Multithreaded C API Python questions

[Svein Seldal]

Hi

A couple of mulithreaded C API python questions:

I) The PyGILState_Ensu re() simply ensures python api call ability, it
doesnt actually lock the GIL, right?

PyGILState_STAT E gstate;
gstate = PyGILState_Ensu re();

II) Am I required to lock the GIL prior to running any python functions
(in a threaded app)?

PyThreadState *pts = PyGILState_GetT hisThreadState( );
PyEval_AcquireT hread(pts);
PyObject_CallOb ject(...);

III) Shouldn't the GIL be released after 100 bytecodes or so to enable
other waiting threads to get the GIL?

I'm unable to get access to python as long as another python call is
executing. The PyEval_AcquireT hread() call blocks until the first call
returns. I was hoping that the python system itself would release the
GIL after some execution, but it itsnt.

I am dependent upon the ability to have to threads executing in python
land at the same time. How can this be done?
Regards,
Svein Seldal

[robert]

Svein Seldal wrote:

Hi

A couple of mulithreaded C API python questions:

I) The PyGILState_Ensu re() simply ensures python api call ability, it
doesnt actually lock the GIL, right?

PyGILState_STAT E gstate;
gstate = PyGILState_Ensu re();

II) Am I required to lock the GIL prior to running any python functions
(in a threaded app)?

PyThreadState *pts = PyGILState_GetT hisThreadState( );
PyEval_AcquireT hread(pts);
PyObject_CallOb ject(...);

III) Shouldn't the GIL be released after 100 bytecodes or so to enable
other waiting threads to get the GIL?

I'm unable to get access to python as long as another python call is
executing. The PyEval_AcquireT hread() call blocks until the first call
returns. I was hoping that the python system itself would release the
GIL after some execution, but it itsnt.

I am dependent upon the ability to have to threads executing in python
land at the same time. How can this be done?

PyGILState_Ensu re/Release guarantees to establish the GIL - even if it was already held (useful if you deal with complex call ordering/dependencies)
PyObject_CallOb ject(...) calls Python code. Thus the interpreter will switch and do as usuall during that.
In your/C/system... code you are responsible to release the GIL or not to enable other Python threads (and prevent from deadlocks)
Usually you'd do this
* if you do time consuming C/system stuff
* or if the code can possibly renter Python through the system (e.g. when you call a Windows function which itself can create Windows messages to be routed back into Python message handlers)
-robert

[Svein Seldal]

robert wrote:

PyGILState_Ensu re/Release guarantees to establish the GIL - even if it
was already held (useful if you deal with complex call
ordering/dependencies)

I understand this to mean that I dont need to explicitly lock the GIL
(with PyEval_AcquireL ock() or PyEval_AcquireT hread()).

Well I cant figure out the PyGILState_Ensu re() because if I only use
this function to establish the GIL, when calling python, python will
shortly after crash with "Fatal Python error: ceval: tstate mix-up".
This happens consistently when the main app and the extra thread has
called python and both are busy executing python code.

PyObject_CallOb ject(...) calls Python code. Thus the interpreter
will switch and do as usuall during that.

BTW What do you mean by switch?

In your/C/system... code you are responsible to release the GIL or not
to enable other Python threads (and prevent from deadlocks) Usually
you'd do this * if you do time consuming C/system stuff
* or if the code can possibly renter Python through the system (e.g.
when you call a Windows function which itself can create Windows
messages to be routed back into Python message handlers)

The main problem is that not done this way, it's the other way around.
My main C app will call a python function which will be a lengthy time
consuming process.

The main C app will call python and it will never return from
PyObject_CallOb ject(...), while the extra thread should call a py
function to deliver occational messages into C.
Svein

[Farshid Lashkari]

Svein Seldal wrote:

I'm unable to get access to python as long as another python call is
executing. The PyEval_AcquireT hread() call blocks until the first call
returns. I was hoping that the python system itself would release the
GIL after some execution, but it itsnt.

