Google Git
Sign in
gem5 / public / gem5-resources / 8ccd059d5dcaaeffe15cd93c17f1e76e92907662 / . / src / parsec / disk-image / parsec / parsec-benchmark / pkgs / libs / tbblib / src / src / rml / server / rml_server.cpp
blob: decf62709ef46634feb5b23fc487b38890b07bac [file] [log] [blame]
/*
Copyright 2005-2010 Intel Corporation. All Rights Reserved.
This file is part of Threading Building Blocks.
Threading Building Blocks is free software; you can redistribute it
and/or modify it under the terms of the GNU General Public License
version 2 as published by the Free Software Foundation.
Threading Building Blocks is distributed in the hope that it will be
useful, but WITHOUT ANY WARRANTY; without even the implied warranty
of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with Threading Building Blocks; if not, write to the Free Software
Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
As a special exception, you may use this file as part of a free software
library without restriction. Specifically, if other files instantiate
templates or use macros or inline functions from this file, or you compile
this file and link it with other files to produce an executable, this
file does not by itself cause the resulting executable to be covered by
the GNU General Public License. This exception does not however
invalidate any other reasons why the executable file might be covered by
the GNU General Public License.
*/
#include "rml_tbb.h"
#define private public /* Sleazy trick to avoid publishing internal names in public header. */
#include "rml_omp.h"
#undef private
#include "tbb/tbb_allocator.h"
#include "tbb/cache_aligned_allocator.h"
#include "tbb/aligned_space.h"
#include "tbb/atomic.h"
#include "tbb/spin_mutex.h"
#include "tbb/tbb_misc.h" // Get DetectNumberOfWorkers() from here.
#if _MSC_VER==1500 && !defined(__INTEL_COMPILER)
// VS2008/VC9 seems to have an issue;
#pragma warning( push )
#pragma warning( disable: 4985 )
#endif
#include "tbb/concurrent_vector.h"
#if _MSC_VER==1500 && !defined(__INTEL_COMPILER)
#pragma warning( pop )
#endif
#if _MSC_VER && defined(_Wp64)
// Workaround for overzealous compiler warnings
#pragma warning (push)
#pragma warning (disable: 4244)
#endif
#include "job_automaton.h"
#include "wait_counter.h"
#include "thread_monitor.h"
#if RML_USE_WCRM
#include <concrt.h>
#include <concrtrm.h>
using namespace Concurrency;
#include <vector>
#include <hash_map>
#define __RML_REMOVE_VIRTUAL_PROCESSORS_DISABLED 0
#endif /* RML_USE_WCRM */
#define STRINGIFY(x) #x
#define TOSTRING(x) STRINGIFY(x)
namespace rml {
namespace internal {
//! Number of hardware contexts
static inline unsigned hardware_concurrency() {
static unsigned DefaultNumberOfThreads = 0;
unsigned n = DefaultNumberOfThreads;
if( !n ) DefaultNumberOfThreads = n = tbb::internal::DetectNumberOfWorkers();
return n;
}
using tbb::internal::rml::tbb_client;
using tbb::internal::rml::tbb_server;
using __kmp::rml::omp_client;
using __kmp::rml::omp_server;
typedef versioned_object::version_type version_type;
#define SERVER_VERSION 2
#define EARLIEST_COMPATIBLE_CLIENT_VERSION 2
static const size_t cache_line_size = tbb::internal::NFS_MaxLineSize;
template<typename Server, typename Client> class generic_connection;
class tbb_connection_v2;
class omp_connection_v2;
#if RML_USE_WCRM
//! State of a server_thread
/** Below are diagrams of legal state transitions.
ts_busy
^ ^
/ \
/ V
ts_done <----- ts_asleep <------> ts_idle
*/
enum thread_state_t {
ts_idle,
ts_asleep,
ts_busy,
ts_done
};
//! Extra state of an omp server thread
enum thread_extra_state_t {
ts_none,
ts_removed,
ts_lent
};
//! Results from try_grab_for()
enum thread_grab_t {
wk_failed,
wk_from_asleep,
wk_from_idle
};
#else /* !RML_USE_WCRM */
//! State of a server_thread
/** Below are diagrams of legal state transitions.
OMP
ts_omp_busy
^ ^
/ \
/ V
ts_asleep <-----------> ts_idle
ts_deactivated
^ ^
/ \
V \
ts_none <--------------> ts_reactivated
TBB
ts_tbb_busy
^ ^
/ \
/ V
ts_asleep <-----------> ts_idle --> ts_done
For TBB only. Extra state transition.
ts_created -> ts_started -> ts_visited
*/
enum thread_state_t {
//! Thread not doing anything useful, but running and looking for work.
ts_idle,
//! Thread not doing anything useful and is asleep */
ts_asleep,
//! Thread is enlisted into OpenMP team
ts_omp_busy,
//! Thread is busy doing TBB work.
ts_tbb_busy,
//! For tbb threads only
ts_done,
ts_created,
ts_started,
ts_visited,
//! For omp threads only
ts_none,
ts_deactivated,
ts_reactivated
};
#endif /* RML_USE_WCRM */
#if TBB_USE_ASSERT
#define PRODUCE_ARG(x) ,x
#else
#define PRODUCE_ARG(x)
#endif /* TBB_USE_ASSERT */
//! Synchronizes dispatch of OpenMP work.
class omp_dispatch_type {
typedef ::rml::job job_type;
omp_client* client;
void* cookie;
omp_client::size_type index;
tbb::atomic<job_type*> job;
#if TBB_USE_ASSERT
omp_connection_v2* server;
#endif /* TBB_USE_ASSERT */
public:
omp_dispatch_type() {job=NULL;}
void consume();
void produce( omp_client& c, job_type& j, void* cookie_, omp_client::size_type index_ PRODUCE_ARG( omp_connection_v2& s )) {
__TBB_ASSERT( &j, NULL );
__TBB_ASSERT( !job, "job already set" );
client = &c;
#if TBB_USE_ASSERT
server = &s;
#endif /* TBB_USE_ASSERT */
cookie = cookie_;
index = index_;
// Must be last
job = &j;
}
};
//! A reference count.
/** No default constructor, because users of ref_count must be very careful about whether the
initial reference count is 0 or 1. */
class ref_count: no_copy {
friend class thread_map;
tbb::atomic<int> my_ref_count;
public:
ref_count(int k ) {my_ref_count=k;}
~ref_count() {__TBB_ASSERT( !my_ref_count, "premature destruction of refcounted object" );}
//! Add one and return new value.
int add_ref() {
int k = ++my_ref_count;
__TBB_ASSERT(k>=1,"reference count underflowed before add_ref");
return k;
}
//! Subtract one and return new value.
int remove_ref() {
int k = --my_ref_count;
__TBB_ASSERT(k>=0,"reference count underflow");
return k;
}
};
#if RML_USE_WCRM
#if USE_UMS_THREAD
#define RML_THREAD_KIND UmsThreadDefault
#define RML_THREAD_KIND_STRING "UmsThread"
#else
#define RML_THREAD_KIND ThreadScheduler
#define RML_THREAD_KIND_STRING "WinThread"
#endif
// Forward declaration
class thread_map;
static const IExecutionResource* c_remove_prepare = (IExecutionResource*)0;
static const IExecutionResource* c_remove_returned = (IExecutionResource*)1;
//! Server thread representation
class server_thread_rep : no_copy {
friend class thread_map;
friend class omp_connection_v2;
friend class server_thread;
friend class tbb_server_thread;
friend class omp_server_thread;
template<typename Connection> friend void make_job( Connection& c, typename Connection::server_thread_type& t );
typedef int thread_state_rep_t;
public:
//! Ctor
server_thread_rep( bool assigned, IScheduler* s, IExecutionResource* r, thread_map& map, rml::client& cl ) :
uid( GetExecutionContextId() ), my_scheduler(s), my_proxy(NULL),
my_thread_map(map), my_client(cl), my_job(NULL)
{
my_state = assigned ? ts_busy : ts_idle;
my_extra_state = ts_none;
terminate = false;
my_execution_resource = r;
}
//! Dtor
~server_thread_rep() {}
//! Synchronization routine
inline rml::job* wait_for_job() {
if( !my_job ) my_job = &my_job_automaton.wait_for_job();
return my_job;
}
// Getters and setters
inline thread_state_t read_state() const { thread_state_rep_t s = my_state; return static_cast<thread_state_t>(s); }
inline void set_state( thread_state_t to ) {my_state = to;}
inline void set_removed() { __TBB_ASSERT( my_extra_state==ts_none, NULL ); my_extra_state = ts_removed; }
inline bool is_removed() const { return my_extra_state==ts_removed; }
inline bool is_lent() const {return my_extra_state==ts_lent;}
inline void set_lent() { my_extra_state=ts_lent; }
inline void set_returned() { my_extra_state=ts_none; }
inline IExecutionResource* get_execution_resource() { return my_execution_resource; }
inline IVirtualProcessorRoot* get_virtual_processor() { return (IVirtualProcessorRoot*)get_execution_resource(); }
//! Enlist the thread for work
inline bool wakeup( thread_state_t to, thread_state_t from ) {
__TBB_ASSERT( from==ts_asleep && (to==ts_idle||to==ts_busy||to==ts_done), NULL );
return my_state.compare_and_swap( to, from )==from;
}
//! Enlist the thread for.
thread_grab_t try_grab_for();
//! Destroy the client job associated with the thread
template<typename Connection> bool destroy_job( Connection* c );
//! Try to re-use the thread
void revive( IScheduler* s, IExecutionResource* r, rml::client& c ) {
// the variables may not have been set before a thread was told to quit
__TBB_ASSERT( my_scheduler==s, "my_scheduler has been altered?\n" );
my_scheduler = s;
__TBB_ASSERT( &my_client==&c, "my_client has been altered?\n" );
if( r ) my_execution_resource = r;
my_client = c;
my_state = ts_idle;
__TBB_ASSERT( my_extra_state==ts_removed, NULL );
my_extra_state = ts_none;
}
protected:
const int uid;
IScheduler* my_scheduler;
IThreadProxy* my_proxy;
tbb::atomic<IExecutionResource*> my_execution_resource; /* for non-masters, it is IVirtualProcessorRoot */
thread_map& my_thread_map;
rml::client& my_client;
job* my_job;
job_automaton my_job_automaton;
tbb::atomic<bool> terminate;
tbb::atomic<thread_state_rep_t> my_state;
tbb::atomic<thread_extra_state_t> my_extra_state;
};
//! Class that implements IExecutionContext
class server_thread : public IExecutionContext, public server_thread_rep {
friend class tbb_connection_v2;
friend class omp_connection_v2;
friend class tbb_server_thread;
friend class omp_server_thread;
friend class thread_map;
template<typename Connection> friend void make_job( Connection& c, typename Connection::server_thread_type& t );
protected:
server_thread( bool is_tbb, bool assigned, IScheduler* s, IExecutionResource* r, thread_map& map, rml::client& cl ) : server_thread_rep(assigned,s,r,map,cl), tbb_thread(is_tbb) {}
~server_thread() {}
/*override*/ unsigned int GetId() const { return uid; }
/*override*/ IScheduler* GetScheduler() { return my_scheduler; }
/*override*/ IThreadProxy* GetProxy() { return my_proxy; }
/*override*/ void SetProxy( IThreadProxy* thr_proxy ) { my_proxy = thr_proxy; }
private:
bool tbb_thread;
};
// Forward declaration
class tbb_connection_v2;
class omp_connection_v2;
//! TBB server thread
class tbb_server_thread : public server_thread {
friend class tbb_connection_v2;
public:
tbb_server_thread( bool assigned, IScheduler* s, IExecutionResource* r, tbb_connection_v2* con, thread_map& map, rml::client& cl ) : server_thread(true,assigned,s,r,map,cl), my_conn(con) {
activation_count = 0;
}
~tbb_server_thread() {}
/*override*/ void Dispatch( DispatchState* );
inline bool initiate_termination();
bool sleep_perhaps();
//! Switch out this thread
bool switch_out();
private:
tbb_connection_v2* my_conn;
public:
tbb::atomic<int> activation_count;
};
//! OMP server thread
class omp_server_thread : public server_thread {
friend class omp_connection_v2;
public:
omp_server_thread( bool assigned, IScheduler* s, IExecutionResource* r, omp_connection_v2* con, thread_map& map, rml::client& cl ) :
server_thread(false,assigned,s,r,map,cl), my_conn(con), my_cookie(NULL), my_index(UINT_MAX) {}
~omp_server_thread() {}
/*override*/ void Dispatch( DispatchState* );
inline void* get_cookie() {return my_cookie;}
inline ::__kmp::rml::omp_client::size_type get_index() {return my_index;}
inline IExecutionResource* get_execution_resource() { return get_execution_resource(); }
inline bool initiate_termination() { return destroy_job( (omp_connection_v2*) my_conn ); }
void sleep_perhaps();
private:
omp_connection_v2* my_conn;
void* my_cookie;
::__kmp::rml::omp_client::size_type my_index;
omp_dispatch_type omp_data;
};
//! Class that implements IScheduler
template<typename Connection>
class scheduler : no_copy, public IScheduler {
public:
/*override*/ unsigned int GetId() const {return uid;}
/*override*/ void Statistics( unsigned int* /*pTaskCompletionRate*/, unsigned int* /*pTaskArrivalRate*/, unsigned int* /*pNumberOfTaskEnqueued*/) {}
/*override*/ SchedulerPolicy GetPolicy() const { __TBB_ASSERT(my_policy,NULL); return *my_policy; }
/*override*/ void AddVirtualProcessors( IVirtualProcessorRoot** vproots, unsigned int count ) { if( !my_conn.is_closing() ) my_conn.add_virtual_processors( vproots, count); }
/*override*/ void RemoveVirtualProcessors( IVirtualProcessorRoot** vproots, unsigned int count );
/*override*/ void NotifyResourcesExternallyIdle( IVirtualProcessorRoot** vproots, unsigned int count ) { __TBB_ASSERT( false, "This call is not allowed for TBB" ); }
/*override*/ void NotifyResourcesExternallyBusy( IVirtualProcessorRoot** vproots, unsigned int count ) { __TBB_ASSERT( false, "This call is not allowed for TBB" ); }
protected:
scheduler( Connection& conn );
virtual ~scheduler() { __TBB_ASSERT( my_policy, NULL ); delete my_policy; }
public:
static scheduler* create( Connection& conn ) {return new scheduler( conn );}
private:
const int uid;
Connection& my_conn;
SchedulerPolicy* my_policy;
};
/*
* --> ts_busy --> ts_done
*/
class thread_scavenger_thread : public IExecutionContext, no_copy {
public:
thread_scavenger_thread( IScheduler* s, IVirtualProcessorRoot* r, thread_map& map ) :
uid( GetExecutionContextId() ), my_scheduler(s), my_virtual_processor_root(r), my_proxy(NULL), my_thread_map(map)
{
my_state = ts_busy;
#if TBB_USE_ASSERT
activation_count = 0;
#endif
}
~thread_scavenger_thread() {}
/*override*/ unsigned int GetId() const { return uid; }
/*override*/ IScheduler* GetScheduler() { return my_scheduler; }
/*override*/ IThreadProxy* GetProxy() { return my_proxy; }
/*override*/ void SetProxy( IThreadProxy* thr_proxy ) { my_proxy = thr_proxy; }
/*override*/ void Dispatch( DispatchState* );
inline thread_state_t read_state() { return my_state; }
inline void set_state( thread_state_t s ) { my_state = s; }
inline IVirtualProcessorRoot* get_virtual_processor() { return my_virtual_processor_root; }
private:
const int uid;
IScheduler* my_scheduler;
IVirtualProcessorRoot* my_virtual_processor_root;
IThreadProxy* my_proxy;
thread_map& my_thread_map;
tbb::atomic<thread_state_t> my_state;
#if TBB_USE_ASSERT
public:
tbb::atomic<int> activation_count;
#endif
};
static const thread_scavenger_thread* c_claimed = reinterpret_cast<thread_scavenger_thread*>(1);
struct garbage_connection_queue {
tbb::atomic<uintptr_t> head;
tbb::atomic<uintptr_t> tail;
static const uintptr_t empty = 0; // connection scavenger thread empty list
static const uintptr_t plugged = 1; // end of use of the list
static const uintptr_t plugged_acked = 2; // connection scavenger saw the plugged flag, and it freed all connections
};
//! Connection scavenger
/** It collects closed connection objects, wait for worker threads belonging to the connection to return to ConcRT RM
* then return the object to the memory manager.
