/sys/doc/ Documentation archive

System and Interface Changes to Inferno

C H Forsyth

Vita Nuova


9 June 2003


This paper describes some of the changes made to Inferno interfaces as they stood in the published Third Edition manuals, to form the current Fourth Edition of the system, and the broad effects on internal and external interfaces. Changes include: extensions to the Limbo language; new instructions in Dis and the virtual machine; extra content in Dis object files; structure of the source tree; configuration of emu; replacement of the window system with changes to the client interface; commands renamed, replaced, and removed; revised support for network booting; 9P2000 becomes the basis for Styx; a graphics model offering alpha-blended compositing and general pixel structure; and improvements to Tk.

1. Limbo

Exceptions and fixed point have been added to the Limbo language. They are described in more detail in separate notes by John Firth, shortly to be available on the Vita Nuova web site www.vitanuova.com. Channels can now be buffered. A form of polymorphism is now available in Limbo.

1.1. Exceptions

Discussion of exceptions will be restricted here to implications for existing source code. The most obvious changes are that Sys->rescue, Sys->rescued, Sys->unrescue and Sys->raise have vanished. Instead the exception handling is expressed using constructions in the Limbo language. Named exceptions can be declared and used (these are described in the note by Firth), and they are declared as part of the type of functions that raise them. There is also a general ‘failure’ exception that effectively subsumes the old Sys->rescue scheme, including run-time errors such as ‘out of memory’ that can happen in almost any function. Unlike named exceptions a ‘failure’ exception can be raised or caught by any function, and its value is a string. The raise statement raises an exception. This is most obvious in commands that wish to produce an ‘exit status’. Instead of


one must now write

raise "fail:usage";

(That is one of the more common source changes required to Third Edition Limbo commands, since that was the most common use of exceptions before.) A block can have an exception handler:


    a := array[128] of byte;


} exception e {

"out of memory:*" =>

    sys->print("i need more space: %s\n", e);

"fail:*" =>

    sys->print("exit status: %s\n", e);

"*" =>

    sys->print("unexpected error: %s\n", e);

    raise;  # propagate it


If an exception is raised during the execution of the block (including functions it calls), execution of the block is abandoned, and control transfers to the appropriate exception handler (which is outside the block). Because the compiler and run-time system know the scope of the exception, values such as a above are correctly reclaimed on exit from the faulty block. Unhandled failures are propagated to callers; unhandled named exceptions (currently) become failures.

A process group can cause unhandled exceptions in any process in the group either to propagate to all members of the group, or to be propagated to the process group leader after destroying the other processes in the group. This makes it easier to program recovery from exceptions within a group of concurrent processes. For instance, if a process is expected to send to another on a channel, but fails unexpectedly instead (eg, because memory was exhausted), instead of leaving the intended recipient blocked on a receive operation, it can be sent an exception to notify it of the failure of the other process, allowing it to take appropriate recovery action. (This could sometimes be programmed using the wait file of prog(3), but not always.)

Exception handling is intended for recovering from disaster. We still think it is better Limbo style to use tuples, channels and processes to make ordinary error handling explicit. The few attempts to use failure exceptions to achieve ‘pretty’ but peculiar control flow have had exactly the usual effect of making the code hard to follow and error-prone.

1.2. Channels

Buffered channels have been added:

c := chan [N] of int;

where N is an integer value, creates a channel that will allow up to N integer values to be sent to it without an intervening receive without blocking the sender. If N is zero, the channel is unbuffered, equivalent to plain chan of int, and synchronises sender and receiver as before.

The restriction that a given channel value could not be sent to or received from in two alt statements simultaneously has been removed.

1.3. Polymorphism

John Firth has implemented a form of parametric polymorphism in Limbo. It too will be described in a separate note. Currently we are still fussing over aspects of the constraint syntax and some other implications of the most general form, and since some aspects are therefore subject to change, including syntax, we have not yet published the details. We think it is possible to use the following subset without having to change the code later:

1. Function declarations can be parametrised by one or more type variables: For example:

reverse[T](l: list of T): list of T


    rl: list of T;

    for(; l != nil; l = tl l)

        rl = hd l :: rl;

    return rl;


Such a function can then be invoked on any compatible set of values. The function invocation does not specify the type (the compiler does type unification on the parameters). Thus the above can be used as:

l1: list of string;

l2: list of ref Item;

l3: list of list of string;

l1 = reverse(l1);

l2 = reverse(l2);

l3 = reverse(l3);

2. ADTs can also be parametrised:

Tree: adt[T] {

    v: T;

    l, r: cyclic ref Tree[T];


allowing declaration of Tree[ref Item] and Tree[string] for instance.

3. Values of the parametrised type can only be declared, assigned, passed as parameters, returned, or sent down channels. The only types that can be used as actual parameter types are reference types (ie, ref ADT, array, chan, list and module), and string (which is a value type but is implemented using a reference). At some point we shall allow a function such as reverse above to be invoked with any compatible type (not just reference types) but that requires changes to Dis and the virtual machine not yet made.

The formal type parameters can be further constrained by listing a set of operations that they must have (which currently implies the actual parameters must be ADT types with compatible operations). We are not completely happy with the current constraint syntax, and some other aspects of the scheme, and so that be described here later once we have settled it.

