Application-level Stackless features

Introduction

PyPy can expose to its user language features similar to the ones present in Stackless Python: the ability to write code in a massively concurrent style. (It does not (any more) offer the ability to run with no recursion depth limit, but the same effect can be achieved indirectly.)

This feature is based on a custom primitive called a continulet. Continulets can be directly used by application code, or it is possible to write (entirely at app-level) more user-friendly interfaces.

Currently PyPy implements greenlets on top of continulets. It also implements (an approximation of) tasklets and channels, emulating the model of Stackless Python.

Continulets are extremely light-weight, which means that PyPy should be able to handle programs containing large amounts of them. However, due to an implementation restriction, a PyPy compiled with --gcrootfinder=shadowstack consumes at least one page of physical memory (4KB) per live continulet, and half a megabyte of virtual memory on 32-bit or a complete megabyte on 64-bit. Moreover, the feature is only available (so far) on x86 and x86-64 CPUs; for other CPUs you need to add a short page of custom assembler to rpython/translator/c/src/stacklet/.

Theory

The fundamental idea is that, at any point in time, the program happens to run one stack of frames (or one per thread, in case of multi-threading). To see the stack, start at the top frame and follow the chain of f_back until you reach the bottom frame. From the point of view of one of these frames, it has a f_back pointing to another frame (unless it is the bottom frame), and it is itself being pointed to by another frame (unless it is the top frame).

The theory behind continulets is to literally take the previous sentence as definition of “an O.K. situation”. The trick is that there are O.K. situations that are more complex than just one stack: you will always have one stack, but you can also have in addition one or more detached cycles of frames, such that by following the f_back chain you run in a circle. But note that these cycles are indeed completely detached: the top frame (the currently running one) is always the one which is not the f_back of anybody else, and it is always the top of a stack that ends with the bottom frame, never a part of these extra cycles.

How do you create such cycles? The fundamental operation to do so is to take two frames and permute their f_back — i.e. exchange them. You can permute any two f_back without breaking the rule of “an O.K. situation”. Say for example that f is some frame halfway down the stack, and you permute its f_back with the f_back of the top frame. Then you have removed from the normal stack all intermediate frames, and turned them into one stand-alone cycle. By doing the same permutation again you restore the original situation.

In practice, in PyPy, you cannot change the f_back of an abitrary frame, but only of frames stored in continulets.

Continulets are internally implemented using stacklets. Stacklets are a bit more primitive (they are really one-shot continuations), but that idea only works in C, not in Python. The basic idea of continulets is to have at any point in time a complete valid stack; this is important e.g. to correctly propagate exceptions (and it seems to give meaningful tracebacks too).

Application level interface

Continulets

A translated PyPy contains by default a module called _continuation exporting the type continulet. A continulet object from this module is a container that stores a “one-shot continuation”. It plays the role of an extra frame you can insert in the stack, and whose f_back can be changed.

To make a continulet object, call continulet() with a callable and optional extra arguments.

Later, the first time you switch() to the continulet, the callable is invoked with the same continulet object as the extra first argument. At that point, the one-shot continuation stored in the continulet points to the caller of switch(). In other words you have a perfectly normal-looking stack of frames. But when switch() is called again, this stored one-shot continuation is exchanged with the current one; it means that the caller of switch() is suspended with its continuation stored in the container, and the old continuation from the continulet object is resumed.

The most primitive API is actually ‘permute()’, which just permutes the one-shot continuation stored in two (or more) continulets.

In more details:

  • continulet(callable, *args, **kwds): make a new continulet. Like a generator, this only creates it; the callable is only actually called the first time it is switched to. It will be called as follows:

    callable(cont, *args, **kwds)
    

    where cont is the same continulet object.

    Note that it is actually cont.__init__() that binds the continulet. It is also possible to create a not-bound-yet continulet by calling explicitly continulet.__new__(), and only bind it later by calling explicitly cont.__init__().

  • cont.switch(value=None, to=None): start the continulet if it was not started yet. Otherwise, store the current continuation in cont, and activate the target continuation, which is the one that was previously stored in cont. Note that the target continuation was itself previously suspended by another call to switch(); this older switch() will now appear to return. The value argument is any object that is carried to the target and returned by the target’s switch().

