tractor/tractor/_shm.py

834 lines
22 KiB
Python

# tractor: structured concurrent "actors".
# Copyright 2018-eternity Tyler Goodlet.
# This program is free software: you can redistribute it and/or modify
# it under the terms of the GNU Affero General Public License as published by
# the Free Software Foundation, either version 3 of the License, or
# (at your option) any later version.
# This program is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
# GNU Affero General Public License for more details.
# You should have received a copy of the GNU Affero General Public License
# along with this program. If not, see <https://www.gnu.org/licenses/>.
"""
SC friendly shared memory management geared at real-time
processing.
Support for ``numpy`` compatible array-buffers is provided but is
considered optional within the context of this runtime-library.
"""
from __future__ import annotations
from sys import byteorder
import time
from typing import Optional
from multiprocessing import shared_memory as shm
from multiprocessing.shared_memory import (
SharedMemory,
ShareableList,
)
from msgspec import Struct
import tractor
from .log import get_logger
_USE_POSIX = getattr(shm, '_USE_POSIX', False)
if _USE_POSIX:
from _posixshmem import shm_unlink
try:
import numpy as np
from numpy.lib import recfunctions as rfn
# import nptyping
except ImportError:
pass
log = get_logger(__name__)
def disable_mantracker():
'''
Disable all ``multiprocessing``` "resource tracking" machinery since
it's an absolute multi-threaded mess of non-SC madness.
'''
from multiprocessing import resource_tracker as mantracker
# Tell the "resource tracker" thing to fuck off.
class ManTracker(mantracker.ResourceTracker):
def register(self, name, rtype):
pass
def unregister(self, name, rtype):
pass
def ensure_running(self):
pass
# "know your land and know your prey"
# https://www.dailymotion.com/video/x6ozzco
mantracker._resource_tracker = ManTracker()
mantracker.register = mantracker._resource_tracker.register
mantracker.ensure_running = mantracker._resource_tracker.ensure_running
mantracker.unregister = mantracker._resource_tracker.unregister
mantracker.getfd = mantracker._resource_tracker.getfd
disable_mantracker()
class SharedInt:
'''
Wrapper around a single entry shared memory array which
holds an ``int`` value used as an index counter.
'''
def __init__(
self,
shm: SharedMemory,
) -> None:
self._shm = shm
@property
def value(self) -> int:
return int.from_bytes(self._shm.buf, byteorder)
@value.setter
def value(self, value) -> None:
self._shm.buf[:] = value.to_bytes(self._shm.size, byteorder)
def destroy(self) -> None:
if _USE_POSIX:
# We manually unlink to bypass all the "resource tracker"
# nonsense meant for non-SC systems.
name = self._shm.name
try:
shm_unlink(name)
except FileNotFoundError:
# might be a teardown race here?
log.warning(f'Shm for {name} already unlinked?')
class NDToken(Struct, frozen=True):
'''
Internal represenation of a shared memory ``numpy`` array "token"
which can be used to key and load a system (OS) wide shm entry
and correctly read the array by type signature.
This type is msg safe.
'''
shm_name: str # this servers as a "key" value
shm_first_index_name: str
shm_last_index_name: str
dtype_descr: tuple
size: int # in struct-array index / row terms
# TODO: use nptyping here on dtypes
@property
def dtype(self) -> list[tuple[str, str, tuple[int, ...]]]:
return np.dtype(
list(
map(tuple, self.dtype_descr)
)
).descr
def as_msg(self):
return self.to_dict()
@classmethod
def from_msg(cls, msg: dict) -> NDToken:
if isinstance(msg, NDToken):
return msg
# TODO: native struct decoding
# return _token_dec.decode(msg)
msg['dtype_descr'] = tuple(map(tuple, msg['dtype_descr']))
return NDToken(**msg)
# _token_dec = msgspec.msgpack.Decoder(NDToken)
# TODO: this api?
# _known_tokens = tractor.ActorVar('_shm_tokens', {})
# _known_tokens = tractor.ContextStack('_known_tokens', )
# _known_tokens = trio.RunVar('shms', {})
# TODO: this should maybe be provided via
# a `.trionics.maybe_open_context()` wrapper factory?
# process-local store of keys to tokens
_known_tokens: dict[str, NDToken] = {}
def get_shm_token(key: str) -> NDToken | None:
'''
Convenience func to check if a token
for the provided key is known by this process.
Returns either the ``numpy`` token or a string for a shared list.
'''
return _known_tokens.get(key)
def _make_token(
key: str,
size: int,
dtype: np.dtype,
) -> NDToken:
'''
Create a serializable token that can be used
to access a shared array.