2. Dis and virtual machine

To make the Limbo changes and extensions some new operators were added to the virtual machine. (We also added a casel operator to allow case statements to work on big values.) Modules that have exception handlers also have a (new) exception table, added to the Dis object format. Furthermore, we moved the import table used by the load operator out of the Dis data space into the object format (which also makes it available for inspection by wm/rt amongst others).

There is now an internal interface to set conditions under which modules must be signed to be loaded, and to check a signature on a module. Appropriate stubs are defined when module signing is not configured; if sign(3) is configured, however, it replaces them by ones that enforce its signing policy.

3. Window manager

The window manager wm(1) has been reimplemented by Roger Peppe. It now multiplexes pointer and keyboard input to applications, and manages windows on the display. Tk(2) no longer manages windows from inside the kernel. In some ways the structure is closer to that of mux(1) and more specifically the design described in Rob Pike’s paper ‘‘A Concurrent Window System’’. It is possible to import and export window system environments between hosts.

This is one of the bigger causes of source file changes, although many of them can be done by global substitutions (eg, using acme(1)). Appendix A gives details. Wmlib is no longer the application’s interface to the window system. Instead that is done through a new Tkclient module; see tkclient(2). (It uses a different Wmlib as an auxiliary module, and also uses a new Titlebar module to allow the look of the window decoration to be changed more easily). An application acquires a window by a call to Tkclient->toplevel; starts pointer or keyboard input if desired by calling Tkclient->startinput; and puts the window on screen (after sending it Tk configuration commands) using Tkclient->onscreen. Nothing appears on screen until that is called (which amongst other things avoids the resizing on start up that afflicted the original scheme). Onscreen gives it a connection to the window manager for pointer, keyboard and control input, with a separate channel for each. When it receives data from any of the channels (typically using alt) it must pass it to Tk using calls to appropriate Tkclient functions.

The toolbar used by the old wm is now provided by a separate program wm/toolbar (see toolbar(1)), and it is toolbar that interprets the /lib/wmsetup file. Wm invokes wm/toolbar by default so most users will see no difference, but it does make it easier to develop alternative interfaces. More visible is that wm/logon is now a client of the window manager, and must be invoked as follows:

wm/wm wm/logon

Applications need not even use tk(2). There is an interface for draw-only clients, wmclient(2).

4. Inferno source tree

The structure of the Inferno source tree has changed in the following ways.

4.1. Library source

The image and memimage directories have gone, replaced by libdraw and libmemdraw. The directories in the Inferno root that contain the source for libraries now always have names starting ‘lib’: libcrypt, libinterp, libkeyring, libmath, etc.

4.2. Emu source

The emu directory now contains a subdirectory structure similar to the os kernels, and uses a similar configuration file (parts list) to say what goes in a given instance of emu. This allows platform-dependent selection of drivers, libraries and even #/ (ie, root(3)) contents to be done easily.

The top directory, /emu, contains: mkfile that simply moves to the platform configured by /mkconfig, allowing builds in the Inferno root as before; a subdirectory port containing portable code (including some code shared by several platforms, such as devfs-posix.c); and a subdirectory for each hosting platform, distinguished by an upper-case initial letter. Current platforms include FreeBSD, Irix, Linux, Nt (for all Windows platforms after 95), Plan9, Solaris, and several others.

4.3. Emu configuration

Each platform-specific directory contains a configuration file with the same structure and indeed similar contents to the ones used for the native kernel. The default configuration file is called emu. Another can be chosen, again in a similar way to the native kernel, by using

mk ’CONF=cfile

where cfile is the name of the configuration file. The name of the resulting executable file contains the configuration file name but depends on the platform: it is cfile.exe on Windows, o.cfile on Unix systems, and 8.cfile on 386 Plan 9 systems. The configuration file format and contents is documented for all types of kernels by conf(10.6).

4.4. Tk source

The Tk implementation in libtk has been made more modular. It allows a significantly different ‘style’ to be implemented, and although that is by no means trivial to do, there is at least an interface to do it. We hope to change various aspects of the standard style further, but that has not yet been done.

5. Commands and modules

There are new commands and library modules, others have become obsolete and been removed, and a few existing ones have been given new names (typically when ones with similar function have been collected together). The biggest change has been to wm(1), which retains the same name but slightly different invocation and completely different implementation, as discussed above. Here I shall simply note the bigger changes, rather than discuss new functionality.

5.1. Renamed commands

As part of a mild reorganisation of the /appl and /dis trees, we have moved commands out of /dis/lib so that it now contains only library modules except for a few commands left there temporarily for compatibility. Commands themselves have sometimes been shuffled to subdirectories, often copying seemingly better structure from Plan 9, so that authentication commands are auth/..., naming service commands are ndb/..., and IP-specific commands are ip/... .

One noticeable change is that lib/cs is now ndb/cs. More dramatically, the command lib/srv (ie, srv(8)) has been replaced by sh(1) scripts, all described by svc(8), that contain appropriate calls to listen(1) after setting up any locally-desired environment.

