Very severe regression in ZFS 0.7.12 in 18.04.3

[Michael Albert]

ZFS version 0.7.12 has a documented very severe regression on sending and receiving datasets with dnodesize=auto set on the dataset (not present in 0.7.9 or 0.8.1). This manifests itself primarily when sending and receiving datasets with lots of files, such as happens if root is on ZFS. This issue has been documented on the ZFSonlinux git repositories in issues:

https://github.com/zfsonlinux/zfs/issues/9147

and

https://github.com/zfsonlinux/zfs/issues/8458

Since 0.7.12 was just brought into 18.04.3, this regression is now affecting LTS. I'm opening this issue in the hopes that the fix is backported to the LTS release

[Halfwalker (deano-ubuntu)]

I can confirm on a Lenovo M920S and Lenovo Thinkpad X1 Carbon, both with 18.04.3 and zfs package version 0.7.5-1ubuntu16.6

See https://github.com/zfsonlinux/zfs/issues/9147#issuecomment-539055281

[Colin Ian King (colin-king)]

The root cause is documented in bug: https://github.com/openzfs/zfs/issues/8458

[Colin Ian King (colin-king)]

This appears to be a relevant fix https://github.com/openzfs/zfs/commit/45f0437912f4688bb15b353c8773b0e37e0b1a89

[Colin Ian King (colin-king)]

Reproducer: Create an fs with dnodesize=auto (also confirmed stall with 4k). Create (touch) a million files in the same directory. send/receive it and it stalls. 'zfs list' shows that the fs uses is 130M. The receive stalls/slowdown at 94M.

[Colin Ian King (colin-king)]

Ubuntu Bionic 18.04.x was released with ZFS 0.7.5 and not 0.7.12. Ubuntu Disco 19.04 contained ZFS 0.7.11 and then 0.7.12 but this is now EOL.

Supported versions of ZFS are as follows:

Xenial: 0.6.5.6
Bionic: 0.7.5
Focal: 0.8.3
Groovy: 0.8.4

0.7.12 is EOL. The bug you are referring to has the fixes in the Focal and Groovy releases. Marking this as Won't Fix.