On Tue, May 27, 2025 at 11:45:26AM -0600, Keith Busch wrote: > Just fyi, the initial user I was planning to target with the block > layer's copy fallback isn't in kernel yet. Just an RFC at this moment on > btrfs: > > https://lore.kernel.org/linux-btrfs/20250515163641.3449017-10-maharmstone@xxxxxx/ > > The blk-lib function could easily replace that patch's "do_copy()" > without to much refactoring on the btrfs side. Well, that code would be much better off using a long living buffer, because the frequent allocations are worse. Also from talking to a btrfs guys sitting next to me at least the classic relocation can actually be in the normal write path, so you probably also don't want it to fail because a large memory allocation fails. What we did for the GC code in XFS that exists to basically do the same is to do a large allocation at mount time, and then just keep reusing it.