    If to is given, it must be another continulet object. In that case, performs a “double switch”: it switches as described above to cont, and then immediately switches again to to. This is different from switching directly to to: the current continuation gets stored in cont, the old continuation from cont gets stored in to, and only then we resume the execution from the old continuation out of to.

  • cont.throw(type, value=None, tb=None, to=None): similar to switch(), except that immediately after the switch is done, raise the given exception in the target.

  • cont.is_pending(): return True if the continulet is pending. This is False when it is not initialized (because we called __new__ and not __init__) or when it is finished (because the callable() returned). When it is False, the continulet object is empty and cannot be switch()-ed to.

  • permute(*continulets): a global function that permutes the continuations stored in the given continulets arguments. Mostly theoretical. In practice, using cont.switch() is easier and more efficient than using permute(); the latter does not on its own change the currently running frame.

Genlets

The _continuation module also exposes the generator decorator:

@generator
def f(cont, a, b):
    cont.switch(a + b)
    cont.switch(a + b + 1)

for i in f(10, 20):
    print i

This example prints 30 and 31. The only advantage over using regular generators is that the generator itself is not limited to yield statements that must all occur syntactically in the same function. Instead, we can pass around cont, e.g. to nested sub-functions, and call cont.switch(x) from there.

The generator decorator can also be applied to methods:

class X:
    @generator
    def f(self, cont, a, b):
        ...

Greenlets

Greenlets are implemented on top of continulets in lib_pypy/greenlet.py. See the official documentation of the greenlets.

Note that unlike the CPython greenlets, this version does not suffer from GC issues: if the program “forgets” an unfinished greenlet, it will always be collected at the next garbage collection.

Unimplemented features

The following features (present in some past Stackless version of PyPy) are for the time being not supported any more:

  • Coroutines (could be rewritten at app-level)

  • Continuing execution of a continulet in a different thread (but if it is “simple enough”, you can pickle it and unpickle it in the other thread).

  • Automatic unlimited stack (must be emulated so far)

  • Support for other CPUs than x86 and x86-64

We also do not include any of the recent API additions to Stackless Python, like set_atomic(). Contributions welcome.

Recursion depth limit

You can use continulets to emulate the infinite recursion depth present in Stackless Python and in stackless-enabled older versions of PyPy.

The trick is to start a continulet “early”, i.e. when the recursion depth is very low, and switch to it “later”, i.e. when the recursion depth is high. Example:

from _continuation import continulet

def invoke(_, callable, arg):
    return callable(arg)

def bootstrap(c):
    # this loop runs forever, at a very low recursion depth
    callable, arg = c.switch()
    while True:
        # start a new continulet from here, and switch to
        # it using an "exchange", i.e. a switch with to=.
        to = continulet(invoke, callable, arg)
        callable, arg = c.switch(to=to)

c = continulet(bootstrap)
c.switch()


def recursive(n):
    if n == 0:
        return ("ok", n)
    if n % 200 == 0:
        prev = c.switch((recursive, n - 1))
    else:
        prev = recursive(n - 1)
    return (prev[0], prev[1] + 1)

print recursive(999999)     # prints ('ok', 999999)

Note that if you press Ctrl-C while running this example, the traceback will be built with all recursive() calls so far, even if this is more than the number that can possibly fit in the C stack. These frames are “overlapping” each other in the sense of the C stack; more precisely, they are copied out of and into the C stack as needed.

(The example above also makes use of the following general “guideline” to help newcomers write continulets: in bootstrap(c), only call methods on c, not on another continulet object. That’s why we wrote c.switch(to=to) and not to.switch(), which would mess up the state. This is however just a guideline; in general we would recommend to use other interfaces like genlets and greenlets.)

Stacklets

Continulets are internally implemented using stacklets, which is the generic RPython-level building block for “one-shot continuations”. For more information about them please see the documentation in the C source at rpython/translator/c/src/stacklet/stacklet.h.