*/
class connection_scavenger_thread {
friend void assist_cleanup_connections();
/*
* connection_scavenger_thread's state
* ts_busy <----> ts_asleep <--
*/
tbb::atomic<thread_state_t> state;
/* We steal two bits from a connection pointer to encode
* whether the connection is for TBB or for OMP.
*
* ----------------------------------
* | | | |
* ----------------------------------
* ^ ^
* / |
* 1 : tbb, 0 : omp |
* if set, terminate
*/
// FIXME: pad these?
thread_monitor monitor;
int default_concurrency;
HANDLE thr_handle;
#if TBB_USE_ASSERT
tbb::atomic<int> n_scavenger_threads;
#endif
public:
connection_scavenger_thread() : thr_handle(NULL) {
state = ts_asleep;
#if TBB_USE_ASSERT
n_scavenger_threads = 0;
#endif
}
~connection_scavenger_thread() {}
void wakeup() {
if( state.compare_and_swap( ts_busy, ts_asleep )==ts_asleep )
monitor.notify();
}
void sleep_perhaps();
void process_requests( uintptr_t conn_ex );
static __RML_DECL_THREAD_ROUTINE thread_routine( void* arg );
void launch( int dc ) {
default_concurrency = dc;
thread_monitor::launch( connection_scavenger_thread::thread_routine, this, NULL );
}
template<typename Server, typename Client>
void add_request( generic_connection<Server,Client>* conn_to_close );
template<typename Server, typename Client>
uintptr_t grab_and_prepend( generic_connection<Server,Client>* last_conn_to_close );
};
void free_all_connections( uintptr_t );
#endif /* RML_USE_WCRM */
#if !RML_USE_WCRM
class server_thread;
//! thread_map_base; we need to make the iterator type available to server_thread
struct thread_map_base {
//! A value in the map
class value_type {
public:
server_thread& thread() {
__TBB_ASSERT( my_thread, "thread_map::value_type::thread() called when !my_thread" );
return *my_thread;
}
rml::job& job() {
__TBB_ASSERT( my_job, "thread_map::value_type::job() called when !my_job" );
return *my_job;
}
value_type() : my_thread(NULL), my_job(NULL) {}
server_thread& wait_for_thread() const {
for(;;) {
server_thread* ptr=const_cast<server_thread*volatile&>(my_thread);
if( ptr )
return *ptr;
__TBB_Yield();
}
}
/** Shortly after when a connection is established, it is possible for the server
to grab a server_thread that has not yet created a job object for that server. */
rml::job& wait_for_job() const {
if( !my_job ) {
my_job = &my_automaton.wait_for_job();
}
return *my_job;
}
private:
server_thread* my_thread;
/** Marked mutable because though it is physically modified, conceptually it is a duplicate of
the job held by job_automaton. */
mutable rml::job* my_job;
job_automaton my_automaton;
// FIXME - pad out to cache line, because my_automaton is hit hard by thread()
friend class thread_map;
};
typedef tbb::concurrent_vector<value_type,tbb::zero_allocator<value_type,tbb::cache_aligned_allocator> > array_type;
};
#endif /* !RML_USE_WCRM */
#if _MSC_VER && !defined(__INTEL_COMPILER)
// Suppress overzealous compiler warnings about uninstantiatble class
#pragma warning(push)
#pragma warning(disable:4510 4610)
#endif
template<typename T>
class padded: public T {
char pad[cache_line_size - sizeof(T)%cache_line_size];
};
#if _MSC_VER && !defined(__INTEL_COMPILER)
#pragma warning(pop)
#endif
// FIXME - should we pad out memory to avoid false sharing of our global variables?
static tbb::atomic<int> the_balance;
static tbb::atomic<int> the_balance_inited;
#if !RML_USE_WCRM
//! Per thread information
/** ref_count holds number of clients that are using this,
plus 1 if a host thread owns this instance. */
class server_thread: public ref_count {
friend class thread_map;
template<typename Server, typename Client> friend class generic_connection;
friend class tbb_connection_v2;
friend class omp_connection_v2;
//! Integral type that can hold a thread_state_t
typedef int thread_state_rep_t;
tbb::atomic<thread_state_rep_t> state;
public:
thread_monitor monitor;
private:
bool is_omp_thread;
tbb::atomic<thread_state_rep_t> my_extra_state;
server_thread* link;
thread_map_base::array_type::iterator my_map_pos;
rml::server *my_conn;
rml::job* my_job;
job_automaton* my_ja;
size_t my_index;
tbb::atomic<bool> terminate;
omp_dispatch_type omp_dispatch;
#if TBB_USE_ASSERT
//! Flag used to check if thread is still using *this.
bool has_active_thread;
#endif /* TBB_USE_ASSERT */
//! Volunteer to sleep.
void sleep_perhaps( thread_state_t asleep );
//! Destroy job corresponding to given client
/** Return true if thread must quit. */
template<typename Connection>
bool destroy_job( Connection& c );
//! Do terminate the thread
/** Return true if thread must quit. */
bool do_termination();
void loop();
static __RML_DECL_THREAD_ROUTINE thread_routine( void* arg );
public:
server_thread();
~server_thread();
//! Read the thread state
thread_state_t read_state() const {
thread_state_rep_t s = state;
__TBB_ASSERT( unsigned(s)<=unsigned(ts_done), "corrupted server thread?" );
return thread_state_t(s);
}
//! Read the tbb-specific extra thread state
thread_state_t read_extra_state() const {
thread_state_rep_t s = my_extra_state;
return thread_state_t(s);
}
//! Launch a thread that is bound to *this.
void launch( size_t stack_size );
//! Attempt to wakeup a thread
/** The value "to" is the new state for the thread, if it was woken up.
Returns true if thread was woken up, false otherwise. */
bool wakeup( thread_state_t to, thread_state_t from );
//! Attempt to enslave a thread for OpenMP/TBB.
/** Returns true if state is successfully changed. 's' takes either ts_omp_busy or ts_tbb_busy */
bool try_grab_for( thread_state_t s );
#if _WIN32||_WIN64
//! Send the worker thread to sleep temporarily
void deactivate();
//! Wake the worker thread up
void reactivate();
#endif /* _WIN32||_WIN64 */
};
//! Bag of threads that are private to a client.
class private_thread_bag {
struct list_thread: server_thread {
list_thread* next;
};
//! Root of atomic linked list of list_thread
/** ABA problem is avoided because items are only atomically pushed, never popped. */
tbb::atomic<list_thread*> my_root;
tbb::cache_aligned_allocator<padded<list_thread> > my_allocator;
public:
//! Construct empty bag
private_thread_bag() {my_root=NULL;}
//! Create a fresh server_thread object.
server_thread& add_one_thread() {
list_thread* t = my_allocator.allocate(1);
new( t ) list_thread;
// Atomically add to list
list_thread* old_root;
do {
old_root = my_root;
t->next = old_root;
} while( my_root.compare_and_swap( t, old_root )!=old_root );
return *t;
}
//! Destroy the bag and threads in it.
~private_thread_bag() {
while( my_root ) {
// Unlink thread from list.
list_thread* t = my_root;
my_root = t->next;
// Destroy and deallocate the thread.
t->~list_thread();
my_allocator.deallocate(static_cast<padded<list_thread>*>(t),1);
}
}
};
//! Forward declaration
void wakeup_some_tbb_threads();
//! Type-independent part of class generic_connection.
/** One to one map from server threads to jobs, and associated reference counting. */
class thread_map : public thread_map_base {
public:
typedef rml::client::size_type size_type;
//! ctor
thread_map( wait_counter& fc, ::rml::client& client ) :
all_visited_at_least_once(false), my_min_stack_size(0), my_server_ref_count(1),
my_client_ref_count(1), my_client(client), my_factory_counter(fc)
{ my_unrealized_threads = 0; }
//! dtor
~thread_map() {}
typedef array_type::iterator iterator;
iterator begin() {return my_array.begin();}
iterator end() {return my_array.end();}
void bind();
void unbind();
void assist_cleanup( bool assist_null_only );
/** Returns number of unrealized threads to create. */
size_type wakeup_tbb_threads( size_type n );
bool wakeup_next_thread( iterator i, tbb_connection_v2& conn );
void release_tbb_threads( server_thread* t );
void adjust_balance( int delta );
//! Add a server_thread object to the map, but do not bind it.
/** Return NULL if out of unrealized threads. */
value_type* add_one_thread( bool is_omp_thread_ );
void bind_one_thread( rml::server& server, value_type& x );
void remove_client_ref();
int add_server_ref() {return my_server_ref_count.add_ref();}
int remove_server_ref() {return my_server_ref_count.remove_ref();}
::rml::client& client() const {return my_client;}
size_type get_unrealized_threads() { return my_unrealized_threads; }
private:
private_thread_bag my_private_threads;
bool all_visited_at_least_once;
array_type my_array;
size_t my_min_stack_size;
tbb::atomic<size_type> my_unrealized_threads;
//! Number of threads referencing *this, plus one extra.
/** When it becomes zero, the containing server object can be safely deleted. */
ref_count my_server_ref_count;
//! Number of jobs that need cleanup, plus one extra.
/** When it becomes zero, acknowledge_close_connection is called. */
ref_count my_client_ref_count;
::rml::client& my_client;
//! Counter owned by factory that produced this thread_map.
wait_counter& my_factory_counter;
};
void thread_map::bind_one_thread( rml::server& server, value_type& x ) {
// Add one to account for the thread referencing this map hereforth.
server_thread& t = x.thread();
my_server_ref_count.add_ref();
my_client_ref_count.add_ref();
#if TBB_USE_ASSERT
__TBB_ASSERT( t.add_ref()==1, NULL );
#else
t.add_ref();
#endif
// Have responsibility to start the thread.
t.my_conn = &server;
t.my_ja = &x.my_automaton;
t.launch( my_min_stack_size );
/* Must wake thread up so it can fill in its "my_job" field in *this.
Otherwise deadlock can occur where wait_for_job spins on thread that is sleeping. */
__TBB_ASSERT( t.state!=ts_tbb_busy, NULL );
t.wakeup( ts_idle, ts_asleep );
}
thread_map::value_type* thread_map::add_one_thread( bool is_omp_thread_ ) {
size_type u;
do {
u = my_unrealized_threads;
if( !u ) return NULL;
} while( my_unrealized_threads.compare_and_swap(u-1,u)!=u );
server_thread& t = my_private_threads.add_one_thread();
t.is_omp_thread = is_omp_thread_;
__TBB_ASSERT( u>=1, NULL );
t.my_index = u - 1;
__TBB_ASSERT( t.state!=ts_tbb_busy, NULL );
t.my_extra_state = t.is_omp_thread ? ts_none : ts_created;
iterator i = t.my_map_pos = my_array.grow_by(1);
value_type& v = *i;
v.my_thread = &t;
return &v;
}
void thread_map::bind() {
++my_factory_counter;
my_min_stack_size = my_client.min_stack_size();
__TBB_ASSERT( my_unrealized_threads==0, "already called bind?" );
my_unrealized_threads = my_client.max_job_count();
}
void thread_map::unbind() {
// Ask each server_thread to cleanup its job for this server.
for( iterator i=begin(); i!=end(); ++i ) {
server_thread& t = i->thread();
t.terminate = true;
t.wakeup( ts_idle, ts_asleep );
}
// Remove extra ref to client.
remove_client_ref();
}
void thread_map::assist_cleanup( bool assist_null_only ) {
// To avoid deadlock, the current thread *must* help out with cleanups that have not started,
// becausd the thread that created the job may be busy for a long time.
for( iterator i = begin(); i!=end(); ++i ) {
rml::job* j=0;
job_automaton& ja = i->my_automaton;
if( assist_null_only ? ja.try_plug_null() : ja.try_plug(j) ) {
if( j ) {
my_client.cleanup(*j);
} else {
// server thread did not get a chance to create a job.
}
remove_client_ref();
}
}
}
thread_map::size_type thread_map::wakeup_tbb_threads( size_type n ) {
__TBB_ASSERT(n>0,"must specify positive number of threads to wake up");
iterator e = end();
for( iterator k=begin(); k!=e; ++k ) {
// If another thread added *k, there is a tiny timing window where thread() is invalid.
server_thread& t = k->wait_for_thread();
thread_state_t thr_s = t.read_state();
if( t.read_extra_state()==ts_created || thr_s==ts_tbb_busy || thr_s==ts_done )
continue;
if( --the_balance>=0 ) { // try to withdraw a coin from the deposit
while( !t.try_grab_for( ts_tbb_busy ) ) {
thr_s = t.read_state();
if( thr_s==ts_tbb_busy || thr_s==ts_done ) {
// we lost; move on to the next.
++the_balance;
goto skip;
}
}
if( --n==0 )
return 0;
} else {
// overdraft.