Other commands have also moved:

lib/plumber is now simply plumber

lib/bootp and lib/tfptd have become ip/bootpd and ip/tftpd, documented in bootpd(8)

lib/virgild has become ip/virgild (see virgild(8))

lib/chatsrv, lib/rdbgsrv and cpuslave have moved to auxi (ie, /dis/auxi and /appl/cmd/auxi)

csquery has become ndb/csquery

5.2. New or newly-documented commands

• an authentication server (signer) can use keyfs(4) to store its keys securely in the encrypted file /keydb/keys (instead of the unencrypted /keydb/password), and run keysrv(4) to offer secure change of password remotely. They are typically started, with other signing services, by svc/auth described in svc(8).

/dis/auth and /appl/cmd/auth contain commands related to authentication; they rely on keyfs(4) in most cases. The older ones that use /keydb/passwd are still in /dis/lib and /appl/lib during the transition

dns(8) has replaced the lib/ipsrv implementation of srv(2); when used, it must be started before ndb/cs. Srv(2) has reverted to being a hosted-only interface to the hosting system’s native DNS resolver. It is automatically used by cs(8) if it cannot find dns(8), and dns(8) will also use it if available before consulting the DNS network.

chgrp(1), cpuview(1), grid(1), 9660srv(4), cpuslave(4), dossrv(4), keyfs(4), keysrv(4), nsslave(4), palmsrv(4), registry(4), rioimport, styxchat(1), styxlisten, wmexport, wmimport, and uniq(1) are new

• the multiplayer games software previously in /appl/games has been replaced by a related but significantly different system in /appl/spree. (Also see spree(2) for supporting modules.)

Registry(4) provides dynamic registration and location of services using sets of attributes/value pairs, through a name space. Registries(2) provides a convenient Limbo interface for registration and query.

5.3. Commands removed

lib/csget (see cs(8) for its replacement csquery)

• the undocumented and obsolete commands lib/isrv and lib/istyxd have been removed, since either the none authentication protocol, or the -A option to mount can be used if no authentication is needed

lib/srv has been replaced by svc(8) as mentioned above.

getenv and setenv have been removed since the Shell provides alternatives

wm/license is no longer needed

5.4. New modules

There are library modules to support: registries and configuration files of attribute/value pairs; Internet address parsing and manipulation; management of windows and subwindows (used by wm(1) itself); timers; Styx; Styx servers; exception handling; memory and performance profiling; Freetype interface; parsing Palm databases; and navigating XML files (without reading them all into memory) and interpreting style sheets.

6. Styx

Styx was derived from the 9P protocol used by Plan 9 in 1995, with changes that reflected the requirements of the Inferno project of the time, mainly by removing features that were thought too closely tied to the Plan 9 environment. Some 9P messages were removed, particularly those that incorporated details of the Plan 9 authentication methods; Styx moved authentication outside the file service protocol. Other changes eliminated file locking and append-only files. Some restrictions that 9P imposed were retained, however, such as limiting file names to 27 bytes. This last restriction is fine for synthetic network services, but has been troublesome when trying to access Unix and Windows systems, amongst others.

A recent revision of 9P adds support for much longer file names and takes the opportunity to improve other aspects of the protocol. It also removes details of authentication algorithms from the protocol. The Styx implementation now uses the new version of 9P as the default file service protocol. (It is possible that for interoperation with older Inferno systems the system will be able to interact with both old and new versions of Styx.)

6.1. Protocol changes

The messages Tauth and Tversion are new to Styx. Tversion includes negotiation (at connection start) of the message size and protocol version; it also introduces a new session. Tauth obtains access to a special authentication file if the server requires authentication within a Styx session. Tclone has been replaced by a more elaborate form of Twalk that allows zero to MAXWELEM (16) elements to be walked, perhaps to a new fid, in a single message, returning a sequence of qid values in Rwalk. (A clone is simply a walk of a fid to a new fid with zero elements.) A walk of several elements can return partial results if the walk of the first element succeeds but subsequent ones fail. A partial walk leaves the state of the fids unchanged. Ropen and Rcreate return a suggested size for atomic I/O on the fid (0 means ‘not given’). All strings are variable length, and consequently Twstat and Rstat data is variable length and formatted differently. Data returned from Tread of a directory is similarly changed, because directory entries are not fixed length. Tnop has gone.

Tags remain 16-bit integers, but fids and counts become 32-bit integers (mainly of interest to large systems), and qids have a different structure. Previously a qid was a pair of 32-bit integers, path and vers, where path had the top bit set for a directory. Now a qid is a triple: a 64-bit path, 32-bit vers, and 8-bit type. The type is defined to be the top 8 bits of the file’s mode. The path does not have the top bit set for a directory, and indeed the path value is not interpreted by the protocol. There are now bits in the file mode for append-only and exclusive-use files (new for Inferno), and for authentication files (new for both Plan 9 and Inferno). The stat information includes the user name that last caused the file’s mtime to be changed. All strings in the protocol are variable length: file names, attach names, user names, and error text.

The message format on the wire is significantly different. The message size is negotiated for a connection by Tversion, and messages can be large, allowing much more data to be sent in single Twrite and Rread messages. The header includes a 32-bit message size, making it easy to find message boundaries without parsing the contents. Strings are represented as a 16-bit size followed by the string’s UTF-8 encoding (without zero byte). R-messages do not carry a copy of the fid from the T-message. Padding bytes have gone. The order of some fields has changed of course to match message parameter changes.

