On 24/07/2024 01:04, Darrick J. Wong wrote:
So why not just enable the per-inode flag with RT right from the start given that this functionality is supposed to work and be globally supported by the rtdev right now? It seems like a whole lot less work to just enable it for RT now than it is to disable it...What needs to be done to the rt allocator, anyway? I think it's mostly turning off the fallback to unaligned allocation, just like what was done for the data device allocator, right? And possibly tweaking whatever this does: /* * Only bother calculating a real prod factor if offset & length are * perfectly aligned, otherwise it will just get us in trouble. */ div_u64_rem(ap->offset, align, &mod); if (mod || ap->length % align) { prod = 1; } else { prod = xfs_extlen_to_rtxlen(mp, align); if (prod > 1) xfs_rtalloc_align_minmax(&raminlen, &ralen, &prod); }
My initial impression is that calling xfs_bmap_rtalloc() -> xfs_rtpick_extent() for XFS_ALLOC_INITIAL_USER_DATA won't always give an aligned extent. However the rest of the allocator paths are giving extents aligned as requested - that is from limited testing.
And we would need to not take the xfs_bmap_rtalloc() retry fallback for -ENOSPC when align > rtextsize, but I have not hit that yet - maybe because xfs_trans_reserve() stops us getting to this point due to lack of free rtextents.
