client: Unbreak standalone build

[empserver] / doc / xdump

Introduction

Empire is designed as a smart server with dumb clients.  An Empire
client need to know nothing about the game.  Even telnet would do.  In
fact, empire-client is little more than a slightly specialized telnet.

In such a design, presentation is in the server, and it is designed
for human consumption.  Ideally, presentation and logic are cleanly
separated, for easy selection between different presentations.
There's no such separation in the Empire server, and separating them
now would be a herculean effort.

Thus, smart clients have to work with output designed for humans.
That's not especially hard, just an awful lot of tedious work, and the
result gets easily broken by minor changes in output format, which
humans hardly notice.

Instead of making smart clients parse output of commands designed for
humans, one can add commands designed for machines.  Such commands can
share the code implementing game rules with their counterparts for
humans.  To do that cleanly means separating logic and presentation.
Implementing them with their own copy of the code is no good --- how
would you ensure that the two copies implement precisely the same game
rules?

Except for commands that actually don't do anything.  There's a useful
class of such commands: commands to show game configuration and state
without altering it.  The only game rules involved are those that
govern who gets to see what.  Ensuring that those are obeyed is
tractable.

Empire has had one such command since the beginning: dump.  Empire
4.0.6 added more: sdump, ldump, pdump and ndump.  4.0.7 added lost and
support for incremental dumps.  These commands have served smart
clients well.  However, they cover just the most important part of the
game state (sectors, ships, planes, land units, nukes), and no game
configuration.  They are not quite complete even for what they attempt
to cover.  Finally, their output is harder to parse than necessary.

The xdump command is designed to be the dump to end all dumps.

Like many good ideas, xdump has secondary uses.  Dumping game state as
text can be one half of a game export/import facility.  Useful to
migrate games to different machines or even, with some text mangling
perhaps, different server versions.  We will see below that xdump
isn't quite sufficient for that, but it's a start.

Means to import game configuration let you customize your game without
recompiling the server.  As we will see, configuration files have
different requirements, which xdump doesn't satisfy without some
extensions.

If game import code can edit everything, then a deity command capable
of editing everything is possible.  Proof-of-concept code exists (not
in the tree yet).


Analysis of the data to dump

Game state consists of a fixed set of table files (sector, ship, ...),
telegram files, and a few miscellaneous files.  Game configuration
consists of a fixed set of configuration tables and scalar parameters.

A table is an ordered set of records (table rows).  All records have
the same fields (table columns), which are statically typed.

Fields may be integers, floating-point numbers, x- or y-coordinates,
symbols and symbol sets encoded as integers in a way specific to the
server version, and character arrays.  Configuration table fields may
be pointers to zero-terminated strings, null pointers allowed.  No
other pointers occur.  Unions do not occur.


Requirements analysis

Requirements:

* Capable to dump all tables.

* Can do incremental dumps.

* Output is text.

* Output is reasonably compact.

* Output is trivial to parse.  Triviality test: if it's easy in AWK, C
  (no lex & yacc, just stdio), Lisp (just reader) and Perl (base
  language, no modules), then it's trivial enough.

* Output identifies itself.

* Output is self-contained; symbol encoding is explicit.

* KISS: Keep it simple, stupid.

Non-requirements:

* Generality.  We're not trying to design a general mechanism for
  dumping C data.

* Completeness.  We're not trying to dump stuff other than tables.

* Abstraction.  We're not trying to hide how things are stored in the
  server.  When storage changes, xdump output will change as well, and
  consumers need to be updated.  This is not because abstraction
  wouldn't be nice to have, just because we don't feel up to the task
  of designing one.


Principles of Design

Traditional dumps have a dump function for every table.  These
functions are simple, but exceedingly dull and repetitive.

The selector code works differently.  Each table has a descriptor,
which among other things defines a dictionary of selector descriptors.
A selector descriptor describes a field (table column) visible to
players.  This is what we call meta-data (data about data).  The
selector code knows nothing about the individual tables, it just
interprets meta-data.  That's smart, as it keeps the dull, repetitive
parts in more easily maintainable meta-data rather than code.

xdump follows the selector design, and uses the existing selector
meta-data.  This requires extending the meta-data to configuration
tables, which weren't previously covered.  It also requires some
generalization of selector descriptors, so that all fields can be
covered.

