Merge pull request #11152 from DeterminateSystems/flip-coroutines

addToStore(): Evaluate on the main stack
This commit is contained in:
Eelco Dolstra 2024-08-20 18:19:49 +02:00 committed by GitHub
commit e1d1eac791
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
5 changed files with 15 additions and 231 deletions

View File

@ -39,8 +39,6 @@
`pkgconfig` and the Boehm garbage collector, and pass the flag
`--enable-gc` to `configure`.
For `bdw-gc` <= 8.2.4 Nix needs a [small patch](https://github.com/NixOS/nix/blob/ac4d2e7b857acdfeac35ac8a592bdecee2d29838/boehmgc-traceable_allocator-public.diff) to be applied.
- The `boost` library of version 1.66.0 or higher. It can be obtained
from the official web site <https://www.boost.org/>.

View File

@ -32,122 +32,6 @@ static void * oomHandler(size_t requested)
throw std::bad_alloc();
}
class BoehmGCStackAllocator : public StackAllocator
{
boost::coroutines2::protected_fixedsize_stack stack{
// We allocate 8 MB, the default max stack size on NixOS.
// A smaller stack might be quicker to allocate but reduces the stack
// depth available for source filter expressions etc.
std::max(boost::context::stack_traits::default_size(), static_cast<std::size_t>(8 * 1024 * 1024))};
// This is specific to boost::coroutines2::protected_fixedsize_stack.
// The stack protection page is included in sctx.size, so we have to
// subtract one page size from the stack size.
std::size_t pfss_usable_stack_size(boost::context::stack_context & sctx)
{
return sctx.size - boost::context::stack_traits::page_size();
}
public:
boost::context::stack_context allocate() override
{
auto sctx = stack.allocate();
// Stacks generally start at a high address and grow to lower addresses.
// Architectures that do the opposite are rare; in fact so rare that
// boost_routine does not implement it.
// So we subtract the stack size.
GC_add_roots(static_cast<char *>(sctx.sp) - pfss_usable_stack_size(sctx), sctx.sp);
return sctx;
}
void deallocate(boost::context::stack_context sctx) override
{
GC_remove_roots(static_cast<char *>(sctx.sp) - pfss_usable_stack_size(sctx), sctx.sp);
stack.deallocate(sctx);
}
};
static BoehmGCStackAllocator boehmGCStackAllocator;
/**
* When a thread goes into a coroutine, we lose its original sp until
* control flow returns to the thread.
* While in the coroutine, the sp points outside the thread stack,
* so we can detect this and push the entire thread stack instead,
* as an approximation.
* The coroutine's stack is covered by `BoehmGCStackAllocator`.
* This is not an optimal solution, because the garbage is scanned when a
* coroutine is active, for both the coroutine and the original thread stack.
* However, the implementation is quite lean, and usually we don't have active
* coroutines during evaluation, so this is acceptable.
*/
void fixupBoehmStackPointer(void ** sp_ptr, void * _pthread_id)
{
void *& sp = *sp_ptr;
