mars-nwe/REDESIGN.md

mars-nwe NCP dispatch redesign notes

This file collects design notes for a possible cleanup of the internal NCP handoff path. It is intentionally separate from TODO.md: the TODO file should track concrete bugs and endpoint audit follow-ups, while this file describes a larger architecture direction that can be implemented gradually.

The goal is not to rewrite MARS-NWE at once. The goal is to make the current handoff behavior explicit, reduce ambiguity around magic return values, and make future endpoint work easier to audit against the Novell/Micro Focus SDK, WebSDK, and NDK Core Protocols PDF.

Current problem

The current NCP path grew around several cooperating processes and handlers:

• nwconn.c owns the connection/session side and receives most packets first.
• nwbind.c handles bindery, queue, some server-management, and some final reply construction.
• Other modules such as semaphore, message, namespace, AFP, file, salvage, and queue code implement individual protocol families or backend actions.
• Some calls are handled completely in nwconn.c.
• Some calls are forwarded to nwbind.c by returning -1 from the nwconn.c dispatcher.
• Some calls are forwarded with saved request state by returning -2, so that nwconn.c can do post-processing after nwbind.c has replied.
• Some forwarded paths mutate request payloads before handoff.
• Some code paths build responses locally, while other paths rely on the target process to build the final completion code and payload.

This works, but it is hard to reason about while auditing endpoint layouts. The same looking value can mean different things depending on which file it appears in. For example, return(-1) in the relevant nwconn.c dispatcher path means "forward this request to nwbind". A disabled return(-1) inside a #if 0 block in nwbind.c does not have that forwarding meaning and should not be copied into active code.

The visible symptoms are:

• endpoint documentation must follow a handoff across files before it can say the request or reply layout is known;
• missing endpoints are difficult to distinguish from forwarded endpoints;
• request parsing, backend behavior, reply encoding, and process routing are often mixed in one switch block;
• byte order differences are easy to miss because parsing and reply writing are open-coded in different places;
• disabled future stubs can look like active dispatch behavior;
• TODO.md can become a dumping ground for architectural observations that are not immediate endpoint bugs.

Desired shape

A cleaner long-term structure would have one small internal NCP dispatch layer:

wire packet
  -> NCP envelope parser
  -> NcpContext
  -> endpoint lookup
  -> endpoint handler / provider
  -> reply encoder
  -> central reply sender

This does not need to be a general-purpose message bus. A full message bus would probably be too large and too abstract for this code base. A typed internal NCP context plus explicit dispatch results would be enough.

The important separation is:

1. decode the packet envelope;
2. identify the endpoint;
3. decode the endpoint request body;
4. execute the backend operation;
5. encode the endpoint reply body;
6. send the response from one well-defined place.

Proposed NCP context

Introduce, in a later functional cleanup, a small context object that represents one NCP request while it moves through the server. The exact field names should fit the existing code style, but the conceptual shape would be:

typedef struct {
  int connection;

  uint16_t request_type;        /* 0x2222, 0x3333, 0x5555, ... */
  uint8_t function;             /* top-level NCP function */
  int has_subfunction;
  uint8_t subfunction;          /* grouped calls such as 23/x, 32/x, 87/x */

  const uint8_t *request;
  int request_len;

  uint8_t *reply;
  int reply_cap;
  int reply_len;

  uint8_t completion;
  uint8_t connection_status;

  uint32_t flags;
} NcpContext;

The context should not replace all old globals in one patch. It can start as a thin wrapper around the existing request and response buffers, then gradually become the preferred handler interface.

The useful property is that endpoint documentation can point to a stable model:

• function and subfunction identify the endpoint;
• request and request_len are the bytes after the already-decoded envelope;
• reply and reply_len are the bytes before the common NCP response envelope;
• completion is set once by the handler or by central error handling.

Replace magic return values with named results

The current 0, -1, and -2 convention should be made explicit before any larger refactor. The first step can be documentation-only or macro-only:

#define NCP_LOCAL_DONE             0
#define NCP_FORWARD_NWBIND        -1
#define NCP_FORWARD_NWBIND_POST   -2

A later cleanup can replace those with an enum:

typedef enum {
  NCP_DISPATCH_DONE,
  NCP_DISPATCH_FORWARD_BIND,
  NCP_DISPATCH_FORWARD_BIND_POST,
  NCP_DISPATCH_NOT_IMPLEMENTED,
  NCP_DISPATCH_BAD_REQUEST,
  NCP_DISPATCH_INTERNAL_ERROR
} NcpDispatchResult;

The important rule is that the meaning must be scoped. A named result returned from a nwconn.c dispatcher may request process handoff. A return statement in nwbind.c should not silently inherit that meaning unless the function is explicitly part of the same dispatch interface.

