stephan 4eabec5b18 JNI SQLTester: fix --new command to force-replace db if it already exists. Add no-op impls for --stmt-cache and --jsonglob commands. 5 kuukautta sitten
..
src 4eabec5b18 JNI SQLTester: fix --new command to force-replace db if it already exists. Add no-op impls for --stmt-cache and --jsonglob commands. 5 kuukautta sitten
GNUmakefile 6b36312e8a JNI: add the @Experimental annotation and mark all java.nio.ByteBuffer-related methods with it. 1 vuosi sitten
README.md bd06d8672d JNI: clear out the sqlite3_context native pointer after calling UDF callbacks which do not have an argv (as was already done for those which have an argv). Add related tests and code commentary. 1 vuosi sitten
jar-dist.make bfdc7ab5a7 Add more JNI docs, tests, and a handful of Java-side overloads. 1 vuosi sitten

README.md

SQLite3 via JNI

This directory houses a Java Native Interface (JNI) binding for the sqlite3 API. If you are reading this from the distribution ZIP file, links to resources in the canonical source tree will note work. The canonical copy of this file can be browsed at:

https://sqlite.org/src/doc/trunk/ext/jni/README.md

Technical support is available in the forum:

https://sqlite.org/forum

FOREWARNING: this subproject is very much in development and subject to any number of changes. Please do not rely on any information about its API until this disclaimer is removed. The JNI bindings released with version 3.43 are a "tech preview." Once finalized, strong backward compatibility guarantees will apply.

Project goals/requirements:

  • A 1-to-1(-ish) mapping of the C API to Java via JNI, insofar as cross-language semantics allow for. A closely-related goal is that the C documentation should be usable as-is, insofar as possible, for the JNI binding.

  • Support Java as far back as version 8 (2014).

  • Environment-independent. Should work everywhere both Java and SQLite3 do.

  • No 3rd-party dependencies beyond the JDK. That includes no build-level dependencies for specific IDEs and toolchains. We welcome the addition of build files for arbitrary environments insofar as they neither interfere with each other nor become a maintenance burden for the sqlite developers.

Non-goals:

  • Creation of high-level OO wrapper APIs. Clients are free to create them off of the C-style API.

  • Virtual tables are unlikely to be supported due to the amount of glue code needed to fit them into Java.

  • Support for mixed-mode operation, where client code accesses SQLite both via the Java-side API and the C API via their own native code. Such cases would be a minefield of potential mis-interactions between this project's JNI bindings and mixed-mode client code.

Hello World

import org.sqlite.jni.*;
import static org.sqlite.jni.CApi.*;

...

final sqlite3 db = sqlite3_open(":memory:");
try {
  final int rc = sqlite3_errcode(db);
  if( 0 != rc ){
    if( null != db ){
      System.out.print("Error opening db: "+sqlite3_errmsg(db));
    }else{
      System.out.print("Error opening db: rc="+rc);
    }
    ... handle error ...
  }
  // ... else use the db ...
}finally{
  // ALWAYS close databases using sqlite3_close() or sqlite3_close_v2()
  // when done with them. All of their active statement handles must
  // first have been passed to sqlite3_finalize().
  sqlite3_close_v2(db);
}

Building

The canonical builds assumes a Linux-like environment and requires:

  • GNU Make
  • A JDK supporting Java 8 or higher
  • A modern C compiler. gcc and clang should both work.

Put simply:

$ export JAVA_HOME=/path/to/jdk/root
$ make
$ make test
$ make clean

The jar distribution can be created with make jar, but note that it does not contain the binary DLL file. A different DLL is needed for each target platform.

One-to-One(-ish) Mapping to C

This JNI binding aims to provide as close to a 1-to-1 experience with the C API as cross-language semantics allow. Interface changes are necessarily made where cross-language semantics do not allow a 1-to-1, and judiciously made where a 1-to-1 mapping would be unduly cumbersome to use in Java. In all cases, this binding makes every effort to provide semantics compatible with the C API documentation even if the interface to those semantics is slightly different. Any cases which deviate from those semantics (either removing or adding semantics) are clearly documented.

Where it makes sense to do so for usability, Java-side overloads are provided which accept or return data in alternative forms or provide sensible default argument values. In all such cases they are thin proxies around the corresponding C APIs and do not introduce new semantics.

In a few cases, Java-specific capabilities have been added in new APIs, all of which have "_java" somewhere in their names. Examples include:

  • sqlite3_result_java_object()
  • sqlite3_column_java_object()
  • sqlite3_value_java_object()

which, as one might surmise, collectively enable the passing of arbitrary Java objects from user-defined SQL functions through to the caller.

Golden Rule: Garbage Collection Cannot Free SQLite Resources

It is important that all databases and prepared statement handles get cleaned up by client code. A database cannot be closed if it has open statement handles. sqlite3_close() fails if the db cannot be closed whereas sqlite3_close_v2() recognizes that case and marks the db as a "zombie," pending finalization when the library detects that all pending statements have been closed. Be aware that Java garbage collection cannot close a database or finalize a prepared statement. Those things require explicit API calls.

Classes for which it is sensible support Java's AutoCloseable interface so can be used with try-with-resources constructs.

