Note: The library is under development and not yet complete. This library should not be used in production code.
Latest release: v0.6.3, 2021-06-18.
A 64-bit C/C++ library that aims to implement robust and efficient multi-prompt delimited control.
The implementation is based on in-place growable light-weight gstacks (called gstack
s) which use
virtual memory to enable growing the gstack (up to 8MiB) but start out using
just 4KiB of committed memory. The library is only available
for 64-bit systems (currently Windows, Linux, macOS, and various BSD's)
as smaller systems do not have enough virtual address space.
There are two libraries provided:
libmprompt
: the primitive library that provides multi-prompt delimited control.
The multi-prompt abstraction has well-defined semantics and is the
minimal control abstraction that can be typed with simple types.
As such, we view the libmprompt
library as a good API that should be provided
by the OS or language runtime to provide generic and sound delimited control.
libmpromptx
is the C++ compiled variant that integrates exception handling
where exceptions are propagated correctly through the gstacks.
libmpeff
: a small example library that uses libmprompt
to implement
efficient algebraic effect handlers (with a similar interface as libhandler).
This is an easier abstraction to program with using multi-prompts directly.
Particular aspects:
The goal is to be fully compatible with C/C++ semantics and to be able to link to this library and use the multi-prompt abstraction as is without special considerations for signals, stack addresses, unwinding etc. In particular, this library has address stability: using the in-place growable gstacks (through virtual memory), these stacks are never moved, which ensures addresses to the stack are always valid (in their lexical scope) -- to the C program it still looks as if there is just one large stack. There is also no special function prologue/epilogue needed as with split stacks for example.
The multi-prompt abstraction has a precise semantics and is well-typed. Moreover, it guarantees there is always just one logical active stack (as a chain of gstacks) which allows exceptions to propagate naturally and also provides natural backtraces for any resumed prompt.
A drawback of our implementation is that it requires 64-bit systems in order to have enough virtual address space. Moreover, at miminum 4KiB of memory is committed per (active) prompt. On systems without "overcommit" we use internal gpools to still be able to commit stack space on-demand using a special signal handler. (as an aside, using this on 32-bit systems is not entirely out of reach -- if one reserves 2GiB virtual memory for the stacks, we could have ~16000 64KiB stacks which may well be enough for almost any application).
We aim to support millions of prompts with fast yielding and resuming. If we run
the mp_async_test1M
test we simulate an asynchronous
service which creates a fresh prompt on each "connection", enters it and then suspends it
(simulating waiting for an async result).
Later it is resumed again where it calls a function that consumes 32KiB stack space,
and finally returns. The test simulates 10 million connections with 10000 suspended
prompts at any time:
run 10M connections with 10000 active at a time, each using 32kb stack...
total stack used: 312500.000mb, count=10000000
elapsed: 0.932s, user: 0.883s, sys: 0.049s, rss: 42mb, main rss: 39mb
This is about 10M "connections" per second (single threaded, Ubuntu 20, AMD5950X), where each connection creates a fresh prompt and context switches 4 times.
Enjoy, Daan Leijen and KC Sivaramakrishnan.
Releases:
v0.6.3
: support for apple M1.v0.6.2
: fix abort yields, add triples test.v0.6.1
: build fixes for freeBSD.v0.6
: improved security further, random initialization of the jump guard.v0.5.4
: improved security against stack buffer overflow, simplify creation of multi-shot resumptions.v0.5.2
: better handling of signals, improved Windows page fault handler.v0.4
: initial support for Linux arm64.v0.3
: better backtraces on Windows, support libunwind.v0.2
: improved debugging on macOS with lldb
.v0.1
: initial release.Tested on Linux (x64 and arm64), macOS (x64, arm64), Windows (x64), and freeBSD (x64).
We use cmake
to build:
> mkdir -p out/debug # or out/release
> cd out/debug
> cmake ../..
> make
> ctest .
This will build the libraries libmpromptx.a
and libmpeffx.a
, and run the tests.
Pass the option cmake ../.. -DMP_USE_C=ON
to build the C versions of the libraries
(but these do not handle- or propagate exceptions).
We use Visual Studio 2019 to develop the library -- open the solution
in ide/vs2019/libmprompt.sln
to build and test.
Some known issues are:
gdb
, lldb
: when debugging on Linux (without overcommit) you may see
segmentation fault errors (SEGV
) which happen when demand paging stack memory;
you need to continue through those or set the debugger to ignore them
(enter handle SIGSEGV nostop noprint
in gdb
).