'''
return NDToken(
shm_name=key,
shm_first_index_name=key + "_first",
shm_last_index_name=key + "_last",
dtype_descr=tuple(np.dtype(dtype).descr),
size=size,
)
class ShmArray:
'''
A shared memory ``numpy.ndarray`` API.
An underlying shared memory buffer is allocated based on
a user specified ``numpy.ndarray``. This fixed size array
can be read and written to by pushing data both onto the "front"
or "back" of a set index range. The indexes for the "first" and
"last" index are themselves stored in shared memory (accessed via
``SharedInt`` interfaces) values such that multiple processes can
interact with the same array using a synchronized-index.
'''
def __init__(
self,
shmarr: np.ndarray,
first: SharedInt,
last: SharedInt,
shm: SharedMemory,
# readonly: bool = True,
) -> None:
self._array = shmarr
# indexes for first and last indices corresponding
# to fille data
self._first = first
self._last = last
self._len = len(shmarr)
self._shm = shm
self._post_init: bool = False
# pushing data does not write the index (aka primary key)
self._write_fields: list[str] | None = None
dtype = shmarr.dtype
if dtype.fields:
self._write_fields = list(shmarr.dtype.fields.keys())[1:]
# TODO: ringbuf api?
@property
def _token(self) -> NDToken:
return NDToken(
shm_name=self._shm.name,
shm_first_index_name=self._first._shm.name,
shm_last_index_name=self._last._shm.name,
dtype_descr=tuple(self._array.dtype.descr),
size=self._len,
)
@property
def token(self) -> dict:
"""Shared memory token that can be serialized and used by
another process to attach to this array.
"""
return self._token.as_msg()
@property
def index(self) -> int:
return self._last.value % self._len
@property
def array(self) -> np.ndarray:
'''
Return an up-to-date ``np.ndarray`` view of the
so-far-written data to the underlying shm buffer.
'''
a = self._array[self._first.value:self._last.value]
# first, last = self._first.value, self._last.value
# a = self._array[first:last]
# TODO: eventually comment this once we've not seen it in the
# wild in a long time..
# XXX: race where first/last indexes cause a reader
# to load an empty array..
if len(a) == 0 and self._post_init:
raise RuntimeError('Empty array race condition hit!?')
# breakpoint()
return a
def ustruct(
self,
fields: Optional[list[str]] = None,
# type that all field values will be cast to
# in the returned view.
common_dtype: np.dtype = float,
) -> np.ndarray:
array = self._array
if fields:
selection = array[fields]
# fcount = len(fields)
else:
selection = array
# fcount = len(array.dtype.fields)
# XXX: manual ``.view()`` attempt that also doesn't work.
# uview = selection.view(
# dtype='<f16',
# ).reshape(-1, 4, order='A')
# assert len(selection) == len(uview)
u = rfn.structured_to_unstructured(
selection,
# dtype=float,
copy=True,
)
# unstruct = np.ndarray(u.shape, dtype=a.dtype, buffer=shm.buf)
# array[:] = a[:]
return u
# return ShmArray(
# shmarr=u,
# first=self._first,
# last=self._last,
# shm=self._shm
# )
def last(
self,
length: int = 1,
) -> np.ndarray:
'''
Return the last ``length``'s worth of ("row") entries from the
array.
'''
return self.array[-length:]
def push(
self,
data: np.ndarray,
field_map: Optional[dict[str, str]] = None,
prepend: bool = False,
update_first: bool = True,
start: int | None = None,
) -> int:
'''
Ring buffer like "push" to append data
into the buffer and return updated "last" index.
NB: no actual ring logic yet to give a "loop around" on overflow
condition, lel.
'''
length = len(data)
if prepend:
index = (start or self._first.value) - length
if index < 0:
raise ValueError(
f'Array size of {self._len} was overrun during prepend.\n'
f'You have passed {abs(index)} too many datums.'
