changes

早期freebsd实现

18. The q_system field of the query structure has been removed in favor of
    a q_type field containing bit definitions.  The old PRIMING_CACHE magic
    cookie is no longer used.  Go to the end of the universe, do not pass go.

17. Converted to ANSI C.  All functions are static unless they are actually
    needed outside the current module ("file" in C terminology); static 
    functions are declared with prototypes if they are forward-referenced.
    Externally visible functions are declared in separate header files, with
    prototypes.  ns.h and db.h have been split into four new header files:
	db.h -> db_defs.h db_glob.h db_func.h
	ns.h -> ns_defs.h ns_glob.h ns_func.h

    The *_defs files contain only structure and type definitions, and macro
    definitions.  Nothing that generates text or data space in the executable
    is declared here.

    The *_glob files contain only global variable declarations, which used to
    be defined in the various *.c files in a more or less random fashion.  The
    declarations are "extern" if included from non-main()-containing files, but
    are defined globally and given initial values in main()-ish files.  This
    reuse of the same declarations insures that the type and size declarations
    match between definitions and external references to them.

    The *_func files contains function prototypes for global ("extern")
    functions.  The prototypes are all optional so will not break non-ANSI
    systems.  Note that I don't have such a system any more so I may be wrong.

16. Removed all remaining references to "short" or "long" that did not
    depend on the vague semantics of those types.  Most uses were actually
    depending on a size of 16 bits for short and 32 bits for long, and there
    are processors/compilers where each of these types is different.  This
    work was begun in 4.9 and is now complete.  Note that some structs that
    are used in large data structures use "char" for 8-bit integers.  It helps.

05Jul93 - ALPHA Released

This is the cleanup release after 4.9.  I'm going to try the TCSH style of
logging the changes; let me know if you think it's a bad way of doing it.

15. the resolver now includes an implementation of RFC 1101, which allows
    network names to be encoded in the DNS tree rather than in /etc/networks.
    this implementation is by [e-mail] (Rui Pedro Mendes Salgueiro)
    i put the test program and original documentation in contrib/rfc1101/.  i
    would like to see their main.c ("nettest") turned into a tools/nettest,
    but i'm not willing to do the work myself.  it needs a man page, etc.

14. as expected the initial HS zone transfer stuff didn't work that well.
    thanks to <[e-mail]>, retries after failed SOA queries will use
    C_IN rather than falling through to C_HS inappropriately.

13. ns_init.c was fcntl(SETFL)'ing in a destructive way.  it now does a
    fcntl(GETFL) to get the old option mask and then |'s in the new flag.
    this patch came from Eduard Vopicka <[e-mail]>.

12. there are two new conf/Info.* files; check 'em out.

11. ultrix (some versions, especially the vax ones) libc.a had some bad
    naming conventions for some resolver routines.  getshort/putshort just
    have to be real functions, not just macros, or you can't link anything
    with this resolver.  patch was sent by <[e-mail]>.

10. sethostent(x) for host files was sticky for nonzero 'x' ([e-mail])

9.  hp9000s700 is now supported in include/arpa/nameser.h ([e-mail])

8.  statistics dumps now print the time in decimal-seconds-since-1970 in
    addition to the old "ctime" format, for ease of debugging.  (Peter Koch).

7.  systems with 14-character filename limitations have apparently been
    having trouble in named-xfer since its temporary file names are bigger
    than they can handle.  ash@hp sent in some patches a while ago, enabled
    with SHORT_FNAMES in conf/options.h, to deal with this appropriately.
    We should probably just generate short names always.

6.  Some security stuff from ISI.  According to Anant Kumar <[e-mail]>:

	The validation procedure is the major change here. Currently, we
	accept anything from a server, as long as we had asked it a question.
	This implies that a malicious server can really send us any data and
	we not only pass it on, we also cache it for as long as the TTL
	holds. This can be really bad for our health and for that of those
	who use the DNS.

