Conversation

Notices

1. Embed this notice
   LisPi (lispi314@udongein.xyz)'s status on Friday, 11-Sep-2026 07:48:06 JST LisPi

   #Gentoo wiki is wrong about btrfs & VMs and it annoys me.

   No, one does not disable CoW on btrfs to improve performance. That nukes not only the CoW properly but also the integrity. The durability of data is also entirely ruined by doing that, btrfs' model of atomicity is entirely built on copy-on-write.

   (swap is effectively the only-use case that makes sense for that file attribute.)

   What one does is adjust commit frequency, dirty page cache ratios and in-VM workloads that unnecessarily request synchronous writes (get rid of systemd-journald persistent/on-disk IO, or the entire program, among others).

   (One could just entirely disable the VMs' ability to do synchronous writes, but then they may truly need them sometimes. So it's best to actually understand one's workload properly.)

   If one wants to use a faster filesystem that doesn't do integrity checking, don't mangle btrfs, configure it properly or just use xfs instead.

   • Embed this notice
     iced depresso (icedquinn@blob.cat)'s status on Friday, 11-Sep-2026 07:48:05 JST iced depresso

     in reply to
     @lispi314 copy on write is settable for a specific directory. it doesn't cripple the whole filesystem.
   • Embed this notice
     iced depresso (icedquinn@blob.cat)'s status on Friday, 11-Sep-2026 08:02:00 JST iced depresso

     in reply to
     @lispi314 it does not "cripple" the VM. it loses some guarantees about how extends are updated, which are often fighting the host program already trying to perform extent management.
     
     snapshotting still forces a new copy on write extent to be made, so the semantics of the filesystem that are public remain fulfilled.
   • Embed this notice
     LisPi (lispi314@udongein.xyz)'s status on Friday, 11-Sep-2026 08:02:02 JST LisPi

     in reply to

     • iced depresso
     @icedquinn It cripples everything in that directory and the entire VMs.
     
     Why assign them persistent storage if you don't care about any of it? An overlayfs atop a tmpfs & a read-only root partition would do even better.