Authentication of the connection itself, and optionally establishing the keys for digesting and encryption, is done before the protocol starts, in both Inferno and Plan 9. Details will follow on the protocol for that, and Limbo interfaces. For now, it can be assumed that the old authentication messages can still be used, even after a more flexible protocol has been implemented. Tauth can be used to authenticate particular accesses within such a session, but implies trust by the server that the client system will not cheat its users. (That trust is typically established by the connection level authentication which is needed anyway for link encryption, and thus for single-user clients further authentication seems extraneous in most cases.) Most Inferno services that run as file servers within a system (eg, 9660srv) will, like Plan 9’s, reply to Tauth with an Rerror stating ‘‘authentication not required’’. Access to them when exported is typically controlled as now by verifying the incoming connection.

6.2. Limbo interface changes

Because Limbo’s interface to file service via Sys and other modules uses Limbo string for names, and that is inherently variable length, there are no interface changes required for that aspect of the protocol change, and consequently no source changes (in contrast to the introduction of 9P2000 in C implementations). Similarly the Inferno directory reading interfaces remain unchanged.

The ‘directory mode’ bit previously called CHDIR is now called DMDIR. It is used only in Dir.mode. CHDIR is no longer defined, partly because it was used both in Dir.mode and Qid.path, and the latter instances must change (discussed below). There are bits (new to Inferno) for DMAPPEND (append-only file), DMEXCL (exclusive-use file), and DMAUTH (authentication file). The protocol can return the user name of the user that caused mtime to be changed on a file; that is now available as Dir.muid.

The structure of Qid has changed. Previously a Qid had a 32-bit path and a 32-bit version number, vers. The top bit (CHDIR) of path was set iff the Qid was that of a directory. The path is now 64 bits (which is big in Limbo and vlong in the kernel), and there is no longer the convention that the top bit of path must be 1 for a directory. Instead, there is a new, separate type field (called qtype in Limbo) that has the value of the top 8 bits of the file’s mode. Each bit DMx in Dir.mode, has got a corresponding bit QTx in Qid.qtype: QTDIR, QTAPPEND, QTEXCL and QTAUTH. The bit QTDIR must be set in the Qid.qtype for a directory, and only then. There is an extra constant QTFILE that is defined to be zero, and is used for clarity when neither QTDIR nor QTAUTH is set.

In Styx file servers, changes are required to reflect the slightly different set of message types and a few new parameters, but the main changes are: handling zero or more name elements at once in Twalk and Rwalk; changing CHDIR to DMDIR in Dir.mode (easy); the use of the new Qid.qtype field and QTDIR instead of CHDIR in Qid.path (a little more effort); and (typically) the insertion of casts to force Qid.path to int and thus ensure the use of 32-bit operations except where 64-bit paths really are needed (hardly ever in synthetic file servers). The new modules for use by file servers are discussed in the next section.

The revised definition of Twstat in stat(5), and thus sys->wstat, provides for ‘‘don’t care’’ values in Dir that are tedious to provide directly; a new adt value Sys->nulldir provides the right initial value for a Dir which is then changed as needed for wstat.


Create a directory:


fd := sys->create(name, Sys->OREAD, Sys->CHDIR | 8r777);


fd := sys->create(name, Sys->OREAD, Sys->DMDIR | 8r777); # not CHDIR

Make Qids for a file and a directory:


Qdir, Qdata: con iota;

qd := Sys->Qid(Sys->CHDIR | Qdir, 0);

qf := Sys->Qid(Qdata, 0);


Qdir, Qdata: con iota;

qd := Sys->Qid(big Qdir, 0, Sys->QTDIR);

qf := Sys->Qid(big Qdata, 0, Sys->QTFILE);

Test if a file is a directory:


isdir(d: Sys->Dir): int


    return (d.mode & Sys->CHDIR) != 0;


    return (d.qid.path & Sys->CHDIR) != 0;



isdir(d: Sys->Dir): int


    return (d.mode & Sys->DMDIR) != 0;


    return (d.qid.qtype & Sys->QTDIR) != 0;


If one wishes to have values big only when required, one can write:

case int dir.qid.path {

Qdir =>


Qdata =>


Qctl =>



Of course with the Dis change mentioned above, case can now be applied to big values, so it is no longer necessary to add the cast (as it once was). Even so, 32-bit operations are faster when they suffice.

6.3. Styx protocol in Limbo: Styx and Styxservers

A new module Styx, defined by styx.m, provides access to the Styx protocol messages, as variants of pick adts Tmsg and Rmsg. (There was an old, undocumented Styx module but this new interface is completely different.) It is used by several file servers, such as dossrv, cdfs, and the new logfs. See the attached manual page. There are several implementations with the same signature, covering different combinations of old and new Inferno and old and new protocols, through the same interface. There are slight differences in the application code for old and new systems because of the changed type and structure of Qid. The versions that talk the old protocol need to store some internal state, and are intended only to meet compatibility requirements during the transition.

