JSON API: Hacker's Guide

(⬑JSON API Index)

Jump to:

This section will only be of interest to those wanting to work on the Fossil/JSON code. That said...

If you happen to hack on the code and find something worth noting here for others, please feel free to expand this section. It will only improve via feedback from those working on the code.

Before Committing Changes...

Because this code lives in the trunk, there are certain guidelines which must be followed before committing any changes:

  1. Read the checkin preparation list.
  2. Changes to the files src/json_*.*, and its related support code (e.g. ajax/*.*), may be made freely without affecting mainline users. Changes to other files, unless they are trivial or made for purposes outside the JSON API (e.g. an unrelated bug fix), must be reviewed carefully before committing. When in doubt, create a branch and post a request for a review.
  3. The Golden Rule is: do not break the trunk build.


libcson, the underlying JSON API, is a separate project, included in fossil in "amalgamation" form: see extsrc/cson_amalgamation.[ch]. It has thorough API docs and a good deal of information is in its wiki:

In particular:

gives an overview of its architecture. Occasionally new versions of it are pulled into the Fossil tree, but other developers generally need not concern themselves with that.

(Trivia: the cson wiki's back-end is fossil using this very JSON API, living on top of a custom JavaScript+HTML5 application.)

Only a small handful of low-level fossil routines actually input or output JSON text (only for reading in POST data and sending the response). In the C code we work with the higher-level JSON value abstractions provided by cson (conceptually similar to an XML DOM). All of the JSON-defined data types are supported, and we can construct JSON output of near arbitrary complexity with the caveat that cyclic data structures are strictly forbidden, and will cause memory corruption, crashes, double free()'s, or other undefined behaviour. Because JSON cannot, without client-specific semantic extensions to JSON, represent cyclic structures, it is not anticipated that this will be a problem/limitation when generating output for fossil.

Architecture of JSON Commands

In order to consolidate CLI/HTTP modes for JSON handling, this code foregoes fossil's conventional command/path dispatching mechanism. Only the top-most "json" command/path is dispatched directly by fossil's core. The disadvantages of this are that we lose fossil's conventional help text mechanism (which is based on code comments in the command/path's dispatcher impl) and the ability to write abbreviated command names in CLI mode ("json" itself may be abbreviated, but not the subcommands). The advantages are that we can handle CLI/HTTP modes almost identically (there are a couple minor differences) by unifying them under the same callback functions much more easily.

The top-level "json" command/path uses its own dispatching mechanism which uses either the path (in HTTP mode) or CLI positional arguments to dispatch commands (stopping at the first "flag option" (e.g. -foo) in CLI mode). The command handlers are simply callback functions which return a cson_value pointer (the C representation of an arbitrary JSON value), representing the "payload" of the response (or NULL - not all responses need a payload). On error these callbacks set the internal JSON error state (detailed in a subsection below) and return NULL. The top-level dispatcher then creates a response envelope and returns the "payload" from the command (if any) to the caller. If a callback sets the error state, the top-level dispatcher takes care to set the error information in the response envelope. In summary:

There is a wrench in all of that, however: the vast majority of fossil's commands "fail fast" - they will exit() if they encounter an error. To handle that, the fossil core error reporting routines have been refactored a small bit to operate differently when we are running in JSON mode. Instead of the conventional output, they generate a JSON error response. In HTTP mode they exit with code 0 to avoid causing an HTTP 500 error, whereas in CLI mode they will exit with a non-0 code. Those routines still exit(), as in the conventional CLI/HTTP modes, but they will exit differently. Because of this, it is perfectly fine for a command handler to exit via one of fossil's conventional mechanisms (e.g. db_prepare() can be fatal, and callbacks may call fossil_panic() if they really want to). One exception is fossil_exit(), which does not generate any extra output and will exit() the app. In the JSON API, as a rule of thumb, fossil_exit() is only used when we want a failed request to cause an HTTP 500 error, and it is reserved for allocation errors and similar truly catostrophic failures. That said... libcson has been hacked to use fossil_alloc() and friends for memory management, and those routines exit on error, so alloc error handling in the JSON command handler code can afford to be a little lax (the majority of potential errors clients get from the cson API have allocation failure as their root cause).

As a side-note: the vast majority (if not all) of the cson API calls are "NULL-safe", meaning that will return an error code (or be a no-op) if passed NULL arguments. e.g. the following chain of calls will not crash if the value we're looking for does not exist, is-not-a String (see cson_value_get_string() for important details), or if myObj is NULL:

const char * str =
 cson_string_cstr( // get the C-string form of a cson_string
   cson_value_get_string( // get its cson_string form
     cson_object_get(myObj,"foo") // search for key in an Object

If "foo" is not found in myObj (or if myObj is NULL) then v will be NULL, as opposed to stepping on a NULL pointer somewhere in that call chain.

