Re: [PATCH] Use uname not sysctl to get the kernel revision

["Albert Cahalan" <acahalan at gmail dot com>]

On 7/13/06, Eric W. Biederman <ebiederm@xmission.com> wrote:
> "Albert Cahalan" <acahalan@gmail.com> writes:
> > Andi Kleen writes:
> >> On Thursday 13 July 2006 01:24, Theodore Tso wrote:
> >
> >>> P.S.  I happen to be one those developers who think the binary
> >>> interface is not so bad, and for compared to reading from /proc/sys,
> >>> the sysctl syscall *is* faster.  But at the same there, there really
> >>> isn't anything where really does require that kind of speed, so that
> >>> point is moot.  But at the same time, what is the cost of leaving
> >>> sys_sysctl in the kernel for an extra 6-12 months, or even longer,
> >>> starting from now?
> >>
> >> The numerical namespace for sysctl is unsalvagable imho. e.g.
> >> distributions regularly break it because there is no central
> >> repository of numbers so it's not very usable anyways in practice.
> >
> > Huh? How exactly is this different from system call numbers,
> > ioctl numbers, fcntl numbers, ptrace command numbers, and every
> > other part of the Linux ABI?
>
> The only practical difference is that what people use is
> /proc/sys so the binary sysctl interface is not seriously maintained
> and bugs crop up.

There is a chicken-and-egg problem here then.
Let's fix it.

I maintain the sysctl program, which most Linux
distributions run at boot. I agree to switch to the
binary sysctl interface if somebody will maintain
the kernel side of things. This will shave a bit of
time off boot on nearly every Linux box out there.
The total time saved is probably a human lifetime,
so it's like saving somebody's life.

> > Normal sysctl works very well for FreeBSD. I'm jealous.
> > They also have a few related calls that are very nice.
> >
> > Here we fight over a few CPU cycles in the syscall entry path,
> > then piss away performance by requiring open-read-close and
> > marshalling everything through decimal ASCII text. WTF? Let's
> > just have one system call (make_XML_SOAP_request) and be done.
>
> There is a cost to open-read-close.  But as a simple benchmark
> against a file will show reading data from /proc/sys is much slower
> than reading data from a file.
>
> From what I have been able to measure so far, open-read-close only
> seems to double the cost over sysctl, and access can do the filename
> resolution about as quickly as sysctl can deal with a binary path.  So
> I suspect it is the allocation of struct file that makes
> open-read-close more expensive.  Reading the data is in the noise.

Eh? A factor of two is not "in the noise".

> sysfs current does a lot better than /proc/sys I think it was only
> 60% heavier than performing the same operation on a real file.

That is still a horrible way to piss away performance.

> Performance wise there does seem to be a problem with the
> implementation.  How to fix it I don't yet know.  But I have
> yet to see ascii text be implicated.

I have more experience with /proc. There, ASCII is
known to be a problem.

Parsing a 64-bit number is horribly slow on i386.

Matching keywords, as is needed for /proc/*/status,
is also horribly slow. I ended up using gperf to make
a perfect hash table, then gcc's computed goto for
jumping to the code, and it still wasn't cheap to do.
(while /sys lacks this, the extra open-read-close is
certain to be far worse)