Many file service applications, however, serve a simple name space, requiring more than can be done with file2chan, but wishing some help in handling the protocol details. Two new modules Styxservers and Nametree are provided to make such applications easier to write. They are closely related and thus both modules are defined by styxservers.m.

Styxservers provides help in handling fids and interpreting the Styx requests for navigating a name space, and provides a reasonable set of default actions, allowing the application to focus on implementing read and write access to the files in the name space. It uses Styx to talk to the Styx client on a connection. It interacts with the application through a channel interface and the Navigator adt to navigate an abstract representation of the application’s name space. The module can be used on its own, with the application doing the work of replying to those queries itself, or it can get extra help in the common cases from Nametree. Nametree provides a Tree adt and operations for the application to build an abstract representation of a name space and maintain it dynamically quite simply, and it exports the channel interface used by Styxservers for navigation, thus connecting the two, but leaving the application in complete control of the name space contents viewed by Styx. See the manual pages styxservers(2) and styxservers-nametree(2), attached. The latter includes a short working example of combining the two modules.

The previous release of the system had a module Styxlib that combined the functions of Styx and Styxservers. It remains for a time for transition, but newer applications should use either Styx or Styxservers.

A new command styxchat(8) exchanges Styx messages with a server, reading a textual representation of T-messages on standard input. It can be helpful when testing a Styx server implementation. (It was originally developed to test the Styx module implementations in several configurations.) See the attached manual page for details. It also supports an option that allows it to act as a server, printing T-messages as they are received from clients, and reading R-messages in a textual form from standard input for replies.

6.4. Device driver changes

Most of the differences for most drivers are relatively minor (in diff terms).

Throughout the hosted and emulated kernels:

Qid now is the structure:

struct Qid {

    vlong   path;

    ulong   vers;

    uchar   type;


The type field has values QTDIR, QTFILE, QTAPPEND, etc. The test previously written

if(qid.path & CHDIR)

is now written

if(qid.type & QTDIR)

Because of that change, the various switch statements in the drivers that previously read



switch(c->qid.path & ~Sys->CHDIR){

now read


to keep operations to 32 bits (except where otherwise required).

∙ The first entry of a driver’s Dirtab must be an entry for ".", if the driver uses devgen to help implement walk, stat, devdirread or open operations.

∙ Offsets passed to the driver’s read and write entry points are 64-bit vlong, not 32-bit ulong.

∙ The stat entry point has an extra buffer size parameter:

int xyzstat(Chan *c, uchar *dp, int n)

It also returns an integer: the size of the result. Devstat accepts the extra parameter and returns an appropriate result:

static int

xyzstat(Chan *c, uchar *dp, int n)


    return devstat(c, dp, n, rtcdir, nelem(xyzdir), devgen);


∙ The biggest change is to walk. It has the signature:

Walkqid *xyzwalk(Chan *c, Chan *nc, char **names, int nname);

and it allows zero or more elements to be walked in a single call, returning its result in a newly-allocated Walkqid structure:

struct Walkqid {

    Chan*   clone;

    int nqid;

    Qid qid[1];


Note that the array Walkqid.qid must actually hold up to nname Qids, and thus is allocated as follows:

wq = smalloc(sizeof(Walkqid)+(nname-1)*sizeof(Qid));

The driver must take care that the space is reclaimed if error is called before its walk function returns, by using waserror as required. Fortunately, devwalk looks after the details of walk and walkqid for most drivers:

static Walkqid*  

xyzwalk(Chan* c, Chan *nc, char** name, int nname)


    return devwalk(c, nc, name, nname, xyzdir,

        nelem(xyzdir), devgen);


∙ The clone entry point has gone, since cloning is seen by a driver as a particular form of call to its walk entry, where the parameter values satisfy:

c != nc && nwname == 0

One difference is that a node can be cloned and walked in a single operation, in other words nwname can be non-zero, and the incoming nc is often nil and a new Chan must be allocated. Note that if the driver found it adequate to call devclone previously, then the new devwalk will generally look after it as well. Devclone remains for use as a utility function for the few drivers that need to clone a channel themselves, in their walk operations or elsewhere.

∙ The detach entry has been renamed shutdown (it was never the opposite of attach). The stub devshutdown can be used by devices that do not need it.

For drivers that serve a simple name space using the functions of dev.c (described in devattach(10.2)), only a handful of simple changes are required. Most are pointed out by the compilers as type clashes. The main exception is the need for a Dirtab to have its first entry be an entry for "." if the Dirtab will be passed to devgen via devwalk, devstat and devdirread.

7. Sys module changes

7.1. Sys: name change(s)

The name ERRLEN has become ERRMAX (since it is the limit to any error string, not its necessary length). NAMELEN has been removed, to allow each instance to be found (by compilation) and either removed (where it was simply limiting the length of a file name), or replaced by NAMEMAX where it was used as a buffer size to read in names such as /dev/sysname or /dev/user.

7.2. Sys: file sizes

The Styx protocol has always supported 64-bit file sizes and file offsets. The Inferno interface has not. Sys has changed so that length and offset values become big, specifically: file size Dir.length, the offset parameter to seek, and seek’s result.