To sum up, meta-data consists of a table of tables, and for each table
a table of selectors (table of columns, so to speak).  It is specific
to the server version and how it is compiled on the host.

To interpret a table xdump, you need its meta-data, because without it
you have no idea what the columns mean.  As meta-data is just a bunch
of tables, xdump can dump it.  But now you need meta-meta-data to make
sense of the meta-data.  Fortunately, meta-meta-data is the same for
all xdumps, and therefore the recursion terminates with a single
meta-meta-table.

xdump dumps symbols and symbol sets as integers.  To decode them, you
need to know what the symbol numbers and symbol set bits mean.  For
this purpose, field meta-data includes the table ID of a symbol table.
A symbol table is a table of value-name pairs, with the value in the
leftmost column.  You decode a symbol value by looking it up in the
symbol table.  You decode a symbol set value by looking up its bits
(powers of two) in the symbol table.

Some integer fields are actually keys in other tables.  For instance,
ship field type is a key in the table of ship types ship-chr, and
plane field ship is a key in the ship table.  Key -1 is special: it's
a null key.  Meta-data encodes these table reference just like for
symbols: the meta-data has the ID of the referenced table, and that
table has the key in the leftmost column.  Obviously, that leftmost
column is a table key as well, referencing the table itself.

A table with its key in the leftmost column can be dumped partially.
Without such a key, you need to count records to find the record
index, and that works only if you can see a prefix of the complete
table.


Syntax of xdump command

See info xdump.


The xdump output Language

Because the output is to be parsed by machines, it needs to be
precisely specified.  We use EBNF (ISO 14977) for syntax, except we
use '-' in meta-identifiers and omit the concatenation symbol ','.

    table = header { record } footer ;
    header = "XDUMP" space [ "meta" space ]
             identifier space timestamp newline ;
    identifier = id-char1 { id-char } ;
    id-char1 = ? ASCII letter ? ;
    id-char = ? ASCII characters 33..126 except '"#()<>=' ? ;
    timestamp = intnum ;
    footer = "/" number newline ;
    record = [ fields ] newline ;
    fields = field { space field } ;
    field = intnum | flonum | string ;
    intnum = ? integer in printf %d format ? ;
    flonum = ? floating-point in printf %g format ? ;
    string = "nil"
           | '"' { str-char } '"' ;
    str-char = "\\" octal-digit octal-digit octal-digit
             | ? ASCII characters 33..126 except '"' and '\\' ? ;
    octal-digit = ? '0'..'7' ? ;
    space = ' ' ;
    newline = ? ASCII character 10 ? ;

Notes:

* The syntax for flonum is debatable.  Precise conversion between
  floating-point and decimal is hard, and C libraries are not required
  to be precise.  Using C99's %a format for flonum would avoid the
  issue, but some programming environments may have trouble converting
  that back to floating-point.  We may change to %a anyway in the
  future.  Clients are advised to accept both.

* Strings syntax could perhaps profit from the remaining C escape
  sequences.  Except for '\"': adding that would complicate regular
  expressions matching the string, and thus violate the `trivial to
  parse' requirement

* Space is to be taken literally: a single space character.  Not a
  non-empty sequence of white-space.

Semantics:

* The table identifier in the header is one of the names in xdump table.

* The timestamp increases monotonically.  It has a noticeable
  granularity: game state may change between an xdump and the next
  timestamp increase.  If the table has a timestamp field, clients can
  xdump incrementally by using a conditional ?timestamp>T, where T is
  one less than the timestamp received with the last xdump of that
  table.

  Timestamp values are currently seconds since the epoch, but this
  might change, and clients are advised not to rely on it.

* The number in the footer matches the number of records.

* Fields match their meta-data (see Meta-Data below).

* "nil" represents a null string (which is not the same as an empty
  string).  Otherwise, fields are to be interpreted just like C
  literals.


Meta-Data

Table meta-data is in xdump table.  Fields:

* uid: The table ID, key for xdump table.  IDs depend on the server
  version; clients should not hard-code them.  This is the leftmost
  field.