	We add this procedure to verify for each RR returned by a server
	that it is indeed authoritative for either that zone, or for a
	parent zone.  We end up trusting the root servers for everything!
	Also, the more rich our cache is the more choosy we become about the
	data we add on to it. This stuff is all ifdef'd with "#ifdef VALIDATE"

	The negative caching stuff adds on a d_rcode field to the databufs.
	Any positive entry now shows a NOERROR there while negative entries
	have either a NXDOMAIN or NOERROR_NODATA. NOERROR_NODATA rcode is
	never returned.  It is used only to differentiate, within the
	internal database, between negative and positive entries. We use the
	regular hash table (hashtab) to store negative entries, too. Only
	authoritative answers are negative cached, for NTTL (parameterized,
	currently 10 minutes) seconds. Non-authoritative NXDOMAINs or
	NOERROR with zero RR count, now generated, are now accepted but
	never cached. This is ifdef'd with "#ifdef NCACHE".

5. "make install" now has a prayer of working for the man pages. an observation
   was made that net2++ systems _require_ formatted "cat" pages and that older
   systems are _able_ to use them, so that's all we install.

4. i wrote man pages for named.reload, named.restart, and named-xfer.  these
   were actually in 4.9.1 for 4.4BSD.

3. unneeded functions in compat/lib will now generate placeholder symbols, to
   make sure that the linker doesn't generate ugly-but-harmless warnings.

2. my ignorance of the true meaning of _POSIX_SOURCE has been corrected,
   along with the ugly-but-working code in conf/portability.h and elsewhere.

1. non-resolver routines moved from res/ to compat/lib/.  this will shorten
   libresolv.a and make it easier to integrate new BIND releases into Net-2
   descendents such as 4.4BSD and BSD/386.

4.9.1 ------------------

This is the integration of the changes that were made for 4.4BSD.  This
release will not be published.  Changes include:

doc/BOG/*: many changes to improve appearance of the output, including
	orphan-avoidance and better tab stops.  Sent to me by someone on
	the net who deserves thanks but I've lost the original mail.  Oops.

include/*: the CSRG people weren't entirely pleased with the interface
	changes i made to the res_*() and inet_*() functions.  in particular,
	the changes from "long" to "u_int32_t" were too sweeping in their
	opinion since Posix is already working on standardizing them and
	might look unkindly on an apparently-still-evolving interface.  also,
	the possibility that all the vendors will change their implementations
	to match the new interface is apparently rather dim.  therefore most
	externally-visible occurances of the int32_t type have been changed
	back to "long" in the resolver interface.  we believe that this should
	still be portable to Cray and AXP machines, but i'll wait to hear from
	someone who can actually try it out and let me know.

tools/*: the "net2" version of "lex" requires some additional flags and libs,
	and this had implications for the Makefiles and the dig.c source file.
	nslookup's man page is now in man/ rather than tools/nslookup, for
	consistency.

named/*: last-minute 4.9-FINAL changes to named-xfer.c and db_load.c resulted
	in corruption of TXT records on zone transfers, and a high number of
	useless syslog(SYS_ERR) messages about zones already being up to date.
	these last-minute changes have been massaged into better shape and are
	now a lot readier for prime time than they were.  a lesson was learned.

	the inet_aton() function is now used where appropriate, rather than the
	old inet_addr().  this is just an evolutionary move that should have no
	practical implications.  bad addresses in the "tcplist", "bogusns", and
	"sortlist" directives (from named.boot) are now syslogged.

	some open files are still inherited by named-xfer from named, but they
	are properly closed now.

	the SIGXFSZ signal is now accepted as an alias for SIGHUP, in support
	of the wierd DEC Hesiod implementation.  no practical significance.

res/*: one important bug fix in the gethostent() stuff, and a whole bunch of
	evolutionary include file changes.

include/*: include/sys was moved to compat/include/sys, since systems that
	do not need it really really really need to get their own instead.
	at some point i'm going to move the res/*.c files that are needed for
	compatibility but not really part of the resolver, into compat/lib.

general: there are more settable parameters in the top-level Makefile, and
	they are propagated downward into the subdirectories' Makefiles.  you
	should not have to edit any Makefile except the top-level one.  Note
	that "make links" still creates local Makefiles in the build directory
	because "mkdep" still edits the Makefiles on most systems.

4.9-FINAL -------------------

Kevin Dunlap sent in some changes for the BOG.  So did a lot of other folks.