These and the Qid changes account for quite a few changes in our own source tree. Typically, applications did things like this:


buf := array[d.length] of byte;

sys->seek(fd, 0, Sys->SEEKSTART);

off := sys->seek(fd, 0, Sys->SEEKRELA); rec := off + HDRLEN;

for(offset := 0; offset < d.length; offset += RECSIZE){

    sys->seek(fd, offset, Sys->SEEKSTART);



The compiler now objects in each case because big values are now appearing where int is required, or conversely. In some cases it is obvious that adding a cast is correct; in others it is worth considering whether the calculation should indeed be big because file sizes for instance can in practice exceed the range of a signed integer without too much trouble today, especially when the ‘file’ is a storage device. The case that some people like and some dislike is:

if(sys->seek(fd, big offset, Sys->SEEKSTART) < big 0) ...

where the big 0 is needed because sys->seek is big, and there are no ‘usual arithmetic conversions’ as in C. (Given the tangle that several languages have made of such conversions, perhaps being strict is correct.)

7.3. Sys: export

Sys->export now has the signature:

export: fn(c: ref Sys->FD, dir: string, flag: int): int;

allowing a directory dir other than "/" to be exported. It replaces the exportdir function of (later) Third Edition.

7.4. Sys: Styx support

The revision of Styx has caused three calls to be added:

fauth:    fn(fd: ref Sys->FD, aname: string): ref Sys->FD;

fversion: fn(fd: ref Sys->FD, msize: int, version: string): (int, string);

iounit:   fn(fd: ref Sys->FD): int;

Fversion initialises a Styx session on connection fd, sending the message size msize and protocol version string version; it returns a tuple giving the message size and version returned by the Styx server. It is rarely called directly; the mount operation does it automatically on an uninitialised connection.

Fauth sends a Styx Tauth message on connection fd, and if successful, returns a file descriptor that refers to an authentication file provided by the file server, which may be read and written by Sys->read and Sys->write to implement the authentication protocol(s) supported by the server. Fauth is needed only when the server requires authentication.

Iounit returns the ‘atomic IO unit’ suggested for the file fd by its file server when it was opened.

7.5. Sys: mount

The mount system call has acquired a second file descriptor parameter:

mount: fn(fd: ref Sys->FD, afd: ref Sys->FD, on: string,

          flags: int, spec: string): int;

Afd is nil if the file server is known not to require authentication within a Styx session. (The connection might itself have been authenticated previously, for instance, and most file servers such as dossrv, ftpfs and dbfs are invoked to provide services to an already-authenticated user, and therefore do not require authentication within a session.) If the server does require authentication, afd refers to a file descriptor returned by a previous fauth on connection fd, on which an authentication protocol has subsequently been executed as required by the file server connected to fd.

7.6. Sys: other new system calls

There are two more new system calls:

fd2path: fn(fd: ref Sys->FD): string;

werrstr: fn(s: string): int;

Fd2path returns the path name under which the file descriptor fd was originally opened (if known). One result is that workdir(2) produces reasonable results for the name of the current directory in the presence of mounts and binds.

Werrstr sets the per-process system error string to s, to allow a Limbo function to save and restore an error string over other system calls, to present a similar interface as the system calls on errors, or to annotate the error from a system call for its own caller.

7.7. Sys: directory reading

The sys-dirread(2) system call’s signature has changed:

dirread: fn(fd: ref Sys->FD): (int, array of Sys->Dir);

Previously it accepted an array of Dir to fill and returned a count; now it returns a tuple containing the count and the array of values read. The change was needed because the representation of directory entries is now variable length, and it is difficult to limit the number returned (it is possible, but all the methods have disadvantages). Dirread still reads a directory incrementally, requesting a block of directory entries of reasonable size from the file server, and unpacking them into the returned array. Use readdir(2) to read whole directories at once.

8. Bufio

There are several changes to Bufio:

Iobuf: adt {


    seek:   fn(b: self ref Iobuf, n: big, where: int): big;

    offset: fn(b: self ref Iobuf): big;


# flush: fn();  # deleted

The module-level function Bufio->flush has been removed (not Iobuf.flush), to allow concurrent use of a single Bufio instance; applications must close or flush each output file explicitly.

As a result of the change to 64-bit offsets for Sys->seek, Iobuf.seek also accepts and returns big offsets. Iobuf.offset is new, and returns the current file offset in bytes, taking account of any buffering.

Iobuf.flush has been extended to flush any data buffered on input files.

9. Draw

The graphics model represented by the draw(3) device and the Draw module is significantly different, including support for a range of pixel formats, and compositing in the drawing operations. Most source code that uses Images directly will require some changes, but the scope of them is limited: needing only extra or different parameter values to individual operations, not radical restructuring. The following changes affect most non-Tk graphics application code:

∙ Pixels in an Image can now be more than 8 bits and have a more flexible structure (eg, several colour channels, and an optional alpha channel, of up to 8 bits each). To support that, the old ldepth field has gone, replaced by a channel descriptor chans of type Chans, which describes the pixel structure, and an integer depth field, which gives the total pixel size (depth) in bits.

∙ The colour parameters are now 32-bit RGBA values (red, green, blue and alpha components, 8-bit each, and big-endian only when an int).