* name: The table name.  Clients may identify tables by name.

Field meta-data for table T is in xdump meta T.  The order of fields
in the xdump T matches the order of records in xdump meta T.  Fields
of xdump meta T are:

* name: The field name.  Matches the selector name.  Clients may
  identify fields by name.  This is the leftmost field.

* type: The field's data type, a symbol.  Clients should use this only
  as key for the symbol table.  Symbols are:
  - "d", field uses intnum syntax
  - "g", field uses flonum syntax
  - "s", field uses string syntax
  - "c", field uses string syntax (only until version 4.3.33)

* flags: The field's flags, a symbol set.  Flags are:
  - "deity", field visible only to deities
  - "bits", field is a symbol set, field type must encode symbol "d",
    field table must not be -1.
  - "hidden", field value is masked for contact when option HIDDEN is
    enabled.  Masked values are replaced by -1.

* len: If non-zero, then the record encodes an array with that many
  elements.  The array is dumped as len fields.

  Only until version 4.3.33: if field type encodes symbol "c", it is a
  character array, and is dumped as a single string field.

* table: Key for xdump table.  Unless -1, it defines the table
  referenced by the field value.  Field type must encode symbol "d"
  then.

Symbol table fields:

* value: The symbol's encoding as integer.  If the symbol can be
  element of a symbol set, this is a power of two.

* name: The symbol's name.


Notes on xdump Implementation

Overall impact on the server code is low.

To keeps xdump simple, storage of game state and game configuration
tables has been unified under the common empfile abstraction, making
nxtitem-iterators and selectors equally applicable to all tables.

xdump required a few extensions to meta-data, which may become useful
in other places as well:

* Selectors can now deal with arrays (revived struct castr member
  ca_len).  Not yet available on the Empire command line.

* Selector meta-data can now express that a selector value is a key
  for another table (new struct castr member ca_table).  The selector
  code doesn't use that, yet.

* Redundant selectors can be marked so that xdump ignores them (new
  struct castr member ca_dump).

Meta-data is in empfile[] (table meta-data), src/lib/global/nsc.c
(selector meta-data), src/lib/global/symbol.c (symbol tables).  The
command is in src/lib/commands/xdump.c, unsurprisingly.


Hints on Using xdump in Clients

Let's explore how to dump a game.  To make sense of a table, we need
its meta-data, and to make sense of that table, we need meta-meta
data.  So we start with that:

    [3:640] Command : xdump meta meta
    XDUMP meta meta 1466920477
    "name" 3 0 0 -1
    "type" 1 0 0 33
    "flags" 1 8 0 32
    "len" 1 0 0 -1
    "table" 1 0 0 26
    /5

To interpret this table, we have to know the field names and their
meanings.  Clients hard-code them.  They should be prepared to accept
and ignore additional fields, and to cope with changes in field order,
except they may rely on "name" coming first.

A word on hard-coding.  Clients hard-code *names*.  The numbers used
for table IDs and to encode symbols are none of the client's business.

The encoding doesn't normally change within a game.  Except when the
game is migrated to a sufficiently different server.  That's a rare
event.  Clients may wish to provide for such changes anyway, by
decoupling the client's encoding from the server's, and dumping fresh
meta-data on login.  Incremental meta-data dump would be nice to have.

So we don't know how symbol type and symbol set flags are encoded.  To
decode them, we need their symbol tables.  However, we need flags and
type only for tables we don't know, and there's one more table we do
know, namely the table of tables.  Let's dump that next, starting with
its meta-data:

    [3:640] Command : xdump meta table
    XDUMP meta table 1466920477
    "uid" 1 0 0 26
    "name" 3 0 0 -1
    /2

Because xdump table is referenced from elsewhere (xdump meta meta
field table), the leftmost field must contain the key.  Thus, the
leftmost field's meta-data field table must be the table ID of xdump
table itself.  Indeed, its value matches the one we got in xdump meta
meta.  Let's try to dump the table:

    [5:640] Command : xdump 26 *
    XDUMP table 1466920477
    0 "sect"
    1 "ship"
[...]
    8 "nat"
[...]
    18 "sect-chr"
    19 "ship-chr"
[...]
    26 "table"
[...]
    /48

It worked!