Golden Rule #2: Never Throw from Callbacks (Unless...)

All routines in this API, barring explicitly documented exceptions, retain C-like semantics. For example, they are not permitted to throw or propagate exceptions and must return error information (if any) via result codes or null. The only cases where the C-style APIs may throw is through client-side misuse, e.g. passing in a null where it may cause a NullPointerException. The APIs clearly mark function parameters which should not be null, but does not generally actively defend itself against such misuse. Some C-style APIs explicitly accept null as a no-op for usability's sake, and some of the JNI APIs deliberately return an error code, instead of segfaulting, when passed a null.

Client-defined callbacks must never throw exceptions unless very explitly documented as being throw-safe. Exceptions are generally reserved for higher-level bindings which are constructed to specifically deal with them and ensure that they do not leak C-level resources. In some cases, callback handlers are permitted to throw, in which cases they get translated to C-level result codes and/or messages. If a callback which is not permitted to throw throws, its exception may trigger debug output but will otherwise be suppressed.

The reason some callbacks are permitted to throw and others not is because all such callbacks act as proxies for C function callback interfaces and some of those interfaces have no error-reporting mechanism. Those which are capable of propagating errors back through the library convert exceptions from callbacks into corresponding C-level error information. Those which cannot propagate errors necessarily suppress any exceptions in order to maintain the C-style semantics of the APIs.

Unwieldy Constructs are Re-mapped

Some constructs, when modelled 1-to-1 from C to Java, are unduly clumsy to work with in Java because they try to shoehorn C's way of doing certain things into Java's wildly different ways. The following subsections cover those, starting with a verbose explanation and demonstration of where such changes are "really necessary"...

Custom Collations

A prime example of where interface changes for Java are necessary for usability is registration of a custom collation:

// C:
int sqlite3_create_collation(sqlite3 * db, const char * name, int eTextRep,
                             void *pUserData,
                             int (*xCompare)(void*,int,void const *,int,void const *));

int sqlite3_create_collation_v2(sqlite3 * db, const char * name, int eTextRep,
                                void *pUserData,
                                int (*xCompare)(void*,int,void const *,int,void const *),
                                void (*xDestroy)(void*));

The pUserData object is optional client-defined state for the xCompare() and/or xDestroy() callback functions, both of which are passed that object as their first argument. That data is passed around "externally" in C because that's how C models the world. If we were to bind that part as-is to Java, the result would be awkward to use (^Yes, we tried this.):

// Java:
int sqlite3_create_collation(sqlite3 db, String name, int eTextRep,
                             Object pUserData, xCompareType xCompare);

int sqlite3_create_collation_v2(sqlite3 db, String name, int eTextRep,
                                Object pUserData,
                                xCompareType xCompare, xDestroyType xDestroy);

The awkwardness comes from (A) having two distinctly different objects for callbacks and (B) having their internal state provided separately, which is ill-fitting in Java. For the sake of usability, C APIs which follow that pattern use a slightly different Java interface:

int sqlite3_create_collation(sqlite3 db, String name, int eTextRep,
                             SomeCallbackType collation);

Where the Collation class has an abstract call() method and no-op xDestroy() method which can be overridden if needed, leading to a much more Java-esque usage:

int rc = sqlite3_create_collation(db, "mycollation", SQLITE_UTF8, new SomeCallbackType(){

  // Required comparison function:
  @Override public int call(byte[] lhs, byte[] rhs){ ... }

  // Optional finalizer function:
  @Override public void xDestroy(){ ... }

  // Optional local state:
  private String localState1 =
    "This is local state. There are many like it, but this one is mine.";
  private MyStateType localState2 = new MyStateType();
  ...
});

Noting that:

  • It is possible to bind in call-scope-local state via closures, if desired, as opposed to packing it into the Collation object.

  • No capabilities of the C API are lost or unduly obscured via the above API reshaping, so power users need not make any compromises.

  • In the specific example above, sqlite3_create_collation_v2() becomes superfluous because the provided interface effectively provides both the v1 and v2 interfaces, the difference being that overriding the xDestroy() method effectively gives it v2 semantics.

User-defined SQL Functions (a.k.a. UDFs)

The sqlite3_create_function() family of APIs make heavy use of function pointers to provide client-defined callbacks, necessitating interface changes in the JNI binding. The Java API has only one core function-registration function:

int sqlite3_create_function(sqlite3 db, String funcName, int nArgs,
                            int encoding, SQLFunction func);

Design question: does the encoding argument serve any purpose in Java? That's as-yet undetermined. If not, it will be removed.

SQLFunction is not used directly, but is instead instantiated via one of its three subclasses:

  • ScalarFunction implements simple scalar functions using but a single callback.
  • AggregateFunction implements aggregate functions using two callbacks.
  • WindowFunction implements window functions using four callbacks.

Search Tester1.java for SQLFunction for how it's used.

Reminder: see the disclaimer at the top of this document regarding the in-flux nature of this API.

And so on...

Various APIs which accept callbacks, e.g. sqlite3_trace_v2() and sqlite3_update_hook(), use interfaces similar to those shown above. Despite the changes in signature, the JNI layer makes every effort to provide the same semantics as the C API documentation suggests.