Page MenuHomeFreeBSD

kern/openzfs: Regulate the ZFS ARC pruning process precisely.
Needs ReviewPublic

Authored by seigo.tanimura_gmail.com on Mar 1 2024, 5:30 PM.
Tags
None
Referenced Files
Unknown Object (File)
Fri, Apr 26, 3:50 AM
Unknown Object (File)
Tue, Apr 23, 6:58 PM
Unknown Object (File)
Mon, Apr 22, 10:58 PM
Unknown Object (File)
Sun, Apr 14, 5:45 PM
Unknown Object (File)
Mon, Apr 8, 10:43 PM
Unknown Object (File)
Mon, Apr 8, 10:02 PM
Unknown Object (File)
Mar 28 2024, 4:23 AM
Unknown Object (File)
Mar 6 2024, 7:22 PM

Details

Reviewers
olce
Summary

This fix essentially eliminates the high CPU usage caused by the ARC pruning
thread, acknowledged by FreeBSD-EN-23:18.openzfs.

  • Key Features
  • Accounting of the ZFS vnodes in use. (v_usecount > 0)
  • Recycling of the ZFS vnodes with the precise target.
  • Acceleration of the ZFS vnode recycling by the ARC eviction waiters.
  • Tested Case
  • poudriere-bulk(8) of 2128 ports by 8 builders.
  • Observed Improvements
  • The CPU usage of kernel{arc_prune} no longer sticks to 100%.
    • There are still occasional surges.
  • The vnode operations are now more smooth.

These improvements are achieved in ZFS and partly VFS.

  • Remaining Issues
  • The ARC dnode size stays high during the build.
  • The ARC evictable sizes fall to zero on all states and types. (data, metadata)

These are due to the use of nullfs over ZFS. They are fixed by a separate
commit.

  • New Sysctl(3) OIDs
  • vfs.zfs.arc.vm_lowmem.event The number of the vm_lowmem kernel events.
  • vfs.zfs.arc.vm_lowmem.kmem The number of the VM_LOW_KMEM vm_lowmem kernel events.
  • vfs.zfs.arc.vm_lowmem.pages The number of the VM_LOW_PAGES vm_lowmem kernel events.
  • vfs.zfs.arc.vm_lowmem.nofree The number of the failures to release the ARC memory.
  • vfs.zfs.arc.vm_lowmem.pagedaemon The number of the ARC memory release run by a pagedaemon kernel thread.
  • vfs.zfs.znode.count The number of the total ZFS vnodes.
  • vfs.zfs.znode.inuse The number of the ZFS vnodes in use.
  • vfs.zfs.znode.prunable The number of the recyclable ZFS vnodes.
  • vfs.zfs.znode.pruning_requested The number of the ARC pruning requests.
  • vfs.zfs.znode.pruning_skipped The number of the ARC pruning requests skipped because the kernel does not have the sufficient recyclable ZFS vnodes.
  • vfs.zfs.znode.pruning_withwaiter The number of the ARC pruning accelerated becasue of the ARC eviction waiters.
  • vfs.zfs.znode.pruning_withwaiter_throttled The number of the ARC pruning throttled despite the ARC eviction waiters.
  • vfs.vnode.vnlru.recycle_bufs_pages The toggle to recycle the vnodes with the clean buffers and clean/dirty VM pages.
    • default: true
  • vfs.vnode.vnlru.recycle_nc_src The toggle to recycle the vnodes acting as namecache source.
    • default: true

PR: 275594
Security: FreeBSD-EN-23:18.openzfs
Signed-off-by: Seigo Tanimura <seigo.tanimura@gmail.com>

Diff Detail

Repository
rG FreeBSD src repository
Lint
Lint Skipped
Unit
Tests Skipped
Build Status
Buildable 56976
Build 53864: arc lint + arc unit

Event Timeline