I am dependent upon the ability to have to threads executing in python
land at the same time. How can this be done?

Are you creating your threads through python or through C code? If you
are only creating it through C code, then make sure you initialize the
GIL with the following call at the beginning of your application:

PyEval_InitThre ads()

From my experience, calling PyGILState_Ensu re() was enough. I didn't
need to call any extra threading functions. But make sure that every
call to PyGILState_Ensu re() is matched with a call to PyGILState_Rele ase()

-Farshid

[robert]

Svein Seldal wrote:

robert wrote:

>PyGILState_Ens ure/Release guarantees to establish the GIL - even if it
was already held (useful if you deal with complex call
ordering/dependencies)

I understand this to mean that I dont need to explicitly lock the GIL
(with PyEval_AcquireL ock() or PyEval_AcquireT hread()).

Well I cant figure out the PyGILState_Ensu re() because if I only use
this function to establish the GIL, when calling python, python will
shortly after crash with "Fatal Python error: ceval: tstate mix-up".
This happens consistently when the main app and the extra thread has
called python and both are busy executing python code.

Do did't tell enough to see what you want to do on the big picture.

usually you create a thread through Python.
If you boot Python at all from C you have to do
PyEval_InitThre ads
or maybe more simple use the high level layer PyRun_...

If you boot a new Python thread manually or a separated interpreter (with separated module state) from outside
the you have to also do:

(PyInterpreterS tate* PyInterpreterSt ate_New() )
PyThreadState* PyThreadState_N ew( PyInterpreterSt ate *interp)
void PyEval_AcquireT hread( PyThreadState *tstate)
If you are already in a Python thread but don't for some reason not know the current thread state:
PyThreadState* PyThreadState_G et( )
but usually you release with PyThreadState* PyEval_SaveThre ad( ) / Py_BEGIN_ALLOW_ THREADS

>PyObject_CallO bject(...) calls Python code. Thus the interpreter
will switch and do as usuall during that.

BTW What do you mean by switch?

the automatic scheduling every sys.getcheckint erval() etc.. you mentioned
You'd only have to take care for that, if your C-code does things (e.g. calling python basic funcs) for a long time and doesn't release the lock with Py_BEGIN_ALLOW_ THREADS ...

The main problem is that not done this way, it's the other way around.
My main C app will call a python function which will be a lengthy time
consuming process.

The main C app will call python and it will never return from
PyObject_CallOb ject(...), while the extra thread should call a py
function to deliver occational messages into C.

Maybe simply boot Python in the main thread (Py_Initialize( )) and run off Python possibly as simple as PyRun_String
and let Python do thread.start_ne w(...)
You'd not have to worry much about states.

( and later you'd probably even more simple do "python myapp.py" and expose your C app as extension module :-) )
-robert

[Hendrik van Rooyen]

"Svein Seldal" <"svein at seldal dot com">wrote:
8<---------------------------------------------------

I am dependent upon the ability to have to threads executing in python
land at the same time. How can this be done?

call time.sleep(0.00 1) in each, as well as the main thread, to politely give the
rest a chance....

Fiddle with the sleep times in your setup to get the best performance - and
don't be misled into thinking that faster is better - its not always true -
there comes a point where the overhead of swapping a thread consumes the total
resource of the process, and no work gets done.

- HTH

- Hendrik

[Svein Seldal]

Hi!

I think I've found the bug, but I need to confirm this behavior.

My findings is that if you use PyEval_InitThre ads(), it is crucial to
release the GIL with PyEval_ReleaseT hread() afterwards.

The API docs states that you can release the GIL with
PyEval_ReleaseL ock() or PyEval_ReleaseT hread() AFAICS.
http://docs.python.org/api/threads.html under docs of void
PyEval_InitThre ads().

However, if I do use PyEval_ReleaseL ock() it will crash shortly after
with "Fatal Python error: ceval: tstate mix-up" in a multithreaded C
environment.