Now dump the two symbol tables we postponed.  Because xdump accepts
table IDs as well as names, we don't have to know their names:

    [5:640] Command : xdump meta 33
    XDUMP meta meta-type 1466920477
    "value" 1 0 0 -1
    "name" 3 0 0 -1
    /2

    [6:640] Command : xdump 33 *
    XDUMP meta-type 1466920477
    1 "d"
    2 "g"
    3 "s"
    /3

    [7:640] Command : xdump meta 32
    XDUMP meta meta-flags 1466920477
    "value" 1 0 0 -1
    "name" 3 0 0 -1
    /2

    [7:640] Command : xdump 32 *
    XDUMP meta-flags 1466920477
    1 "deity"
    8 "bits"
    16 "hidden"
    /3

We now have complete meta-meta information:

    name  type         flags  len  table
    -----------------------------------------
    name     s       (const)    0
    type     d       (const)    0  meta-type
    flags    d  (bits const)    0  meta-flags
    len      d       (const)    0
    table    d       (const)    0  table

Dumping the remaining tables is easy: just walk the table of tables.
Here's the first one:

    [7:640] Command : xdump meta 0
    XDUMP meta sect 1466920477
    "owner" 1 0 0 8
    "xloc" 1 0 0 -1
    "yloc" 1 0 0 -1
    "des" 1 0 0 18
[...]
    /78

A whole load of tables referenced!  Only one of them (not shown above)
is a symbol table.

owner references table nat.  No surprise.

xloc and yloc together reference the sector table, but that's not
expressed in meta-data (yet).

Let's stop here before this gets too long and boring.  Experiment
yourself!  Check out example Perl code scripts/xdump.pl.


Analysis of xdump as Configuration File Format

xdump makes a lousy configuration format because it is unwieldy to
edit for humans.  That's because configuration files have different
requirements than dumps:

* Can be edited by humans with common tools, including text editors
  and spreadsheets.

  Using text editors requires a nice fixed-width table layout.
  Spreadsheet import requires trivial field separation.  Tab character
  field separator or fixed width columns should do.  The syntax should
  allow all that, but not require it.

  Trouble spots:

  - xdump's rigid horizontal and vertical spacing makes it impossible
    to align things visually.

  - xdump uses one line per record, which can lead to excessively long
    lines.

  - xdump's string syntax requires octal escape for space.

  - No comment syntax.

* Each table is self-contained.  You don't have to look into other
  tables to make sense of it.

  This conflicts with xdump's separation of data and meta-data.  You
  need the table's meta-data to identify fields, and the referenced
  symbol tables to decode symbols.

* Easy to parse.  Don't compromise legibility just to please some dumb
  tool, though.

Since we're trying to apply xdump to the configuration file problem,
we get an additional requirement:

* Reasonably close to xdump.  Translation between machine-readable and
  human-readable should be straightforward, if meta-data is available.

This leads to a human-readable dialect of the xdump language.


Human-Readable xdump Language

Fundamental difference to basic, machine-readable xdump: the rigid
single space between fields is replaced by the rule known from
programming languages: white-space (space and tab) separates tokens
and is otherwise ignored.  The space non-terminal is no longer needed.

Rationale: This allows visual alignment of columns and free mixing of
space and tab characters.

Comments start with "#" and extend to the end of the line.  They are
equivalent to a newline.

Rationale: Follow econfig syntax.

Tables with a record UID in the leftmost field can be `split
vertically' into multiple parts.  Each part must contain the same set
of records.  The leftmost field must be repeated in each part.  Other
fields may be repeated.  Repeated fields must be the same in all
parts.  Naturally, the parts together must provide the same fields as
a table that is not split.

Rationale: This is to let you avoid long lines.  Line continuation
syntax would be simpler, but turns out to be illegible.  Requiring
record UID is not technically necessary, as counting records works the
same whether a table is split or not.  Except humans can't count.
Perhaps this should be a recommendation for use rather than part of
the language.