Endpoint table as audit index first

Before replacing switch statements, add an endpoint inventory table as a non-invasive audit aid. It can be compiled only for debug builds or kept as a source-level documentation table.

Conceptual form:

typedef struct {
  uint16_t request_type;
  uint8_t function;
  int has_subfunction;
  uint8_t subfunction;
  const char *name;
  const char *provider;
  uint32_t flags;
} NcpEndpointDoc;

Example entries:

{ 0x2222, 23, 1, 109, "Change Queue Job Entry old", "nwbind/queue", NCPDOC_FORWARDED },
{ 0x2222, 32, 1,   0, "Open Semaphore old",         "sema",         NCPDOC_LOCAL },
{ 0x2222, 33, 0,   0, "Negotiate Buffer Size",      "nwconn",       NCPDOC_LOCAL },

This table would help with the ongoing endpoint audit:

• SDK/PDF/WebSDK listed and implemented;
• SDK/PDF/WebSDK listed and forwarded;
• SDK/PDF/WebSDK listed but disabled as a future stub;
• SDK/PDF/WebSDK listed but absent from the current compatibility target;
• later NetWare 4.x/OES/MOAB endpoint, not part of the default NetWare 3.x compatibility target.

The first version should not drive runtime dispatch. It should only make review and missing-endpoint checks less error-prone.

Handler structure

For newly touched endpoint families, prefer the following logical split even if it remains in one C function at first:

request decode
  -> validation
  -> backend operation
  -> reply encode

For complex endpoints this could become explicit helper functions:

static int decode_foo(NcpContext *ctx, FooRequest *out);
static int exec_foo(NcpContext *ctx, const FooRequest *req, FooReply *reply);
static void encode_foo(NcpContext *ctx, const FooReply *reply);

This is especially useful for endpoint families where the audit has already found old/new layout differences:

• 16-bit old queue job numbers versus newer 32-bit job numbers;
• big-endian versus little-endian SDK notation;
• old short replies versus newer long replies;
• connection-side prehandling that inserts or rewrites fields;
• bindery or queue paths that build final replies in a different process.

Small endpoints do not need three separate helper functions if that would make the code noisier. The rule is that request bytes and reply bytes should be easy to identify and compare with the SDK documents.

Make handoff explicit

Forwarded calls should say exactly what is handed off. A good comment should answer:

• which bytes are forwarded;
• whether the subfunction byte is preserved or stripped;
• whether nwconn.c mutates the request before forwarding;
• whether nwbind.c or another provider builds the final reply;
• whether nwconn.c expects post-processing after the provider reply.

Examples of handoff cases that need this clarity:

• Queue calls where nwconn.c expands paths or inserts job file handles before nwbind.c sees the request.
• Quota/bindery prehandling where the destination handler receives an already transformed request.
• Semaphore and message groups that are grouped in the SDK but routed through local helper modules.
• Direct lifecycle calls such as End Of Job and Logout where local cleanup and final success reply are split across files.

The preferred future style is not "nwbind must do the rest" but something like:

Forward to nwbind with the original subfunction byte and payload unchanged.
No nwconn post-processing is expected; nwbind builds the completion-only reply.

or:

Forward to nwbind after saving the original request.  nwbind validates bindery
state and returns the bindery result; nwconn then performs the file-handle
post-processing in handle_after_bind().

Response building rule

Every endpoint audit should identify the reply builder, not only the request parser. A handler is not fully documented until the response path is known.

For each endpoint family, record:

• completion-only reply;
• fixed-size payload reply;
• variable-length payload reply;
• provider-built reply;
• nwconn.c post-processed reply;
• intentionally unsupported reply status.

Long-term, response sending should become centralized enough that endpoint code only encodes payload bytes and a completion code. This reduces off-by-one reply length bugs and makes the logs easier to normalize.

