At 08:25 AM 11/6/01 +0100, Jakob Østergaard wrote:
>It seems to me that the sysctl interface does not have any type checking,
>so if for example I want to read the jiffies counter and supply a 32-bit
>field, sysctl will happily give me the first/last 32 bits of a field that
>could as well be 64 bits (bit widths do sometimes change, even on
>architectures
>that do not change). How am I to know ?
This fault isn't isolated to the sysctl interface. Look sometime at the
ioctl and fcntl interfaces and you will see that they have the same
problem. The issue is that the original Unix implementation of the
special-function interface assumed that only "ints" would be passed around,
and the need for special interfaces outgrew that assumption.
These functions have been so abused that POSIX refuses to "standardize"
them; instead, special APIs such as TERMIOS have been devised to put a
fairly-well defined shell around the most needed of these interfaces. (How
the implementer decides to bridge the userland/kernel barrier is not part
of the specification -- and doesn't need to be.)
The /proc API was developed to solve a specific problem. Now, people have
proposed and The Powers That Be have accepted extensions to the /proc
interface as a superior way to tune the kernel, particularly from shell
scripts, and to monitor the kernel, again from shell scripts. It's a good
thing, actually, in that it preserves the best of the Unix
mentality: don't re-invent, reuse.
What this whole discussion boils down to is people who want to tackle the
symptoms instead of the disease. The PROBLEM is that we have inadequate
standards and documentation of the /proc interface in its original ASCII
form. The proposed solution does NOTHING to address the real problem --
and I understand that because too many people here are so used to using a
hammer (code) that all problems start looking like nails.
The RIGHT tool to fix the problem is the pen, not the coding pad. I hereby
pick up that pen and put forth version 0.0.0.0.0.0.0.0.1 of the Rules of /Proc:
1) IT SHOULD NOT BE PRETTY. No tabs to line up columns. No "progress
bars." No labels except as "proc comments" (see later). No in-line labelling.
2) All signed decimal values shall be preceded by the "+" or "-" character
-- no exceptions. Implementers: this is available with *printf formats
with the + modifier, so the cost of this rule is one character per signed
value.
3) All integral decimal values shall be assumed by both programs and
humans to consist of any number of bits. [C'mon, people, dealing with
64-bit or 128-bit numbers is NOT HARD. If you don't know how,
LEARN. bc(1) can provide hints on how to do this -- use the Source,
Luke.] Numbers shall contain decimal digits [0-9] only. Zero-padding is
allowed.
4) All floating-point values shall contain a leading sign ("+" or "-") and
a decimal point (US) or comma (Europe). This rule assumes that the locale
for the kernel can be set; if this isn't true, then a period shall be used
to separate the integral part and the fractional part. Floating point
values may also contain exponents (using the *printf format %E or %G, NOT
%e -- the exponent must be preceded by the letter "e" or "E"). The
specification of a zero precision (which suppresses the output of the
decimal point or comma) is prohibited.
5) All string values matching the regular expression [!"$-+--~]* shall be
output as they are. Strings that do not match the above regular expression
shall be escaped in a standard manner, using a single routine provided in
the kernel's /proc interface to provide the proper escape sequences. The
output of that routine shall output standard backslash-character
representation of standard C-language control characters, and 3-digit octal
representation of any other character encountered. Output of the octal
representation may be truncated when such truncation would not cause
confusion -- see strace(1) for examples.
6) If you are wanting to display octal data, display it byte at a time
with a backslash. If you want to display hexadecimal data, use the "\x"
introduction, but include all bits so that the using program knows how long
the damn element is supposed to be -- NO leading -zero suppression should
be done. (Use the %x.xX format item in *printf, where "x" is the number of
hexadecimal digits.)
7) The /proc data may include comments. Comments start when an unescaped
hash character "#" is seen, and end at the next newline \n. Comments may
appear on a line of data, and the unescaped # shall be treated as end of
data for that line.
8) The regular expression ^#!([A-Za-z0-9_.-]+ )*[A-Za-z0-9_.-]$ defines a
special form of comment, which may be used to introduce header labels to an
application. As shown in the regular expression, each label is defined by
the regular subexpression [A-Za-z0-9_.-]+ and are separated by a single
space. The final (or only) label is terminated by a newline \n. No data
may appear on the header comment line. This line may only appear at the
beginning of the /proc pseudo file, and appears only ONCE.
9) The regular expression ^#=[0-9]+$ shall be used to output a optional
"version number" comment line If this appears in the /proc output, it
precedes the header comment line, and appears only ONCE.
10) Network addresses are defined as strings, either in their name form,
in dot quad notation for IPV4 numeric addresses, or in the numeric
equivalent for IPV6. Parsers can recognize the difference between a
dot-quad IP address and a floating-point number by the presence of the
second dot in the number. Network information output on /proc shall not
use the base/mask notation (123.456.789.012/255.255.255.0) and instead use
the bit-length notation (123.456.789.012/24).
11) IPX network addresses are a problem. In their normal form, they are
indistinguishable from a %F-format floating-point number with leading zeros
(which is allowed). Therefore the dot that usually appears in an IPX
network number must be replaced with the hyphen or dash "-"
character. Parsers can then differentiate an IPX network address from a
floating point number by noticing the embedded dash without the leading "e"
or "E" character. Flex handles this just fine.
-end-
This represents my first cut into a specification for the /proc interface
to deal with some of the issues that have come up in this thread. It's not
going to satisfy the "performance for me at all costs, DAMMIT" people and
it's not going to satisfy the "I like it PRETTY, DAMMIT" crowd either, but
it would provide a means for coming up with some standard tools to deal
with /proc, and a way to reign in the madness.
In particular, it means that a single tool could be developed to take a
/proc file and, in userland, make it a little more pretty. Those that
don't like table presentations can use the source of the tool to make a
display more to their liking.
The spec has a number of things missing. One issue missing is how to make
a predictable /proc subtree, so that people can find the goodies more
easily. Another issue is specifying how /proc can be used to set
parameters. (We seem to have less confusion in this area, so I didn't want
to spend any time on this aspect of the specification.)
OK, I'm clear of the firing range, start shooting holes in it.
Stephen Satchell
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
This archive was generated by hypermail 2b29 : Wed Nov 07 2001 - 21:00:31 EST