Part 5 in my rmap patch queue against 2.5.26:
http://loke.as.arizona.edu/~ckulesa/kernel/rmap-vm/2.5.26/
[ 2.5.26-rmap-5-slablru 18-Jul-2002 22:35 40k ]
This is a port of Ed Tomlinson's ([email protected]) really nifty patch to
let the (full) rmap VM do the work of freeing (inode/dentry/dquot) slab
pages based on page aging.
http://mail.nl.linux.org/linux-mm/2002-06/msg00001.html
Summary: The slab pages are moved into the active list, where they are
aged and pruned. The rate of pruning is simply the rate we see entries
for these slabs in refill_inactive_zone (while scanning the active list
for pages to deactivate). This has the significant advantage of being
wholly self-tuning! If we can count a slab page as fully pruned and no
longer in use, it is freed. For non-lru slab caches, we occasionally call
kmem_cache_shrink to keep those caches in check. A final sweep though
kmem_cache_reap() is our last protection against OOM.
---
The only significant structural change to the patch is 2.5's invocation
of the inode caches per-filesystem. I have given each fs's inode
cache a pruner method during its allocation (a one-line change).
Filesystems edited:
fs/adfs/super.c | 1
fs/affs/super.c | 1
fs/bfs/inode.c | 1
fs/coda/inode.c | 2
fs/efs/super.c | 1
fs/ext2/super.c | 1
fs/ext3/super.c | 1
fs/fat/inode.c | 1
fs/freevxfs/vxfs_super.c | 4
fs/hfs/super.c | 1
fs/hpfs/super.c | 1
fs/isofs/inode.c | 1
fs/jffs2/super.c | 1
fs/jfs/super.c | 1
fs/minix/inode.c | 1
fs/ncpfs/inode.c | 1
fs/nfs/inode.c | 1
fs/ntfs/super.c | 2
fs/proc/inode.c | 1
fs/qnx4/inode.c | 1
fs/reiserfs/super.c | 1
fs/romfs/inode.c | 1
fs/smbfs/inode.c | 1
fs/sysv/inode.c | 1
fs/udf/super.c | 1
fs/ufs/super.c | 1
Intermezzo has a funky dentry cache that may need a pruner method (??),
but I didn't touch it. If there was a better way to do this, I was too
blind to see it.
This patch has been well tested with ext3 and the generic slab caches. I
expect the other filesystems' inode caches to behave correctly, but it
wouldn't hurt to test! ... ;)
I haven't benchmarked the patch quantitatively, but the balance of
eviction between nonslab and slab pages seems to scale with the age of the
pages. Slocate runs don't evict the desktop, but do evict cold daemons
and other cruft. New allocations for applications shrink the older slabs
accordingly. Seems sane. By comparison, plain rmap-13b plunders the slab
caches rather ruthlessly.
The overhead for putting these pages on the lru lists, and the hopeful
benefits for properly aging and evicting them, should be benchmarked
sometime.
Give it a try! :)
Craig Kulesa
Steward Observatory
Univ. of Arizona