Date:	Thu, 12 Apr 2012 22:05:33 +0100
From:	Ben Hutchings <bhutchings@...arflare.com>
To:	Ben Greear <greearb@...delatech.com>
CC:	Florian Fainelli <florian@...nwrt.org>,
	<linux-wireless@...r.kernel.org>, <netdev@...r.kernel.org>
Subject: Re: [PATCH 4/5] mac80211: Add more ethtools stats: survey, rates,
 etc

On Thu, 2012-04-12 at 13:46 -0700, Ben Greear wrote:
> On 04/12/2012 12:30 PM, Ben Hutchings wrote:
> > On Thu, 2012-04-12 at 09:51 -0700, Ben Greear wrote:
> >> On 04/12/2012 09:42 AM, Florian Fainelli wrote:
> >>> Hi,
> >>>
> >>> Le 04/12/12 18:32, greearb@...delatech.com a écrit :
> >>>> From: Ben Greear<greearb@...delatech.com>
> >>>>
> >>>> The signal and noise are forced to be positive since ethtool
> >>>> deals in unsigned 64-bit values and this number should be human
> >>>> readable. This gives easy access to some of the data formerly
> >>>> exposed in the deprecated /proc/net/wireless file.
> >>>
> >>> Uh, that's misleading, the signal and noise values are typically negative, so one needs to think about mentally adding a minus sign if he/she wants to
> >>> understand it. Does not ethtool know about 32-bits signed integers?
> >>
> >> Ethtool stats only supports u64.  I think it's easy enough for
> >> humans or programs to add the negative sign.  Can signal or noise
> >> ever be>  0?  If so, that could actually break something that depends
> >> on flipping the value to negative....
> >
> > So far as I can see, the ethtool stats were expected to be counters,
> > which obviously cannot become negative (or fractional).  Maybe it's time
> > to define another command and string-set to cover other types of status
> > information.  The ethtool utility could ask for those typed statistics
> > as well, so 'ethtool -S' would get all of them.
> 
> One nice thing about ethtool stats API is that it is backwards and forwards
> compatible for a long while.

Agreed.

> Also, adding a new API means a second call
> into the kernel (most likely) for the other set of strings, which effectively
> doubles the cost of getting stats (and allows the stats to be updated
> out-of-sync with each other more easily).

It's generally not possible to take an atomic snapshot of all statistics
for a NIC, so that shouldn't be a consideration.

> I wonder if instead we could add a convention where we add a short
> prefix (or suffix) to a string to denote it's type (and cast the value
> into u64).  So, an old ethtool might see:
> 
> foo:s32: 4294967295

Actually it would be:

foo:s32: 18446744073709551615

> while a new one understand that the s32: prefix is special, do
> some casting, and could show:
> foo: -1
>
> Both are still at least somewhat human readable,

I don't think many humans can mentally substract 2^64.

> and probably wouldn't confuse anyone that is parsing the output of
> existing ethtool output.

I think you have this backwards: printing numbers in two different ways
(old and new versions of ethtool) is likely to confuse scripts that are
parsing and doing calculations with these numbers.  While I try to avoid
gratuitous changes in output formatting, scripts should use the ethtool
API if they really want interface stability.  It's not difficult (there
are at least Python and Perl bindings available) and it's a lot more
efficient.

Ben.

-- 
Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.

--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Powered by blists - more mailing lists