Provider boundaries

A clean design would treat the existing modules as providers instead of hidden fallback paths:

nwconn       connection/session, packet IO, top-level envelope
ncpdispatch  endpoint lookup, handoff policy, common errors
nwbind       bindery database and bindery-backed services
queue        queue metadata and print/backend adapter
sema         semaphore state
message      station/message/broadcast state
namespace    path, directory handle, name-space operations
file         file handle and read/write/open/close operations
salvage      deleted-file scan/recover/purge backend
AFP          AFP metadata and AFP namespace adapter

This is a design target, not a demand to move files immediately. The important part is that future code should avoid making nwbind a catch-all sink for unrelated NCPs just because it already has an IPC path.

Provider boundary versus process boundary

A provider boundary is not the same thing as a Unix process boundary. This is an important distinction because splitting every NCP family into a separate process would make the server harder to debug and could introduce new ordering, locking, and reply-ownership bugs.

The preferred rule is:

first define logical providers;
only later promote the few large stateful providers to separate processes.

A logical provider can start as an ordinary C module called from the existing process path. It becomes valuable as soon as the dispatch table can say "this endpoint belongs to the queue provider" or "this endpoint belongs to the connection-local provider", even if no new process exists yet. A process split should be treated as an implementation detail that is only justified when the provider has enough independent state and lifecycle to benefit from isolation.

This keeps the redesign incremental:

now:
  nwconn switch -> existing local code or nwbind handoff

first cleanup:
  nwconn switch -> provider-named helper/module

later, only where useful:
  nwconn/dispatcher -> IPC -> provider process

Good process candidates

Bindery

Bindery is already a natural service boundary. It owns long-lived server state: objects, properties, sets, security, password/login/key handling, and object lookup. Keeping bindery behind a clear provider boundary is appropriate, and the existing nwbind process can remain that boundary while the dispatch layer is cleaned up.

The main cleanup is not to remove nwbind, but to stop treating it as a generic catch-all for unrelated forwarded requests. A future endpoint table should mark true bindery calls as bindery, and queue or management calls should not be classified as bindery merely because their current implementation lives in nwbind.c.

Queue / possible nwqueue

Queue management is the strongest candidate for a future separate process after bindery. Queue handling has its own domain state:

• queue objects and queue metadata;
• queue job lifecycle;
• queue server attach/detach state;
• service, finish, and abort state;
• job position and priority;
• client-rights transitions during job servicing;
• queue directories and spool/job files.

That is large enough to deserve a logical queue provider even before any runtime split. A future nwqueue process can be considered once request/reply ownership and bindery access are explicit.

The first step should only be a provider split:

0x2222/23 queue subfunctions -> queue provider
queue provider -> bindery provider/library for object/security/property checks
queue provider -> file/path helpers for queue job files

A real nwqueue process should not be created by simply moving the current queue cases out of nwbind.c. It needs an explicit contract for:

• which process owns the final NCP reply;
• how queue calls read bindery objects and properties;
• how queue job files are opened and handed back to the connection process;
• how connection cleanup affects attached queue servers and in-service jobs;
• how old 16-bit job-number calls and newer 32-bit job-number calls are kept compatible.

Until those contracts are clear, nwqueue should remain a design target, not an immediate functional change.

Possible but risky process candidates

File and volume subsystem

The file/volume/name-space area is large and stateful, so it can look like a candidate for a separate process. It owns or touches directory handles, file handles, locks, trustee evaluation, volume information, name spaces, salvage and purge operations, and Unix filesystem mapping.

However, this area is also tightly coupled to connection state and existing file descriptor ownership. Moving it behind IPC too early could create more problems than it solves. The safer path is:

first:  file/volume/name-space provider modules inside the current process model
later:  consider a process split only after handle ownership is explicit

A file provider boundary is useful for documentation and dispatch cleanup. A separate file process is optional and should be considered high-risk.

Accounting

Accounting is a maybe. It has a separate protocol domain, but in many setups it may be small enough to stay as an in-process provider. A process boundary only makes sense if accounting grows into a real persistent service with charges, holds, notes, audit records, and recovery behavior that should be isolated from connection handlers.

Poor process candidates

Semaphore

Semaphore calls should have a clean provider boundary, but a dedicated process is probably overkill. The old semaphore group is small: open, examine, wait, signal, and close. It needs shared state, but not necessarily a standalone process. A sema provider module with clear request/reply ownership should be enough unless later testing shows that cross-connection semaphore state cannot be managed safely in the existing process model.

