Add the decision records

docs/architecture.md

@@ -129,7 +129,8 @@ same context are dispatched immediately because they come from that process.
 
 Every message and frame, with the rules for changing them, is in
 [protocols](protocols.md); the states each process and thread moves through
-are in [state machines](state-machines.md).
+are in [state machines](state-machines.md); the reasons behind the design
+are in the [decision records](decisions/overview.md).
 
 ## asyncio
 

docs/contributing.md

@@ -205,6 +205,8 @@ supervisor to carry it.
   clean locally.
 - `CHANGELOG.md` updated under the unreleased version: `Added`, `Changed`,
   `Removed` or `Fixed`. Removing a public function is a major version.
+- A change that reverses or extends a decision in `docs/decisions/` gets a
+  new record there; the old one is not edited.
 - The PR text says what the change intends and which path it takes; the
   diff already lists the files.
 - One squashed commit per PR.

docs/decisions/0001-one-thread-per-context.md

@@ -0,0 +1,36 @@
+# 0001: One pthread per context, NIFs only enqueue
+
+Since 3.0.0. Code: `worker_context_thread_main`, `owngil_context_thread_main`,
+the request queue on `py_context_t` (`c_src/py_nif.c`, `c_src/py_nif.h`).
+
+## Situation
+
+Before 3.0.0 Python ran on dirty schedulers: a NIF took the GIL, swapped
+in the context's thread state and executed the call. That pinned a dirty
+scheduler for the whole call with a 30 s cap, gave numpy, torch and
+tensorflow a different OS thread on every call (they keep per-thread
+state and some refuse to work across threads), and needed a single-slot
+request buffer that raced with concurrent callers.
+
+## Decision
+
+Every context owns one pthread that runs all of its Python. NIFs called
+from Erlang processes allocate a `ctx_request_t`, append it to the
+context's queue and return at once; the thread dequeues, executes with
+the GIL, and sends `{py_result, Ref, Result}` back. The same thread
+serves worker mode (main interpreter, shared GIL) and owngil mode (its
+own interpreter and GIL).
+
+## Consequences
+
+- Stable thread affinity: a context's Python always runs on the same OS
+  thread, and no dirty scheduler is held during a call.
+- Erlang processes never touch a context's Python objects; the lock and
+  ownership contract on `py_context_t` follows from this.
+- Shutdown must join the thread. A thread stuck in a C call cannot be
+  joined, so the context is leaked on purpose rather than freed under a
+  running thread.
+- Interrupts have to reach the thread from outside: `interrupt_mutex`,
+  `exec_thread_id` and `PyThreadState_SetAsyncExc`.
+- One request at a time per context is a property of the design, not a
+  limitation to work around; parallelism comes from more contexts.

docs/decisions/0002-callback-delivery-paths.md

@@ -0,0 +1,43 @@
+# 0002: Three callback delivery paths, chosen by the calling thread
+
+Since 3.0.0. Code: `erlang_call_impl` in `c_src/py_callback.c`,
+`c_src/py_thread_worker.c`, `src/py_thread_handler.erl`, the callback
+handling in `src/py_context.erl`.
+
+## Situation
+
+`erlang.call` must work from the context thread, from a Python thread the
+user started (executors, `threading.Thread`), and while the context
+thread is already waiting for Erlang (nested calls). A single blocking
+handler process deadlocks as soon as the callback calls Python again on
+the same context.
+
+## Decision
+
+The path is chosen by who is calling:
+
+1. Suspension, on a context thread with suspension enabled: the call
+   raises `SuspensionRequired`, the context returns
+   `{suspended, ...}` to its process, which runs the function, serves
+   nested requests meanwhile, and resumes. Inspired by PyO3's model.
+2. The context callback pipe, on a context thread with a handler
+   registered: the thread blocks on a pipe while the handler process
+   runs the function.
+3. The thread worker, for every other Python thread: a per-thread worker
+   with its own pipe and handler process, coordinated by
+   `py_thread_handler`.
+
+All three answer with the same response body (`<<2, ETF>>` or
+`<<1, Message>>`).
+
+## Consequences
+
+- Nested callbacks of any depth work on the suspension path because the
+  Erlang process is never blocked in a NIF while it waits.
+- Three writers and two parsers of the response body must stay in step;
+  the protocols page lists them.
+- A Python thread that calls Erlang is not a context thread and cannot
+  reach the calling context's Python objects; re-entrant calls into the
+  same owngil context from a spawned thread are not supported.
+- Adding a mode means adding a path (isolated mode has a fourth: the
+  socket frame), not extending one of these.