Another workaround is to use overcommit on Linux when debugging which does not use
a signal handler:
mp_config_t cfg = mp_config_default(); cfg.stack_use_overcommit = true; mp_init(&cfg);
lldb
: when debugging on macOS we use an extra thread
to handle Mach exceptions (to avoid a long standing bug in lldb
).
Debug stack traces in Visual Studio (and windbg
) work well most of the time, but sometimes the debugger stops
a backtrace too soon when libmprompt is unable to put a gstack at a lower address than its parent.
// Types
typedef struct mp_prompt_s mp_prompt_t; // resumable "prompts"
typedef struct mp_resume_s mp_resume_t; // abstract resumption
// Function types
typedef void* (mp_start_fun_t)(mp_prompt_t*, void* arg);
typedef void* (mp_yield_fun_t)(mp_resume_t*, void* arg);
// Continue with `fun(p,arg)` under a fresh prompt `p`.
void* mp_prompt(mp_start_fun_t* fun, void* arg);
// Yield back up to a parent prompt `p` and run `fun(r,arg)`
void* mp_yield(mp_prompt_t* p, mp_yield_fun_t* fun, void* arg);
// Resume back to the yield point with a result; can be used at most once.
void* mp_resume(mp_resume_t* resume, void* arg);
void* mp_resume_tail(mp_resume_t* resume, void* arg);
void mp_resume_drop(mp_resume_t* resume);
// Multi-shot resumptions; use with care in combination with linear resources.
mp_resume_t* mp_resume_multi(mp_resume_t* r); // create a fresh multi-shot resumption
mp_resume_t* mp_resume_dup(mp_resume_t* r); // increase ref-count on a multi-shot resumption
// Portable backtrace
int mp_backtrace(void** backtrace, int len);
A nice property of muli-prompts is that there is always a single strand of execution, together with suspended prompts. In contrast to lower level abstractions, like fibers, there is no arbitrary switching between stacks: one can only yield up to a parent prompt (capturing all gstacks up to that prompt) or resume a suspended prompt chain (and restoring all gstacks in that context). As a consequence, the active chain of prompts always form a logical stack and we can have natural propagation of exceptions with proper backtraces.
Here is an example of a backtrace on Linux:
Here a breakpoint was set in code that was resumed where the backtrace continues into the main stack. This is quite nice for debugging compared to callback based programming for example.
Here is a backtrace in the Visual Studio debugger:
(Unfortunately, on Windows, in rare cases a backtrace can still be cut short when libmprompt is unable to place a gstack at a lower address as its parent.)
The semantics of delimited multi-prompt control can be described precisely:
Syntax:
e ::= v ; value
| e e ; application
| yield m v ; yield to a prompt identified by marker `m`
| prompt v ; start a new prompt (passing its marker to `v`)
| @prompt m e ; internal: a prompt frame identified by marker `m`
v ::= x ; variables
| \x. e ; function with parameter `x` (lambda expression)
| ... ; integer constants, primitives (e.g. addition), etc.
Evaluation context:
E ::= [] ; hole
| E e ; evaluate function first
| v E ; and then the argument
| @prompt m E ; we can evaluate under a prompt frame
An evaluation context is an expression with a hole at the current point in the
evalution. It essentially describes the stack+registers where the hole is the current instruction pointer.
We can apply a context to an expression using square bracets E[x]
;
for example, (@prompt m (f (g [])))[x]
becomes @prompt m (f (g x))
.
Operational semantics:
e ----> e'
(STEP) -----------------
E[e] |----> E[e']
We can now keep evaluating inside an expression context using small step transitions:
(APP) (\x. e) v ----> e[x := v] ; beta-rule, application
(PROMPT) prompt v ----> @prompt m (v m) ; install a new prompt with a fresh marker `m`
(RETURN) @prompt m v ----> v ; returning a result discards the prompt frame
(YIELD) @prompt m E[yield m f] ----> f (\x. @prompt m E[x]) ; yield to prompt frame `m`, capturing context `E`
Note how in (YIELD)
we yield with a function f
to a prompt m
. This
continues with executing f
(popping the prompt) but with the argument
\x. @prompt m E[x]
which is the resumption function: calling it will
restore the prompt and the original execution
context E
(!), and resume execution at the original yield location.
