Shared Libraries and shared_preload_libraries

A PostgreSQL extension is a shared library that PostgreSQL loads at runtime. There are two loading paths:

• On-demand loading -- the library is loaded the first time one of its functions is called. This is the default for most extensions.
• Preloading -- the library is loaded when the postmaster starts up, by listing it in shared_preload_libraries in postgresql.conf. Preloading is mandatory for any extension that installs a global hook or registers a background worker, because those happen during postmaster initialisation and cannot be retro-fitted.

ProvSQL must be preloaded. Its _PG_init function installs a planner hook and registers the mmap background worker (see memory); both require running before any backend has started executing queries. Forgetting to add provsql to shared_preload_libraries is the single most common installation mistake.

Hooks

A hook is a global function pointer that PostgreSQL exposes at a well-defined point in its query-processing pipeline. An extension overwrites the pointer in _PG_init, chains to the previous value (if any), and runs its own logic before delegating to the default implementation.

ProvSQL uses one hook:

• planner_hook -- called for every query just before planning. ProvSQL installs provsql_planner, which inspects the Query tree and rewrites it to track provenance.

The pattern is always the same:

prev_planner = planner_hook;
planner_hook = my_planner;

...

PlannedStmt *my_planner(Query *q, ...) {
  // do work
  return prev_planner ? prev_planner(q, ...) : standard_planner(q, ...);
}

Other extensions may install their own planner_hook after ProvSQL, which is why chaining via prev_planner matters.

The Query Node Tree

PostgreSQL represents a parsed-and-analysed SQL query as a Query node defined in nodes/parsenodes.h. The fields ProvSQL touches most are:

• commandType -- CMD_SELECT, CMD_INSERT, etc.
• rtable -- the range table, a list of RangeTblEntry (RTE) nodes describing every relation referenced in the query (base tables, subqueries, joins, function calls, VALUES lists...).
• targetList -- the SELECT list, a list of TargetEntry nodes wrapping Expr trees.
• jointree -- the FROM and WHERE clauses encoded as a tree of FromExpr and JoinExpr nodes.
• groupClause -- the GROUP BY clause.
• havingQual -- the HAVING predicate.
• sortClause, distinctClause, setOperations... -- the rest of the SELECT clause.
• hasAggs, hasSubLinks, hasDistinctOn -- boolean flags used by the planner to skip work.

A Var node references a column by (varno, varattno): a 1-based index into rtable and a 1-based attribute number within that RTE. Most of the rewriting in query-rewriting is about inserting, deleting, and renumbering Var nodes correctly.

Extension code rarely constructs Query trees by hand; instead it uses the helper constructors in nodes/makefuncs.h (makeVar, makeNode...).

Background Workers

A background worker is a separate server process, distinct from any client backend. Workers are described by a BackgroundWorker struct that the extension fills in and passes to RegisterBackgroundWorker from _PG_init (or to RegisterDynamicBackgroundWorker from a running backend, for on-demand workers). The fields that matter most:

• bgw_library_name -- the name of the extension shared object.
• bgw_function_name -- the name of the C function to invoke as the worker's entry point.
• bgw_start_time -- when the postmaster should launch the worker (BgWorkerStart_PostmasterStart, BgWorkerStart_ConsistentState, or BgWorkerStart_RecoveryFinished). ProvSQL's mmap worker uses BgWorkerStart_PostmasterStart, so it is up before any client backend connects.

When the postmaster decides to start the worker, it dynamically loads bgw_library_name and looks up bgw_function_name by name -- effectively a dlsym() call. The entry function must therefore have default ELF visibility; on some build configurations, -fvisibility=hidden is in effect and a plain void f() declaration is not exported. ProvSQL works around this by declaring its background worker entry point with PGDLLEXPORT, which expands to __attribute__((visibility("default"))) on GCC/Clang. See provsql_mmap_worker and RegisterProvSQLMMapWorker.

Workers communicate with normal backends through PostgreSQL shared memory and (in ProvSQL's case) anonymous pipes -- see memory.

Memory Management: palloc and Memory Contexts

PostgreSQL replaces malloc / free with a memory context system: palloc(size) allocates from the current context, and contexts can be reset or destroyed wholesale. Most allocations happen in a per-query or per-transaction context that PostgreSQL frees automatically when the query (or transaction) ends, so a lot of extension code never explicitly frees anything.

C++ code that allocates with new or stores objects in STL containers does not participate in this scheme. ProvSQL's C++ side uses ordinary heap allocation; the boundary between the two worlds is the C Datum interface (see below).

Error Reporting: ereport / elog

PostgreSQL extensions report errors with the ereport() and elog() macros. ereport is the modern, preferred form for user-facing diagnostics (it lets you attach an SQL error code via errcode(), a primary message via errmsg(), optional errdetail() / errhint(), etc.); elog is the older single-argument shortcut and is best reserved for internal / debug messages. At ERROR level (or higher) both abort execution of the current query and do not return to the caller; at WARNING, NOTICE, or LOG they emit a message and return normally.

ProvSQL wraps both forms in convenience macros declared in provsql_error.h:

• provsql_error -- aborts with "ProvSQL: ..." prefix.
• provsql_warning, provsql_notice, provsql_log -- non-aborting levels.

Because provsql_error ultimately raises an ERROR, control never returns from it. Inside C++ code that needs to release resources on the way out, use exceptions and let the SQL-callable wrapper catch them and re-raise as a PostgreSQL ERROR. Throwing exceptions across the C/C++ boundary is undefined behaviour, so the catch must happen in the C++ side.

SQL-Callable C Functions

PostgreSQL exposes C functions to SQL through a fixed calling convention: every function takes a FunctionCallInfo (which holds the arguments) and returns a Datum. A function is registered for SQL via the PG_FUNCTION_INFO_V1 macro:

PG_FUNCTION_INFO_V1(my_function);

Datum my_function(PG_FUNCTION_ARGS) {
  int32 arg = PG_GETARG_INT32(0);
  // ...
  PG_RETURN_INT32(result);
}

A Datum is a fixed-width opaque value that can hold any PostgreSQL type, with conversion macros (PG_GETARG_*, PG_RETURN_*, DatumGet*, *GetDatum) on each side. The CREATE FUNCTION SQL statement then declares the function with LANGUAGE C and the appropriate argument and return types.

The PG_FUNCTION_INFO_V1 macro emits a PGDLLEXPORT info struct so the function is reachable through dlsym(). This is why most ProvSQL SQL-callable C functions do not need an explicit PGDLLEXPORT -- the macro takes care of it.

OIDs and Catalog Lookups

Almost everything in PostgreSQL has an Object Identifier (OID): tables, columns, types, functions, operators, schemas... When ProvSQL builds expression trees by hand, it must reference its own types and functions by OID, not by name. The OIDs are not stable across installations, so ProvSQL caches them at first use. See constants_t and get_constants (covered in architecture).

GUCs (Configuration Parameters)

PostgreSQL exposes server- and session-scoped settings as Grand Unified Configuration (GUC) variables, registered via DefineCustom*Variable from _PG_init. ProvSQL exposes four:

• provsql.active -- master switch.
• provsql.where_provenance -- enable where-provenance tracking (see where-provenance).
• provsql.update_provenance -- enable data-modification tracking (see data-modification).
• provsql.verbose_level -- diagnostic verbosity (see debugging).

GUCs can be set in postgresql.conf, with ALTER SYSTEM, per-session with SET, or per-transaction with SET LOCAL.