On Tue, Apr 28, 2020 at 6:25 PM Zygo Blaxell <ce3g8jdj@xxxxxxxxxxxxxxxxxxxxx> wrote: > Doing this with scrub is not reliable with the crc32c csum method--every > ~16TB of updates, you'll get a crc32c collision, so you'll have the > wrong data on disk and no csum failure to detect it with. Any of the > other csum options will solve this. Use SHA256 if (and only if) you > are worried about crypto collision attacks in your data; otherwise, > xxhash64 is fine. What about blake2b? Hash benchmark on x86_64 shows it's quite a lot faster than SHA256 (and yet still way slower than xxhash64 or crc32c). But I have no idea if this actually affects overall file system read/write performance when under load. I've started to migrate to xxhash64. It'd be nice to have a convert option. Rewriting 100% of the metadata is still a fraction of having to rewrite out TBs of data. But this is not a complaint. Btrfs is still badass. -- Chris Murphy
