#asahi · 2021-05-28 09:59
pinskiaThe harder bugs have been global TLBI related and/or interrupt related#asahi · 2021-05-28 09:54
pinskiaDC ZVA was broken too in ThunderX1 but was not found out until much latter when debugging OcteonTX2 (96xx) where a GCC bootstrap was failing; tar would produce a corrupt file. That was one of the easier bugs to debug really#asahi · 2021-05-28 09:47
pinskialet me go look up the errata from the bug database#asahi · 2021-05-28 09:46
pinskiawell and VHE#asahi · 2021-05-28 09:46
pinskiacasp was the biggest thing they fucked up#asahi · 2021-05-28 09:46
pinskiaI debugged a lot of it#asahi · 2021-05-28 09:45
pinskiayes I know#asahi · 2021-05-28 09:43
pinskiaAlso I don't know how many cores implement 16k pagesize;is it still an optional feature, I have not looked into the spec in this area in the last 3 or so years#asahi · 2021-05-28 09:34
pinskiaand the reason why CRC is an optional feature of ARMv8.0 was because xgene's core was completed before crc was added to the spec too#asahi · 2021-05-28 09:31
pinskia@maz I do know that Cavium was the one of the drivers for LSE and v8.1#asahi · 2021-05-28 08:08
pinskiaI always wondered why ARM added 16k page size to the spec as a (late) optional feature and I guess Apple is the answer#asahi · 2021-05-28 08:06
pinskiaI should say arm64 Linux#asahi · 2021-05-28 08:06
pinskiaI missed that M1 requires 16k page sizes. The last time I tested 16k pages sizes in arm Linux (long time ago) on the Cavium ThunderX1, it failed if there was more than 32G installed. I don't know if this was solved yet.