[OpenWrt-Devel] uci commit bumping amount of dentry active slab objects

[Felix Fietkau]

On 2016-01-11 13:19, Felix Fietkau wrote:
> On 2016-01-11 12:57, Rafał Miłecki wrote:
>> I was looking at slabinfo and noticed something about uci. Executing
>> "uci commit" increases amount of "dentry" active objects.
>> 
>> # echo 2 > /proc/sys/vm/drop_caches
>> 
>> # egrep "dentry|kmalloc-64" /proc/slabinfo
>> dentry               559   2040    136   30    1 : tunables  120   60    8 : sla
>> kmalloc-64          5357   9324     64   63    1 : tunables  120   60    8 : sla
>> 
>> # for N in `seq 1 1 1000`; do uci commit network; done
>> 
>> # egrep "dentry|kmalloc-64" /proc/slabinfo
>> dentry              3540   3540    136   30    1 : tunables  120   60    8 : sla
>> kmalloc-64          6457   9324     64   63    1 : tunables  120   60    8 : sla
>> 
>> # for N in `seq 1 1 1000`; do uci commit network; done
>> 
>> # egrep "dentry|kmalloc-64" /proc/slabinfo
>> dentry              6570   6570    136   30    1 : tunables  120   60    8 : sla
>> kmalloc-64          7587   9324     64   63    1 : tunables  120   60    8 : sla
>> 
>> This is not critical. As you know Linux kernel loves caching. When
>> amount of free RAM goes below ~2 MiB I can see kernel freeing dentry
>> objects and it doesn't OOM.
>> 
>> But I'm still wondering if everything is OK. Is this normal kernel
>> keeps increasing pool of dentry objects as the result of "uci commit"
>> command?
>> 
>> I expected "uci command" first to bump amount of active and total
>> dentry objects. Then when uci exits I expected amount of active
>> objects to go down. Kernel could reuse inactive objects during next
>> "uci commit" call. In such scenario amount of active objects would
>> vary, but amount of total object would be more-or-less stable.
>> 
>> Can we be missing some (f)close in UCI? I checked uci_file_commit and
>> it looks alright. Or is this just how the filesystem works? I'm using
>> 3.18.21 (OpenWrt 15.05 branch) on device with serial flash (so JFFS2).
> User space programs are not required to close all their file descriptor,
> such a close happens implicitly on exit.
> I did an strace on uci commit without any delta, and I believe the
> likely cause is cached negative dentries for temporary files.
> On commit it checks a temp file before bailing out due to the lack of a
> config delta.
Yep, that was it. Fixed in r48200

- Felix
_______________________________________________
openwrt-devel mailing list
openwrt-devel at lists.openwrt.org
https://lists.openwrt.org/cgi-bin/mailman/listinfo/openwrt-devel