++the_balance;
break;
}
skip:
;
}
return n<my_unrealized_threads ? n : my_unrealized_threads;
}
#else /* RML_USE_WCRM */
class thread_map : no_copy {
friend class omp_connection_v2;
typedef ::std::hash_map<uintptr_t,server_thread*> hash_map_type;
size_t my_min_stack_size;
size_t my_unrealized_threads;
::rml::client& my_client;
//! Counter owned by factory that produced this thread_map.
wait_counter& my_factory_counter;
//! Ref counters
ref_count my_server_ref_count;
ref_count my_client_ref_count;
// FIXME: pad this?
hash_map_type my_map;
bool shutdown_in_progress;
std::vector<IExecutionResource*> original_exec_resources;
tbb::cache_aligned_allocator<padded<tbb_server_thread> > my_tbb_allocator;
tbb::cache_aligned_allocator<padded<omp_server_thread> > my_omp_allocator;
tbb::cache_aligned_allocator<padded<thread_scavenger_thread> > my_scavenger_allocator;
IResourceManager* my_concrt_resource_manager;
IScheduler* my_scheduler;
ISchedulerProxy* my_scheduler_proxy;
tbb::atomic<thread_scavenger_thread*> my_thread_scavenger_thread;
#if TBB_USE_ASSERT
tbb::atomic<int> n_add_vp_requests;
tbb::atomic<int> n_thread_scavengers_created;
#endif
public:
thread_map( wait_counter& fc, ::rml::client& client ) :
my_min_stack_size(0), my_client(client), my_factory_counter(fc),
my_server_ref_count(1), my_client_ref_count(1), shutdown_in_progress(false),
my_concrt_resource_manager(NULL), my_scheduler(NULL), my_scheduler_proxy(NULL)
{
my_thread_scavenger_thread = NULL;
#if TBB_USE_ASSERT
n_add_vp_requests = 0;
n_thread_scavengers_created;
#endif
}
~thread_map() {
__TBB_ASSERT( n_thread_scavengers_created<=1, "too many scavenger thread created" );
// if thread_scavenger_thread is launched, wait for it to complete
if( my_thread_scavenger_thread ) {
__TBB_ASSERT( my_thread_scavenger_thread!=c_claimed, NULL );
while( my_thread_scavenger_thread->read_state()==ts_busy )
__TBB_Yield();
thread_scavenger_thread* tst = my_thread_scavenger_thread;
my_scavenger_allocator.deallocate(static_cast<padded<thread_scavenger_thread>*>(tst),1);
}
// deallocate thread contexts
for( hash_map_type::const_iterator hi=my_map.begin(); hi!=my_map.end(); ++hi ) {
server_thread* thr = hi->second;
if( thr->tbb_thread ) {
while( ((tbb_server_thread*)thr)->activation_count>1 )
__TBB_Yield();
((tbb_server_thread*)thr)->~tbb_server_thread();
my_tbb_allocator.deallocate(static_cast<padded<tbb_server_thread>*>(thr),1);
} else {
((omp_server_thread*)thr)->~omp_server_thread();
my_omp_allocator.deallocate(static_cast<padded<omp_server_thread>*>(thr),1);
}
}
if( my_scheduler_proxy ) {
my_scheduler_proxy->Shutdown();
my_concrt_resource_manager->Release();
__TBB_ASSERT( my_scheduler, NULL );
delete my_scheduler;
} else {
__TBB_ASSERT( !my_scheduler, NULL );
}
}
typedef hash_map_type::key_type key_type;
typedef hash_map_type::value_type value_type;
typedef hash_map_type::iterator iterator;
iterator begin() {return my_map.begin();}
iterator end() {return my_map.end();}
iterator find( key_type k ) {return my_map.find( k );}
iterator insert( key_type k, server_thread* v ) {
std::pair<iterator,bool> res = my_map.insert( value_type(k,v) );
return res.first;
}
void bind( IScheduler* s ) {
++my_factory_counter;
if( s ) {
my_unrealized_threads = s->GetPolicy().GetPolicyValue( MaxConcurrency );
__TBB_ASSERT( my_unrealized_threads>0, NULL );
my_scheduler = s;
my_concrt_resource_manager = CreateResourceManager(); // reference count==3 when first created.
my_scheduler_proxy = my_concrt_resource_manager->RegisterScheduler( s, CONCRT_RM_VERSION_1 );
my_scheduler_proxy->RequestInitialVirtualProcessors( false );
}
}
bool is_closing() { return shutdown_in_progress; }
void unbind( rml::server& server, ::tbb::spin_mutex& mtx );
void add_client_ref() { my_server_ref_count.add_ref(); }
void remove_client_ref();
void add_server_ref() {my_server_ref_count.add_ref();}
int remove_server_ref() {return my_server_ref_count.remove_ref();}
int get_server_ref_count() { int k = my_server_ref_count.my_ref_count; return k; }
void assist_cleanup( bool assist_null_only );
void adjust_balance( int delta );
int current_balance() const {int k = the_balance; return k;}
::rml::client& client() const {return my_client;}
void register_as_master( server::execution_resource_t& v ) const { (IExecutionResource*&)v = my_scheduler_proxy ? my_scheduler_proxy->SubscribeCurrentThread() : NULL; }
// Rremove() should be called from the same thread that subscribed the current h/w thread (i.e., the one that
// called register_as_master() ).
void unregister( server::execution_resource_t v ) const {if( v ) ((IExecutionResource*)v)->Remove( my_scheduler );}
void add_virtual_processors( IVirtualProcessorRoot** vprocs, unsigned int count, tbb_connection_v2& conn, ::tbb::spin_mutex& mtx );
void add_virtual_processors( IVirtualProcessorRoot** vprocs, unsigned int count, omp_connection_v2& conn, ::tbb::spin_mutex& mtx );
void remove_virtual_processors( IVirtualProcessorRoot** vproots, unsigned count, ::tbb::spin_mutex& mtx );
void mark_virtual_processors_as_lent( IVirtualProcessorRoot** vproots, unsigned count, ::tbb::spin_mutex& mtx );
void create_oversubscribers( unsigned n, std::vector<server_thread*>& thr_vec, omp_connection_v2& conn, ::tbb::spin_mutex& mtx );
void wakeup_tbb_threads( int c, ::tbb::spin_mutex& mtx );
void mark_virtual_processors_as_returned( IVirtualProcessorRoot** vprocs, unsigned int count, tbb::spin_mutex& mtx );
inline void addto_original_exec_resources( IExecutionResource* r, ::tbb::spin_mutex& mtx ) {
::tbb::spin_mutex::scoped_lock lck(mtx);
__TBB_ASSERT( !is_closing(), "try to regster master while connection is being shutdown?" );
original_exec_resources.push_back( r );
}
#if !__RML_REMOVE_VIRTUAL_PROCESSORS_DISABLED
void allocate_thread_scavenger( IExecutionResource* v );
#endif
inline thread_scavenger_thread* get_thread_scavenger() { return my_thread_scavenger_thread; }
};
garbage_connection_queue connections_to_reclaim;
connection_scavenger_thread connection_scavenger;
#endif /* !RML_USE_WCRM */
//------------------------------------------------------------------------
// generic_connection
//------------------------------------------------------------------------
template<typename Server, typename Client>
struct connection_traits {};
// head of the active tbb connections
static tbb::atomic<uintptr_t> active_tbb_connections;
static tbb::atomic<int> current_tbb_conn_readers;
static size_t current_tbb_conn_reader_epoch;
static tbb::atomic<size_t> close_tbb_connection_event_count;
#if RML_USE_WCRM
template<typename Connection>
void make_job( Connection& c, server_thread& t );
#endif
template<typename Server, typename Client>
class generic_connection: public Server, no_copy {
/*override*/ version_type version() const {return SERVER_VERSION;}
/*override*/ void yield() {thread_monitor::yield();}
/*override*/ void independent_thread_number_changed( int delta ) { my_thread_map.adjust_balance( -delta ); }
/*override*/ unsigned default_concurrency() const {return hardware_concurrency()-1;}
friend void wakeup_some_tbb_threads();
friend class connection_scavenger_thread;
protected:
thread_map my_thread_map;
generic_connection* next_conn;
size_t my_ec;
#if RML_USE_WCRM
// FIXME: pad it?
tbb::spin_mutex map_mtx;
IScheduler* my_scheduler;
void do_open( IScheduler* s ) {
my_scheduler = s;
my_thread_map.bind( s );
}
bool is_closing() { return my_thread_map.is_closing(); }
void request_close_connection( bool existing );
#else
void do_open() {my_thread_map.bind();}
void request_close_connection( bool );
#endif /* RML_USE_WCRM */
//! Make destructor virtual
virtual ~generic_connection() {}
#if !RML_USE_WCRM
generic_connection( wait_counter& fc, Client& c ) : my_thread_map(fc,c), next_conn(NULL), my_ec(0) {}
#else
generic_connection( wait_counter& fc, Client& c ) :
my_thread_map(fc,c), next_conn(NULL), my_ec(0), map_mtx(), my_scheduler(NULL) {}
void add_virtual_processors( IVirtualProcessorRoot** vprocs, unsigned int count );
void remove_virtual_processors( IVirtualProcessorRoot** vprocs, unsigned int count );
void notify_resources_externally_busy( IVirtualProcessorRoot** vprocs, unsigned int count ) { my_thread_map.mark_virtual_processors_as_lent( vprocs, count, map_mtx ); }
void notify_resources_externally_idle( IVirtualProcessorRoot** vprocs, unsigned int count ) {
my_thread_map.mark_virtual_processors_as_returned( vprocs, count, map_mtx );
}
#endif /* !RML_USE_WCRM */
public:
typedef Server server_type;
typedef Client client_type;
Client& client() const {return static_cast<Client&>(my_thread_map.client());}
void set_scratch_ptr( job& j, void* ptr ) { ::rml::server::scratch_ptr(j) = ptr; }
#if RML_USE_WCRM
template<typename Connection>
friend void make_job( Connection& c, server_thread& t );
void add_server_ref () {my_thread_map.add_server_ref();}
void remove_server_ref() {if( my_thread_map.remove_server_ref()==0 ) delete this;}
void add_client_ref () {my_thread_map.add_client_ref();}
void remove_client_ref() {my_thread_map.remove_client_ref();}
#else /* !RML_USE_WCRM */
int add_server_ref () {return my_thread_map.add_server_ref();}
void remove_server_ref() {if( my_thread_map.remove_server_ref()==0 ) delete this;}
void remove_client_ref() {my_thread_map.remove_client_ref();}
void make_job( server_thread& t, job_automaton& ja );
#endif /* RML_USE_WCRM */
static generic_connection* get_addr( uintptr_t addr_ex ) {
return reinterpret_cast<generic_connection*>( addr_ex&~(uintptr_t)3 );
}
};
//------------------------------------------------------------------------
// TBB server
//------------------------------------------------------------------------
template<>
struct connection_traits<tbb_server,tbb_client> {
static const bool assist_null_only = true;
static const bool is_tbb = true;
};
//! Represents a server/client binding.
/** The internal representation uses inheritance for the server part and a pointer for the client part. */
class tbb_connection_v2: public generic_connection<tbb_server,tbb_client> {
/*override*/ void adjust_job_count_estimate( int delta );
#if !RML_USE_WCRM
#if _WIN32||_WIN64
/*override*/ void register_master ( rml::server::execution_resource_t& /*v*/ ) {}
/*override*/ void unregister_master ( rml::server::execution_resource_t /*v*/ ) {}
#endif
#else
/*override*/ void register_master ( rml::server::execution_resource_t& v ) {
my_thread_map.register_as_master(v);
if( v ) ++nesting;
}
/*override*/ void unregister_master ( rml::server::execution_resource_t v ) {
if( v ) {
__TBB_ASSERT( nesting>0, NULL );
if( --nesting==0 ) {
#if !__RML_REMOVE_VIRTUAL_PROCESSORS_DISABLED
my_thread_map.allocate_thread_scavenger( (IExecutionResource*)v );
#endif
}
}
my_thread_map.unregister(v);
}
IScheduler* create_scheduler() {return( scheduler<tbb_connection_v2>::create( *this ) );}
friend void free_all_connections( uintptr_t );
friend class scheduler<tbb_connection_v2>;
friend class execution_context;
friend class connection_scavenger_thread;
#endif /* RML_USE_WCRM */
friend void wakeup_some_tbb_threads();
//! Estimate on number of jobs without threads working on them.
tbb::atomic<int> my_slack;
friend class dummy_class_to_shut_up_gratuitous_warning_from_gcc_3_2_3;
#if TBB_USE_ASSERT
tbb::atomic<int> my_job_count_estimate;
#endif /* TBB_USE_ASSERT */
tbb::atomic<int> n_adjust_job_count_requests;
#if RML_USE_WCRM
tbb::atomic<int> nesting;
#endif
// dtor
~tbb_connection_v2();
public:
#if RML_USE_WCRM
typedef tbb_server_thread server_thread_type;
#endif
//! True if there is slack that try_process can use.
bool has_slack() const {return my_slack>0;}
#if RML_USE_WCRM
bool try_process( job& job )
#else
bool try_process( server_thread& t, job& job )
#endif
{
bool visited = false;
// No check for my_slack>0 here because caller is expected to do that check.
int k = --my_slack;
if( k>=0 ) {
#if !RML_USE_WCRM
t.my_extra_state = ts_visited; // remember the thread paid a trip to process() at least once
#endif
client().process(job);
visited = true;
}
++my_slack;
return visited;
}
tbb_connection_v2( wait_counter& fc, tbb_client& client ) : generic_connection<tbb_server,tbb_client>(fc,client)
{
my_slack = 0;
#if RML_USE_WCRM
nesting = 0;
#endif
#if TBB_USE_ASSERT
my_job_count_estimate = 0;
#endif /* TBB_USE_ASSERT */
__TBB_ASSERT( !my_slack, NULL );
#if RML_USE_WCRM
do_open( client.max_job_count()>0 ? create_scheduler() : NULL );
#else
do_open();
#endif /* !RML_USE_WCRM */
n_adjust_job_count_requests = 0;
// Acquire head of active_tbb_connections & push the connection into the list
uintptr_t conn;
do {
for( ; (conn=active_tbb_connections)&1; )
__TBB_Yield();
} while( active_tbb_connections.compare_and_swap( conn|1, conn )!=conn );
this->next_conn = generic_connection<tbb_server,tbb_client>::get_addr(conn);
// Update and release head of active_tbb_connections
active_tbb_connections = (uintptr_t) this; // set and release
}
inline void wakeup_tbb_threads( unsigned n ) {
my_thread_map.wakeup_tbb_threads( n
#if RML_USE_WCRM
, map_mtx
#endif
);
}
#if RML_USE_WCRM
inline int get_nesting_level() { return nesting; }
#else
inline bool wakeup_next_thread( thread_map::iterator i ) {return my_thread_map.wakeup_next_thread( i, *this );}
inline thread_map::size_type get_unrealized_threads () {return my_thread_map.get_unrealized_threads();}
#endif /* !RML_USE_WCRM */
};
//------------------------------------------------------------------------
// OpenMP server
//------------------------------------------------------------------------
template<>
struct connection_traits<omp_server,omp_client> {
static const bool assist_null_only = false;
static const bool is_tbb = false;
};
class omp_connection_v2: public generic_connection<omp_server,omp_client> {
#if !RML_USE_WCRM
/*override*/ int current_balance() const {return the_balance;}
#else
friend void free_all_connections( uintptr_t );
friend class scheduler<omp_connection_v2>;
/*override*/ int current_balance() const {return my_thread_map.current_balance();}
#endif /* !RML_USE_WCRM */
/*override*/ int try_increase_load( size_type n, bool strict );
/*override*/ void decrease_load( size_type n );
/*override*/ void get_threads( size_type request_size, void* cookie, job* array[] );
#if !RML_USE_WCRM
#if _WIN32||_WIN64
/*override*/ void register_master ( rml::server::execution_resource_t& /*v*/ ) {}
/*override*/ void unregister_master ( rml::server::execution_resource_t /*v*/ ) {}
#endif
#else
/*override*/ void register_master ( rml::server::execution_resource_t& v ) {
my_thread_map.register_as_master( v );
my_thread_map.addto_original_exec_resources( (IExecutionResource*)v, map_mtx );
}
/*override*/ void unregister_master ( rml::server::execution_resource_t v ) { my_thread_map.unregister(v); }
#endif /* !RML_USE_WCRM */
#if _WIN32||_WIN64
/*override*/ void deactivate( rml::job* j );
/*override*/ void reactivate( rml::job* j );
#endif /* _WIN32||_WIN64 */
#if RML_USE_WCRM
public:
typedef omp_server_thread server_thread_type;
private:
IScheduler* create_scheduler() {return( scheduler<omp_connection_v2>::create( *this ) );}
#endif /* RML_USE_WCRM */
public:
#if TBB_USE_ASSERT
//! Net change in delta caused by this connection.