)
else:
index = start if start is not None else self._last.value
end = index + length
if field_map:
src_names, dst_names = zip(*field_map.items())
else:
dst_names = src_names = self._write_fields
try:
self._array[
list(dst_names)
][index:end] = data[list(src_names)][:]
# NOTE: there was a race here between updating
# the first and last indices and when the next reader
# tries to access ``.array`` (which due to the index
# overlap will be empty). Pretty sure we've fixed it now
# but leaving this here as a reminder.
if (
prepend
and update_first
and length
):
assert index < self._first.value
if (
index < self._first.value
and update_first
):
assert prepend, 'prepend=True not passed but index decreased?'
self._first.value = index
elif not prepend:
self._last.value = end
self._post_init = True
return end
except ValueError as err:
if field_map:
raise
# should raise if diff detected
self.diff_err_fields(data)
raise err
def diff_err_fields(
self,
data: np.ndarray,
) -> None:
# reraise with any field discrepancy
our_fields, their_fields = (
set(self._array.dtype.fields),
set(data.dtype.fields),
)
only_in_ours = our_fields - their_fields
only_in_theirs = their_fields - our_fields
if only_in_ours:
raise TypeError(
f"Input array is missing field(s): {only_in_ours}"
)
elif only_in_theirs:
raise TypeError(
f"Input array has unknown field(s): {only_in_theirs}"
)
# TODO: support "silent" prepends that don't update ._first.value?
def prepend(
self,
data: np.ndarray,
) -> int:
end = self.push(data, prepend=True)
assert end
def close(self) -> None:
self._first._shm.close()
self._last._shm.close()
self._shm.close()
def destroy(self) -> None:
if _USE_POSIX:
# We manually unlink to bypass all the "resource tracker"
# nonsense meant for non-SC systems.
shm_unlink(self._shm.name)
self._first.destroy()
self._last.destroy()
def flush(self) -> None:
# TODO: flush to storage backend like markestore?
...
def open_shm_ndarray(
size: int,
key: str | None = None,
dtype: np.dtype | None = None,
append_start_index: int | None = None,
readonly: bool = False,
) -> ShmArray:
'''
Open a memory shared ``numpy`` using the standard library.
This call unlinks (aka permanently destroys) the buffer on teardown
and thus should be used from the parent-most accessor (process).
'''
# create new shared mem segment for which we
# have write permission
a = np.zeros(size, dtype=dtype)
a['index'] = np.arange(len(a))
shm = SharedMemory(
name=key,
create=True,
size=a.nbytes
)
array = np.ndarray(
a.shape,
dtype=a.dtype,
buffer=shm.buf
)
array[:] = a[:]
array.setflags(write=int(not readonly))
token = _make_token(
key=key,
size=size,
dtype=dtype,
)
# create single entry arrays for storing an first and last indices
first = SharedInt(
shm=SharedMemory(
name=token.shm_first_index_name,
create=True,
size=4, # std int
)
)
last = SharedInt(
shm=SharedMemory(
name=token.shm_last_index_name,
create=True,
size=4, # std int
)
)
# Start the "real-time" append-updated (or "pushed-to") section
# after some start index: ``append_start_index``. This allows appending
# from a start point in the array which isn't the 0 index and looks
# something like,
# -------------------------
# | | i
# _________________________
# <-------------> <------->
# history real-time
#
# Once fully "prepended", the history section will leave the
# ``ShmArray._start.value: int = 0`` and the yet-to-be written
# real-time section will start at ``ShmArray.index: int``.
# this sets the index to nearly 2/3rds into the the length of
# the buffer leaving at least a "days worth of second samples"
# for the real-time section.
if append_start_index is None:
append_start_index = round(size * 0.616)
last.value = first.value = append_start_index
shmarr = ShmArray(
array,
first,
last,
shm,
)
assert shmarr._token == token
_known_tokens[key] = shmarr.token
# "unlink" created shm on process teardown by
# pushing teardown calls onto actor context stack
stack = tractor.current_actor().lifetime_stack
stack.callback(shmarr.close)
stack.callback(shmarr.destroy)
return shmarr
def attach_shm_ndarray(
token: tuple[str, str, tuple[str, str]],
readonly: bool = True,
) -> ShmArray:
'''
Attach to an existing shared memory array previously
created by another process using ``open_shared_array``.
No new shared mem is allocated but wrapper types for read/write
access are constructed.
'''
token = NDToken.from_msg(token)
key = token.shm_name
if key in _known_tokens:
assert NDToken.from_msg(_known_tokens[key]) == token, "WTF"
# XXX: ugh, looks like due to the ``shm_open()`` C api we can't
# actually place files in a subdir, see discussion here:
# https://stackoverflow.com/a/11103289
# attach to array buffer and view as per dtype
_err: Optional[Exception] = None
for _ in range(3):
try:
shm = SharedMemory(
name=key,
create=False,
)
break
except OSError as oserr:
_err = oserr
time.sleep(0.1)
else:
if _err:
raise _err
shmarr = np.ndarray(
(token.size,),
dtype=token.dtype,
buffer=shm.buf
)
shmarr.setflags(write=int(not readonly))
first = SharedInt(
shm=SharedMemory(
name=token.shm_first_index_name,
create=False,
size=4, # std int
),
)
last = SharedInt(
shm=SharedMemory(
name=token.shm_last_index_name,
create=False,
size=4, # std int
),
)
# make sure we can read
first.value
sha = ShmArray(
shmarr,
first,
last,
shm,
)