The module rpython.rlib.rstacklet is a thin wrapper around the above functions. The key point is that new() and switch() always return a fresh stacklet handle (or an empty one), and switch() additionally consumes one. It makes no sense to have code in which the returned handle is ignored, or used more than once. Note that stacklet.c is written assuming that the user knows that, and so no additional checking occurs; this can easily lead to obscure crashes if you don’t use a wrapper like PyPy’s ‘_continuation’ module.

Theory of composability

Although the concept of coroutines is far from new, they have not been generally integrated into mainstream languages, or only in limited form (like generators in Python and iterators in C#). We can argue that a possible reason for that is that they do not scale well when a program’s complexity increases: they look attractive in small examples, but the models that require explicit switching, for example by naming the target coroutine, do not compose naturally. This means that a program that uses coroutines for two unrelated purposes may run into conflicts caused by unexpected interactions.

To illustrate the problem, consider the following example (simplified code using a theorical coroutine class). First, a simple usage of coroutine:

main_coro = coroutine.getcurrent()    # the main (outer) coroutine
data = []

def data_producer():
    for i in range(10):
        # add some numbers to the list 'data' ...
        data.append(i)
        data.append(i * 5)
        data.append(i * 25)
        # and then switch back to main to continue processing
        main_coro.switch()

producer_coro = coroutine()
producer_coro.bind(data_producer)

def grab_next_value():
    if not data:
        # put some more numbers in the 'data' list if needed
        producer_coro.switch()
    # then grab the next value from the list
    return data.pop(0)

Every call to grab_next_value() returns a single value, but if necessary it switches into the producer function (and back) to give it a chance to put some more numbers in it.

Now consider a simple reimplementation of Python’s generators in term of coroutines:

def generator(f):
    """Wrap a function 'f' so that it behaves like a generator."""
    def wrappedfunc(*args, **kwds):
        g = generator_iterator()
        g.bind(f, *args, **kwds)
        return g
    return wrappedfunc

class generator_iterator(coroutine):
    def __iter__(self):
        return self
    def next(self):
        self.caller = coroutine.getcurrent()
        self.switch()
        return self.answer

def Yield(value):
    """Yield the value from the current generator."""
    g = coroutine.getcurrent()
    g.answer = value
    g.caller.switch()

def squares(n):
    """Demo generator, producing square numbers."""
    for i in range(n):
        Yield(i * i)
squares = generator(squares)

for x in squares(5):
    print x       # this prints 0, 1, 4, 9, 16

Both these examples are attractively elegant. However, they cannot be composed. If we try to write the following generator:

def grab_values(n):
    for i in range(n):
        Yield(grab_next_value())
grab_values = generator(grab_values)

then the program does not behave as expected. The reason is the following. The generator coroutine that executes grab_values() calls grab_next_value(), which may switch to the producer_coro coroutine. This works so far, but the switching back from data_producer() to main_coro lands in the wrong coroutine: it resumes execution in the main coroutine, which is not the one from which it comes. We expect data_producer() to switch back to the grab_next_values() call, but the latter lives in the generator coroutine g created in wrappedfunc, which is totally unknown to the data_producer() code. Instead, we really switch back to the main coroutine, which confuses the generator_iterator.next() method (it gets resumed, but not as a result of a call to Yield()).

Thus the notion of coroutine is not composable. By opposition, the primitive notion of continulets is composable: if you build two different interfaces on top of it, or have a program that uses twice the same interface in two parts, then assuming that both parts independently work, the composition of the two parts still works.

A full proof of that claim would require careful definitions, but let us just claim that this fact is true because of the following observation: the API of continulets is such that, when doing a switch(), it requires the program to have some continulet to explicitly operate on. It shuffles the current continuation with the continuation stored in that continulet, but has no effect outside. So if a part of a program has a continulet object, and does not expose it as a global, then the rest of the program cannot accidentally influence the continuation stored in that continulet object.

In other words, if we regard the continulet object as being essentially a modifiable f_back, then it is just a link between the frame of callable() and the parent frame — and it cannot be arbitrarily changed by unrelated code, as long as they don’t explicitly manipulate the continulet object. Typically, both the frame of callable() (commonly a local function) and its parent frame (which is the frame that switched to it) belong to the same class or module; so from that point of view the continulet is a purely local link between two local frames. It doesn’t make sense to have a concept that allows this link to be manipulated from outside.