Connection lifecycle and session-local calls

Connection lifecycle operations should stay with nwconn or a connection-local provider. Calls such as Logout, End Of Job, watchdog handling, buffer negotiation, and connection-state cleanup are fundamentally tied to the session that received the packet. Moving them into another process would make cleanup ordering and error handling harder.

Simple server-management calls

Simple management and information calls should not become their own process. Examples include login-status queries, server description strings, server time, console-privilege checks, and small broadcast/control helpers. These can be represented as a servermgmt provider for dispatch clarity, but they should stay in-process unless a specific call requires an existing backend service.

Suggested provider map

The endpoint audit table should be able to use provider names like these:

local          packet/session-local handling in nwconn
bindery        object/property/security/login backend
queue          queue objects, jobs, queue servers, spool/job lifecycle
filesystem     file, directory, volume, namespace, trustee, salvage helpers
semaphore      semaphore state and old 0x2222/32 calls
message        station messaging and broadcast helpers
servermgmt     small server-management and information calls
accounting     account status, charges, holds, notes
AFP            AFP namespace and metadata helpers
unknown        documented but not yet mapped

Only some providers should ever become processes:

already process-like:  bindery / nwbind
likely future process: queue / possible nwqueue
maybe, high risk:     filesystem
usually in-process:   semaphore, message, servermgmt, accounting, AFP helpers

The practical design rule is:

Use provider names everywhere in documentation and endpoint tables.
Use new processes only where shared state, isolation, and lifecycle justify the
extra IPC complexity.

Logging connection

The dispatch redesign also supports the desired log cleanup. If every request has a context, logs can consistently include:

INFO NCP 23/109 DISPATCH type=0x2222 fn=0x17 sub=0x6d provider=nwbind/queue
INFO NCP 32/0 REPLY type=0x2222 fn=0x20 sub=0x00 result=0x00 len=4
WARN NCP 23/130 LAYOUT-MISMATCH sdk="32-bit JobNumber" code="16-bit parser"

The logging cleanup should still reuse existing mars-nwe logging functions. Do not add a second logging subsystem just to support the dispatch cleanup.

Migration plan

Phase 1: Name the existing conventions

Low risk. No behavior change.

• Add named constants or comments for the current 0, -1, and -2 dispatcher results.
• Keep existing control flow unchanged.
• Update comments so return(-1) is never described ambiguously outside the exact dispatcher where it is meaningful.

Phase 2: Add an endpoint audit table

Low risk. Mostly documentation/debug.

• Add a table of known endpoints by request type, function, and subfunction.
• Mark provider, generation bucket, and implementation state.
• Use it to compare SDK/PDF/WebSDK coverage against actual handlers.
• Do not switch runtime dispatch to the table yet.

Phase 3: Introduce a thin NcpContext

Moderate risk if kept small.

• Wrap existing request and reply buffers without changing ownership.
• Use the context only in newly audited or newly implemented handlers.
• Keep old handlers callable until they are touched for another reason.

Phase 4: Convert small endpoint families first

Moderate risk, easy to test.

Good candidates:

• 0x2222/32 old Semaphore calls;
• direct calls such as End Of Job, Logout, and Negotiate Buffer Size;
• small message/station groups once their handoff has been audited.

Avoid converting queue and bindery first because they have more process coupling and more old/new layout variants.

Phase 5: Move runtime dispatch to tables gradually

Higher risk. Do this only after enough endpoint families have stable audit coverage and tests.

• Keep switch wrappers during the transition.
• Convert one family at a time.
• Preserve exact completion codes and reply lengths.
• Add targeted smoke tests for any family whose dispatch path changes.

Non-goals

This redesign should not:

• change protocol behavior merely to match a cleaner abstraction;
• remove NetWare 1.x/2.x/3.x compatibility paths;
• enable NetWare 4.x/OES/MOAB-only endpoints by default;
• replace existing mars-nwe path, bindery, queue, AFP, trustee, or salvage backends with parallel databases;
• add a large external message bus dependency;
• rewrite all handlers in one patch;
• turn documentation-only endpoint audit patches into functional refactors.

Practical rule for future patches

For the ongoing endpoint documentation pass, keep doing the conservative thing:

1. enumerate SDK/PDF/WebSDK/include endpoints for the family;
2. compare them with actual case labels and forwarded destination handlers;
3. document missing, disabled, implemented, and later-generation slots;
4. document request parser/handoff and response builder;
5. record real layout differences, but do not change behavior in the same patch.

Functional cleanup should come later in small patches with tests.