Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

FIX memoryview leaks and retrofit memory-manager as context-managers #33

Open
wants to merge 19 commits into
base: master
Choose a base branch
from
Open
Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
19 commits
Select commit Hold shift + click to select a range
2347f3d
feat(enc): test with unicode tmpdir
ankostis Oct 23, 2016
883217b
fix(leaks): attempt to plug the leaks & filter dead regions
ankostis Oct 24, 2016
43c5f23
chore(ver): bump 2.0.1-->2.1.0.dev0
ankostis Oct 24, 2016
f10196f
fix(regs): fix/rename scream_if_closed()-->collect_closed_regions()
ankostis Oct 24, 2016
88e2769
chore(ver): bump 2.0.1-->2.1.0.dev1
ankostis Oct 24, 2016
133dd1c
style(listuple): pep8, literals for empty lists/tuples
ankostis Oct 24, 2016
7c1eac7
refact(region): rename offset `_b --> _ofs`
ankostis Oct 24, 2016
bba086a
refact(minor): use region.priv-func, close fd on same condition
ankostis Oct 25, 2016
a2bc2d2
feat(mman): BREAKING API `mman` as context-manager to release regions
ankostis Oct 25, 2016
01df7f3
chore(ver): bump 2.1.0.dev1-->2.1.0.dev3
ankostis Oct 25, 2016
bf68f77
feat(mman-contxt): opt-out not to scream if mman not entered
ankostis Oct 25, 2016
4598966
fix(leaks): attempt to gc-collect before region-collect
ankostis Oct 25, 2016
d0bd74e
fix(mman): exit log-msg were missing left-overs arg, log as debug
ankostis Oct 25, 2016
e33235a
refact(buf): also use SlidingWindowMapBuffer as optional context-manager
ankostis Oct 25, 2016
33f12e6
refact(TCs): unittestize assertions
ankostis Oct 25, 2016
8489c31
refact(buf): simplify API - no begin/end after construct
ankostis Oct 25, 2016
d81dc1d
style(mman): move managed_mmaps() closer to 2 mmans
ankostis Oct 25, 2016
9ba1649
fix(leaks): FIX memoryview leak in Windows
ankostis Oct 27, 2016
144891b
chore(ver): bump 2.1.0.dev3-->2.1.0.dev4
ankostis Oct 27, 2016
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 4 additions & 0 deletions .travis.yml
Original file line number Diff line number Diff line change
Expand Up @@ -8,8 +8,12 @@ python:
- 3.3
- 3.4
- 3.5
env:
# test with paths containing the unicode characters
- TMPDIR="/tmp/καλημέρα"
install:
- pip install coveralls
- ln -s /tmp "$TMPDIR"
script:
- ulimit -n 48
- ulimit -n
Expand Down
44 changes: 31 additions & 13 deletions doc/source/changes.rst
Original file line number Diff line number Diff line change
Expand Up @@ -2,34 +2,52 @@
Changelog
#########

**********
2.1.0
======

- **BREAKING API:** retrofit ``git.util.mman`` as context-manager,
to release memory-mapped regions held.

The *mmap-manager(s)* are re-entrant, but not thread-safe **context-manager(s)**,
to be used within a ``with ...:`` block, ensuring any left-overs cursors are cleaned up.
If not entered, :meth:`StaticWindowMapManager.make_cursor()` and/or
:meth:`WindowCursor.use_region()` will scream.

Get them from ``smmap.managed_mmaps()``.

- FIX ``memoryview`` leak in Windows; now all *gitdb* TCs now pass without explicit release!

- Simplify :class:`SlidingWindowMapBuffer` as create/close context-manager
(no ``begin_access()``, or ``end_access()``).


v0.9.0
**********
========
- Fixed issue with resources never being freed as mmaps were never closed.
- Client counting is now done manually, instead of relying on pyton's reference count

**********

v0.8.5
**********
========
- Fixed Python 3.0-3.3 regression, which also causes smmap to become about 3 times slower depending on the code path. It's related to this bug (http://bugs.python.org/issue15958), which was fixed in python 3.4

**********

v0.8.4
**********
========
- Fixed Python 3 performance regression

**********

v0.8.3
**********
========
- Cleaned up code and assured it works sufficiently well with python 3

**********

v0.8.1
**********
========
- A single bugfix

**********
v0.8.0
**********

v0.8.0
========

- Initial Release
204 changes: 114 additions & 90 deletions doc/source/tutorial.rst
Original file line number Diff line number Diff line change
Expand Up @@ -5,114 +5,138 @@ Usage Guide
###########
This text briefly introduces you to the basic design decisions and accompanying classes.