/** Should be zero when connection is broken */
tbb::atomic<int> net_delta;
#endif /* TBB_USE_ASSERT */
omp_connection_v2( wait_counter& fc, omp_client& client ) : generic_connection<omp_server,omp_client>(fc,client) {
#if TBB_USE_ASSERT
net_delta = 0;
#endif /* TBB_USE_ASSERT */
#if RML_USE_WCRM
do_open( create_scheduler() );
#else
do_open();
#endif /* RML_USE_WCRM */
}
~omp_connection_v2() {__TBB_ASSERT( net_delta==0, "net increase/decrease of load is nonzero" );}
};
#if !RML_USE_WCRM
/* to deal with cases where the machine is oversubscribed; we want each thread to trip to try_process() at least once */
/* this should not involve computing the_balance */
bool thread_map::wakeup_next_thread( thread_map::iterator this_thr, tbb_connection_v2& conn ) {
if( all_visited_at_least_once )
return false;
iterator e = end();
retry:
bool exist = false;
iterator k=this_thr;
for( ++k; k!=e; ++k ) {
// If another thread added *k, there is a tiny timing window where thread() is invalid.
server_thread& t = k->wait_for_thread();
if( t.my_extra_state!=ts_visited )
exist = true;
if( t.read_state()!=ts_tbb_busy && t.my_extra_state==ts_started )
if( t.try_grab_for( ts_tbb_busy ) )
return true;
}
for( k=begin(); k!=this_thr; ++k ) {
server_thread& t = k->wait_for_thread();
if( t.my_extra_state!=ts_visited )
exist = true;
if( t.read_state()!=ts_tbb_busy && t.my_extra_state==ts_started )
if( t.try_grab_for( ts_tbb_busy ) )
return true;
}
if( exist )
if( conn.has_slack() )
goto retry;
else
all_visited_at_least_once = true;
return false;
}
void thread_map::release_tbb_threads( server_thread* t ) {
for( ; t; t = t->link ) {
while( t->read_state()!=ts_asleep )
__TBB_Yield();
t->my_extra_state = ts_started;
}
}
#endif /* !RML_USE_WCRM */
void thread_map::adjust_balance( int delta ) {
int new_balance = the_balance += delta;
if( new_balance>0 && 0>=new_balance-delta /*== old the_balance*/ )
wakeup_some_tbb_threads();
}
void thread_map::remove_client_ref() {
int k = my_client_ref_count.remove_ref();
if( k==0 ) {
// Notify factory that thread has crossed back into RML.
--my_factory_counter;
// Notify client that RML is done with the client object.
my_client.acknowledge_close_connection();
}
}
#if RML_USE_WCRM
/** Not a member of generic_connection because we need Connection to be the derived class. */
template<typename Connection>
void make_job( Connection& c, typename Connection::server_thread_type& t ) {
if( t.my_job_automaton.try_acquire() ) {
rml::job& j = *t.my_client.create_one_job();
__TBB_ASSERT( &j!=NULL, "client:::create_one_job returned NULL" );
__TBB_ASSERT( (intptr_t(&j)&1)==0, "client::create_one_job returned misaligned job" );
t.my_job_automaton.set_and_release( j );
c.set_scratch_ptr( j, (void*) &t );
}
}
#endif /* RML_USE_WCRM */
#if _MSC_VER && !defined(__INTEL_COMPILER)
// Suppress "conditional expression is constant" warning.
#pragma warning( push )
#pragma warning( disable: 4127 )
#endif
#if RML_USE_WCRM
template<typename Server, typename Client>
void generic_connection<Server,Client>::request_close_connection( bool exiting ) {
// for TBB connections, exiting should always be false
if( connection_traits<Server,Client>::is_tbb )
__TBB_ASSERT( !exiting, NULL);
#if TBB_USE_ASSERT
else if( exiting )
reinterpret_cast<omp_connection_v2*>(this)->net_delta = 0;
#endif
if( exiting ) {
uintptr_t tail = connections_to_reclaim.tail;
while( connections_to_reclaim.tail.compare_and_swap( garbage_connection_queue::plugged, tail )!=tail )
__TBB_Yield();
my_thread_map.unbind( *this, map_mtx );
my_thread_map.assist_cleanup( connection_traits<Server,Client>::assist_null_only );
// It is assumed that the client waits for all other threads to terminate before
// calling request_close_connection with true. Thus, it is safe to return all
// outstanding connection objects that are reachable. It is possible that there may
// be some unreachable connection objects lying somewhere.
free_all_connections( connection_scavenger.grab_and_prepend( this ) );
return;
}
#else /* !RML_USE_WCRM */
template<typename Server, typename Client>
void generic_connection<Server,Client>::request_close_connection( bool ) {
#endif /* RML_USE_WCRM */
if( connection_traits<Server,Client>::is_tbb ) {
// acquire the head of active tbb connections
uintptr_t conn;
do {
for( ; (conn=active_tbb_connections)&1; )
__TBB_Yield();
} while( active_tbb_connections.compare_and_swap( conn|1, conn )!=conn );
// Locate the current connection
generic_connection* pred_conn = NULL;
generic_connection* curr_conn = (generic_connection*) conn;
for( ; curr_conn && curr_conn!=this; curr_conn=curr_conn->next_conn )
pred_conn = curr_conn;
__TBB_ASSERT( curr_conn==this, "the current connection is not in the list?" );
// Remove this from the list
if( pred_conn ) {
pred_conn->next_conn = curr_conn->next_conn;
active_tbb_connections = reinterpret_cast<uintptr_t>(generic_connection<tbb_server,tbb_client>::get_addr(active_tbb_connections)); // release it
} else
active_tbb_connections = (uintptr_t) curr_conn->next_conn; // update & release it
curr_conn->next_conn = NULL;
// Increment the tbb connection close event count
my_ec = ++close_tbb_connection_event_count;
// Wait happens in tbb_connection_v2::~tbb_connection_v2()
}
#if RML_USE_WCRM
my_thread_map.unbind( *this, map_mtx );
my_thread_map.assist_cleanup( connection_traits<Server,Client>::assist_null_only );
connection_scavenger.add_request( this );
#else
my_thread_map.unbind();
my_thread_map.assist_cleanup( connection_traits<Server,Client>::assist_null_only );
// Remove extra reference
remove_server_ref();
#endif
}
#if _MSC_VER && !defined(__INTEL_COMPILER)
#pragma warning( pop )
#endif
#if RML_USE_WCRM
template<typename Server, typename Client>
void generic_connection<Server,Client>::add_virtual_processors( IVirtualProcessorRoot** vproots, unsigned int count )
{}
template<>
void generic_connection<tbb_server,tbb_client>::add_virtual_processors( IVirtualProcessorRoot** vproots, unsigned int count )
{
my_thread_map.add_virtual_processors( vproots, count, (tbb_connection_v2&)*this, map_mtx );
}
template<>
void generic_connection<omp_server,omp_client>::add_virtual_processors( IVirtualProcessorRoot** vproots, unsigned int count )
{
// For OMP, since it uses ScheudlerPolicy of MinThreads==MaxThreads, this is called once when
// RequestInitialVirtualProcessors() is called.
my_thread_map.add_virtual_processors( vproots, count, (omp_connection_v2&)*this, map_mtx );
}
template<typename Server, typename Client>
void generic_connection<Server,Client>::remove_virtual_processors( IVirtualProcessorRoot** vproots, unsigned int count )
{
__TBB_ASSERT( false, "should not be called" );
}
/* For OMP, RemoveVirtualProcessors() will never be called. */
template<>
void generic_connection<tbb_server,tbb_client>::remove_virtual_processors( IVirtualProcessorRoot** vproots, unsigned int count )
{
my_thread_map.remove_virtual_processors( vproots, count, map_mtx );
}
void tbb_connection_v2::adjust_job_count_estimate( int delta ) {
#if TBB_USE_ASSERT
my_job_count_estimate += delta;
#endif /* TBB_USE_ASSERT */
// Atomically update slack.
int c = my_slack+=delta;
if( c>0 ) {
++n_adjust_job_count_requests;
my_thread_map.wakeup_tbb_threads( c, map_mtx );
--n_adjust_job_count_requests;
}
}
#endif /* RML_USE_WCRM */
tbb_connection_v2::~tbb_connection_v2() {
#if TBB_USE_ASSERT
if( my_job_count_estimate!=0 ) {
fprintf(stderr, "TBB client tried to disconnect with non-zero net job count estimate of %d\n", int(my_job_count_estimate ));
abort();
}
__TBB_ASSERT( !my_slack, "attempt to destroy tbb_server with nonzero slack" );
__TBB_ASSERT( this!=static_cast<tbb_connection_v2*>(generic_connection<tbb_server,tbb_client >::get_addr(active_tbb_connections)), "request_close_connection() must be called" );
#endif /* TBB_USE_ASSERT */
#if !RML_USE_WCRM
// If there are other threads ready for work, give them coins
if( the_balance>0 )
wakeup_some_tbb_threads();
#endif
// Someone might be accessing my data members
while( current_tbb_conn_readers>0 && (ptrdiff_t)(my_ec-current_tbb_conn_reader_epoch)>0 )
__TBB_Yield();
}
#if !RML_USE_WCRM
template<typename Server, typename Client>
void generic_connection<Server,Client>::make_job( server_thread& t, job_automaton& ja ) {
if( ja.try_acquire() ) {
rml::job& j = *client().create_one_job();
__TBB_ASSERT( &j!=NULL, "client:::create_one_job returned NULL" );
__TBB_ASSERT( (intptr_t(&j)&1)==0, "client::create_one_job returned misaligned job" );
ja.set_and_release( j );
__TBB_ASSERT( t.my_conn && t.my_ja && t.my_job==NULL, NULL );
t.my_job = &j;
set_scratch_ptr( j, (void*) &t );
}
}
void tbb_connection_v2::adjust_job_count_estimate( int delta ) {
#if TBB_USE_ASSERT
my_job_count_estimate += delta;
#endif /* TBB_USE_ASSERT */
// Atomically update slack.
int c = my_slack+=delta;
if( c>0 ) {
++n_adjust_job_count_requests;
// The client has work to do and there are threads available
thread_map::size_type n = my_thread_map.wakeup_tbb_threads(c);
server_thread* new_threads_anchor = NULL;
thread_map::size_type i;
for( i=0; i<n; ++i ) {
// Obtain unrealized threads
thread_map::value_type* k = my_thread_map.add_one_thread( false );
if( !k )
// No unrealized threads left.
break;
// Eagerly start the thread off.
my_thread_map.bind_one_thread( *this, *k );
server_thread& t = k->thread();
__TBB_ASSERT( !t.link, NULL );
t.link = new_threads_anchor;
new_threads_anchor = &t;
}
thread_map::size_type j=0;
for( ; the_balance>0 && j<i; ++j ) {
if( --the_balance>=0 ) {
// Withdraw a coin from the bank
__TBB_ASSERT( new_threads_anchor, NULL );
server_thread* t = new_threads_anchor;
new_threads_anchor = t->link;
while( !t->try_grab_for( ts_tbb_busy ) )
__TBB_Yield();
t->my_extra_state = ts_started;
} else {
// Overdraft. return it to the bank
++the_balance;
break;
}
}
__TBB_ASSERT( i-j!=0||new_threads_anchor==NULL, NULL );
// Mark the ones that did not get started as eligible for being snatched.
if( new_threads_anchor )
my_thread_map.release_tbb_threads( new_threads_anchor );
--n_adjust_job_count_requests;
}
}
#endif /* RML_USE_WCRM */
#if RML_USE_WCRM
int omp_connection_v2::try_increase_load( size_type n, bool strict ) {
__TBB_ASSERT(int(n)>=0,NULL);
if( strict ) {
the_balance -= int(n);
} else {
int avail, old;
do {
avail = the_balance;
if( avail<=0 ) {
// No atomic read-write-modify operation necessary.
return avail;
}
// Don't read the_system_balance; if it changes, compare_and_swap will fail anyway.
old = the_balance.compare_and_swap( int(n)<avail ? avail-n : 0, avail );
} while( old!=avail );
if( int(n)>avail )
n=avail;
}
#if TBB_USE_ASSERT
net_delta += n;
#endif /* TBB_USE_ASSERT */
return n;
}
void omp_connection_v2::decrease_load( size_type /*n*/ ) {}
void omp_connection_v2::get_threads( size_type request_size, void* cookie, job* array[] ) {
unsigned index = 0;
std::vector<omp_server_thread*> enlisted(request_size);
std::vector<thread_grab_t> to_activate(request_size);
if( request_size==0 ) return;
{
tbb::spin_mutex::scoped_lock lock(map_mtx);
__TBB_ASSERT( !is_closing(), "try to get threads while connection is being shutdown?" );
for( int scan=0; scan<2; ++scan ) {
for( thread_map::iterator i=my_thread_map.begin(); i!=my_thread_map.end(); ++i ) {
omp_server_thread* thr = (omp_server_thread*) (*i).second;
// in the first scan, skip VPs that are lent
if( scan==0 && thr->is_lent() ) continue;
thread_grab_t res = thr->try_grab_for();
if( res!=wk_failed ) {// && if is not busy by some other scheduler
to_activate[index] = res;
enlisted[index] = thr;
if( ++index==request_size )
goto activate_threads;
}
}
}
}
activate_threads:
for( unsigned i=0; i<index; ++i ) {
omp_server_thread* thr = enlisted[i];
if( to_activate[i]==wk_from_asleep )
thr->get_virtual_processor()->Activate( thr );
job* j = thr->wait_for_job();
array[i] = j;
thr->omp_data.produce( client(), *j, cookie, i PRODUCE_ARG(*this) );
}
if( index==request_size )
return;
// If we come to this point, it must be becuase dynamic==false
// Create Oversubscribers..
// Note that our policy is such that MinConcurrency==MaxConcurrency.
// RM will deliver MaxConcurrency of VirtualProcessors and no more.
__TBB_ASSERT( request_size>index, NULL );
unsigned n = request_size - index;
std::vector<server_thread*> thr_vec(n);
typedef std::vector<server_thread*>::iterator iterator_thr;
my_thread_map.create_oversubscribers( n, thr_vec, *this, map_mtx );
for( iterator_thr ti=thr_vec.begin(); ti!=thr_vec.end(); ++ti ) {
omp_server_thread* thr = (omp_server_thread*) *ti;
__TBB_ASSERT( thr, "thread not created?" );
// Thread is already grabbed; since it is nrewly created, we need to activate it.
thr->get_virtual_processor()->Activate( thr );
job* j = thr->wait_for_job();
array[index] = j;
thr->omp_data.produce( client(), *j, cookie, index PRODUCE_ARG(*this) );
++index;
}
}
#if _WIN32||_WIN64
void omp_connection_v2::deactivate( rml::job* j )
{
my_thread_map.adjust_balance(1);
#if TBB_USE_ASSERT
net_delta -= 1;
#endif
omp_server_thread* thr = (omp_server_thread*) scratch_ptr( *j );
(thr->get_virtual_processor())->Deactivate( thr );
}
void omp_connection_v2::reactivate( rml::job* j )
{
// Should not adjust the_balance because OMP client is supposed to
// do try_increase_load() to reserve the threads to use.
omp_server_thread* thr = (omp_server_thread*) scratch_ptr( *j );
(thr->get_virtual_processor())->Activate( thr );
}
#endif /* !_WIN32||_WIN64 */
#endif /* RML_USE_WCRM */
//! Wake up some available tbb threads
void wakeup_some_tbb_threads()
{
/* First, atomically grab the connection, then increase the server ref count to keep
it from being released prematurely. Second, check if the balance is available for TBB
and the tbb conneciton has slack to exploit. If the answer is true, go ahead and
try to wake some up. */
if( generic_connection<tbb_server,tbb_client >::get_addr(active_tbb_connections)==0 )
// the next connection will see the change; return.
return;
start_it_over:
int n_curr_readers = ++current_tbb_conn_readers;
if( n_curr_readers>1 ) // I lost
return;
// if n_curr_readers==1, i am the first one, so I will take responsibility for waking tbb threads up.