∙ The graphics subsystem supports Porter-Duff compositing, combining a destination image with a source image (within an optional matte) according to a compositing operator. The interpretation of the old ‘mask’ Image parameter to draw and gendraw has changed. Previously it provided a simple binary mask; it now provides a ‘matte’, and its alpha channel shapes the source image and adds partial transparencies. If the matte parameter is nil, the source image is used unmodified. If it lacks an alpha channel, one is computed from the matte image colour channels. The drawing operations draw, gendraw, line, text, and so on, have all got variants drawop, gendrawop, lineop, textop, and so on, each taking an extra final parameter that specifies a Porter-Duff compositing operator from a set predefined by Draw: SoverD, SinD, DatopS, and so on. In each case, S refers to the source image (within a matte, if provided), and D refers to the destination image. Most of them are useful only when either or both source or destination images have got alpha channels (or a matte is used to shape the source). The old function names without the op suffix use the most common compositing operation Draw->SoverD, drawing the source image over the destination, taking account of the shaping of the source and destination images by their alpha channels, with the source further shaped by the optional matte. Thus Image.draw continues to do the ‘obvious’ thing.

∙ There are new colour map conversion functions.

The Chans adt is the following:

Chans: adt


    # interpret standard channel string

    mk:    fn(s: string): Chans;

    # standard printable form

    text:  fn(c: self Chans): string;

    # equality

    eq:    fn(c: self Chans, d: Chans): int;

    # bits per pixel

    depth: fn(c: self Chans): int;


Values are created by Chans.mk, which accepts a string that is a sequence of channel descriptors, each being a letter representing a channel type followed by an integer giving the channel’s size (depth, width) in bits. The letters include: r, g and b for red, green and blue; a for alpha; k (!) for greyscale; and x for padding (‘‘unspecified’’, ‘‘don’t care’’). Thus Chans.mk("r8g8b8a8") produces a descriptor for a 32-bit pixel with 8-bit colour and alpha components. The same descriptor is used in the revised image(6) format, although the older image file format with ldepth only is still recognised. Given a Chans value c, c.text() returns such a descriptor for it as a string.

When newimage previously was called with a specific value for ldepth, an appropriate Chans value must replace it. A few common variants are defined as constants of type Chans in Draw. (We extended the Limbo compiler last year to support the use of con with adt and tuple constants with this use in mind.) For example, the value Draw->CMAP8 is the descriptor for the 8-bit deep rgbv colour-mapped Image format previously used by Inferno. The list of predefined channels includes:

The use of Chans instead of ldepth means that calls to Display.newimage must be changed. For instance:


buffer := display.newimage(r.inset(3), t.image.ldepth, 0, Draw->White);



buffer := display.newimage(r.inset(3), t.image.chans, 0, Draw->White);

There is an obvious difference: the use of t.image.chans instead of t.image.ldepth to create a buffer Image with the same pixel structure as t. There is, however, another difference. The final colour parameter to newimage is also different in structure: in the new graphics model, it is a 32-bit integer value giving RGBA components, not a colour map index, and the name Draw->White has the value 16rFFFFFFFF not 0. Because a symbolic name was used, however, the source need not change. As another example, Draw->Palegreyblue is int 16r4993DDFF. Note the final FF for the alpha component (creating a fully opaque colour). When the top bit is set, the int cast shown here is needed to force the otherwise big value to 32 bits.

The values of colour components are now uniformly expressed as intensity, so that a pixel with all zero colour components is black and one with all colour components at maximum (all ones, full intensity) is white. The rgbv map has therefore been reversed. Given a map index, Display.cmap2rgba returns the 32-bit RGBA format used as a parameter in other calls. All colour components are linear values, as required for compositing to work properly; gamma correction is done as required by the display subsystem.

The colour components of a pixel with an alpha component are always pre-multiplied by the alpha value, following Porter and Duff, as further justified by Alvy Ray Smith and Jim Blinn. ‘‘Thus a 50% red is 16r7F00007F not 16rFF00007F.’’ The function Draw->setalpha does the computation.

Because of the changes to colours and the replacement of simple masks by mattes, the Images Display.ones and Display.zeros are no longer defined. Instead, when they were intended to represent colours, the new Images Display.black and Display.white provide the obvious colours. When ones and zeros were used as masks, the new predefined Images Display.opaque and Display.transparent are used instead as constant mattes, with alpha channels (fully opaque and fully transparent, respectively). As noted above, where Display.ones was used as a mask parameter in drawing operations, one can simply specify a nil Image as a matte (‘no matte’) instead. (That has been allowed for quite some time and is in use but might not be widely known.)

For example, Charon allocated a mask using:

dpicmask = display.newimage(pic.r, 0, 0, Draw->White);

which becomes

dpicmask = display.newimage(pic.r, Draw->GREY1, 0, Draw->Opaque);

where GREY1 is a constant value of the Chans adt type, predefined by Draw, for Images that have a single 1-bit deep grey channel (ie, a bitmap). (Note that to form a fully-opaque matte, Draw->Opaque was used for clarity, not Draw->White; Draw->Transparent could also be used, as the basis for building a matte with transparency.)

