Re: Documentation about sysfs/procfs entries

["Peter Teoh" <htmldeveloper@gmail.com>]

On Feb 12, 2008 2:17 PM, Scott Lovenberg <scott.lovenberg@gmail.com> wrote:
>
> On Feb 12, 2008 12:23 AM, Peter Teoh <htmldeveloper@gmail.com> wrote:
> >
> > I  was looking for documentation on the kstack_depth_to_print under
> /proc/sys/kernel, and I found it in Documentation/sysctl.txt (written by
> Rik).
> >
> > How about /proc/sys/net?   or all other directories under /sys or /proc
> fs?
> >
> > Wouldn't it be useful to have a centralized store located in Documentation
> - properly structured,  just a one-liner or two - on the reasons and
> explanations for each of these /proc or /sys parameters?   And something to
> look for further details?
> >
> > Something of the following:
> >
> > |-- netfilter
> > |   |-- nf_conntrack_buckets:your explanation here
> > |   |-- nf_conntrack_checksum:your explanation here...
> > |   |-- nf_conntrack_count
> > |   |-- nf_conntrack_expect_max
> > |   |-- nf_conntrack_generic_timeout
> > |   |-- nf_conntrack_icmp_timeout
> > |   |-- nf_conntrack_log_invalid
> > |   |-- nf_conntrack_max
> > |   |-- nf_conntrack_tcp_be_liberal
> > |   |-- nf_conntrack_tcp_loose
> > |   |-- nf_conntrack_tcp_max_retrans
> > |   |-- nf_conntrack_tcp_timeout_close
> > |   |-- nf_conntrack_tcp_timeout_close_wait
> > |   |-- nf_conntrack_tcp_timeout_established
> > |   |-- nf_conntrack_tcp_timeout_fin_wait
> > |   |-- nf_conntrack_tcp_timeout_last_ack
> > |   |-- nf_conntrack_tcp_timeout_max_retrans
> > |   |-- nf_conntrack_tcp_timeout_syn_recv
> > |   |-- nf_conntrack_tcp_timeout_syn_sent
> > |   |-- nf_conntrack_tcp_timeout_time_wait
> > |   |-- nf_conntrack_udp_timeout
> > |   `-- nf_conntrack_udp_timeout_stream
> > |-- nf_conntrack_max
> > |-- token-ring
> > |   `-- rif_timeout
> > `-- unix
> >     `-- max_dgram_qlen
> >
> >
> > Alternatively, we can write a script to extract out the partial
> explanations/details from existing source code, based on some coding
> convention/style structure, and further hand-modification from there.
> (given the dynamic nature of the kernel code, this may be preferred?)
> >
> >
>
> I was looking for exactly this about 2 weeks ago; I needed to look up the
> knobs for net and disk elevators, and had to jump all over the place.  I
> assumed that such a thing should exist, but I didn't find it.  I know I
> would personally benefit from this, and I'm sure many others would, too.

My suggestion is that since the parameters are:

a.   divide the symbols into two group:   well-documented vs
undocumented (prefixed by "__" eg, __xxxx).   this supposedly for
sysadmin vs kernel developer.

b.   for each of the official documented /sys control, provide a
one-line description when passed via /sys/help (eg, echo
kstack_depth_to_print > /sys/help) and the helpstring will go to the
console.

On the other hand, this may potentially blow up the vmlinux image size.