******
Design
******
Per application, there is *MemoryManager* which is held as static instance and used throughout the application. It can be configured to keep your resources within certain limits.
======
Per application, there must be a *MemoryManager* to be used throughout the application.
It can be configured to keep your resources within certain limits (see :func:`smmap.managed_mmaps()`).

To access mapped regions, you require a cursor. Cursors point to exactly one file and serve as handles into it. As long as it exists, the respective memory region will remain available.
To access mapped regions, you require a cursor. Cursors point to exactly one file
and serve as handles into it.
As long as it exists, the respective memory region will remain available.

For convenience, a buffer implementation is provided (:class:`smmap.SlidingWindowMapBuffer`)
which handles cursors and resource allocation behind its simple buffer like interface.

For convenience, a buffer implementation is provided which handles cursors and resource allocation behind its simple buffer like interface.

***************
Memory Managers
***************
There are two types of memory managers, one uses *static* windows, the other one uses *sliding* windows. A window is a region of a file mapped into memory. Although the names might be somewhat misleading as technically windows are always static, the *sliding* version will allocate relatively small windows whereas the *static* version will always map the whole file.
================
There are two types of memory managers, one uses *static* windows, the other one uses *sliding* windows.
A window is a region of a file mapped into memory. Although the names might be somewhat misleading,
as technically windows are always static, the *sliding* version will allocate relatively small windows
whereas the *static* version will always map the whole file.

The *static* memory-manager does nothing more than keeping a client count on the respective memory maps
which always map the whole file, which allows to make some assumptions that can lead to simplified
data access and increased performance, but reduces the compatibility to 32 bit systems or giant files.

The *sliding* memory-manager therefore should be the default manager when preparing an application
for handling huge amounts of data on 32 bit and 64 bit platforms

The *static* manager does nothing more than keeping a client count on the respective memory maps which always map the whole file, which allows to make some assumptions that can lead to simplified data access and increased performance, but reduces the compatibility to 32 bit systems or giant files.
.. Note::
The *mmap-manager(s)* are re-entrant, but not thread-safe **context-manager(s)**,
to be used within a ``with ...:`` block, ensuring any left-overs cursors are cleaned up.
If not entered, :meth:`StaticWindowMapManager.make_cursor()` and/or
:meth:`WindowCursor.use_region()` will scream.

The *sliding* memory manager therefore should be the default manager when preparing an application for handling huge amounts of data on 32 bit and 64 bit platforms::

Use the :math:`smmap.managed_mmaps()` to take care of all this::

import smmap
# This instance should be globally available in your application
# It is configured to be well suitable for 32-bit or 64 bit applications.
mman = smmap.SlidingWindowMapManager()
# the manager provides much useful information about its current state
# like the amount of open file handles or the amount of mapped memory
mman.num_file_handles()
mman.mapped_memory_size()
# and many more ...
with smmap.managed_mmaps() as mman:

# the manager provides much useful information about its current state
# like the amount of open file handles or the amount of mapped memory
mman.num_file_handles()
mman.mapped_memory_size()
# and many more ...


Cursors
*******
========
*Cursors* are handles that point onto a window, i.e. a region of a file mapped into memory. From them you may obtain a buffer through which the data of that window can actually be accessed::

import smmap.test.lib
fc = smmap.test.lib.FileCreator(1024*1024*8, "test_file")

# obtain a cursor to access some file.
c = mman.make_cursor(fc.path)
# the cursor is now associated with the file, but not yet usable
assert c.is_associated()
assert not c.is_valid()
# before you can use the cursor, you have to specify a window you want to
# access. The following just says you want as much data as possible starting
# from offset 0.
# To be sure your region could be mapped, query for validity
assert c.use_region().is_valid() # use_region returns self
# once a region was mapped, you must query its dimension regularly
# to assure you don't try to access its buffer out of its bounds
assert c.size()
c.buffer()[0] # first byte
c.buffer()[1:10] # first 9 bytes
c.buffer()[c.size()-1] # last byte
# its recommended not to create big slices when feeding the buffer
# into consumers (e.g. struct or zlib).
# Instead, either give the buffer directly, or use pythons buffer command.
buffer(c.buffer(), 1, 9) # first 9 bytes without copying them
# you can query absolute offsets, and check whether an offset is included
# in the cursor's data.
assert c.ofs_begin() < c.ofs_end()
assert c.includes_ofs(100)
# If you are over out of bounds with one of your region requests, the
# cursor will be come invalid. It cannot be used in that state
assert not c.use_region(fc.size, 100).is_valid()
# map as much as possible after skipping the first 100 bytes
assert c.use_region(100).is_valid()
# You can explicitly free cursor resources by unusing the cursor's region
c.unuse_region()
assert not c.is_valid()