docs/decisions/0003-callback-results-as-etf.md

@@ -0,0 +1,32 @@
+# 0003: Callback results cross as external term format
+
+Since 4.0.0. Code: `handle_blocking_callback/3` and friends in
+`src/py_context.erl`, `parse_callback_response` in `c_src/py_callback.c`,
+`priv/_erlang_impl/_etf.py` for the child.
+
+## Situation
+
+Results of an Erlang callback used to reach Python as the Python repr of
+the term, parsed with `ast.literal_eval`. Binaries with backslashes,
+quotes or newlines produced unparseable literals and were handed over as
+raw text, `[]` arrived as `''`, floats lost precision, and pids and
+references went through a base64 marker that had to be decoded with an
+unsafe `binary_to_term`.
+
+## Decision
+
+Results are encoded with `term_to_binary` and decoded by the same
+`term_to_py` converter that call arguments use, with
+`ERL_NIF_BIN2TERM_SAFE`. Pids and references cross as native objects.
+The one visible change is accepted as breaking: an Erlang string
+(`"abc"`) reaches Python as `[97, 98, 99]`, exactly as it does for
+arguments; return a binary for a `str`.
+
+## Consequences
+
+- One type mapping in both directions, documented once
+  (`docs/type-conversion.md`, `c_src/py_convert.c`).
+- The `__etf__:` marker is data, never re-interpreted;
+  `test_etf_decode_safe` guards that no atoms are minted.
+- The child needs a Python ETF codec (`_etf.py`) that produces the same
+  terms as the C converter; keeping them in step is a maintenance duty.

docs/decisions/0004-isolated-mode-child-process.md

@@ -0,0 +1,40 @@
+# 0004: Isolation is a child OS process over a Unix socket
+
+Since 5.0.0. Code: `src/py_isolated.erl`, `priv/py_isolated_child.py`,
+`priv/_erlang_impl/_isolated.py`.
+
+## Situation
+
+Embedded modes cannot stop a Python call stuck in C, cannot bound its
+memory or CPU, and a segfault in an extension kills the node. Running
+untrusted or unknown Python code needs those guarantees, with the public
+API unchanged so a pool can mix modes.
+
+## Decision
+
+An isolated context is one child process per context, started with
+`open_port` so the VM reaps it, talking to its `py_context` process over
+a Unix socket with the frame format of the callback pipe
+(`<<Id:64, Len:32, Body>>`, body `<<Status, ETF>>`). Interrupts are a
+signal to the child's main thread with a `SIGKILL` backstop; limits are
+rlimits, cgroups v2 on Linux and an RSS watchdog on macOS; a crash
+restarts the child within a budget. The child uses the standard asyncio
+loop; there is no C code of its own in the VM beyond `os_kill`.
+
+Not chosen: a seccomp or Capsicum sandbox (a later hardening step, the
+process boundary is the first one), a NIF-side sub-process pool, and a
+new wire protocol (the existing frame and ETF conventions are enough).
+
+## Consequences
+
+- Everything crossing the socket is a term. NIF resources (channels,
+  native buffers, object references) do not cross; the API says so with
+  `{error, not_supported_in_isolated}`.
+- The child can create atoms (`binary_to_term` is not called with
+  `safe` because handles and control terms need atoms); untrusted code
+  must not mint unbounded distinct atoms.
+- Each call copies arguments and results through the socket; bulk data
+  needs shared memory (0006).
+- Python state is lost on restart; the context stays usable.
+- Platform code lives in the child launcher: parent-death signal
+  (`prctl` on Linux, `procctl` on FreeBSD), memory limits per OS.

docs/decisions/0005-py-isolated-gen-statem.md

@@ -0,0 +1,39 @@
+# 0005: The isolated context process is a gen_statem
+
+Since 5.0.0. Code: `src/py_isolated.erl`, entered from `init/4` in `py_context`
+with `gen_statem:enter_loop/5`.
+
+## Situation
+
+The embedded context process is a hand-written receive loop in
+`py_context`. The isolated process has more to track: a request in
+flight, requests to hold while the child restarts, a running loop with a
+grace period, an interrupt with a kill backstop bound to one request id,
+and callback processes whose nested requests must pass while others
+wait. A first version as a receive loop mirrored `py_context` but every
+wait needed its own selective receive and its own timer bookkeeping.
+
+## Decision
+
+`py_isolated` is a `gen_statem` (`handle_event_function`, state enter
+calls) with states `idle`, `{busy, Id}`, `looping`, `stopping_loop` and
+`{restarting, Reason}`. Requests that must wait are `postpone`d and
+replayed by the behaviour in arrival order; timers are `state_timeout`
+and named generic timeouts (`{timeout, kill}`) cancelled by the state
+change that makes them moot. It is spawned with `proc_lib` so it keeps
+the process identity and message protocol of `py_context`.
+
+Not chosen: sharing the receive loop with `py_context` (the two have
+different failure models: a child can die and restart, a thread cannot),
+or a `gen_server` with a state field (postpone and state timeouts would
+have to be reimplemented).
+
+## Consequences
+
+- `sys:get_state/1` shows what a context is doing and `sys:trace/2`
+  prints every event; the state machines page is a transcription of the
+  callback module.
+- Callers do not see the behaviour: messages and replies are those of
+  `py_context`.
+- The two context processes are different code. A change to the message
+  protocol touches both.