EBNF changes:

* Header and footer:

    header = "config" identifier newline colhdr newline ;
    colhdr = { identifier [ "(" ( intnum | identifier ) ")" ] } [ "..." ] ;
    footer = "/config" newline ;

  If colhdr ends with "...", the table is continued in another part,
  which shall follow immediately.

  Rationale:

  - The xdump needs to identify itself as human-readable, hence change
    from "XDUMP" to "config".

  - The timestamp in the header is useless for the applications we
    have in mind for human-readable xdumps.  The number of records in
    the footer is of marginal value at best, and a pain for humans to
    update.

  - The column header is due to the self-containedness requirement.
    It contains just the essential bit of meta-data: the column names.

* Symbolic fields:

    field = intnum | flonum | string | symbol | symset ;

  Rationale:

  - Syntax for symbols and sets of symbols is due to the
    self-containedness requirement.  Machine-readable xdump gets away
    with just numbers, which have to be decoded using meta-data.

* Friendlier numbers and strings:

    flonum = ? floating-point in scanf %g format ? ;
    str-char = "\\" octal-digit octal-digit octal-digit
             | ? ASCII characters 32..126 except '"' and '\\' ? ;

  Rationale:

  - Machine-readable floating-point syntax is too rigid.  Accept
    everything that scanf does.  Could also change intnum to %i
    format, which accepts octal and hexadecimal in C syntax, but that
    seems not worth the documentation bother.

  - Machine-readable syntax requires \040 instead of space in strings
    to allow trivial splitting into fields.  This is unacceptable here
    due to the legibility requirement, hence the change to str-char.

* Parse nil as symbol:

    string = '"' { str-char } '"' ;

  Rationale: This is a technicality required to keep the parse
  unambiguous.

* Symbols:

    symbol = identifier ;
    symset = "(" { symbol } ")" ;

  The special symbol "nil" is to be interpreted as null string.

  Rationale:

  - The symbol set syntax is the simplest that could work.  We need to
    allow space between the symbols for legibility anyway, so why not
    make it the delimiter.  A stop token is required to find the end
    of the field, and a start token is useful for distinguishing
    between symbol and symset.  Bracketing with some kind of
    parenthesis is an obvious solution.

The resulting sub-language for records is a superset of
machine-readable sub-language for records.

See src/lib/global/*.config for examples.


Notes on Table Configuration Implementation

econfig key custom_tables lists table configuration files.  At this
time, reading a custom table merges it with the built-in table, then
truncates the result after the last record read from the custom table.

Some of the tables are rather ugly in C, and cumbersome to edit.  We
thus moved them to configuration files (src/lib/global/*.config).  The
server reads them from builtindir before reading custom tables.

The code dealing with these files is in src/lib/common/conftab.c.

Actual work is done by src/lib/common/xundump.c, which accepts both
human-readable and machine-readable input.  The parser is not precise;
it accepts human-readable syntax even within tables whose header marks
them machine-readable.

Symbolic index values in column headers are not implemented.  They
occur in item selector pkg, which is an array indexed by values in
symbol table packing.

Configuration tables contain values that are not meant to be
customized.  For instance, meta-data and symbol tables reflect the
encoding of C language constructs in the server.  Such selectors are
marked (struct castr member ca_dump), so that the code can prohibit
changes.

All tables are checked against meta-data on server startup by
ef_verify().  More elaborate checking would be nice, and probably
requires additional meta-data.


Appendix: Empire 3 C_SYNC --- A Cautionary Tale

Clients are just as important as the server, and it's too darn hard to
write a good client.  In 1995, Ken Stevens decided to do something
about it.

Ken cast the problem as a data synchronization problem.  Quote C_SYNC
RFC 5.1, section `Abstract':

  This is a specification for a new method of synchronizing game data
  in the Empire client with data in the server.

and section `Objectives':

  This new mode of communication between the server and the client will
  be called C_SYNC communication and will satisfy the following 6
  criterea:

  (1) Output format will be version independent.  So if someone is
  using an old EmpireToolkit, then it will still work with a newer
  version of the server.

  (2) Every C_SYNC message will be a self-contained packet.  i.e. the
  client will not need to depend on previous messages (header messages)
  to determine the meaning of a C_SYNC message.