// update the current epoch
current_tbb_conn_reader_epoch = close_tbb_connection_event_count;
// read and clear
// Newly added connection will not invalidate the pointer, and it will
// compete with the current one to claim coins.
// One that is about to close the connection increments the event count
// after it removes the connection from the list. But it will keep around
// the connection until all readers including this one catch up. So, reading
// the head and clearing the lock bit should be o.k.
generic_connection<tbb_server,tbb_client>* next_conn_wake_up = generic_connection<tbb_server,tbb_client>::get_addr( active_tbb_connections );
for( ; next_conn_wake_up; ) {
/* some threads are creating tbb server threads; they may not see my changes made to the_balance */
/* When a thread is in adjust_job_count_estimate() to increase the slack
RML tries to activate worker threads on behalf of the requesting thread
by repeatedly drawing a coin from the bank optimistically and grabbing a
thread. If it finds the bank overdrafted, it returns the coin back to
the bank and returns the control to the thread (return from the method).
There lies a tiny timing hole.
When the overdraft occurs (note that multiple masters may be in
adjust_job_count_estimate() so the_balance can be any negative value) and
a worker returns from the TBB work at that moment, its returning the coin
does not bump up the_balance over 0, so it happily returns from
wakeup_some_tbb_threads() without attempting to give coins to worker threads
that are ready.
*/
while( ((tbb_connection_v2*)next_conn_wake_up)->n_adjust_job_count_requests>0 )
__TBB_Yield();
int bal = the_balance;
n_curr_readers = current_tbb_conn_readers; // get the snapshot
if( bal<=0 ) break;
// if the connection is deleted, the following will immediately return because its slack would be 0 or less.
tbb_connection_v2* tbb_conn = (tbb_connection_v2*)next_conn_wake_up;
int my_slack = tbb_conn->my_slack;
if( my_slack>0 ) tbb_conn->wakeup_tbb_threads( my_slack );
next_conn_wake_up = next_conn_wake_up->next_conn;
}
int delta = current_tbb_conn_readers -= n_curr_readers;
//if delta>0, more threads entered the routine since this one took the snapshot
if( delta>0 ) {
current_tbb_conn_readers = 0;
if( the_balance>0 && generic_connection<tbb_server,tbb_client >::get_addr(active_tbb_connections)!=0 )
goto start_it_over;
}
// Signal any connection that is waiting for me to complete my access that I am done.
current_tbb_conn_reader_epoch = close_tbb_connection_event_count;
}
#if !RML_USE_WCRM
int omp_connection_v2::try_increase_load( size_type n, bool strict ) {
__TBB_ASSERT(int(n)>=0,NULL);
if( strict ) {
the_balance -= int(n);
} else {
int avail, old;
do {
avail = the_balance;
if( avail<=0 ) {
// No atomic read-write-modify operation necessary.
return avail;
}
// don't read the_balance; if it changes, compare_and_swap will fail anyway.
old = the_balance.compare_and_swap( int(n)<avail ? avail-n : 0, avail );
} while( old!=avail );
if( int(n)>avail )
n=avail;
}
#if TBB_USE_ASSERT
net_delta += n;
#endif /* TBB_USE_ASSERT */
return n;
}
void omp_connection_v2::decrease_load( size_type n ) {
__TBB_ASSERT(int(n)>=0,NULL);
my_thread_map.adjust_balance(int(n));
#if TBB_USE_ASSERT
net_delta -= n;
#endif /* TBB_USE_ASSERT */
}
void omp_connection_v2::get_threads( size_type request_size, void* cookie, job* array[] ) {
if( !request_size )
return;
unsigned index = 0;
for(;;) { // don't return until all request_size threads are grabbed.
// Need to grab some threads
thread_map::iterator k_end=my_thread_map.end();
for( thread_map::iterator k=my_thread_map.begin(); k!=k_end; ++k ) {
// If another thread added *k, there is a tiny timing window where thread() is invalid.
server_thread& t = k->wait_for_thread();
if( t.try_grab_for( ts_omp_busy ) ) {
// The preincrement instead of post-increment of index is deliberate.
job& j = k->wait_for_job();
array[index] = &j;
t.omp_dispatch.produce( client(), j, cookie, index PRODUCE_ARG(*this) );
if( ++index==request_size )
return;
}
}
// Need to allocate more threads
for( unsigned i=index; i<request_size; ++i ) {
__TBB_ASSERT( index<request_size, NULL );
thread_map::value_type* k = my_thread_map.add_one_thread( true );
#if TBB_USE_ASSERT
if( !k ) {
// Client erred
__TBB_ASSERT(false, "server::get_threads: exceeded job_count\n");
}
#endif
my_thread_map.bind_one_thread( *this, *k );
server_thread& t = k->thread();
if( t.try_grab_for( ts_omp_busy ) ) {
job& j = k->wait_for_job();
array[index] = &j;
// The preincrement instead of post-increment of index is deliberate.
t.omp_dispatch.produce( client(), j, cookie, index PRODUCE_ARG(*this) );
if( ++index==request_size )
return;
} // else someone else snatched it.
}
}
}
#endif /* !RML_USE_WCRM */
//------------------------------------------------------------------------
// Methods of omp_dispatch_type
//------------------------------------------------------------------------
void omp_dispatch_type::consume() {
job_type* j = job;
// Wait for short window between when master sets state of this thread to ts_omp_busy
// and master thread calls produce.
if( !j ) {
tbb::internal::atomic_backoff bo;
do {
bo.pause();
j = job;
} while( !j );
}
job = static_cast<job_type*>(NULL);
client->process(*j,cookie,index);
#if TBB_USE_ASSERT
// Return of method process implies "decrease_load" from client's viewpoint, even though
// the actual adjustment of the_balance only happens when this thread really goes to sleep.
--server->net_delta;
#endif /* TBB_USE_ASSERT */
}
#if !RML_USE_WCRM
#if _WIN32||_WIN64
void omp_connection_v2::deactivate( rml::job* j )
{
#if TBB_USE_ASSERT
net_delta -= 1;
#endif
__TBB_ASSERT( j, NULL );
server_thread* thr = (server_thread*) scratch_ptr( *j );
thr->deactivate();
}
void omp_connection_v2::reactivate( rml::job* j )
{
// Should not adjust the_balance because OMP client is supposed to
// do try_increase_load() to reserve the threads to use.
__TBB_ASSERT( j, NULL );
server_thread* thr = (server_thread*) scratch_ptr( *j );
thr->reactivate();
}
#endif /* _WIN32||_WIN64 */
//------------------------------------------------------------------------
// Methods of server_thread
//------------------------------------------------------------------------
server_thread::server_thread() :
ref_count(0),
link(NULL),
my_map_pos(),
my_conn(NULL), my_job(NULL), my_ja(NULL)
{
state = ts_idle;
terminate = false;
#if TBB_USE_ASSERT
has_active_thread = false;
#endif /* TBB_USE_ASSERT */
}
server_thread::~server_thread() {
__TBB_ASSERT( !has_active_thread, NULL );
}
#if _MSC_VER && !defined(__INTEL_COMPILER)
// Suppress overzealous compiler warnings about an initialized variable 'sink_for_alloca' not referenced
#pragma warning(push)
#pragma warning(disable:4189)
#endif
__RML_DECL_THREAD_ROUTINE server_thread::thread_routine( void* arg ) {
server_thread* self = static_cast<server_thread*>(arg);
AVOID_64K_ALIASING( self->my_index );
#if TBB_USE_ASSERT
__TBB_ASSERT( !self->has_active_thread, NULL );
self->has_active_thread = true;
#endif /* TBB_USE_ASSERT */
self->loop();
return 0;
}
#if _MSC_VER && !defined(__INTEL_COMPILER)
#pragma warning(pop)
#endif
void server_thread::launch( size_t stack_size ) {
thread_monitor::launch( thread_routine, this, stack_size );
}
void server_thread::sleep_perhaps( thread_state_t asleep ) {
if( terminate ) return;
__TBB_ASSERT( asleep==ts_asleep, NULL );
thread_monitor::cookie c;
monitor.prepare_wait(c);
if( state.compare_and_swap( asleep, ts_idle )==ts_idle ) {
if( !terminate ) {
monitor.commit_wait(c);
// Someone else woke me up. The compare_and_swap further below deals with spurious wakeups.
} else {
monitor.cancel_wait();
}
thread_state_t s = read_state();
if( s==ts_asleep ) {
state.compare_and_swap( ts_idle, ts_asleep );
// I woke myself up, either because I cancelled the wait or suffered a spurious wakeup.
} else {
// Someone else woke me up; there the_balance is decremented by 1. -- tbb only
if( !is_omp_thread ) {
__TBB_ASSERT( s==ts_tbb_busy||s==ts_idle, NULL );
}
}
} else {
// someone else made it busy ; see try_grab_for when state==ts_idle.
__TBB_ASSERT( state==ts_omp_busy||state==ts_tbb_busy, NULL );
monitor.cancel_wait();
}
__TBB_ASSERT( read_state()!=asleep, "a thread can only put itself to sleep" );
}
bool server_thread::wakeup( thread_state_t to, thread_state_t from ) {
bool success = false;
__TBB_ASSERT( from==ts_asleep && (to==ts_idle||to==ts_omp_busy||to==ts_tbb_busy), NULL );
if( state.compare_and_swap( to, from )==from ) {
if( !is_omp_thread ) __TBB_ASSERT( to==ts_idle||to==ts_tbb_busy, NULL );
// There is a small timing window that permits balance to become negative,
// but such occurrences are probably rare enough to not worry about, since
// at worst the result is slight temporary oversubscription.
monitor.notify();
success = true;
}
return success;
}
//! Attempt to change a thread's state to ts_omp_busy, and waking it up if necessary.
bool server_thread::try_grab_for( thread_state_t target_state ) {
bool success = false;
switch( read_state() ) {
case ts_asleep:
success = wakeup( target_state, ts_asleep );
break;
case ts_idle:
success = state.compare_and_swap( target_state, ts_idle )==ts_idle;
break;
default:
// Thread is not available to be part of an OpenMP thread team.
break;
}
return success;
}
#if _WIN32||_WIN64
void server_thread::deactivate() {
thread_state_t es = (thread_state_t) my_extra_state.fetch_and_store( ts_deactivated );
__TBB_ASSERT( my_extra_state==ts_deactivated, "someone else tampered with my_extra_state?" );
if( es==ts_none )
state = ts_idle;
else
__TBB_ASSERT( es==ts_reactivated, "Cannot call deactivate() while in ts_deactivated" );
// only the thread can transition itself from ts_deactivted to ts_none
__TBB_ASSERT( my_extra_state==ts_deactivated, "someone else tampered with my_extra_state?" );
my_extra_state = ts_none; // release the critical section
int bal = ++the_balance;
if( bal>0 )
wakeup_some_tbb_threads();
if( es==ts_none )
sleep_perhaps( ts_asleep );
}
void server_thread::reactivate() {
thread_state_t es;
do {
while( (es=read_extra_state())==ts_deactivated )
__TBB_Yield();
if( es==ts_reactivated ) {
__TBB_ASSERT( false, "two Reactivate() calls in a row. Should not happen" );
return;
}
__TBB_ASSERT( es==ts_none, NULL );
} while( (thread_state_t)my_extra_state.compare_and_swap( ts_reactivated, ts_none )!=ts_none );
if( state!=ts_omp_busy ) {
my_extra_state = ts_none;
while( !try_grab_for( ts_omp_busy ) )
__TBB_Yield();
}
}
#endif /* _WIN32||_WIN64 */
template<typename Connection>
bool server_thread::destroy_job( Connection& c ) {
__TBB_ASSERT( !is_omp_thread||(state==ts_idle||state==ts_omp_busy), NULL );
__TBB_ASSERT( is_omp_thread||(state==ts_idle||state==ts_tbb_busy), NULL );
if( !is_omp_thread ) {
__TBB_ASSERT( state==ts_idle||state==ts_tbb_busy, NULL );
if( state==ts_idle )
state.compare_and_swap( ts_done, ts_idle );
// 'state' may be set to ts_tbb_busy by another thread.
if( state==ts_tbb_busy ) { // return the coin to the deposit
// need to deposit first to let the next connection see the change
++the_balance;
state = ts_done; // no other thread changes the state when it is ts_*_busy
}
}
if( job_automaton* ja = my_ja ) {
rml::job* j;
if( ja->try_plug(j) ) {
__TBB_ASSERT( j, NULL );
c.client().cleanup(*j);
c.remove_client_ref();
} else {
// Some other thread took responsibility for cleaning up the job.
}
}
// Must do remove client reference first, because execution of
// c.remove_ref() can cause *this to be destroyed.
int k = remove_ref();
__TBB_ASSERT_EX( k==0, "more than one references?" );
#if TBB_USE_ASSERT
has_active_thread = false;
#endif /* TBB_USE_ASSERT */
c.remove_server_ref();
return true;
}
bool server_thread::do_termination() {
if( is_omp_thread )
return destroy_job( *static_cast<omp_connection_v2*>(my_conn) );
else
return destroy_job( *static_cast<tbb_connection_v2*>(my_conn) );
}
//! Loop that each thread executes
void server_thread::loop() {
if( is_omp_thread )
static_cast<omp_connection_v2*>(my_conn)->make_job( *this, *my_ja );
else
static_cast<tbb_connection_v2*>(my_conn)->make_job( *this, *my_ja );
for(;;) {
__TBB_Yield();
if( state==ts_idle )
sleep_perhaps( ts_asleep );
// Check whether I should quit.
if( terminate )
if( do_termination() )
return;
// read the state
thread_state_t s = read_state();
__TBB_ASSERT( s==ts_idle||s==ts_omp_busy||s==ts_tbb_busy, NULL );
if( s==ts_omp_busy ) {
// Enslaved by OpenMP team.
omp_dispatch.consume();
/* here wake tbb threads up if feasible */
if( ++the_balance>0 )
wakeup_some_tbb_threads();
state = ts_idle;
} else if( s==ts_tbb_busy ) {
// do some TBB work.
__TBB_ASSERT( my_conn && my_job, NULL );
tbb_connection_v2& conn = *static_cast<tbb_connection_v2*>(my_conn);
// give openmp higher priority
bool has_coin = true;
if( conn.has_slack() ) {
// it has the coin, it should trip to the scheduler at least once as long as its slack is positive
do {
if( conn.try_process( *this, *my_job ) )
if( conn.has_slack() && the_balance>=0 )
has_coin = !conn.wakeup_next_thread( my_map_pos );
} while( has_coin && conn.has_slack() && the_balance>=0 );
}
state = ts_idle;
if( has_coin ) {
++the_balance; // return the coin back to the deposit
if( conn.has_slack() ) { // a new adjust_job_request_estimate() is in progress
// it may have missed my changes to state and/or the_balance
if( --the_balance>=0 ) { // try to grab the coin back
// I got the coin
if( state.compare_and_swap( ts_tbb_busy, ts_idle )!=ts_idle )
++the_balance; // someone else enlisted me.
} else {
// overdraft. return the coin
++the_balance;
}
} // else the new request will see my changes to state & the_balance.