docs/decisions/0006-shared-memory-over-iommap.md

@@ -0,0 +1,42 @@
+# 0006: Bulk data through iommap regions, handles as plain tuples
+
+Since 5.0.0. Code: `src/py_shm.erl`, `src/py_buffer.erl` (shared variant),
+`priv/_erlang_impl/_shm.py`, the tagged-tuple case in `c_src/py_convert.c`.
+
+## Situation
+
+An isolated call copies its arguments and result through the socket:
+1.3 ms per MB, 300 ms for 64 MB. Request bodies, arrays and model inputs
+need a path that does not copy, and it must work the same in a pool that
+mixes embedded and isolated contexts.
+
+## Decision
+
+A region is a file mapped `MAP_SHARED` by the VM through iommap
+(`region_binary/3` gives a refcounted binary with no copy) and by any
+interpreter through `mmap`. Its handle is the plain term
+`{'$py_shm', Id, Path, Size}`, so it travels inside any argument or
+result with no special encoding and becomes a `SharedMemory` on arrival
+in every mode (the C converter and the child both call `_shm.from_term`).
+A shared `py_buffer` is a region used as a ring, with flow control
+through registered callbacks (`_py_buffer_wait`, `_py_buffer_consumed`),
+the mechanism channels already use, so no new control frames exist.
+iommap is optional: `py_shm:new/1` returns `{error, iommap_not_available}`
+without it.
+
+Not chosen: passing the region's fd (iommap exposes none; a path in a
+0700 directory is enough), a NIF of our own for mapping, and channels
+over shared memory (small terms stay on the socket).
+
+## Consequences
+
+- Python-produced data is zero-copy both ways; Erlang-produced data
+  costs one `pwrite` because a binary cannot be written in place.
+- Sharing memory weakens isolation for that region only; read-only
+  handles keep a callee from writing. Sealing and syscall filtering are
+  separate work.
+- In embedded contexts a shared buffer costs a callback round trip per
+  blocking read where the native buffer costs a pointer; the guide says
+  when to use which.
+- The region file must never be truncated (`SIGBUS`); sizes are fixed at
+  creation and verified with `fstat` before mapping.

docs/decisions/0007-remove-legacy-execution-paths.md

@@ -0,0 +1,31 @@
+# 0007: One execution path per mode; the legacy API is removed
+
+Since 5.0.0. Code: the removal in `c_src/py_nif.c`, `c_src/py_exec.c`,
+`c_src/py_callback.c`, `src/py_nif.erl` (PR #75).
+
+## Situation
+
+After 0001 every context had a thread, but the code still carried the
+paths from before: an executor thread with a worker pool, blocking NIF
+variants that ran Python on dirty schedulers, suspended-state resources
+for a resume protocol nothing used, and `py_nif` stubs for all of it.
+Roughly 4 500 lines were reachable only from suites or from nothing, and
+every reader had to work out which of two paths was live.
+
+## Decision
+
+Remove them. A context created today has exactly one path per mode:
+the queue and context thread for worker and owngil, the socket for
+isolated. NIFs that needed a thread now answer
+`{error, context_has_no_thread}` instead of falling back to a scheduler.
+Because public functions disappeared, the release that carries this is a
+major version (5.0.0), not a minor one.
+
+## Consequences
+
+- `docs/code-map.md` can say "live" for everything in `src/` and
+  `c_src/` except the test helpers, and mean it.
+- There is no fallback when a context has no thread; that state is a
+  bug, and it is reported as one.
+- Anyone on the removed functions upgrades through the changelog's
+  Removed section.