If I use PyEval_ReleaseT hread() to release the GIL, my app seems stable.
I release the GIL like this:

PyThreadState *pts = PyGILState_GetT hisThreadState( );
PyEval_ReleaseT hread(pts);

Now, is this a feature or expected behavior in python (i.e. unclear API
documentation), or is this a bug in python itself?
Regards,
Svein
PS:

For reference I did something like this in pseudo-code:

Py_Initialize() ;
PyEval_InitThre ads();

other_py_inits( ); // Load py modules, etc.

PyEval_ReleaseL ock(); // <-- MAKES THE APP UNSTABLE

create_c_thread ();

PyGILState_STAT E gstate;
gstate = PyGILState_Ensu re();

call_py_functio n_main(); // Py main() wont return

PyGILState_Rele ase(gstate);
And the "main" of the C thread function looks like this:
while(1)
{
PyGILState_STAT E gstate;
gstate = PyGILState_Ensu re();

call_py_functio n_send();

PyGILState_Rele ase(gstate);
}

[robert]

Svein Seldal wrote:

You seem to use the functions in very rude manner - wanting to force the GIL around at lowest level.
Better forget the word GIL and think about acquiring and releasing Threads.
For each thread wanting to execute Python stuff there has to be a thread state (ts). And then you have to enter/leave for executing Python by using the thread state.
PyEval_InitThre ads enters Python (incl. the GIL) first time.
To let Python free in main thread do
ts=PyEval_SaveT hread() // or more easy: Py_BEGIN_ALLOW_ THREADS

to reenter Python do simply:

PyEval_RestoreT hread(ts) // or: Py_END_ALLOW_TH READS

Forget all the low level PyGIL... functions.

when you really want another c thread to enter Python first time with no thread state existing through Python itsel (better make the thread in Python with thread.start_ne w ?), then you have to do once

ts_cthread = PyThreadState_N ew(interp)

then enter Python in this thread:

PyEval_AcquireT hread( ts_cthread )

to leave again:

PyEval_ReleaseT hread( ts_cthread )
If you just loop the call_py_functio n_send() mainly in this thread you probably better create the thread at all in Python and make the loop there. You probably stick too much to C-level at any price :-) Probably you just do a PyRun_xxxx in main thread and then everything else in Python, and expose C-parts for the thread-loop to Python as function (in other_py_inits) - where in the c-function you probably have the usual Py_BEGIN_ALLOW_ THREADS/Py_END_ALLOW_TH READS bracket during time consuming/system/reenter-endangered message stuff.
Robert

I think I've found the bug, but I need to confirm this behavior.

My findings is that if you use PyEval_InitThre ads(), it is crucial to
release the GIL with PyEval_ReleaseT hread() afterwards.

The API docs states that you can release the GIL with
PyEval_ReleaseL ock() or PyEval_ReleaseT hread() AFAICS.
http://docs.python.org/api/threads.html under docs of void
PyEval_InitThre ads().

However, if I do use PyEval_ReleaseL ock() it will crash shortly after
with "Fatal Python error: ceval: tstate mix-up" in a multithreaded C
environment.

If I use PyEval_ReleaseT hread() to release the GIL, my app seems stable.
I release the GIL like this:

PyThreadState *pts = PyGILState_GetT hisThreadState( );
PyEval_ReleaseT hread(pts);

Now, is this a feature or expected behavior in python (i.e. unclear API
documentation), or is this a bug in python itself?
Regards,
Svein
PS:

For reference I did something like this in pseudo-code:

Py_Initialize() ;
PyEval_InitThre ads();

other_py_inits( ); // Load py modules, etc.

PyEval_ReleaseL ock(); // <-- MAKES THE APP UNSTABLE

create_c_thread ();

PyGILState_STAT E gstate;
gstate = PyGILState_Ensu re();

call_py_functio n_main(); // Py main() wont return

PyGILState_Rele ase(gstate);
And the "main" of the C thread function looks like this:
while(1)
{
PyGILState_STAT E gstate;
gstate = PyGILState_Ensu re();

call_py_functio n_send();

PyGILState_Rele ase(gstate);
}

[Svein Seldal]

robert wrote:

Forget all the low level PyGIL... functions.'

