#2581 F35 Change: POWER 4KiB page size
Closed: Rejected 5 months ago by zbyszek. Opened 6 months ago by bcotton.

On ppc64le, the kernel is currently compiled for 64 KiB page size.

This change proposes using the more common 4 KiB page size.

Some HPC workloads may be disadvantaged slightly. Workstation users are likely to encounter fewer bugs.

Some things, like the AMD Radeon GPU drivers, firmware or related code, appear to be completely non-functional on the 64k page size. Insufficient upstream developers are testing such issues on this architecture.


Metadata Update from @bcotton:
- Issue set to the milestone: Fedora 35
- Issue tagged with: system wide change

6 months ago

For what it's worth, the lead Fedora kernel maintainer has flatly opposed this proposal.

For what it's worth, the lead Fedora kernel maintainer has flatly opposed this proposal.

With no argument what so ever :/

Anyway, consider me -1 to prevent any auto-approval. In fact, I don't really know if we want this or not, yet anyway.

After one week, the vote is (0,0,-1). Tagging for meeting.

Metadata Update from @bcotton:
- Issue tagged with: meeting

5 months ago

I will unfortunately be at a Dr. appointment when the meeting starts, but as I said on the threads before the proposal was created, and on the proposal thread itself, I am opposed to this change.

The issues with 64k page sizes as they stand are really btrfs, which should be fixed upstream before F35 anyway, and a few graphics drivers, which really need to be fixed upstream. For instance, Radeon RX580 works, but the 5700XT supposedly does not. From the thread, it seemed that upstream was unaware, and thought there should not be any issues. Might be difficult to debug since it would take a miracle to purchase a Radeon 5700XT or newer right now.

On the flip side, there are still things that expect a 64K page size, as that has been the default on "things that matter" for a while now. For instance, CONFIG_PPC_VAS which is what enables the accelerators in co-processors like NX-GZIP and NX-842 to be accessible to kernel subsystems and user processes in Power 9 CPUs. This is dependent on CONFIG_PPC_64K_PAGES. And let's be honest, the PPC workstation market is very small. The workstation market with cards which are known to have issues with 64K page sizes is even smaller, and those drivers should be fixed upstream. PPC is not the first arch to try to leverage larger page sizes, and as memory footprints grow, it certainly will not be the last. I simply do not find it productive or worthwhile to cripple some systems and workloads in an effort to paper over a couple of issues for a small subset of users. I might feel differently if we were currently using 4K pages and trying to move to 64K, perhaps wait until a couple more issues are worked out, but Fedora has been using 64K pages for years, and changing back would cause more problems than it solves.

Using a 4K page size in Fedora for Aarch64 does make a bit more sense, as the typical Fedora user is running the aarch64 kernel on small devices, very few of them even having 8GB of memory. PPC is a different story, there are no small devices, but there can be extremely large devices.

I'm -1 for this change as well (heavily influenced by @jforbes "as per my previous email" comment above).

This was discussed during today's FESCo meeting:
REJECTED (0, 0, -5)

Metadata Update from @zbyszek:
- Issue untagged with: meeting
- Issue close_status updated to: Rejected
- Issue status updated to: Closed (was: Open)

5 months ago

Login to comment on this ticket.

Metadata