Convert sqlite3 to multi-phase initialisation (PEP 489)

[erlend-aasland]

BPO	42064
Nosy	@encukou, @berkerpeksag, @corona10, @miss-islington, @erlend-aasland, @DiddiLeija
PRs	bpo-42064: Move sqlite3 types to global state #26537 bpo-42064: Move sqlite3 exceptions to global state, part 1 of 2 #26745 bpo-42064: Remove stale extern declarations in sqlite3 headers #26840 bpo-42064: Move sqlite3 exceptions to global state, part 2 of 2 #26884 bpo-42064: Finalise establishing sqlite3 global state #27155 bpo-42064: Migrate to sqlite3_create_collation_v2 #27156 bpo-42064: Optimise sqlite3 state access, part 1 #27273 bpo-42064: Pass module state to sqlite3 UDF callbacks #27456 bpo-42064: Offset arguments for PyObject_Vectorcall in the _sqlite module #27931 bpo-42064: Pass module state to trace, progress, and authorizer callbacks #27940 bpo-42064: Add module backref to sqlite3 callback context #28242 bpo-42064: Convert sqlite3 global state to module state #29073 bpo-42064: Adapt sqlite3 to multi-phase init (PEP 489) #29234

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields:

assignee = None
closed_at = <Date 2021-11-02.18:10:54.692>
created_at = <Date 2020-10-17.19:37:05.423>
labels = ['type-feature', 'library', '3.10']
title = 'Convert sqlite3 to multi-phase initialisation (PEP 489)'
updated_at = <Date 2021-11-03.11:51:24.093>
user = 'https://github.com/erlend-aasland'

bugs.python.org fields:

activity = <Date 2021-11-03.11:51:24.093>
actor = 'erlendaasland'
assignee = 'none'
closed = True
closed_date = <Date 2021-11-02.18:10:54.692>
closer = 'erlendaasland'
components = ['Library (Lib)']
creation = <Date 2020-10-17.19:37:05.423>
creator = 'erlendaasland'
dependencies = []
files = []
hgrepos = []
issue_num = 42064
keywords = ['patch']
message_count = 34.0
messages = ['378825', '378829', '378832', '378905', '379029', '379172', '379326', '379476', '380299', '395047', '395876', '396411', '396416', '397477', '397866', '398298', '398470', '400205', '400206', '400208', '400712', '400720', '400725', '400726', '400733', '400774', '401245', '401259', '401423', '404301', '405086', '405512', '405583', '405609']
nosy_count = 6.0
nosy_names = ['petr.viktorin', 'berker.peksag', 'corona10', 'miss-islington', 'erlendaasland', 'DiddiLeija']
pr_nums = ['26537', '26745', '26840', '26884', '27155', '27156', '27273', '27456', '27931', '27940', '28242', '29073', '29234']
priority = 'normal'
resolution = 'fixed'
stage = 'resolved'
status = 'closed'
superseder = None
type = 'enhancement'
url = 'https://bugs.python.org/issue42064'
versions = ['Python 3.10']

[erlend-aasland]

Porting sqlite3 to multi-phase initialisation is non-trivial and must be broken into smaller bpo's, often with multiple PR's per bpo. As far as I can see, this is the task list:

1. Heap types (PEP-384): bpo-41861
2. Argument Clinic: bpo-40956
3. Establish global module state
4. Pass self iso. callback object to trace, auth, and progress handlers (required for item 7)
5. Use sqlite3_create_collation_v2 iso. sqlite3_create_collation (see next list item)
6. For sqlite3_create_* functions, establish a context struct that contains the target user callback and the state. Modify callers to allocate memory, and the destructor to deallocate
7. Module state (PEP-573)
8. Multi-phase initialisation (PEP-489)

The list may be incomplete.

Separate bpo's for 1) and 2) are opened. Pr. 2020-10-17, 1) is almost done, and 2), part 1 of 5 is awaiting review. It may be convenient to open up sub-bpo's for some of the other tasks as well. For instance, using sqlite3_create_collation_v2() iso. sqlite3_create_collation(). (FYI, this imposes no new SQLite version requirements.)

I'd wait until AC is implemented with moving forward with the rest of the PR's, in order to avoid merge mess. The exception might be item 4, preparing the trace/progress/auth callbacks, which is easy to isolate from the rest.

I've prepared branches for most of these things locally, some split in multiple PR's to ease reviewing.

I've tagged Victor Stinner and Dong-hee Na (hope that's ok), since they've been helpful with reviewing and providing helpful hints thus far. Any input from Berker Peksag would be highly appreciated; after all, he is the maintainer of this module :)

Ref. bpo-1635741 (the grand multi-phase issue)

[erlend-aasland]

Instead of using a custom context struct for sqlite3_create_funcion_v2(), we may just use sqlite3_context_db_handle() to get self, and then fetch the state using PyType_GetModuleState(Py_TYPE(self)).

[erlend-aasland]

No, unless we can add self as a private context to a sqlite3 database handle, we can't use that shortcut. However, using a custom context struct works, and it's a minor code change.

Proof-of-concept implementation diffstat:
2 files changed, 54 insertions(+), 10 deletions(-)

[erlend-aasland]

FYI, WIP branch for this bpo is here: https://github.com/erlend-aasland/cpython/commits/bpo-42064/all
This currently includes items 1 though 6. I'll update the WIP branch with a draft of items 7 and 8, hopefully tomorrow.

[erlend-aasland]

Work with item 7 is progressing; see WIP branch for details. Current showstopper is how to fetch module state from Connection.__init__ when it has been overloaded (for example in /Lib/sqlite3/test/factory.py). Guessing that Py_TYPE(self) does not return the sqlite3.Connection type.

[erlend-aasland]

Connection.__init__() fixed by iterating though tp_bases in order to find the correct base class to load the state from. There should be a better way to do this than to dig through type member internals.

[erlend-aasland]

Sorted out item 7 (module state). Item 8 is mostly in place, although I get ref. count errors if I add traverse/clear callbacks, so I've temporarily disabled those.

Proof-of-concept of multi-phase init can be found here: https://github.com/erlend-aasland/cpython/commits/bpo-42064/all
I'll try to break items 3 and 7 up in smaller pieces for the reviewer's convenience.

[erlend-aasland]

I've narrowed down the traverse problems to Py_VISIT(state->CacheType). Suspecting this has to do with the cyclic references in cache/connection. Preliminary tests with GC tracking cache->factory look promising, but this is a part of the C API that's completely unknown to me (yet), so I'm on thin ice here.

[erlend-aasland]

FYI, rebased https://github.com/erlend-aasland/cpython/tree/bpo-42064/all onto master, added Petr Viktorin's _PyType_GetModuleByDef() for use in item 7 (module state). I still run into problems in item 8, but I haven't devoted much time for this yet:
Modules/gcmodule.c:116: gc_decref: Assertion "gc_get_refs(g) > 0" failed: refcount is too small
Enable tracemalloc to get the memory block allocation traceback

object address : 0x7f861140c6b0
object refcount : 2
object type : 0x109666678
object type name: type
object repr : <class 'sqlite3.InterfaceError'>

Fatal Python error: _PyObject_AssertFailed: _PyObject_AssertFailed
Python runtime state: finalizing (tstate=0x7f8631409000)