xref: /oneTBB/src/tbb/private_server.cpp (revision 2eccd5f9)
1 /*
2     Copyright (c) 2005-2022 Intel Corporation
3 
4     Licensed under the Apache License, Version 2.0 (the "License");
5     you may not use this file except in compliance with the License.
6     You may obtain a copy of the License at
7 
8         http://www.apache.org/licenses/LICENSE-2.0
9 
10     Unless required by applicable law or agreed to in writing, software
11     distributed under the License is distributed on an "AS IS" BASIS,
12     WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13     See the License for the specific language governing permissions and
14     limitations under the License.
15 */
16 
17 #include "oneapi/tbb/cache_aligned_allocator.h"
18 #include "oneapi/tbb/mutex.h"
19 
20 #include "rml_tbb.h"
21 #include "rml_thread_monitor.h"
22 
23 #include "scheduler_common.h"
24 #include "governor.h"
25 #include "misc.h"
26 
27 #include <atomic>
28 
29 
30 namespace tbb {
31 namespace detail {
32 namespace r1 {
33 namespace rml {
34 
35 using rml::internal::thread_monitor;
36 typedef thread_monitor::handle_type thread_handle;
37 
38 class private_server;
39 
40 class private_worker: no_copy {
41 private:
42     //! State in finite-state machine that controls the worker.
43     /** State diagram:
44         init --> starting --> normal
45           |         |           |
46           |         V           |
47           \------> quit <------/
48       */
49     enum state_t {
50         //! *this is initialized
51         st_init,
52         //! *this has associated thread that is starting up.
53         st_starting,
54         //! Associated thread is doing normal life sequence.
55         st_normal,
56         //! Associated thread has ended normal life sequence and promises to never touch *this again.
57         st_quit
58     };
59     std::atomic<state_t> my_state;
60 
61     //! Associated server
62     private_server& my_server;
63 
64     //! Associated client
65     tbb_client& my_client;
66 
67     //! index used for avoiding the 64K aliasing problem
68     const std::size_t my_index;
69 
70     //! Monitor for sleeping when there is no work to do.
71     /** The invariant that holds for sleeping workers is:
72         "my_slack<=0 && my_state==st_normal && I am on server's list of asleep threads" */
73     thread_monitor my_thread_monitor;
74 
75     //! Handle of the OS thread associated with this worker
76     thread_handle my_handle;
77 
78     //! Link for list of workers that are sleeping or have no associated thread.
79     private_worker* my_next;
80 
81     friend class private_server;
82 
83     //! Actions executed by the associated thread
84     void run() noexcept;
85 
86     //! Wake up associated thread (or launch a thread if there is none)
87     void wake_or_launch();
88 
89     //! Called by a thread (usually not the associated thread) to commence termination.
90     void start_shutdown();
91 
92     static __RML_DECL_THREAD_ROUTINE thread_routine( void* arg );
93 
94     static void release_handle(thread_handle my_handle, bool join);
95 
96 protected:
97     private_worker( private_server& server, tbb_client& client, const std::size_t i ) :
98         my_state(st_init), my_server(server), my_client(client), my_index(i),
99         my_handle(), my_next()
100     {}
101 };
102 
103 static const std::size_t cache_line_size = tbb::detail::max_nfs_size;
104 
105 #if _MSC_VER && !defined(__INTEL_COMPILER)
106     // Suppress overzealous compiler warnings about uninstantiable class
107     #pragma warning(push)
108     #pragma warning(disable:4510 4610)
109 #endif
110 class padded_private_worker: public private_worker {
111     char pad[cache_line_size - sizeof(private_worker)%cache_line_size];
112 public:
113     padded_private_worker( private_server& server, tbb_client& client, const std::size_t i )
114     : private_worker(server,client,i) { suppress_unused_warning(pad); }
115 };
116 #if _MSC_VER && !defined(__INTEL_COMPILER)
117     #pragma warning(pop)
118 #endif
119 
120 class private_server: public tbb_server, no_copy {
121 private:
122     tbb_client& my_client;
123     //! Maximum number of threads to be created.