Quite the contrary, I would say!

From the Py C API doc, I interpret the PyGIL... functions to be higher
leveled than eg. PyEval_SaveThre ad(). I've checked into the source of
python to find out what's really going on, and I've learnt a couple of
interesting facts. The PyGIL... are helper functions build over the
PyEval... functions, not the other way around.

A real beauty about the PyGILState_Ensu re() is the fact that if it's
called from a thread that is (yet) unknown to python, it will actually
create the thread state object for you (with PyThreadState_N ew) and
acquire the lock. Hence, I dont have to worry about the process of
swapping thread states with PyEval_xxxThrea d() functions. I ONLY need to
use PyGILState_Ensu re() prior to any py-ops.

From my previous post, I asked about the difference between
PyEval_ReleaseL ock() and PyEval_ReleaseT hread() and I've found one major
difference. Both release the GIL as the docs states, but the
PyEval_ReleaseT hread() saves the current thread state as well. When I
used PyEval_ReleaseL ock() this didnt happen, preventing proper saving of
the current thread, causing py crash when control were handed back to
this thread.

By using:

PyEval_InitThre ads();
py_ops();
PyThreadState *pts = PyGILState_GetT hisThreadState( );
PyEval_ReleaseT hread(pts);

And for each py-op later on (from arbitrary thread):

PyGILState_STAT E gstate;
gstate = PyGILState_Ensu re();
py_ops();
PyGILState_Rele ase(gstate);

Then you're home free. Everything related to threading is handled by
Py/API itself.

Probably you just do a PyRun_xxxx in main thread and then everything
else in Python, and expose C-parts for the thread-loop to Python as
function (in other_py_inits) - where in the c-function you probably have
the usual Py_BEGIN_ALLOW_ THREADS/Py_END_ALLOW_TH READS bracket during
time consuming/system/reenter-endangered message stuff.

The app i'm building is a plugin to a server system, and I cant control
nor remove any threads that the server uses. I am given one main thread
to run python forever, and messages that are to be delivered into python
are called from another.

I could do it like you propose: The data coming from my server will
arrive into a c-function called by a server thread. The c-part of the
thread-loop would then be run as another thread (started from python).
Its fully feasible, yet the challenge is to make proper data sync-ing
between these two threads.

Well, I think I've achieved what I wanted. Python seems apparently
stable, and reading from the python sources, I cant see any immediate
reasons why it shouldn't.

Thanks for letting me closer to a working solution!
Regads,
Svein

[robert]

Svein Seldal wrote:

A real beauty about the PyGILState_Ensu re() is the fact that if it's
called from a thread that is (yet) unknown to python, it will actually
create the thread state object for you (with PyThreadState_N ew) and
acquire the lock. Hence, I dont have to worry about the process of
swapping thread states with PyEval_xxxThrea d() functions. I ONLY need to
use PyGILState_Ensu re() prior to any py-ops.

Interesting - the PyGILState_Ensu re was new to py2.3. I didn't notice so far, that it will even auto-create a fresh tread state (from default InterpreterStat e only). From the docs its not that clear.

The confusion in usage is that one has to do

Py_Initialize()
PyEval_InitThre ads()
ts=PyEval_SaveT hread()

in order to start clean for further pure "meta-high-level" PyGILState_Ensu re/Release.

They obviously meant PyGILState_Ensu re to make live easier, but the discussion shows, it adds additional worries probably forcing each user into investigating the Python C-code :-)

There should probably a extra init function for this meta-high-level usage like ..

PyGILState_PyIn itThreads() // which would not pre-acquire the lock! - as anyone expects to enter then with PyGILState_Ensu re

.. and these 3 Functions grouped together in the docs for an easy-to-use meta-high-level thread state API.

The current thread state API doc, as you read it from top to bottom now, is in fact totally confusing for anyone who didn't develop Python himself :-)
Robert