docs/decisions/0008-pipe-io-rules.md

@@ -0,0 +1,31 @@
+# 0008: Pipe I/O is non-blocking, deadlined and waited with poll
+
+Since 3.1.0 (deadlines), 5.0.0 (poll). Code: `read_with_timeout` and
+`write_all_with_deadline` in `c_src/py_nif.h`, the pipe setup in
+`c_src/py_thread_worker.c` and `c_src/py_callback.c`.
+
+## Situation
+
+Callback responses are written by Erlang processes on dirty I/O
+schedulers into pipes read by Python threads. A blocking write to a
+stalled reader pinned a dirty scheduler for good; a short read or write
+left the framed protocol out of phase with no way back. Later, a CI run
+with more than 1024 open files showed that `select()` cannot watch such a
+descriptor at all, and the first thread worker created past that point
+took every later thread callback down with it.
+
+## Decision
+
+Write ends are `O_NONBLOCK`; every write is `write_all_with_deadline`
+and every read `read_with_timeout`, both waiting with `poll()`. A
+partial frame is never recovered in band: the thread worker is poisoned
+and replaced, and the context pipe is closed only when its thread has
+been joined. The coordinator reports a failed ready signal instead of
+leaving Python to time out.
+
+## Consequences
+
+- No scheduler thread waits on Python without a bound.
+- A desynchronised pipe costs one worker, not a hang.
+- `select()` and `<sys/select.h>` do not appear in the NIF; a review
+  that sees them come back should ask why.

docs/decisions/overview.md

@@ -0,0 +1,18 @@
+# Decision records
+
+Why the code is the way it is, one decision per file, in the order they
+were taken. Read the record before changing what it decided; if the
+reasons no longer hold, write a new record that supersedes it rather than
+editing the old one. Each record has the same four parts: the situation,
+what was decided, what it costs, and where the code is.
+
+| # | Decision | Since |
+|---|---|---|
+| [0001](0001-one-thread-per-context.md) | One pthread per context, NIFs only enqueue | 3.0.0 |
+| [0002](0002-callback-delivery-paths.md) | Three callback delivery paths, chosen by the calling thread | 3.0.0 |
+| [0003](0003-callback-results-as-etf.md) | Callback results cross as external term format | 4.0.0 |
+| [0004](0004-isolated-mode-child-process.md) | Isolation is a child OS process over a Unix socket | 5.0.0 |
+| [0005](0005-py-isolated-gen-statem.md) | The isolated context process is a gen_statem | 5.0.0 |
+| [0006](0006-shared-memory-over-iommap.md) | Bulk data through iommap regions, handles as plain tuples | 5.0.0 |
+| [0007](0007-remove-legacy-execution-paths.md) | One execution path per mode; the legacy API is removed | 5.0.0 |
+| [0008](0008-pipe-io-rules.md) | Pipe I/O is non-blocking, deadlined and waited with poll | 3.1.0, 5.0.0 |

rebar.config

@@ -82,6 +82,15 @@
         <<"docs/protocols.md">>,
         <<"docs/state-machines.md">>,
         <<"docs/contributing.md">>,
+        <<"docs/decisions/overview.md">>,
+        <<"docs/decisions/0001-one-thread-per-context.md">>,
+        <<"docs/decisions/0002-callback-delivery-paths.md">>,
+        <<"docs/decisions/0003-callback-results-as-etf.md">>,
+        <<"docs/decisions/0004-isolated-mode-child-process.md">>,
+        <<"docs/decisions/0005-py-isolated-gen-statem.md">>,
+        <<"docs/decisions/0006-shared-memory-over-iommap.md">>,
+        <<"docs/decisions/0007-remove-legacy-execution-paths.md">>,
+        <<"docs/decisions/0008-pipe-io-rules.md">>,
         <<"docs/preload.md">>,
         <<"docs/owngil_internals.md">>,
         <<"docs/event_loop_architecture.md">>
@@ -126,6 +135,17 @@
             <<"docs/preload.md">>,
             <<"docs/owngil_internals.md">>,
             <<"docs/event_loop_architecture.md">>
+        ]},
+        {<<"Decisions">>, [
+            <<"docs/decisions/overview.md">>,
+            <<"docs/decisions/0001-one-thread-per-context.md">>,
+            <<"docs/decisions/0002-callback-delivery-paths.md">>,
+            <<"docs/decisions/0003-callback-results-as-etf.md">>,
+            <<"docs/decisions/0004-isolated-mode-child-process.md">>,
+            <<"docs/decisions/0005-py-isolated-gen-statem.md">>,
+            <<"docs/decisions/0006-shared-memory-over-iommap.md">>,
+            <<"docs/decisions/0007-remove-legacy-execution-paths.md">>,
+            <<"docs/decisions/0008-pipe-io-rules.md">>
         ]}
     ]}
 ]}.