Someone asked about AXP-OSF, so I did a trivial 64-bit port.  Porting to
other 64-bit systems should be simple now.  Someone also sent in some MIPS
RISCOS portability changes, which were simple and therefore were put in.  
Note that some type names have been added to BSD 4.4 as a result of this
work; they are going to be in <sys/types.h> in BSD 4.4 but they are in a
local include file called <sys/bitypes.h> in this distribution, with 
appropriate #ifdef's in the include files that depend on them.  Those of you
who are porting to 64-bit platforms where "long" isn't 32 bits should be
using these new names for your types; there was no standard before this,
but the names we've added for BIND 4.9 and BSD 4.4 are going to be proposed
to Posix at some point.  Sometimes it's just not OK for "int" to be the
"natural integer size of the machine" and you just _have_ to tell the compiler
how many bits you want.

The NIC added a new root server, thus pushing the size of a nonauthoritative
root server response (which includes the root server list in the answer as
well as the authority sections) over the 512-byte limit.  This showed up a
long-term BIND bug wherein it failed to set the TC ("truncation occurred")
bit if truncation occurred anywhere but the answer section.  Since truncation
was occuring at the end of the packet, in the additional data section, this
meant that BIND was generating truncated responses without setting TC in the
response header.  Upon further investigation, I found that BIND ignored TC
on responses it received from other name servers.  RFC 1035 states that RR's
from truncated responses should not be cached; with creative interpretation
of the exact 1035 wording, I found a way to reach this goal while still 
caching the answer section (as long as the truncation occurred in some other
section, which 1035 gives no definitive way to determine but I'm happy with
my guess).

While researching the above, I finally broke down and added credibility
output to the zone dump files.  They are in the comments so should cause
no trouble.  There's more work to be done on the dump output; in particular,
Phil Almquist proposed and even prototyped a "tagging" of all RR's with the
A RR of the nameserver that sent them to us; this feature should be added
and the dump output should include it.  This would add a lot to our ability
to track down corrupt data.

Don Lewis and I had more discussions about TC and ended up agreeing that the
right thing to do is to set TC on responses that overflow in the answer or
authority section, truncating at an RR boundary, but do not set TC on responses
that overflow in the additional-data section (truncating at a {name,type}
boundary).  This actually solves the root server problem pretty well, since
BIND 4.9 will, when it tries to use an NS whose A isn't in the cache, generate
a sysquery() for the missing A.  (Heck, additional data TTL's are depreciated
at the rate of 5% per use, so this would end up happening pretty quickly even
if we did cache a partial {name,type} -- but now we won't have to.)

While trying to fix all of this stuff I ended up moving some functions around
to avoid duplicating them in different source files, and I reformatted some
source lines that went over 80 characters.  I also made a few things "static"
that used to be unneccessarily global.  More of that will happen in 4.9.1.

DEC's product version of MIT Hesiod uses SIGXFSZ for what we do with SIGHUP;
since the default for SIGXFSZ is to exit, it seemed prudent to wire it up to
do what SIGHUP does instead, so that this BIND can run on DEC Hesiod servers.

At the request of several people, I integrated the USC "dig" and Rutgers
"host" tools into the distribution.  This required some changes to the
resolver library's debugging output formats, which will be visible in
nslookup, nsquery, and any other tool that sets the RES_DEBUG option.
Note that there is no support for "DEFNAMES" in this version of dig, due
to design changes between 4.8 (from which "dig" is derived) and 4.9.  there
is no reason in principle why it can't be made to work, but it doesn't work
now.  therefore only fully-qualified names can be looked up with this "dig".

I had to change the name of the resolver "state" structure to be "__res_state"
for standards conformance (really, it is not reasonable to expect that because
a program includes <resolv.h> it will never define its own structure called
"state".  This change highlights the imperative that any application which is
relinked against this resolver must first be recompiled against these include
files (notably <resolv.h>).  This is true for almost all versions of libresolv.

I asked for items for the "TODO" list and got quite a few.  Check them out
before you hack; someone else may already have started doing what you want to
do.  I also asked for tools for the "contrib" subdirectory and got 650KB worth.
They make the BIND 4.9 distribution a lot larger than 4.8.3 was, but the extra
bytes are well worth their weight.