Note that all cson JSON values except Arrays and Objects are immutable

In short, after creating a new value you must eventually either add it to a container (Object or Array) to transfer ownership or call cson_value_free() to clean it up (exception: the Fossil/JSON command callbacks return a value to transfer ownership to the dispatcher). Sometimes it's more complex than that, but not normally. Any given value may legally be stored in any number of containers (or multiple times within one container), as long as no cycles are introduced (cycles will cause undefined behaviour). Ownership is shared using reference counting and the value will eventually be freed up when its last remaining reference is freed (e.g. when the last container holding it is cleaned up). For many examples of using cson in the context of fossil, see the existing json_page_xxx() functions in json_*.c.

Reporting Errors

To report an error from a command callback, one abstractly needs to:

That normally looks something like this:

  json_set_err(FSL_JSON_E_DENIED, "Requires 'o' permissions.");
  return NULL;

json_set_err() is a variadic printf-like function, and can use the printf extensions supported by mprintf() and friends (e.g. %Q and %q) (but they are normally not needed in the context of JSON). If the error string is NULL or empty then json_err_cstr(errorCode) is used to fetch the standard/generic error string for the given code.

When control returns to the top-level dispatching function it will check g.json.resultCode and, if it is not 0, create an error response using the g.json.resultCode and g.zErrMsg to construct the response's resultCode and resultText properties.

If a function wants to output an error and exit by itself, as opposed to returning to the dispatcher, then it must behave slightly differently. See the docs for json_err() (in src/json.c) for details, and search that file for various examples of its usage. It is also used by fossil's core error-reporting APIs, e.g. fossil_panic() (defined in src/main.c). That said, it would be "highly unusual" for a callback to need to do this - it is far simpler (and more consistent/reliable) to set the error state and return to the dispatcher.

Getting Command Arguments

Positional parameters can be fetched usinig json_command_arg(N), where N is the argument position, with position 0 being the "json" command/path. In CLI mode positional arguments have their obvious meaning. In HTTP mode the request path (or the "command" request property) is used to build up the "command path" instead. For example:

CLI: fossil json a b c

HTTP: /json/a/b/c

HTTP POST or CLI with --json-input: /json with POSTed envelope {"command": "a/b/c" …}

Those will have identical "command paths," and json_command_path(2) would return the "b" part.

Caveat: a limitation of this support is that all CLI flags must come after all non-flag positional arguments (e.g. file names or subcommand names). Any argument starting with a dash ("-") is considered by this code to be a potential "flag" argument, and all arguments after it are ignored (because the generic handling cannot know if a flag requires an argument, which changes how the rest of the arguments need to be interpreted).

To get named parameters, there are several approaches (plus some special cases). Named parameters can normally come from any of the following sources:

To try to simplify the guessing process the API has a number of functions which behave ever so slightly differently. A summary:

For example code: the existing command callbacks demonstrate all kinds of uses and the various styles of parameter/option inspection. Check out any of the functions named json_page_SOMETHING().

Creating JSON Data

Creating JSON Values

cson has a fairly rich API for creating and manipulating the various JSON-defined value types. For a detailed overview and demonstration i recommend reading:

That said, the Fossil/JSON API has several convenience wrappers to save a few bytes of typing:

It also provides several lower-level JSON features which aren't of general utility but provide necessary functionality for some of the framework-level code (e.g. cson_data_dest_cgi()), which is only used by the deepest of the JSON internals).

Converting SQL Query Results to JSON

The cson_sqlite3_xxx() family of functions convert sqlite3_stmt rows to Arrays or Objects, or convert single columns to a JSON-compatible form. See json_stmt_to_array_of_obj(), json_stmt_to_array_of_array() (both in src/json.c), and cson_sqlite3_column_to_value() and friends (in extsrc/cson_amalgamation.h). They work in an intuitive way for numeric types, but they optimistically/natively assume that any fields of type TEXT or BLOB are actually UTF8 data, and treat them as such. cson's string class only handles UTF8 data and it is semantically illegal to feed them anything but UTF8. Violating this will likely result in down-stream errors (e.g. when emiting the JSON string output). The moral of this story is: do not use these APIs to fetch binary data. JSON doesn't do binary and the cson_string class does not protect itself against clients feeding it non-UTF8 data.

Here's a basic example of using these features:

Stmt q = empty_Stmt;
cson_value * rows = NULL;
db_prepare(&q, "SELECT a AS a, b AS b, c AS c FROM foo");
rows = json_stmt_to_array_of_obj( &sql, NULL );
// side note: if db_prepare()/finalize() fail (==they exit())
// then a JSON-format error reponse will be generated.

On success (and if there were results), rows is now an Array value, each entry of which contains an Object containing the fields (key/value pairs) of each row. json_stmt_to_array_of_array() returns each row as an Array containing the column values (with no column name information).

Note the seemingly superfluous use of the "AS" clause in the above SQL. Having them is actually significant! If a query does not use AS clauses, the row names returned by the db driver might be different than they appear in the query (this is documented behaviour of sqlite3). Because the JSON API needs to return stable field names, we need to use AS clauses to be guaranteed that the db driver will return the column names we want. Note that the AS clause is often used to translate column names into something more JSON-conventional or user-friendly, e.g. "SELECT cap AS capabilities...". Alternately, we can convert the individual sqlite3_stmt column values to JSON using cson_sqlite3_column_to_value(), without refering directly to the db-reported column name.