auto pthread_id = reinterpret_cast<pthread_t>(_pthread_id);
# ifndef __APPLE__
pthread_attr_t pattr;
# endif
size_t osStackSize;
// The low address of the stack, which grows down.
void * osStackLimit;
void * osStackBase;
# ifdef __APPLE__
osStackSize = pthread_get_stacksize_np(pthread_id);
osStackLimit = pthread_get_stackaddr_np(pthread_id);
# else
if (pthread_attr_init(&pattr)) {
throw Error("fixupBoehmStackPointer: pthread_attr_init failed");
}
# ifdef HAVE_PTHREAD_GETATTR_NP
if (pthread_getattr_np(pthread_id, &pattr)) {
throw Error("fixupBoehmStackPointer: pthread_getattr_np failed");
}
# elif HAVE_PTHREAD_ATTR_GET_NP
if (!pthread_attr_init(&pattr)) {
throw Error("fixupBoehmStackPointer: pthread_attr_init failed");
}
if (!pthread_attr_get_np(pthread_id, &pattr)) {
throw Error("fixupBoehmStackPointer: pthread_attr_get_np failed");
}
# else
# error "Need one of `pthread_attr_get_np` or `pthread_getattr_np`"
# endif
if (pthread_attr_getstack(&pattr, &osStackLimit, &osStackSize)) {
throw Error("fixupBoehmStackPointer: pthread_attr_getstack failed");
}
if (pthread_attr_destroy(&pattr)) {
throw Error("fixupBoehmStackPointer: pthread_attr_destroy failed");
}
# endif
osStackBase = (char *) osStackLimit + osStackSize;
// NOTE: We assume the stack grows down, as it does on all architectures we support.
// Architectures that grow the stack up are rare.
if (sp >= osStackBase || sp < osStackLimit) { // sp is outside the os stack
sp = osStackLimit;
}
}
/* Disable GC while this object lives. Used by CoroutineContext.
*
* Boehm keeps a count of GC_disable() and GC_enable() calls,
* and only enables GC when the count matches.
*/
class BoehmDisableGC
{
public:
BoehmDisableGC()
{
GC_disable();
};
~BoehmDisableGC()
{
GC_enable();
};
};
static inline void initGCReal()
{
/* Initialise the Boehm garbage collector. */
@ -168,24 +52,6 @@ static inline void initGCReal()
GC_set_oom_fn(oomHandler);
StackAllocator::defaultAllocator = &boehmGCStackAllocator;
// TODO: Remove __APPLE__ condition.
// Comment suggests an implementation that works on darwin and windows
// https://github.com/ivmai/bdwgc/issues/362#issuecomment-1936672196
# if GC_VERSION_MAJOR >= 8 && GC_VERSION_MINOR >= 2 && GC_VERSION_MICRO >= 4 && !defined(__APPLE__)
GC_set_sp_corrector(&fixupBoehmStackPointer);
if (!GC_get_sp_corrector()) {
printTalkative("BoehmGC on this platform does not support sp_corrector; will disable GC inside coroutines");
/* Used to disable GC when entering coroutines on macOS */
create_coro_gc_hook = []() -> std::shared_ptr<void> { return std::make_shared<BoehmDisableGC>(); };
}
# else
# warning \
"BoehmGC version does not support GC while coroutine exists. GC will be disabled inside coroutines. Consider updating bdw-gc to 8.2.4 or later."
# endif
/* Set the initial heap size to something fairly big (25% of
physical RAM, up to a maximum of 384 MiB) so that in most cases
we don't need to garbage collect at all. (Collection has a