# read test
sha.array
# Stash key -> token knowledge for future queries
# via `maybe_opepn_shm_array()` but only after we know
# we can attach.
if key not in _known_tokens:
_known_tokens[key] = token
# "close" attached shm on actor teardown
tractor.current_actor().lifetime_stack.callback(sha.close)
return sha
def maybe_open_shm_ndarray(
key: str, # unique identifier for segment
size: int,
dtype: np.dtype | None = None,
append_start_index: int = 0,
readonly: bool = True,
) -> tuple[ShmArray, bool]:
'''
Attempt to attach to a shared memory block using a "key" lookup
to registered blocks in the users overall "system" registry
(presumes you don't have the block's explicit token).
This function is meant to solve the problem of discovering whether
a shared array token has been allocated or discovered by the actor
running in **this** process. Systems where multiple actors may seek
to access a common block can use this function to attempt to acquire
a token as discovered by the actors who have previously stored
a "key" -> ``NDToken`` map in an actor local (aka python global)
variable.
If you know the explicit ``NDToken`` for your memory segment instead
use ``attach_shm_array``.
'''
try:
# see if we already know this key
token = _known_tokens[key]
return (
attach_shm_ndarray(
token=token,
readonly=readonly,
),
False, # not newly opened
)
except KeyError:
log.warning(f"Could not find {key} in shms cache")
if dtype:
token = _make_token(
key,
size=size,
dtype=dtype,
)
else:
try:
return (
attach_shm_ndarray(
token=token,
readonly=readonly,
),
False,
)
except FileNotFoundError:
log.warning(f"Could not attach to shm with token {token}")
# This actor does not know about memory
# associated with the provided "key".
# Attempt to open a block and expect
# to fail if a block has been allocated
# on the OS by someone else.
return (
open_shm_ndarray(
key=key,
size=size,
dtype=dtype,
append_start_index=append_start_index,
readonly=readonly,
),
True,
)
class ShmList(ShareableList):
'''
Carbon copy of ``.shared_memory.ShareableList`` with a few
enhancements:
- readonly mode via instance var flag `._readonly: bool`
- ``.__getitem__()`` accepts ``slice`` inputs
- exposes the underlying buffer "name" as a ``.key: str``
'''
def __init__(
self,
sequence: list | None = None,
*,
name: str | None = None,
readonly: bool = True
) -> None:
self._readonly = readonly
self._key = name
return super().__init__(
sequence=sequence,
name=name,
)
@property
def key(self) -> str:
return self._key
@property
def readonly(self) -> bool:
return self._readonly
def __setitem__(
self,
position,
value,
) -> None:
# mimick ``numpy`` error
if self._readonly:
raise ValueError('assignment destination is read-only')
return super().__setitem__(position, value)
def __getitem__(
self,
indexish,
) -> list:
# NOTE: this is a non-writeable view (copy?) of the buffer
# in a new list instance.
if isinstance(indexish, slice):
return list(self)[indexish]
return super().__getitem__(indexish)
# TODO: should we offer a `.array` and `.push()` equivalent
# to the `ShmArray`?
# currently we have the following limitations:
# - can't write slices of input using traditional slice-assign
# syntax due to the ``ShareableList.__setitem__()`` implementation.
# - ``list(shmlist)`` returns a non-mutable copy instead of
# a writeable view which would be handier numpy-style ops.
def open_shm_list(
key: str,
sequence: list | None = None,
size: int = int(2 ** 10),
dtype: float | int | bool | str | bytes | None = float,
readonly: bool = True,
) -> ShmList:
if sequence is None:
default = {
float: 0.,
int: 0,
bool: True,
str: 'doggy',
None: None,
}[dtype]
sequence = [default] * size
shml = ShmList(
sequence=sequence,
name=key,
readonly=readonly,
)
# "close" attached shm on actor teardown
try:
actor = tractor.current_actor()
actor.lifetime_stack.callback(shml.shm.close)
actor.lifetime_stack.callback(shml.shm.unlink)
except RuntimeError:
log.warning('tractor runtime not active, skipping teardown steps')
return shml
def attach_shm_list(
key: str,
readonly: bool = False,
) -> ShmList:
return ShmList(
name=key,
readonly=readonly,
)