For example:
prompt (\x. 1 + yield x (\k. k 41))
|----> @prompt m ((\x. 1 + yield x (\k. k 41)) m) ; fresh marker `m`
|----> @prompt m (1 + yield m (\k. k 41)) ; note: `\k. k 41` is the function that is yielded up
== @prompt m ((1 + [])[yield m (\k. k 41)]) ; yield back up to `m`, capturing E
|----> (\k. k 41) (\x. @prompt m (1 + [])[x]) ; continue with the function of the yield
|----> (\x. @prompt m (1 + [])[x]) 41 ; resume by applying `k`
|----> @prompt m ((1 + [])[41]) ; resumed to the yield with result 41
== @prompt m (1 + 41)
|----> @prompt m 42
|----> 42
In the C implementation, the unique markers m
are simply
represented directly by a mp_prompt_t*
.
At runtime, yielding to a prompt that is not an ancestor (i.e. no
longer in scope), is an error -- just like an unhandled exception.
(Effect type systems, like in Koka, can prevent this situation
statically at compile-time but in our library this is a runtime error).
These primitives are very expressive but can still be typed in in simply typed lambda calculus, and are sound and composable:
prompt :: (Marker a -> a) -> a
yield :: Marker a -> ((b -> a) -> a) -> b
The action given to prompt
gets a marker that has
a type a
, corresponding to the type of the current context a
(the answer type).
When yielding to a marker of type a
, the yielded function has type (b -> a) -> a
,
and must return results of type a
(corresponding to the marker context).
Meanwhile, the passed in resumption function (b -> a)
expects an argument
of type b
to resume back to the yield point. Such simple types cannot be
given for example to any of shift
/reset
, call/cc
, fibers, or co-routines,
which is one aspect why we believe multi-prompt delimited control is preferable.
The growable gstacks are used to make capturing- and resuming
evaluation contexts efficient. Each @prompt m
frame sits
on the top a gstack from which it can yield and resume
in constant time. This can for example be used to create
green thread scheduling, exceptions, iterators, async-await, etc.
For a more in-depth explanation of multi-prompt delimited control, see "Evidence Passing Semantics for Effect Handler", Ningning Xie and Daan Leijen, MSR-TR-2021-5 (pdf).
Each prompt starts a growable gstack and executes from there. For example, we can have:
(gstack 1) (gstack 2) (gstack 3)
|-------------|
| @prompt A |
|-------------|
| |
| ... |
| prompt <------------> |-------------|
| | | @prompt B |
. . |-------------|
. . | |
. . | ... |
| prompt <------------> |------------|
. . | @prompt C |
. . |------------|
. . | 1+ |
| yield B f |<<<
. .
. .
where <<<
is the currently executing statement.
The yield B f
can yield directly to prompt B
by
just switching stacks. The resumption r
is also
just captured as a pointer and execution continues
with f(r)
: (rule (YIELD)
with r
= \x. @prompt B(... @prompt C. 1+[])[x]
)
(gstack 1) (gstack 2) (gstack 3)
|-------------|
| @prompt A |
|-------------|
| |
| ... | (suspended)
| resume_t* r ~~~~~~~~> |-------------|
| f(r) |<<< | @prompt B |
. . |-------------|
. . | |
. . | ... |
| prompt <------------> |------------|
. . | @prompt C |
. . |------------|
. . | 1+ |
| [] |
. .
. .
Later we may want to resume the resumption r
again with
the result 42
: (r(42)
)
(gstack 1) (gstack 2) (gstack 3)
|-------------|
| @prompt A |
|-------------|
| |
| ... | (suspended)
| resume_t* r ~~~~~~~~> |-------------|
| | | @prompt B |
| ... | |-------------|
| resume(r,42)|<<< | |
. . | ... |
. . | prompt <------------> |------------|
. . . . | @prompt C |
. . |------------|
. . | 1+ |
| [] |
. .
Note how we grew the gstack 1 without moving gstack 2 and 3. If we have just one stack, an implementation needs to copy and restore fragments of the stack (which is what libhandler does), but that leads to trouble in C and C++ where stack addresses can temporarily become invalid. With the in-place growable gstacks, objects on the stack are never moved and their addresses stay valid (in their lexical scope).
Again, we can just switch stacks to resume at the original yield location:
(gstack 1) (gstack 2) (gstack 3)
|-------------|
| @prompt A |
|-------------|
| |
| ... |
| resume_t* r ~~~~~+--> |-------------|
| | | | @prompt B |
| ... | | |-------------|
| resume <---------+ | |
| | | ... |
. . | prompt <------------> |------------|
. . . . | @prompt C |
. . |------------|
. . | 1+ |
| 42 |<<<
. .
. .
Suppose, gstack 3 now returns normally with a result 43:
(gstack 1) (gstack 2) (gstack 3)
|-------------|
| @prompt A |
|-------------|
| |
| ... |
| resume_t* r ~~~~~+--> |-------------|
| | | | @prompt B |
| ... | | |-------------|
| resume <---------+ | |
| | | ... |
. . | prompt <------------> |------------|
. . . . | @prompt C |
. . |------------|
. . | 43 |<<<
. .