}
/* here wake tbb threads up if feasible */
if( the_balance>0 )
wakeup_some_tbb_threads();
}
}
}
#endif /* !RML_USE_WCRM */
#if RML_USE_WCRM
class tbb_connection_v2;
class omp_connection_v2;
#define CREATE_SCHEDULER_POLICY(policy,min_thrs,max_thrs,stack_size) \
try { \
policy = new SchedulerPolicy (7, \
SchedulerKind, RML_THREAD_KIND, /*defined in _rml_serer_msrt.h*/ \
MinConcurrency, min_thrs, \
MaxConcurrency, max_thrs, \
TargetOversubscriptionFactor, 1, \
ContextStackSize, stack_size/1000, /*ConcRT:kB, iRML:bytes*/ \
ContextPriority, THREAD_PRIORITY_NORMAL, \
DynamicProgressFeedback, ProgressFeedbackDisabled ); \
} catch ( invalid_scheduler_policy_key & ) { \
__TBB_ASSERT( false, "invalid scheduler policy key exception caught" );\
} catch ( invalid_scheduler_policy_value & ) { \
__TBB_ASSERT( false, "invalid scheduler policy value exception caught" );\
}
static unsigned int core_count;
static tbb::atomic<int> core_count_inited;
static unsigned int get_processor_count()
{
if( core_count_inited!=2 ) {
if( core_count_inited.compare_and_swap( 1, 0 )==0 ) {
core_count = GetProcessorCount();
core_count_inited = 2;
} else {
tbb::internal::spin_wait_until_eq( core_count_inited, 2 );
}
}
return core_count;
}
template<typename Connection>
scheduler<Connection>::scheduler( Connection& conn ) : uid(GetSchedulerId()), my_conn(conn) {}
template<>
scheduler<tbb_connection_v2>::scheduler( tbb_connection_v2& conn ) : uid(GetSchedulerId()), my_conn(conn)
{
rml::client& cl = my_conn.client();
unsigned max_job_count = cl.max_job_count();
unsigned count = get_processor_count();
__TBB_ASSERT( max_job_count>0, "max job count must be positive" );
__TBB_ASSERT( count>1, "The processor count must be greater than 1" );
if( max_job_count>count-1) max_job_count = count-1;
CREATE_SCHEDULER_POLICY( my_policy, 0, max_job_count, cl.min_stack_size() );
}
#if __RML_REMOVE_VIRTUAL_PROCESSORS_DISABLED
template<>
void scheduler<tbb_connection_v2>::RemoveVirtualProcessors( IVirtualProcessorRoot**, unsigned int)
{
}
#else
template<>
void scheduler<tbb_connection_v2>::RemoveVirtualProcessors( IVirtualProcessorRoot** vproots, unsigned int count )
{
if( !my_conn.is_closing() )
my_conn.remove_virtual_processors( vproots, count );
}
#endif
template<>
void scheduler<tbb_connection_v2>::NotifyResourcesExternallyIdle( IVirtualProcessorRoot** /*vproots*/, unsigned int /*count*/)
{
__TBB_ASSERT( false, "NotifyResourcesExternallyIdle() is not allowed for TBB" );
}
template<>
void scheduler<tbb_connection_v2>::NotifyResourcesExternallyBusy( IVirtualProcessorRoot** /*vproots*/, unsigned int /*count*/ )
{
__TBB_ASSERT( false, "NotifyResourcesExternallyBusy() is not allowed for TBB" );
}
template<>
scheduler<omp_connection_v2>::scheduler( omp_connection_v2& conn ) : uid(GetSchedulerId()), my_conn(conn)
{
unsigned count = get_processor_count();
rml::client& cl = my_conn.client();
__TBB_ASSERT( count>1, "The processor count must be greater than 1" );
CREATE_SCHEDULER_POLICY( my_policy, count-1, count-1, cl.min_stack_size() );
}
template<>
void scheduler<omp_connection_v2>::RemoveVirtualProcessors( IVirtualProcessorRoot** /*vproots*/, unsigned int /*count*/ ) {
__TBB_ASSERT( false, "RemoveVirtualProcessors() is not allowed for OMP" );
}
template<>
void scheduler<omp_connection_v2>::NotifyResourcesExternallyIdle( IVirtualProcessorRoot** vproots, unsigned int count ){
if( !my_conn.is_closing() )
my_conn.notify_resources_externally_idle( vproots, count );
}
template<>
void scheduler<omp_connection_v2>::NotifyResourcesExternallyBusy( IVirtualProcessorRoot** vproots, unsigned int count ){
if( !my_conn.is_closing() )
my_conn.notify_resources_externally_busy( vproots, count );
}
/* ts_idle, ts_asleep, ts_busy */
void tbb_server_thread::Dispatch( DispatchState* ) {
// Activate() will resume a thread right after Deactivate() as if it returns from the call
tbb_connection_v2* tbb_conn = static_cast<tbb_connection_v2*>(my_conn);
make_job( *tbb_conn, *this );
for( ;; ) {
// Try to wake some tbb threads if the balance is positive.
// When a thread is added by ConcRT and enter here for the first time,
// the thread may wake itself up (i.e., atomically change its state to ts_busy.
if( the_balance>0 )
wakeup_some_tbb_threads();
if( read_state()!=ts_busy )
if( sleep_perhaps() )
return;
if( terminate )
if( initiate_termination() )
return;
if( read_state()==ts_busy ) {
// this thread has a coin (i.e., state=ts_busy; it should trip to the scheduler at least once
if ( tbb_conn->has_slack() ) {
do {
tbb_conn->try_process( *wait_for_job() );
} while( tbb_conn->has_slack() && the_balance>=0 && !is_removed() );
}
__TBB_ASSERT( read_state()==ts_busy, "thread is not in busy state after returning from process()" );
// see remove_virtual_processors()
if( my_state.compare_and_swap( ts_idle, ts_busy )==ts_busy ) {
int bal = ++the_balance;
if( tbb_conn->has_slack() ) {
// slack is positive, volunteer to help
bal = --the_balance; // try to grab the coin back
if( bal>=0 ) { // got the coin back
if( my_state.compare_and_swap( ts_busy, ts_idle )!=ts_idle )
++the_balance; // someone else enlisted me.
// else my_state is ts_busy, I will come back to tbb_conn->try_process().
} else {
// overdraft. return the coin
++the_balance;
}
} // else the new request will see my changes to state & the_balance.
} else {
__TBB_ASSERT( false, "someone tampered with my state" );
}
} // someone else might set the state to somthing other than ts_idle
}
}
void omp_server_thread::Dispatch( DispatchState* ) {
// Activate() will resume a thread right after Deactivate() as if it returns from the call
make_job( *static_cast<omp_connection_v2*>(my_conn), *this );
for( ;; ) {
if( read_state()!=ts_busy )
sleep_perhaps();
if( terminate ) {
if( initiate_termination() )
return;
}
if( read_state()==ts_busy ) {
omp_data.consume();
__TBB_ASSERT( read_state()==ts_busy, "thread is not in busy state after returning from process()" );
my_thread_map.adjust_balance( 1 );
set_state( ts_idle );
}
// someone else might set the state to somthing other than ts_idle
}
}
//! Attempt to change a thread's state to ts_omp_busy, and waking it up if necessary.
thread_grab_t server_thread_rep::try_grab_for() {
thread_grab_t res = wk_failed;
thread_state_t s = read_state();
switch( s ) {
case ts_asleep:
if( wakeup( ts_busy, ts_asleep ) )
res = wk_from_asleep;
__TBB_ASSERT( res==wk_failed||read_state()==ts_busy, NULL );
break;
case ts_idle:
if( my_state.compare_and_swap( ts_busy, ts_idle )==ts_idle )
res = wk_from_idle;
// At this point a thread is grabbed (i.e., its state has changed to ts_busy.
// It is possible that the thread 1) processes the job, returns from process() and
// sets its state ts_idle again. In some cases, it even sets its state to ts_asleep.
break;
default:
break;
}
return res;
}
bool tbb_server_thread::switch_out() {
thread_state_t s = read_state();
__TBB_ASSERT( s==ts_asleep||s==ts_busy, NULL );
// This thread comes back from the TBB scheduler, and changed its state to ts_asleep successfully.
// The master enlisted it and woke it up by Activate()'ing it; now it is emerging from Deactivated().
// ConcRT requested for removal of the vp associated with the thread, and RML marks it removed.
// Now, it has ts_busy, and removed. -- we should remove it.
IExecutionResource* old_vp = my_execution_resource;
if( s==ts_busy ) {
++the_balance;
my_state = ts_asleep;
}
IThreadProxy* proxy = my_proxy;
__TBB_ASSERT( proxy, NULL );
my_execution_resource = (IExecutionResource*) c_remove_prepare;
old_vp->Remove( my_scheduler );
my_execution_resource = (IExecutionResource*) c_remove_returned;
int cnt = --activation_count;
__TBB_ASSERT_EX( cnt==0||cnt==1, "too many activations?" );
proxy->SwitchOut();
if( terminate ) {
bool activated = activation_count==1;
#if TBB_USE_ASSERT
/* In a rare sequence of events, a thread comes out of SwitchOut with activation_count==1.
* 1) The thread is SwitchOut'ed.
* 2) AddVirtualProcessors() arrived and the thread is Activated.
* 3) The thread is coming out of SwitchOut().
* 4) request_close_connection arrives and inform the thread that it is time to terminate.
* 5) The thread hits the check and falls into the path with 'activated==true'.
* In that case, do the clean-up but do not switch to the thread scavenger; rather simply return to RM.
*/
if( activated ) {
// thread is 'revived' in add_virtual_processors after being Activated().
// so, if the thread extra state is still marked 'removed', it will shortly change to 'none'
// i.e., !is_remove(). The thread state is changed to ts_idle before the extra state, so
// the thread's state should be either ts_idle or ts_done.
while( is_removed() )
__TBB_Yield();
thread_state_t s = read_state();
__TBB_ASSERT( s==ts_idle || s==ts_done, NULL );
}
#endif
__TBB_ASSERT( my_state==ts_asleep||my_state==ts_idle, NULL );
// it is possible that in make_job() the thread may not have a chance to create a job.
// my_job may not be set if the thread did not get a chance to process client's job (i.e., call try_process())
rml::job* j;
if( my_job_automaton.try_plug(j) ) {
__TBB_ASSERT( j, NULL );
my_client.cleanup(*j);
my_conn->remove_client_ref();
}
// Must do remove client reference first, because execution of
// c.remove_ref() can cause *this to be destroyed.
if( !activated )
proxy->SwitchTo( my_thread_map.get_thread_scavenger(), Idle );
my_conn->remove_server_ref();
return true;
}
// We revive a thread in add_virtual_processors() after we Activate the thread on a new virtual processor.
// So briefly wait until the thread's my_execution_resource gets set.
while( get_virtual_processor()==c_remove_returned )
__TBB_Yield();
return false;
}
bool tbb_server_thread::sleep_perhaps () {
if( terminate ) return false;
thread_state_t s = read_state();
if( s==ts_idle ) {
if( my_state.compare_and_swap( ts_asleep, ts_idle )==ts_idle ) {
// If a thread is between read_state() and compare_and_swap(), and the master tries to terminate,
// the master's compare_and_swap() will fail because the thread's state is ts_idle.
// We need to check if terminate is true or not before letting the thread go to sleep oetherwise
// we will miss the terminate signal.
if( !terminate ) {
if( !is_removed() ) {
--activation_count;
get_virtual_processor()->Deactivate( this );
}
if( is_removed() ) {
if( switch_out() )
return true;
__TBB_ASSERT( my_execution_resource>c_remove_returned, NULL );
}
// in add_virtual_processors(), when we revive a thread, we change its state after Activate the thread
// in that case the state may be ts_asleep for a short period
while( read_state()==ts_asleep )
__TBB_Yield();
} else {
if( my_state.compare_and_swap( ts_done, ts_asleep )!=ts_asleep ) {
--activation_count;
// unbind() changed my state. It will call Activate(). So issue a matching Deactivate()
get_virtual_processor()->Deactivate( this );
}
}
}
} else {
__TBB_ASSERT( s==ts_busy, NULL );
}
return false;
}
void omp_server_thread::sleep_perhaps () {
if( terminate ) return;
thread_state_t s = read_state();
if( s==ts_idle ) {
if( my_state.compare_and_swap( ts_asleep, ts_idle )==ts_idle ) {
// If a thread is between read_state() and compare_and_swap(), and the master tries to terminate,
// the master's compare_and_swap() will fail because the thread's state is ts_idle.
// We need to check if terminate is true or not before letting the thread go to sleep oetherwise
// we will miss the terminate signal.
if( !terminate ) {
get_virtual_processor()->Deactivate( this );
__TBB_ASSERT( !is_removed(), "OMP threads should not be deprived of a virtual processor" );
__TBB_ASSERT( read_state()!=ts_asleep, NULL );
} else {
if( my_state.compare_and_swap( ts_done, ts_asleep )!=ts_asleep )
// unbind() changed my state. It will call Activate(). So issue a matching Deactivate()
get_virtual_processor()->Deactivate( this );
}
}
} else {
__TBB_ASSERT( s==ts_busy, NULL );
}
}
bool tbb_server_thread::initiate_termination() {
if( read_state()==ts_busy ) {
int bal = ++the_balance;
if( bal>0 ) wakeup_some_tbb_threads();
}
return destroy_job( (tbb_connection_v2*) my_conn );
}
template<typename Connection>
bool server_thread_rep::destroy_job( Connection* c ) {
__TBB_ASSERT( my_state!=ts_asleep, NULL );
rml::job* j;
if( my_job_automaton.try_plug(j) ) {
__TBB_ASSERT( j, NULL );
my_client.cleanup(*j);
c->remove_client_ref();
}
// Must do remove client reference first, because execution of
// c.remove_ref() can cause *this to be destroyed.
c->remove_server_ref();
return true;
}
void thread_map::assist_cleanup( bool assist_null_only ) {
// To avoid deadlock, the current thread *must* help out with cleanups that have not started,
// becausd the thread that created the job may be busy for a long time.
for( iterator i = begin(); i!=end(); ++i ) {
rml::job* j=0;
server_thread* thr = (*i).second;
job_automaton& ja = thr->my_job_automaton;
if( assist_null_only ? ja.try_plug_null() : ja.try_plug(j) ) {
if( j ) {
my_client.cleanup(*j);
} else {
// server thread did not get a chance to create a job.
