On Saturday 16 April 2005 18:59, Johannes Formann wrote:
> --On 16. April 2005 18:57:49 +0200 Blaisorblade <blaisorblade@...>
> > On Saturday 16 April 2005 18:33, Johannes Formann wrote:
> >> Hi,
> >> I've testing your uml-2.4.27-bs2-pre10-patch with 2.4.30, it compiles
> >> fine, but there seems to be a memory-leak.
Hmm, there should not be any difference with 2.4.26-3um/2.4.27-1um (apart the
hostfs code, which is the one from 2.4.24-1um).
> >> After running a few days a not very heavily loaded guest (only sshd,
> >> exim and rsyn daemon) needs 80 MB of RAM.
> >> $ free -m
> >> total used free shared buffers
> >> cached Mem: 95 93 1 0 1
> >> 8 -/+ buffers/cache: 83 11
> >> Swap: 131 0 131
> > Post the content of /proc/slabinfo and /proc/meminfo on a such situation
> > for verification.
> $ cat /proc/meminfo
> total: used: free: shared: buffers: cached:
> Mem: 99897344 98086912 1810432 0 1740800 8069120
> Swap: 138403840 643072 137760768
> cat /proc/slabinfo
> blkdev_requests 1024 1040 96 26 26 1
This is just a bit high.
Now, here are the culprits
> inode_cache 117046 123102 512 17586 17586 1
> dentry_cache 61074 80580 128 2686 2686 1
This is a lot of allocated memory by these two caches. If this increases over
time, it's this the cause of the leak. In this case, try not using hostfs (or
HPPFS, if you use it; it's even less stable than hostfs) and the leak should
disappear; in this case, I'll know
it's a hostfs bug.
To calculate the total size, read man slabinfo - I don't remember well which
columns gives the interesting numbers. It should be like the 2nd * the 3rd
number... so about 64M for inode_cache and 8M for dentry_cache (which could
be normal if rsync is used a lot, maybe).
However, before talking about "memory leak", maybe you should trying flushing
out these caches (by running some kind of memory-intensive process). Even a
program mallocing and using lots of memory is ok...
> buffer_head 7236 8440 96 211 211 1
A bit high, too (800k). That's related with not flushed datas, IIRC.
> size-32 4796 5537 32 49 49 1
I think this item should be normal... It's just 160k
> >> Have you tested/expericend something like that?
> > No, even because I have not time to do a lot of testing on the 2.4
> > branch... I'm working mostly on 2.6, and it actually feels better most
> > of the times.
> hmm, sounds like I should consider switching to 2.6 in the near future.
Probably yes, unless you have a very confortable setup currently and problems
However, there are also the famous security problems which have not yet been
fixed well in 2.4.
Paolo Giarrusso, aka Blaisorblade
Linux registered user n. 292729