. .
Then the gstacks can unwind like a regular stack (this is
also how exceptions are propagated): (rule (RETURN)
)
(gstack 1) (gstack 2) (gstack 3)
|-------------|
| @prompt A |
|-------------|
| |
| ... |
| resume_t* r ~~~~~+--> |-------------|
| | | | @prompt B |
| ... | | |-------------|
| resume <---------+ | |
| | | ... |
. . | 43 |<<< (cached to reuse)
. . . . |------------|
. . | |
. . | |
. .
. .
See mprompt.c
for the implementation of this.
Here is an example of running N
"async" workers over M
requests
using resumptions as first-class values stored in the workers
array:
#include <stdlib.h>
#include <stdio.h>
#include <stdint.h>
#include <mprompt.h>
#define N 1000 // max active async workers
#define M 1000000 // total number of requests
static void* await_result(mp_resume_t* r, void* arg) {
(void)(arg);
return r; // instead of resuming ourselves, we return the resumption as a "suspended async computation" (A)
}
static void* async_worker(mp_prompt_t* parent, void* arg) {
(void)(arg);
// start a fresh worker
// ... do some work
intptr_t partial_result = 0;
// and await some request; we do this by yielding up to our prompt and running `await_result` (in the parent context!)
mp_yield( parent, &await_result, NULL );
// when we are resumed at some point, we do some more work
// ... do more work
partial_result++;
// and return with the result (B)
return (void*)(partial_result);
}
static void async_workers(void) {
mp_resume_t** workers = (mp_resume_t**)calloc(N,sizeof(mp_resume_t*)); // allocate array of N resumptions
intptr_t count = 0;
for( int i = 0; i < M+N; i++) { // perform M connections
int j = i % N; // pick an active worker
// if the worker is actively waiting (suspended), resume it
if (workers[j] != NULL) {
count += (intptr_t)mp_resume(workers[j], NULL); // (B)
workers[j] = NULL;
}
// and start a fresh worker and wait for its first yield (suspension).
// the worker returns its own resumption as a result.
if (i < M) {
workers[j] = (mp_resume_t*)mp_prompt( &async_worker, NULL ); // (A)
}
}
printf("ran %zd workers\n", count);
}
int main() {
async_workers();
return 0;
}
A small library on top of libmprompt
that implements
algebraic effect handlers. Effect handlers give more structure
than basic multi-prompts and are a better abstraction for
programming. In particular,
You do not need the particular prompt marker, but always yield to the innermost handler for a particular effect. This is much more convenient and is essential for example model dynamically bound state (much like implicit parameters).
All potential operations are bound statically at the handler and you always yield to a particular operation providing arguments (where the handler definition is basically a v-table with a slot for every operation). This makes it easier to reason about than using a multi-prompt yield which can yield with any arbitrary function.
As effect handlers are linked on the stack, this abstraction can be used across libraries/languages and is thus more composable than using multi-prompts directly.
See effects.c
for many examples of common
effect patterns.
// handle an effect
void* mpe_handle(const mpe_handlerdef_t* hdef, void* local, mpe_actionfun_t* body, void* arg);
// perform an operation
void* mpe_perform(mpe_optag_t optag, void* arg);
// resume from an operation clause (in a mp_handler_def_t)
void* mpe_resume(mpe_resume_t* resume, void* local, void* arg);
void* mpe_resume_final(mpe_resume_t* resume, void* local, void* arg);
void* mpe_resume_tail(mpe_resume_t* resume, void* local, void* arg);
void mpe_resume_release(mpe_resume_t* resume);
Handler definitions:
// An action executing under a handler
typedef void* (mpe_actionfun_t)(void* arg);
// A function than handles an operation receives a resumption
typedef void* (mpe_opfun_t)(mpe_resume_t* r, void* local, void* arg);
// Operation kinds can make resuming more efficient
typedef enum mpe_opkind_e {
...
MPE_OP_TAIL,
MPE_OP_GENERAL
} mpe_opkind_t;
// Operation definition
typedef struct mpe_operation_s {
mpe_opkind_t opkind;
mpe_optag_t optag;
mpe_opfun_t* opfun;
} mpe_operation_t;
// Handler definition
typedef struct mpe_handlerdef_s {
mpe_effect_t effect;
mpe_resultfun_t* resultfun;
...
mpe_operation_t operations[8];
} mpe_handlerdef_t;