View File

@ -210,14 +210,16 @@ StorePath Store::addToStore(
fsm = FileSerialisationMethod::NixArchive;
break;
}
auto source = sinkToSource([&](Sink & sink) {
dumpPath(path, sink, fsm, filter);
std::optional<StorePath> storePath;
auto sink = sourceToSink([&](Source & source) {
LengthSource lengthSource(source);
storePath = addToStoreFromDump(lengthSource, name, fsm, method, hashAlgo, references, repair);
if (lengthSource.total >= settings.warnLargePathThreshold)
warn("copied large path '%s' to the store (%s)", path, renderSize(lengthSource.total));
});
LengthSource lengthSource(*source);
auto storePath = addToStoreFromDump(lengthSource, name, fsm, method, hashAlgo, references, repair);
if (lengthSource.total >= settings.warnLargePathThreshold)
warn("copied large path '%s' to the store (%s)", path, renderSize(lengthSource.total));
return storePath;
dumpPath(path, *sink, fsm, filter);
sink->finish();
return storePath.value();
}
void Store::addMultipleToStore(

View File

@ -171,55 +171,6 @@ size_t StringSource::read(char * data, size_t len)
#error Coroutines are broken in this version of Boost!
#endif
/* A concrete datatype allow virtual dispatch of stack allocation methods. */
struct VirtualStackAllocator {
StackAllocator *allocator = StackAllocator::defaultAllocator;
boost::context::stack_context allocate() {
return allocator->allocate();
}
void deallocate(boost::context::stack_context sctx) {
allocator->deallocate(sctx);
}
};
/* This class reifies the default boost coroutine stack allocation strategy with
a virtual interface. */
class DefaultStackAllocator : public StackAllocator {
boost::coroutines2::default_stack stack;
boost::context::stack_context allocate() override {
return stack.allocate();
}
void deallocate(boost::context::stack_context sctx) override {
stack.deallocate(sctx);
}
};
static DefaultStackAllocator defaultAllocatorSingleton;
StackAllocator *StackAllocator::defaultAllocator = &defaultAllocatorSingleton;
std::shared_ptr<void> (*create_coro_gc_hook)() = []() -> std::shared_ptr<void> {
return {};
};
/* This class is used for entry and exit hooks on coroutines */
class CoroutineContext {
/* Disable GC when entering the coroutine without the boehm patch,
* since it doesn't find the main thread stack in this case.
* std::shared_ptr<void> performs type-erasure, so it will call the right
* deleter. */
const std::shared_ptr<void> coro_gc_hook = create_coro_gc_hook();
public:
CoroutineContext() {};
~CoroutineContext() {};
};
std::unique_ptr<FinishSink> sourceToSink(std::function<void(Source &)> fun)
{
struct SourceToSink : FinishSink
@ -241,14 +192,12 @@ std::unique_ptr<FinishSink> sourceToSink(std::function<void(Source &)> fun)
cur = in;
if (!coro) {
CoroutineContext ctx;
coro = coro_t::push_type(VirtualStackAllocator{}, [&](coro_t::pull_type & yield) {
LambdaSource source([&](char *out, size_t out_len) {
coro = coro_t::push_type([&](coro_t::pull_type & yield) {
LambdaSource source([&](char * out, size_t out_len) {
if (cur.empty()) {
yield();
if (yield.get()) {
return (size_t)0;
}
if (yield.get())
throw EndOfFile("coroutine has finished");
}
size_t n = std::min(cur.size(), out_len);
@ -263,20 +212,14 @@ std::unique_ptr<FinishSink> sourceToSink(std::function<void(Source &)> fun)
if (!*coro) { unreachable(); }
if (!cur.empty()) {
CoroutineContext ctx;
(*coro)(false);
}
}
void finish() override
{
if (!coro) return;
if (!*coro) unreachable();
{
CoroutineContext ctx;
if (coro && *coro)
(*coro)(true);
}
if (*coro) unreachable();
}
};
@ -307,8 +250,7 @@ std::unique_ptr<Source> sinkToSource(
size_t read(char * data, size_t len) override
{
if (!coro) {
CoroutineContext ctx;
coro = coro_t::pull_type(VirtualStackAllocator{}, [&](coro_t::push_type & yield) {
coro = coro_t::pull_type([&](coro_t::push_type & yield) {
LambdaSink sink([&](std::string_view data) {
if (!data.empty()) yield(std::string(data));
});
@ -320,7 +262,6 @@ std::unique_ptr<Source> sinkToSource(
if (pos == cur.size()) {
if (!cur.empty()) {
CoroutineContext ctx;
(*coro)();
}
cur = coro->get();

View File

@ -557,27 +557,4 @@ struct FramedSink : nix::BufferedSink
};
};
/**
* Stack allocation strategy for sinkToSource.
* Mutable to avoid a boehm gc dependency in libutil.
*
* boost::context doesn't provide a virtual class, so we define our own.
*/
struct StackAllocator {
virtual boost::context::stack_context allocate() = 0;
virtual void deallocate(boost::context::stack_context sctx) = 0;
/**
* The stack allocator to use in sinkToSource and potentially elsewhere.
* It is reassigned by the initGC() method in libexpr.
*/
static StackAllocator *defaultAllocator;
};
/* Disabling GC when entering a coroutine (without the boehm patch).
mutable to avoid boehm gc dependency in libutil.
*/
extern std::shared_ptr<void> (*create_coro_gc_hook)();
}