}
remove_client_ref();
}
}
}
void thread_map::add_virtual_processors( IVirtualProcessorRoot** vproots, unsigned int count, tbb_connection_v2& conn, ::tbb::spin_mutex& mtx )
{
#if TBB_USE_ASSERT
int req_cnt = ++n_add_vp_requests;
__TBB_ASSERT( req_cnt==1, NULL );
#endif
std::vector<thread_map::iterator> vec(count);
std::vector<tbb_server_thread*> tvec(count);
iterator end;
{
tbb::spin_mutex::scoped_lock lck( mtx );
__TBB_ASSERT( my_map.size()==0||count==1, NULL );
end = my_map.end(); //remember 'end' at the time of 'find'
// find entries in the map for those VPs that were previosly added and then removed.
for( size_t i=0; i<count; ++i ) {
vec[i] = my_map.find( (key_type) vproots[i] );
#if TBB_USE_DEBUG
if( vec[i]!=end ) {
tbb_server_thread* t = (tbb_server_thread*) (*vec[i]).second;
IVirtualProcessorRoot* v = t->get_virtual_processor();
__TBB_ASSERT( v==c_remove_prepare||v==c_remove_returned, NULL );
}
#endif
}
iterator nxt = my_map.begin();
for( size_t i=0; i<count; ++i ) {
if( vec[i]!=end ) {
#if TBB_USE_ASSERT
tbb_server_thread* t = (tbb_server_thread*) (*vec[i]).second;
__TBB_ASSERT( t->read_state()==ts_asleep, NULL );
IVirtualProcessorRoot* r = t->get_virtual_processor();
__TBB_ASSERT( r==c_remove_prepare||r==c_remove_returned, NULL );
#endif
continue;
}
if( my_unrealized_threads>0 ) {
--my_unrealized_threads;
} else {
__TBB_ASSERT( nxt!=end, "nxt should not be thread_map::iterator::end" );
// find a removed thread context for i
for( ; nxt!=end; ++nxt ) {
tbb_server_thread* t = (tbb_server_thread*) (*nxt).second;
if( t->is_removed() && t->read_state()==ts_asleep && t->get_virtual_processor()==c_remove_returned ) {
vec[i] = nxt++;
break;
}
}
// break target
if( vec[i]==end ) // ignore excessive VP.
vproots[i] = NULL;
}
}
}
for( size_t i=0; i<count; ++i ) {
__TBB_ASSERT( !tvec[i], NULL );
if( vec[i]==end ) {
if( vproots[i] ) {
tvec[i] = my_tbb_allocator.allocate(1);
new ( tvec[i] ) tbb_server_thread( false, my_scheduler, (IExecutionResource*)vproots[i], &conn, *this, my_client );
}
#if TBB_USE_ASSERT
} else {
tbb_server_thread* t = (tbb_server_thread*) (*vec[i]).second;
__TBB_ASSERT( t->GetProxy(), "Proxy is cleared?" );
#endif
}
}
{
tbb::spin_mutex::scoped_lock lck( mtx );
bool closing = is_closing();
for( size_t i=0; i<count; ++i ) {
if( vec[i]==end ) {
if( vproots[i] ) {
thread_map::key_type key = (thread_map::key_type) vproots[i];
vec[i] = insert( key, (server_thread*) tvec[i] );
my_client_ref_count.add_ref();
my_server_ref_count.add_ref();
}
} else if( !closing ) {
tbb_server_thread* t = (tbb_server_thread*) (*vec[i]).second;
if( (*vec[i]).first!=(thread_map::key_type)vproots[i] ) {
my_map.erase( vec[i] );
thread_map::key_type key = (thread_map::key_type) vproots[i];
__TBB_ASSERT( key, NULL );
vec[i] = insert( key, t );
}
__TBB_ASSERT( t->read_state()==ts_asleep, NULL );
// We did not decrement server/client ref count when a thread is removed.
// So, don't increment server/client ref count here.
}
}
// we could check is_closing() earlier. That requires marking the newly allocated server_thread objects
// that are not inserted into the thread_map, and deallocate them. Doing so seems more cumbersome
// than simply adding these to the thread_map and let thread_map's destructor take care of reclamation.
__TBB_ASSERT( closing==is_closing(), NULL );
if( closing ) return;
}
for( size_t i=0; i<count; ++i ) {
if( vproots[i] ) {
tbb_server_thread* t = (tbb_server_thread*) (*vec[i]).second;
__TBB_ASSERT( tvec[i]!=NULL||t->GetProxy(), "Proxy is cleared?" );
if( t->is_removed() )
__TBB_ASSERT( t->get_virtual_processor()==c_remove_returned, NULL );
int cnt = ++t->activation_count;
__TBB_ASSERT_EX( cnt==0||cnt==1, NULL );
vproots[i]->Activate( t );
if( t->is_removed() )
t->revive( my_scheduler, vproots[i], my_client );
}
}
#if TBB_USE_ASSERT
req_cnt = --n_add_vp_requests;
__TBB_ASSERT( req_cnt==0, NULL );
#endif
}
void thread_map::remove_virtual_processors( IVirtualProcessorRoot** vproots, unsigned count, ::tbb::spin_mutex& mtx ) {
if( my_map.size()==0 )
return;
tbb::spin_mutex::scoped_lock lck( mtx );
if( is_closing() ) return;
for( unsigned int c=0; c<count; ++c ) {
iterator i = my_map.find( (key_type) vproots[c] );
if( i==my_map.end() ) {
thread_scavenger_thread* tst = my_thread_scavenger_thread;
if( !tst ) {
// Remove unknown vp from my scheduler;
vproots[c]->Remove( my_scheduler );
} else {
while( (tst=my_thread_scavenger_thread)==c_claimed )
__TBB_Yield();
if( vproots[c]!=tst->get_virtual_processor() )
vproots[c]->Remove( my_scheduler );
}
continue;
}
tbb_server_thread* thr = (tbb_server_thread*) (*i).second;
__TBB_ASSERT( thr->tbb_thread, "incorrect type of server_thread" );
thr->set_removed();
if( thr->read_state()==ts_asleep ) {
while( thr->activation_count>0 ) {
if( thr->get_virtual_processor()<=c_remove_returned )
break;
__TBB_Yield();
}
if( thr->get_virtual_processor()>c_remove_returned ) {
// the thread is in Deactivated state
++thr->activation_count;
// wake the thread up so that it Switches Out itself.
thr->get_virtual_processor()->Activate( thr );
} // else, it is Switched Out
} // else the thread will see that it is removed and proceed to switch itself out without Deactivation
}
}
void thread_map::add_virtual_processors( IVirtualProcessorRoot** vproots, unsigned int count, omp_connection_v2& conn, ::tbb::spin_mutex& mtx )
{
std::vector<thread_map::iterator> vec(count);
std::vector<server_thread*> tvec(count);
iterator end;
{
tbb::spin_mutex::scoped_lock lck( mtx );
// read the map
end = my_map.end(); //remember 'end' at the time of 'find'
for( size_t i=0; i<count; ++i )
vec[i] = my_map.find( (key_type) vproots[i] );
}
for( size_t i=0; i<count; ++i ) {
__TBB_ASSERT( !tvec[i], NULL );
if( vec[i]==end ) {
tvec[i] = my_omp_allocator.allocate(1);
new ( tvec[i] ) omp_server_thread( false, my_scheduler, (IExecutionResource*)vproots[i], &conn, *this, my_client );
}
}
{
tbb::spin_mutex::scoped_lock lck( mtx );
for( size_t i=0; i<count; ++i ) {
if( vec[i]==my_map.end() ) {
thread_map::key_type key = (thread_map::key_type) vproots[i];
vec[i] = insert( key, tvec[i] );
my_client_ref_count.add_ref();
my_server_ref_count.add_ref();
}
}
// we could check is_closing() earlier. That requires marking the newly allocated server_thread objects
// that are not inserted into the thread_map, and deallocate them. Doing so seems more cumbersome
// than simply adding these to the thread_map and let thread_map's destructor take care of reclamation.
if( is_closing() ) return;
}
for( size_t i=0; i<count; ++i )
vproots[i]->Activate( (*vec[i]).second );
{
tbb::spin_mutex::scoped_lock lck( mtx );
for( size_t i=0; i<count; ++i )
original_exec_resources.push_back( vproots[i] );
}
}
void thread_map::mark_virtual_processors_as_lent( IVirtualProcessorRoot** vproots, unsigned count, ::tbb::spin_mutex& mtx ) {
tbb::spin_mutex::scoped_lock lck( mtx );
if( is_closing() ) return;
iterator end = my_map.end();
for( unsigned int c=0; c<count; ++c ) {
iterator i = my_map.find( (key_type) vproots[c] );
if( i==end ) {
// The vproc has not been added to the map in create_oversubscribers()
my_map.insert( hash_map_type::value_type( (key_type) vproots[c], (server_thread*)1 ) );
} else {
server_thread* thr = (*i).second;
if( ((uintptr_t)thr)&~(uintptr_t)1 ) {
__TBB_ASSERT( !thr->is_removed(), "incorrectly removed" );
((omp_server_thread*)thr)->set_lent();
}
}
}
}
void thread_map::create_oversubscribers( unsigned n, std::vector<server_thread*>& thr_vec, omp_connection_v2& conn, ::tbb::spin_mutex& mtx ) {
std::vector<IExecutionResource*> curr_exec_rsc;
{
tbb::spin_mutex::scoped_lock lck( mtx );
curr_exec_rsc = original_exec_resources; // copy construct
}
typedef std::vector<IExecutionResource*>::iterator iterator_er;
typedef ::std::vector<std::pair<hash_map_type::key_type, hash_map_type::mapped_type> > hash_val_vector_t;
hash_val_vector_t v_vec(n);
iterator_er begin = curr_exec_rsc.begin();
iterator_er end = curr_exec_rsc.end();
iterator_er i = begin;
for( unsigned c=0; c<n; ++c ) {
IVirtualProcessorRoot* vpr = my_scheduler_proxy->CreateOversubscriber( *i );
omp_server_thread* t = new ( my_omp_allocator.allocate(1) ) omp_server_thread( true, my_scheduler, (IExecutionResource*)vpr, &conn, *this, my_client );
thr_vec[c] = t;
v_vec[c] = hash_map_type::value_type( (key_type) vpr, t );
if( ++i==end ) i = begin;
}
{
tbb::spin_mutex::scoped_lock lck( mtx );
if( is_closing() ) return;
iterator end = my_map.end();
unsigned c = 0;
for( hash_val_vector_t::iterator vi=v_vec.begin(); vi!=v_vec.end(); ++vi, ++c ) {
iterator i = my_map.find( (key_type) (*vi).first );
if( i==end ) {
my_map.insert( *vi );
} else {
// the vproc has not been added to the map in mark_virtual_processors_as_returned();
unsigned lent = (unsigned) (*i).second;
__TBB_ASSERT( lent<=1, "vproc map entry added incorrectly?");
(*i).second = thr_vec[c];
if( lent )
((omp_server_thread*)thr_vec[c])->set_lent();
else
((omp_server_thread*)thr_vec[c])->set_returned();
}
my_client_ref_count.add_ref();
my_server_ref_count.add_ref();
}
}
}
void thread_map::wakeup_tbb_threads( int c, ::tbb::spin_mutex& mtx ) {
std::vector<tbb_server_thread*> vec(c);
size_t idx = 0;
{
tbb::spin_mutex::scoped_lock lck( mtx );
if( is_closing() ) return;
// only one RML thread is in here to wake worker threads up.
int bal = the_balance;
int cnt = c<bal ? c : bal;
if( cnt<=0 ) { return; }
for( iterator i=begin(); i!=end(); ++i ) {
tbb_server_thread* thr = (tbb_server_thread*) (*i).second;
// ConcRT RM should take threads away from TBB scheduler instead of lending them to another scheduler
if( thr->is_removed() )
continue;
if( --the_balance>=0 ) {
thread_grab_t res;
while( (res=thr->try_grab_for())!=wk_from_idle ) {
if( res==wk_from_asleep ) {
vec[idx++] = thr;
break;
} else {
thread_state_t s = thr->read_state();
if( s==ts_busy ) {// failed because already assigned. move on.
++the_balance;
goto skip;
}
}
}
thread_state_t s = thr->read_state();
__TBB_ASSERT_EX( s==ts_busy, "should have set the state to ts_busy" );
if( --cnt==0 )
break;
} else {
// overdraft
++the_balance;
break;
}
skip:
;
}
}
for( size_t i=0; i<idx; ++i ) {
tbb_server_thread* thr = vec[i];
__TBB_ASSERT( thr, NULL );
thread_state_t s = thr->read_state();
__TBB_ASSERT_EX( s==ts_busy, "should have set the state to ts_busy" );
++thr->activation_count;
thr->get_virtual_processor()->Activate( thr );
}
}
void thread_map::mark_virtual_processors_as_returned( IVirtualProcessorRoot** vprocs, unsigned int count, tbb::spin_mutex& mtx ) {
{
tbb::spin_mutex::scoped_lock lck( mtx );
if( is_closing() ) return;
iterator end = my_map.end();
for(unsigned c=0; c<count; ++c ) {
iterator i = my_map.find( (key_type) vprocs[c] );
if( i==end ) {
// the vproc has not been added to the map in create_oversubscribers()
my_map.insert( hash_map_type::value_type( (key_type) vprocs[c], static_cast<server_thread*>(0) ) );
} else {
omp_server_thread* thr = (omp_server_thread*) (*i).second;
if( ((uintptr_t)thr)&~(uintptr_t)1 ) {
__TBB_ASSERT( !thr->is_removed(), "incorrectly removed" );
// we shoud not make any assumption on the initial state of an added vproc.
thr->set_returned();
}
}
}
}
}
void thread_map::unbind( rml::server& /*server*/, tbb::spin_mutex& mtx ) {
{
tbb::spin_mutex::scoped_lock lck( mtx );
shutdown_in_progress = true; // ignore any callbacks from ConcRT RM
// Ask each server_thread to cleanup its job for this server.
for( iterator i = begin(); i!=end(); ++i ) {
server_thread* t = (*i).second;
t->terminate = true;
if( t->is_removed() ) {
// This is for TBB only as ConcRT RM does not request OMP schedulers to remove virtual processors
if( t->read_state()==ts_asleep ) {
__TBB_ASSERT( my_thread_scavenger_thread, "this is TBB connection; thread_scavenger_thread must be allocated" );
// thread is on its way to switch_out; see remove_virtual_processors() where
// the thread is Activated() to bring it back from 'Deactivated' in sleep_perhaps()
// now assume that the thread will go to SwitchOut()
#if TBB_USE_ASSERT
while( t->get_virtual_processor()>c_remove_returned )
__TBB_Yield();
#endif
// A removed thread is supposed to proceed to SwithcOut.
// There, we remove client&server references.
}
} else {
if( t->wakeup( ts_done, ts_asleep ) ) {
if( t->tbb_thread )
++((tbb_server_thread*)t)->activation_count;
t->get_virtual_processor()->Activate( t );
// We mark in the thread_map such that when termination sequence started, we ignore
// all notification from ConcRT RM.
}
}
}
}
// Remove extra ref to client.
remove_client_ref();
if( my_thread_scavenger_thread ) {
thread_scavenger_thread* tst;
while( (tst=my_thread_scavenger_thread)==c_claimed )
__TBB_Yield();
#if TBB_USE_ASSERT
++my_thread_scavenger_thread->activation_count;
#endif
tst->get_virtual_processor()->Activate( tst );
}
}
#if !__RML_REMOVE_VIRTUAL_PROCESSORS_DISABLED
void thread_map::allocate_thread_scavenger( IExecutionResource* v )
{
if( my_thread_scavenger_thread>c_claimed ) return;
thread_scavenger_thread* c = my_thread_scavenger_thread.fetch_and_store((thread_scavenger_thread*)c_claimed);
if( c==NULL ) { // successfully claimed
add_server_ref();
#if TBB_USE_ASSERT
++n_thread_scavengers_created;
#endif
__TBB_ASSERT( v, NULL );
IVirtualProcessorRoot* vpr = my_scheduler_proxy->CreateOversubscriber( v );
my_thread_scavenger_thread = c = new ( my_scavenger_allocator.allocate(1) ) thread_scavenger_thread( my_scheduler, vpr, *this );
#if TBB_USE_ASSERT
++c->activation_count;
#endif
vpr->Activate( c );
} else if( c>c_claimed ) {
my_thread_scavenger_thread = c;
}
}
#endif
void thread_scavenger_thread::Dispatch( DispatchState* )
{
__TBB_ASSERT( my_proxy, NULL );
#if TBB_USE_ASSERT
--activation_count;
#endif
get_virtual_processor()->Deactivate( this );
for( thread_map::iterator i=my_thread_map.begin(); i!=my_thread_map.end(); ++i ) {
tbb_server_thread* t = (tbb_server_thread*) (*i).second;
if( t->read_state()==ts_asleep && t->is_removed() ) {
while( t->get_execution_resource()!=c_remove_returned )
__TBB_Yield();
my_proxy->SwitchTo( t, Blocking );
}
}
get_virtual_processor()->Remove( my_scheduler );
my_thread_map.remove_server_ref();
// signal to the connection scavenger that i am done with the map.