A small if obscure change is that Display.newwindow has a new parameter:

newwindow:  fn(screen: self ref Screen, r: Rect,

    backing: int, color: int): ref Image;

The backing parameter should usually be Draw->Refbackup, except for windows allocated on an image that already has got backing store assigned, for instance because it is an image on a screen on an existing window image, in which case it should be Draw->Refnone, because the parent window already provides the backing.

As a small but helpful change, the adt Draw->Pointer has a new element msec that reports a relative time stamp in milliseconds.

The Draw->Context content is significantly different, for the benefit of the new window system implementation.

10. Tk module

There is a new function in Tk:

quote:   fn(s: string): string;

Quote returns string s quoted according to Tk’s ‘{}’ quoting conventions. It replaces Wmlib->tkquote.

There is a new widget type: panel(9). A panel instance can be packed and otherwise manipulated in the same way as any other Tk widget. An image is associated with it by calling Tk->putimage defined in tk(2). The associated images can be drawn on directly by the application, using all the operations provided by Draw. The coordinates of the changed rectangle must be given to Tk using the panel widget command dirty; that part of the image will be redrawn if necessary at the next Tk update. A panel has no default bindings. See panel(9) for details.

For example, wm/coffee now uses the following:

r := Rect((0, 0), (400, 300));

buffer := display.newimage(r, t.image.chans, 0, Draw->Black);

tk->cmd(t, "panel .f.p -bd 3 -relief flat");

tk->cmd(t, "pack .f.p -fill both -expand 1");

tk->cmd(t, "update");

org := buffer.r.min;

tk->putimage(t, ".f.p", buffer, nil);

When it has updated the buffer, it tells Tk:

tk->cmd(t, ".f.p dirty; update");

In this case the whole image is marked dirty, but dirty can be given an optional rectangle parameter to restrict redrawing.

Tk->putimage and Tk->getimage replace imageput and imageget.

11. Selectfile, Tabs and Dialog

The functions filename, mktabs (and tabsctl), dialog and getstring have been moved to separate new modules, to allow those aspects of the user interface to be changed by replacing the implementations, and to allow standard modules to be provided for picking colours (for instance). Selectfile acquires filename, Tabs acquires the ‘tabs’ Tk pseudo-widget, and Dialog acquires dialog, which is renamed prompt, and getstring. In cases where the functions took a Tk->Toplevel as a parameter to specify a parent window, they now take a Draw->Context and (parent) Image parameter; given a Toplevel t, use t.image. See dialog(2), selectfile(2) and tabs(2).

Appendix A: Tk client conversion

Wm(1) applications now have to feed their own pointer and keyboard input to Tk. The window manager is now kept informed about the placement of windows.

A Tk toplevel now holds a window manager context:

Wmcontext: adt


    kbd:        chan of int;        # incoming characters from keyboard

    ptr:        chan of ref Pointer;    # incoming stream of mouse positions

    ctl:        chan of string;     # commands from wm to application

    wctl:       chan of string;     # commands from application to wm

    images: chan of ref Image;  # exchange of images

    connfd: ref Sys->FD;        # connection control

    ctxt:       ref Context;


It contains some channels on which the window manager sends information to the application, and a file descriptor that can be used to write requests to the window manager. The channels used directly by the application are:

kbd characters typed by the user (pass them to Tk->pointer)

ptr pointer events (pass them to Tk->keyboard)

ctl application control requests. Passing these to Tkclient->wmctl will do the default action. Requests starting with an exclamation mark (!) can cause the application’s image to change.

The toplevel also holds a channel wreq on which it sends application control requests; these have the same form as those sent on Wmcontext.ctl, and can be forwarded to Tkclient->wmctl in the same way.

Control requests currently understood by wm(1) are:

!reshape tag reqid minx miny maxx maxy [how]

Reshape the window referenced by tag, creating a new image if tag did not previously exist. Reqid is ignored. How can be one of:

place Wm attempts to find a suitable patch of screen real estate on which to place the window; the size of the given rectangle is taken to be the minimum size for that window.

exact Reshape to the exact rectangle requested. This is the default if how is not given.

onscreen The given rectangle is adjusted so that it is no bigger than the available screen space, and is entirely on screen.

delete tag

Delete the image associated with tag.


Raise the window


Lower the window

!move tag reqid startx starty

Request the user to move the window to a new place. Startx and starty are the coordinates of the pointer when the request was initiated.

!size tag

Request the user to resize the window.

To convert a typical Tk application, do the following.

1. Use an editor to make the following changes:

2. Insert the following code at the top of the central alt statement. The names ‘wmctl’ and ’top‘ will need changing to the appropriate variables in the program:

s := <-top.ctxt.kbd =>

    tk->keyboard(top, s);

s := <-top.ctxt.ptr =>

    tk->pointer(top, *s);

s := <-top.ctxt.ctl or

s = <-top.wreq or

s = <-wmctl =>

    tkclient->wmctl(top, s);

3. Add the following just after the Tk configuration code and before the main processing starts:

tkclient->onscreen(top, nil);

tkclient->startinput(top, "kbd"::"ptr"::nil);

This is possibly the easiest part to forget.

Be careful of cases where a blocking function is called from the main loop that relies on keyboard/mouse input. The easiest solution can be to spawn a thread to handle the keyboard and mouse independently.