <pgwipeout@xxxxxxxxx> wrote: > > On Thu, Apr 23, 2020 at 11:49 AM Robin Murphy <robin.murphy@xxxxxxx> wrote: > > > > On 2020-04-23 4:09 pm, Peter Geis wrote: > > > On Thu, Apr 23, 2020 at 11:05 AM Peter Geis <pgwipeout@xxxxxxxxx> wrote: > > >> > > >> The rk3399 is capable of operating at PCIe gen 2 as per the TRM. > > >> The device-tree incorrectly limits us to gen 1. > > >> > > >> Correctly set the maximum link speed to <2>. > > >> > > >> Tested on the rockpro64. > > > > > > Note, this was tested on the rockpro64 after I performed the hardware > > > fixes as delineated at > > > https://forum.pine64.org/showthread.php?tid=8374 Are you going to fix everyone's board? > > > > > > We probably will have to drop this back to <1> on board specific dts > > > files if issues are discovered. > > > > I'd say commit 712fa1777207 and the fact that the current rev 1.8 > > datasheet only mentions 2.5GT/s rather weaken that argument. It would > > seem safer to leave the default as-is, and only override it for boards > > where Gen2 really is proven to work reliably. Which, er, is already the > > case ;) > > Do we have a copy of this errata? > I can't seem to find it. > The write up in that commit is extremely vague. > > As the tegra mailing list often points out, the device-tree describes > the hardware as it is. I think that's DT describes the h/w not settings for the Linux kernel which is different from what's discussed here. > As: > The rk3399 itself supports PCIe gen 2. > The board specific implementations determine if we need to limit that to gen 1. > The rk3399 should be set to 2, and any board that requires that to be > redefined should do that via an override in their device-tree. > This is similar to the gmac overrides for timing. > > Do we have a list of the boards that require pulling back down to gen 1? > > > > > That said, "proven to work reliably" is itself a bit doubtful - my > > NanoPC-T4 has always been rock-solid at Gen2 with a Samsung Evo 960 > > NVMe, yet I've seen plenty of reports of other NVMe models being > > unusable with mainline due to failing link training ~90% of the time. > > It's a grey area for sure. That seems pretty clear to me as to what the default should be. What's most likely to work OOTB for users. Rob _______________________________________________ Linux-rockchip mailing list Linux-rockchip@xxxxxxxxxxxxxxxxxxx http://lists.infradead.org/mailman/listinfo/linux-rockchip