124     /** Threads are created lazily, so maximum might not actually be reached. */
125     const tbb_client::size_type my_n_thread;
126 
127     //! Stack size for each thread. */
128     const std::size_t my_stack_size;
129 
130     //! Number of jobs that could use their associated thread minus number of active threads.
131     /** If negative, indicates oversubscription.
132         If positive, indicates that more threads should run.
133         Can be lowered asynchronously, but must be raised only while holding my_asleep_list_mutex,
134         because raising it impacts the invariant for sleeping threads. */
135     std::atomic<int> my_slack;
136 
137     //! Counter used to determine when to delete this.
138     std::atomic<int> my_ref_count;
139 
140     padded_private_worker* my_thread_array;
141 
142     //! List of workers that are asleep or committed to sleeping until notified by another thread.
143     std::atomic<private_worker*> my_asleep_list_root;
144 
145     //! Protects my_asleep_list_root
146     typedef mutex asleep_list_mutex_type;
147     asleep_list_mutex_type my_asleep_list_mutex;
148 
149 #if TBB_USE_ASSERT
150     std::atomic<int> my_net_slack_requests;
151 #endif /* TBB_USE_ASSERT */
152 
153     //! Wake up to two sleeping workers, if there are any sleeping.
154     /** The call is used to propagate a chain reaction where each thread wakes up two threads,
155         which in turn each wake up two threads, etc. */
156     void propagate_chain_reaction() {
157         // First test of a double-check idiom.  Second test is inside wake_some(0).
158         if( my_asleep_list_root.load(std::memory_order_relaxed) )
159             wake_some(0);
160     }
161 
162     //! Try to add t to list of sleeping workers
163     bool try_insert_in_asleep_list( private_worker& t );
164 
165     //! Equivalent of adding additional_slack to my_slack and waking up to 2 threads if my_slack permits.
166     void wake_some( int additional_slack );
167 
168     ~private_server() override;
169 
170     void remove_server_ref() {
171         if( --my_ref_count==0 ) {
172             my_client.acknowledge_close_connection();
173             this->~private_server();
174             tbb::cache_aligned_allocator<private_server>().deallocate( this, 1 );
175         }
176     }
177 
178     friend class private_worker;
179 public:
180     private_server( tbb_client& client );
181 
182     version_type version() const override {
183         return 0;
184     }
185 
186     void request_close_connection( bool /*exiting*/ ) override {
187         for( std::size_t i=0; i<my_n_thread; ++i )
188             my_thread_array[i].start_shutdown();
189         remove_server_ref();
190     }
191 
192     void yield() override { d0::yield(); }
193 
194     void independent_thread_number_changed( int ) override {__TBB_ASSERT(false, nullptr);}
195 
196     unsigned default_concurrency() const override { return governor::default_num_threads() - 1; }
197 
198     void adjust_job_count_estimate( int delta ) override;
199 
200 #if _WIN32 || _WIN64
201     void register_external_thread ( ::rml::server::execution_resource_t& ) override {}
202     void unregister_external_thread ( ::rml::server::execution_resource_t ) override {}
203 #endif /* _WIN32||_WIN64 */
204 };
205 
206 //------------------------------------------------------------------------
207 // Methods of private_worker
208 //------------------------------------------------------------------------
209 #if _MSC_VER && !defined(__INTEL_COMPILER)
210     // Suppress overzealous compiler warnings about an initialized variable 'sink_for_alloca' not referenced
211     #pragma warning(push)
212     #pragma warning(disable:4189)
213 #endif
214 #if __MINGW32__ && __GNUC__==4 &&__GNUC_MINOR__>=2 && !__MINGW64__
215 // ensure that stack is properly aligned for TBB threads
216 __attribute__((force_align_arg_pointer))
217 #endif
218 __RML_DECL_THREAD_ROUTINE private_worker::thread_routine( void* arg ) {
219     private_worker* self = static_cast<private_worker*>(arg);
220     AVOID_64K_ALIASING( self->my_index );
221     self->run();
222     // return 0 instead of nullptr due to the difference in the type __RML_DECL_THREAD_ROUTINE on various OSs
223     return 0;
224 }
225 #if _MSC_VER && !defined(__INTEL_COMPILER)
226     #pragma warning(pop)
227 #endif
228 
229 void private_worker::release_handle(thread_handle handle, bool join) {
230     if (join)
231         thread_monitor::join(handle);
232     else
233         thread_monitor::detach_thread(handle);
234 }
235 
236 void private_worker::start_shutdown() {
237     __TBB_ASSERT(my_state.load(std::memory_order_relaxed) != st_quit, "The quit state is expected to be set only once");
238 
239     // `acq` to acquire my_handle
240     // `rel` to release market state
241     state_t prev_state = my_state.exchange(st_quit, std::memory_order_acq_rel);
242 
243     if (prev_state == st_init) {
244         // Perform action that otherwise would be performed by associated thread when it quits.