Now you would have to write your algorithms around this interface to properly slide through huge amounts of data.
import smmap.test.lib as tlib

with smmap.managed_mmaps() as mman, tlib.FileCreator(1024*1024*8, "test_file") as fc:
# obtain a cursor to access some file.
with mman.make_cursor(fc.path) as c:

# the cursor is now associated with the file, but not yet usable
assert c.is_associated()
assert not c.is_valid()

# before you can use the cursor, you have to specify a window you want to
# access. The following just says you want as much data as possible starting
# from offset 0.
# To be sure your region could be mapped, query for validity
assert c.use_region().is_valid() # use_region returns self

# once a region was mapped, you must query its dimension regularly
# to assure you don't try to access its buffer out of its bounds
assert c.size()
c.buffer()[0] # first byte
c.buffer()[1:10] # first 9 bytes
c.buffer()[c.size()-1] # last byte

# its recommended not to create big slices when feeding the buffer
# into consumers (e.g. struct or zlib).
# Instead, either give the buffer directly, or use pythons buffer command.
buffer(c.buffer(), 1, 9) # first 9 bytes without copying them

# you can query absolute offsets, and check whether an offset is included
# in the cursor's data.
assert c.ofs_begin() < c.ofs_end()
assert c.includes_ofs(100)

# If you are over out of bounds with one of your region requests, the
# cursor will be come invalid. It cannot be used in that state
assert not c.use_region(fc.size, 100).is_valid()
# map as much as possible after skipping the first 100 bytes
assert c.use_region(100).is_valid()

# You must explicitly free cursor resources by unusing the cursor's region
c.unuse_region()
assert not c.is_valid()


Now you would have to write your algorithms around this interface to properly slide through huge amounts of data.

Alternatively you can use a convenience interface.

*******

========
Buffers
*******
To make first use easier, at the expense of performance, there is a Buffer implementation which uses a cursor underneath.

With it, you can access all data in a possibly huge file without having to take care of setting the cursor to different regions yourself::

# Create a default buffer which can operate on the whole file
buf = smmap.SlidingWindowMapBuffer(mman.make_cursor(fc.path))

# you can use it right away
assert buf.cursor().is_valid()

buf[0] # access the first byte
buf[-1] # access the last ten bytes on the file
buf[-10:]# access the last ten bytes

# If you want to keep the instance between different accesses, use the
# dedicated methods
buf.end_access()
assert not buf.cursor().is_valid() # you cannot use the buffer anymore
assert buf.begin_access(offset=10) # start using the buffer at an offset

# it will stop using resources automatically once it goes out of scope

========
To make first use easier, at the expense of performance, there is a Buffer implementation
which uses a cursor underneath.

With it, you can access all data in a possibly huge file
without having to take care of setting the cursor to different regions yourself::

## Create a default buffer which can operate on the whole file
cur = mman.make_cursor(fc.path)
with smmap.SlidingWindowMapBuffer(cur) as buf:
# you can use it right away
assert buf.cursor().is_valid()

buf[0] # access the first byte
buf[-1] # access the last ten bytes on the file
buf[-10:]# access the last ten bytes

## You cannot use the buffer anymore.
assert not buf.cursor().is_valid()

## If you want to keep the instance between different accesses,
# use another instance.
with smmap.SlidingWindowMapBuffer(cur, offset=10) as buf:
assert buf.cursor().is_valid()


Disadvantages
*************
Buffers cannot be used in place of strings or maps, hence you have to slice them to have valid input for the sorts of struct and zlib. A slice means a lot of data handling overhead which makes buffers slower compared to using cursors directly.
--------------
Buffers cannot be used in place of strings or maps, hence you have to slice them to have valid
input for the sorts of struct and zlib.
A slice means a lot of data handling overhead which makes buffers slower compared to using cursors directly.

6 changes: 3 additions & 3 deletions smmap/__init__.py
Original file line number Diff line number Diff line change
Expand Up @@ -3,9 +3,9 @@
__author__ = "Sebastian Thiel"
__contact__ = "[email protected]"
__homepage__ = "https://github.com/Byron/smmap"
version_info = (2, 0, 1)
version_info = (2, 1, 0, 'dev4')
__version__ = '.'.join(str(i) for i in version_info)

# make everything available in root package for convenience
from .mman import *
from .buf import *
from .mman import * # @IgnorePep8
from .buf import * # @IgnorePep8
Loading