The bad news is that I was little quick claiming complete success. There is a dramatic change and I can certainly transfer a 10GB file with 100% success the first time after I boot. If I continue transfering 10GB files eventually the system will hang. However, the number of files I can transfer successfully is variable. My record to date is 5 files transfered.Amazing! The 16K pages somehow interacted with something in your environment to trigger bad behaviour in the Linux kernel. Since there is a factor 4 difference in page size I'd want to check that 6.5*4=26GB of file transfers are successful before claiming victory.
In my opinion faults like this should be followed up because they can illustrate bugs in the kernel that are difficult to find but affect longtime reliability for many people.
Some time ago I noticed a problem in one of my computers. In that case the root cause turned out to be an error in how the BIOS initialised the northbridge for a certain class of motherboards. That error was obvious and possible to debug when running an optimised memory copy but likely affected others to a lesser degree who did not use that exact sequence of instructions.
It's good to make use of an opportunity. Can you think of anything distinctive in your setup? What NAS are you connecting to? What filesystem are you running on the SSD attached to the Pi?
Although I doubt it is a common setup, NVME on a Pi5 to NAS is not unique. The NAS is a synology DSM1819+ running DSM7.2.2 with disks mounted using CIFS . The NVME is running ext4.
Not sure how to confirm the stepping. I am reluctant to lever the cooler off the board to read the marking. However, it is my understanding that ALL 2GB Pi5's are D0. The D0 stepping is supposed to reduce power consumption from C1's 3.3W to 2W and as I mentioned earlier the quiescent power consumption on my boards is 2W. While D0 is only supposed to remove dead code, to get everything onto a smaller silicon area, it seems likely some of the live areas were moved around.I think it would be good to know the exact HW revision of the used Pi5. It should the new D0 variant which uses a different device tree. So although 'Pi5', in combination with that RP1 it is quite possible some address mapping or so is wrong. There is a lot detailed info on https://www.raspberrypi.com/documentation/, what is missing is a brief architectural blocklevel overview. That would help pin-pointing the problem area.
There does seem to be a shortage of date on the Pi5 compared with earlier versions. In particular no datasheet or schematic. The RP1 bootloader is dated 23rd September 2024.
Statistics: Posted by TrevorG — Sat Nov 16, 2024 11:50 am