__TBB_ASSERT( activation_count==1, NULL );
set_state( ts_done );
}
//! Windows "DllMain" that handles startup and shutdown of dynamic library.
extern "C" bool WINAPI DllMain( HINSTANCE /*hinstDLL*/, DWORD fwdReason, LPVOID lpvReserved ) {
void assist_cleanup_connections();
if( fwdReason==DLL_PROCESS_DETACH ) {
// dll is being unloaded
if( !lpvReserved ) // if FreeLibrary has been called
assist_cleanup_connections();
}
return true;
}
void free_all_connections( uintptr_t conn_ex ) {
while( conn_ex ) {
bool is_tbb = (conn_ex&2)>0;
//clear extra bits
uintptr_t curr_conn = conn_ex & ~(uintptr_t)3;
__TBB_ASSERT( curr_conn, NULL );
// Wait for worker threads to return
if( is_tbb ) {
tbb_connection_v2* tbb_conn = reinterpret_cast<tbb_connection_v2*>(curr_conn);
conn_ex = reinterpret_cast<uintptr_t>(tbb_conn->next_conn);
while( tbb_conn->my_thread_map.remove_server_ref()>0 )
__TBB_Yield();
delete tbb_conn;
} else {
omp_connection_v2* omp_conn = reinterpret_cast<omp_connection_v2*>(curr_conn);
conn_ex = reinterpret_cast<uintptr_t>(omp_conn->next_conn);
while( omp_conn->my_thread_map.remove_server_ref()>0 )
__TBB_Yield();
delete omp_conn;
}
}
}
void assist_cleanup_connections()
{
//signal to connection_scavenger_thread to terminate
uintptr_t tail = connections_to_reclaim.tail;
while( connections_to_reclaim.tail.compare_and_swap( garbage_connection_queue::plugged, tail )!=tail ) {
__TBB_Yield();
tail = connections_to_reclaim.tail;
}
__TBB_ASSERT( connection_scavenger.state==ts_busy || connection_scavenger.state==ts_asleep, NULL );
// Scavenger thread may be busy freeing connections
DWORD thr_exit_code = STILL_ACTIVE;
while( connection_scavenger.state==ts_busy ) {
if( GetExitCodeThread( connection_scavenger.thr_handle, &thr_exit_code )>0 )
if( thr_exit_code!=STILL_ACTIVE )
break;
__TBB_Yield();
thr_exit_code = STILL_ACTIVE;
}
if( connection_scavenger.state==ts_asleep && thr_exit_code==STILL_ACTIVE )
connection_scavenger.wakeup(); // wake the connection scavenger thread up
// it is possible that the connection scavenger thread already exited. Take over its responsibility.
if( tail && connections_to_reclaim.tail!=garbage_connection_queue::plugged_acked ) {
// atomically claim the head of the list.
uintptr_t head = connections_to_reclaim.head.fetch_and_store( garbage_connection_queue::empty );
if( head==garbage_connection_queue::empty )
head = tail;
connection_scavenger.process_requests( head );
}
__TBB_ASSERT( connections_to_reclaim.tail==garbage_connection_queue::plugged||connections_to_reclaim.tail==garbage_connection_queue::plugged_acked, "someone else added a request after termination has initiated" );
__TBB_ASSERT( the_balance==connection_scavenger.default_concurrency, NULL );
}
void connection_scavenger_thread::sleep_perhaps() {
uintptr_t tail = connections_to_reclaim.tail;
// connections_to_reclaim.tail==garbage_connection_queue::plugged --> terminate,
// connections_to_reclaim.tail>garbage_connection_queue::plugged : we got work to do
if( tail>=garbage_connection_queue::plugged ) return;
__TBB_ASSERT( !tail, NULL );
thread_monitor::cookie c;
monitor.prepare_wait(c);
if( state.compare_and_swap( ts_asleep, ts_busy )==ts_busy ) {
if( connections_to_reclaim.tail!=garbage_connection_queue::plugged ) {
monitor.commit_wait(c);
// Someone else woke me up. The compare_and_swap further below deals with spurious wakeups.
} else {
monitor.cancel_wait();
}
thread_state_t s = state;
if( s==ts_asleep ) // if spurious wakeup.
state.compare_and_swap( ts_busy, ts_asleep );
// I woke myself up, either because I cancelled the wait or suffered a spurious wakeup.
} else {
__TBB_ASSERT( false, "someone else tampered with my state" );
}
__TBB_ASSERT( state==ts_busy, "a thread can only put itself to sleep" );
}
void connection_scavenger_thread::process_requests( uintptr_t conn_ex )
{
__TBB_ASSERT( conn_ex>1, NULL );
__TBB_ASSERT( n_scavenger_threads==1||connections_to_reclaim.tail==garbage_connection_queue::plugged, "more than one connection_scavenger_thread being active?" );
bool done = false;
while( !done ) {
bool is_tbb = (conn_ex&2)>0;
//clear extra bits
uintptr_t curr_conn = conn_ex & ~(uintptr_t)3;
// no contention. there is only one connection_scavenger_thread!!
uintptr_t next_conn;
tbb_connection_v2* tbb_conn = NULL;
omp_connection_v2* omp_conn = NULL;
// Wait for worker threads to return
if( is_tbb ) {
tbb_conn = reinterpret_cast<tbb_connection_v2*>(curr_conn);
next_conn = reinterpret_cast<uintptr_t>(tbb_conn->next_conn);
while( tbb_conn->my_thread_map.get_server_ref_count()>1 )
__TBB_Yield();
} else {
omp_conn = reinterpret_cast<omp_connection_v2*>(curr_conn);
next_conn = reinterpret_cast<uintptr_t>(omp_conn->next_conn);
while( omp_conn->my_thread_map.get_server_ref_count()>1 )
__TBB_Yield();
}
//someone else may try to write into this connection object.
//So access next_conn field first before remove the extra server ref count.
if( next_conn==0 ) {
uintptr_t tail = connections_to_reclaim.tail;
if( tail==garbage_connection_queue::plugged ) {
tail = garbage_connection_queue::plugged_acked; // connection scavenger saw the flag, and it freed all connections.
done = true;
} else if( tail==conn_ex ) {
if( connections_to_reclaim.tail.compare_and_swap( garbage_connection_queue::empty, tail )==tail ) {
__TBB_ASSERT( !connections_to_reclaim.head, NULL );
done = true;
}
}
if( !done ) {
// A new connection to close is added to connections_to_reclaim.tail;
// Wait for curr_conn->next_conn to be set.
if( is_tbb ) {
while( !tbb_conn->next_conn )
__TBB_Yield();
conn_ex = reinterpret_cast<uintptr_t>(tbb_conn->next_conn);
} else {
while( !omp_conn->next_conn )
__TBB_Yield();
conn_ex = reinterpret_cast<uintptr_t>(omp_conn->next_conn);
}
}
} else {
conn_ex = next_conn;
}
__TBB_ASSERT( conn_ex, NULL );
if( is_tbb )
// remove extra srever ref count; this will trigger Shutdown/Release of ConcRT RM
tbb_conn->remove_server_ref();
else
// remove extra srever ref count; this will trigger Shutdown/Release of ConcRT RM
omp_conn->remove_server_ref();
}
}
__RML_DECL_THREAD_ROUTINE connection_scavenger_thread::thread_routine( void* arg ) {
connection_scavenger_thread* thr = (connection_scavenger_thread*) arg;
thr->state = ts_busy;
thr->thr_handle = GetCurrentThread();
#if TBB_USE_ASSERT
++thr->n_scavenger_threads;
#endif
for(;;) {
__TBB_Yield();
thr->sleep_perhaps();
if( connections_to_reclaim.tail==garbage_connection_queue::plugged || connections_to_reclaim.tail==garbage_connection_queue::plugged_acked ) {
thr->state = ts_asleep;
return 0;
}
__TBB_ASSERT( connections_to_reclaim.tail!=garbage_connection_queue::plugged_acked, NULL );
__TBB_ASSERT( connections_to_reclaim.tail>garbage_connection_queue::plugged && (connections_to_reclaim.tail&garbage_connection_queue::plugged)==0 , NULL );
while( connections_to_reclaim.head==garbage_connection_queue::empty )
__TBB_Yield();
uintptr_t head = connections_to_reclaim.head.fetch_and_store( garbage_connection_queue::empty );
thr->process_requests( head );
wakeup_some_tbb_threads();
}
}
template<typename Server, typename Client>
void connection_scavenger_thread::add_request( generic_connection<Server,Client>* conn_to_close )
{
uintptr_t conn_ex = (uintptr_t)conn_to_close | (connection_traits<Server,Client>::is_tbb<<1);
__TBB_ASSERT( !conn_to_close->next_conn, NULL );
uintptr_t old_tail_ex = connections_to_reclaim.tail;
__TBB_ASSERT( old_tail_ex==0||old_tail_ex>garbage_connection_queue::plugged_acked, "Unloading DLL called while this connection is being closed?" );
tbb::internal::atomic_backoff backoff;
while( connections_to_reclaim.tail.compare_and_swap( conn_ex, old_tail_ex )!=old_tail_ex ) {
backoff.pause();
old_tail_ex = connections_to_reclaim.tail;
}
if( old_tail_ex==garbage_connection_queue::empty )
connections_to_reclaim.head = conn_ex;
else {
bool is_tbb = (old_tail_ex&2)>0;
uintptr_t old_tail = old_tail_ex & ~(uintptr_t)3;
if( is_tbb )
reinterpret_cast<tbb_connection_v2*>(old_tail)->next_conn = reinterpret_cast<tbb_connection_v2*>(conn_ex);
else
reinterpret_cast<omp_connection_v2*>(old_tail)->next_conn = reinterpret_cast<omp_connection_v2*>(conn_ex);
}
if( state==ts_asleep )
wakeup();
}
template<>
uintptr_t connection_scavenger_thread::grab_and_prepend( generic_connection<tbb_server,tbb_client>* /*last_conn_to_close*/ ) { return 0;}
template<>
uintptr_t connection_scavenger_thread::grab_and_prepend( generic_connection<omp_server,omp_client>* last_conn_to_close )
{
uintptr_t conn_ex = (uintptr_t)last_conn_to_close;
uintptr_t head = connections_to_reclaim.head.fetch_and_store( garbage_connection_queue::empty );
reinterpret_cast<omp_connection_v2*>(last_conn_to_close)->next_conn = reinterpret_cast<omp_connection_v2*>(head);
return conn_ex;
}
extern "C" ULONGLONG NTAPI VerSetConditionMask( ULONGLONG, DWORD, BYTE);
bool is_windows7_or_later ()
{
try {
return GetOSVersion()>=IResourceManager::Win7OrLater;
} catch( ... ) {
return false;
}
}
#endif /* RML_USE_WCRM */
template<typename Connection, typename Server, typename Client>
static factory::status_type connect( factory& f, Server*& server, Client& client ) {
server = new Connection(*static_cast<wait_counter*>(f.scratch_ptr),client);
return factory::st_success;
}
extern "C" factory::status_type __RML_open_factory( factory& f, version_type& server_version, version_type client_version ) {
// Hack to keep this library from being closed by causing the first client's dlopen to not have a corresponding dlclose.
// This code will be removed once we figure out how to do shutdown of the RML perfectly.
static tbb::atomic<bool> one_time_flag;
if( one_time_flag.compare_and_swap(true,false)==false) {
__TBB_ASSERT( (size_t)f.library_handle!=factory::c_dont_unload, NULL );
#if _WIN32||_WIN64
f.library_handle = reinterpret_cast<HMODULE>(factory::c_dont_unload);
#else
f.library_handle = reinterpret_cast<void*>(factory::c_dont_unload);
#endif
}
// End of hack
// Initialize the_balance only once
if( the_balance_inited!=2 ) {
if( the_balance_inited.compare_and_swap( 1, 0 )==0 ) {
the_balance = hardware_concurrency()-1;
the_balance_inited = 2;
#if RML_USE_WCRM
connection_scavenger.launch( the_balance );
#endif
} else {
tbb::internal::spin_wait_until_eq( the_balance_inited, 2 );
}
}
server_version = SERVER_VERSION;
f.scratch_ptr = 0;
if( client_version==0 ) {
return factory::st_incompatible;
#if RML_USE_WCRM
} else if ( !is_windows7_or_later() ) {
#if TBB_USE_DEBUG
fprintf(stderr, "This version of the RML library requires Windows 7 to run on.\nConnection request denied.\n");
#endif
return factory::st_incompatible;
#endif
} else {
#if TBB_USE_DEBUG
if( client_version<EARLIEST_COMPATIBLE_CLIENT_VERSION )
fprintf(stderr, "This client library is too old for the current RML server.\nThe connection request is granted but oversubscription/undersubscription may occur.\n");
#endif
f.scratch_ptr = new wait_counter;
return factory::st_success;
}
}
extern "C" void __RML_close_factory( factory& f ) {
if( wait_counter* fc = static_cast<wait_counter*>(f.scratch_ptr) ) {
f.scratch_ptr = 0;
fc->wait();
size_t bal = the_balance;
f.scratch_ptr = (void*)bal;
delete fc;
}
}
void call_with_build_date_str( ::rml::server_info_callback_t cb, void* arg );
}} // rml::internal
namespace tbb {
namespace internal {
namespace rml {
extern "C" tbb_factory::status_type __TBB_make_rml_server( tbb_factory& f, tbb_server*& server, tbb_client& client ) {
return ::rml::internal::connect< ::rml::internal::tbb_connection_v2>(f,server,client);
}
extern "C" void __TBB_call_with_my_server_info( ::rml::server_info_callback_t cb, void* arg ) {
return ::rml::internal::call_with_build_date_str( cb, arg );
}
}}}
namespace __kmp {
namespace rml {
extern "C" omp_factory::status_type __KMP_make_rml_server( omp_factory& f, omp_server*& server, omp_client& client ) {
return ::rml::internal::connect< ::rml::internal::omp_connection_v2>(f,server,client);
}
extern "C" void __KMP_call_with_my_server_info( ::rml::server_info_callback_t cb, void* arg ) {
return ::rml::internal::call_with_build_date_str( cb, arg );
}
}}
/*
* RML server info
*/
#include "version_string.tmp"
#ifndef __TBB_VERSION_STRINGS
#pragma message("Warning: version_string.tmp isn't generated properly by version_info.sh script!")
#endif
// We use the build time as the RML server info. TBB is required to build RML, so we make it the same as the TBB build time.
#ifndef __TBB_DATETIME
#define __TBB_DATETIME __DATE__ " " __TIME__
#endif
#if !RML_USE_WCRM
#define RML_SERVER_BUILD_TIME "Intel(R) RML library built: " __TBB_DATETIME
#define RML_SERVER_VERSION_ST "Intel(R) RML library version: v" TOSTRING(SERVER_VERSION)
#else
#define RML_SERVER_BUILD_TIME "Intel(R) RML library built: " __TBB_DATETIME
#define RML_SERVER_VERSION_ST "Intel(R) RML library version: v" TOSTRING(SERVER_VERSION) " on ConcRT RM with " RML_THREAD_KIND_STRING
#endif
namespace rml {
namespace internal {
void call_with_build_date_str( ::rml::server_info_callback_t cb, void* arg )
{
(*cb)( arg, RML_SERVER_BUILD_TIME );
(*cb)( arg, RML_SERVER_VERSION_ST );
}
}} // rml::internal