#
✕ nick:arnd_
#asahi · 2022-07-23 16:10
arnd_found the patches from gcliement: this was for Armada XP, so early ARMv7 that lack LPAE and large page support https://lore.kernel.org/linux-arm-kernel/20200611134914.765827-1-gregory.clement@bootlin.com/
#asahi · 2022-07-23 16:07
arnd_Annapurna Labs (now Amazon) worked around the issue in their NAS SoCs by using a 64-bit pgoff_t, which is a better way to handle this IMO, but this also never went upstream
#asahi · 2022-07-23 16:05
arnd_the patches never went upstream but were in vendor kernels that shipped with NAS boxes
#asahi · 2022-07-23 16:05
arnd_In this case itwas done to avoid the limitations of a 32-bit pgoff_t that ultimately limits the the maximum disk partition size to 16GB
#asahi · 2022-07-23 16:04
arnd_janrinze: Marvell did a hack to use 64KB page sizes on arm32 on CPUs that did not actually support the larger TLB. Not sure if this was on the Kirkwood (ARMv5E) or Armada XP (ARMv7-a) generation. It very wasteful though, and you don't normally want to do it on a machine that is already memory limited
#asahi · 2022-07-23 15:45
arnd_j`ey: this was an Amazon EC2 Graviton2/C6g
#asahi · 2021-11-16 12:27
arnd_I think you can have any combination of the two on LPDDR5, but only the "link ECC" would lead to the CPU noticing a problem
#asahi · 2021-11-16 12:26
arnd_LPDDR5 also adds a "link ECC" option, which would comput a checksum only for data that gets sent between CPU and RAM and checked on the other end.
#asahi · 2021-11-16 12:25
arnd_more).
#asahi · 2021-11-16 12:25
arnd_marcan: regarding ECC, it appears that LPDDR4 and LPDDR5 chips advertised with "integrated ECC" do this by having all the error correction logic in the package and no additional data lines to the host, so they can automatically correct any single-bit errors but would not raise an exception to the host for a double-bit error. I can see announcements of this from LPDDR4 manufacturers but there is no indication that these are being sold (any