  (3) A C_SYNC message will be able to represent any of the
  player-accessible data that is contained in the server database (e.g.
  enemy ships, nations).

  (4) Bandwidth will be minimized (i.e. the format will be as
  concise as possible) while remaining human-readable (i.e. no
  binary messages).  [Note that data compression may be added at a later
  date, but if it is added, it will be added on a separate port to
  maintain backwards compatability.]

  (5) The client will be able to tell the server whether it wants
  to receive C_SYNC messages and whether these messages can be sent
  asynchroniously (via "toggle sync" and "toggle async" respectively).

  (6) A portable ANSI C EmpireToolkit will be made available for
  parsing C_SYNC messages and managing the data they contain.

C_SYNC worked by hooking into ef_write() & friends so it could
`synchronize' the client on game state changes.

Sounds jolly good, doesn't it?

Well, it was a failure, and Wolfpack ripped it out right away.  Quote
the change log:

  Changes to Empire 4.0.0 - Initial release
   * Initial Wolfpack release - Long live the Wolfpack!!!!
[...]
   * Removed C_SYNC.  This is done for 2 reasons.  1) None of us like it or
      wish to support it.  2) We envision a better scheme for doing similar
      things will come along.

But *why* did it fail?  Just because Steve McClure hated it?  Nope.
C_SYNC failed for several different reasons, each of them bad, but
only the last one is truly fundamental.

a. Lack of a rigorous and complete definition.  The RFC is long on
   syntax, but short on semantics.  For instance, the unit type was
   encoded as a number.  Unit characteristics happened to be dumped in
   an order that matched these numbers, but that wasn't defined
   anywhere.

b. Overly complicated syntax.  Trouble with encoding of strings.

c. Buggy implementation.  Malformed C_SYNC messages, duplicate
   messages, missing messages, semantically incorrect messages, you
   name it.

d. Change of crew before it was finished.  Wolfpack took over and
   understandable wasn't interested in this half-finished mess.

None of the above is a fundamental, inherent flaw of the idea.  The
next one is more serious:

e. It failed to achieve objective (4), and therefore slowed down
   clients too much to be of use in real-time combat.  When you fired
   from a bunch of ships, C_SYNC would push complete records for all
   the ships and the target to you.  Most of that data is redundant.

   That's because C_SYNC didn't transmit state changes, it
   resynchronized state, and the pieces of state it could transmit
   were too large.

   The network was slower then.  But let's not be complacent.  I/O is
   slow.  Always was, most likely ever will be.

   Maybe sending the messages out of band (separate TCP stream) would
   help.  Maybe not.

And here comes the killer:

f. The data to sync is not readily available on the server.

   Yup.  Think about it.  The game state on the server is *not* the
   same as on the client.  The server grants the client a carefully
   limited view on certain parts of server game state on certain
   events.

   To be complete, a machine-readable protocol must disclose as much
   information as the human-readable output.  Tracking server game
   state changes cannot do that alone.  For instance, lookout tells
   you ship#, owner and location.  That event does not trigger any
   state change on the server!

   To be correct, a machine-readable protocol must disclose no more
   information than the human-readable output.  When you observe a
   server game state change, you can only guess what event triggered
   it, and what it disclosed to which player.  You're stuck with
   conservative assumptions.  That's the death knell for completeness.
   Correct assumptions will be non-obvious, so correctness is
   non-obvious, too, hence hard to achieve and maintain.

   Bottom line: tracking server state change cannot support a complete
   client protocol for hard theoretical reasons, and I believe it
   cannot support a correct one for practical reasons.

Oddly enough, people criticized C_SYNC for all the flaws it had (and
some it hadn't), except for f.

What now?  Throw up our hands in despair and give up?  Nah.  Ken tried
a shortcut, and it didn't work.  That doesn't mean there's no way at
all.  I believe the only way to get this done right is by tracking
*events*.  Whenever something is printed to a player, be it live
connection or telegram, we need to transmit precisely the same
information in machine-readable form.  Much more work.

xdump shares valuable ideas with C_SYNC, e.g. using selector
meta-data.  It is, however, much more modest in scope.  We're pretty
sure we can get it right and get it done in a reasonable time frame.