245         my_server.remove_server_ref();
246     } else {
247         __TBB_ASSERT(prev_state == st_normal || prev_state == st_starting, nullptr);
248         // May have invalidated invariant for sleeping, so wake up the thread.
249         // Note that the notify() here occurs without maintaining invariants for my_slack.
250         // It does not matter, because my_state==st_quit overrides checking of my_slack.
251         my_thread_monitor.notify();
252         // Do not need release handle in st_init state,
253         // because in this case the thread wasn't started yet.
254         // For st_starting release is done at launch site.
255         if (prev_state == st_normal)
256             release_handle(my_handle, governor::does_client_join_workers(my_client));
257     }
258 }
259 
260 void private_worker::run() noexcept {
261     my_server.propagate_chain_reaction();
262 
263     // Transiting to st_normal here would require setting my_handle,
264     // which would create race with the launching thread and
265     // complications in handle management on Windows.
266 
267     ::rml::job& j = *my_client.create_one_job();
268     // memory_order_seq_cst to be strictly ordered after thread_monitor::wait on the next iteration
269     while( my_state.load(std::memory_order_seq_cst)!=st_quit ) {
270         if( my_server.my_slack.load(std::memory_order_acquire)>=0 ) {
271             my_client.process(j);
272         } else if( my_server.try_insert_in_asleep_list(*this) ) {
273             my_thread_monitor.wait();
274             __TBB_ASSERT(my_state.load(std::memory_order_relaxed) == st_quit || !my_next, "Thread monitor missed a spurious wakeup?" );
275             my_server.propagate_chain_reaction();
276         }
277     }
278     my_client.cleanup(j);
279 
280     ++my_server.my_slack;
281     my_server.remove_server_ref();
282 }
283 
284 inline void private_worker::wake_or_launch() {
285     state_t state = my_state.load(std::memory_order_relaxed);
286 
287     switch (state) {
288     case st_starting:
289         __TBB_fallthrough;
290     case st_normal:
291         __TBB_ASSERT(!my_next, "Should not wake a thread while it's still in asleep list");
292         my_thread_monitor.notify();
293         break;
294     case st_init:
295         if (my_state.compare_exchange_strong(state, st_starting)) {
296             // after this point, remove_server_ref() must be done by created thread
297 #if __TBB_USE_WINAPI
298             // Win thread_monitor::launch is designed on the assumption that the workers thread id go from 1 to Hard limit set by TBB market::global_market
299             const std::size_t worker_idx = my_server.my_n_thread - this->my_index;
300             my_handle = thread_monitor::launch(thread_routine, this, my_server.my_stack_size, &worker_idx);
301 #elif __TBB_USE_POSIX
302             {
303                 affinity_helper fpa;
304                 fpa.protect_affinity_mask( /*restore_process_mask=*/true);
305                 my_handle = thread_monitor::launch(thread_routine, this, my_server.my_stack_size);
306                 // Implicit destruction of fpa resets original affinity mask.
307             }
308 #endif /* __TBB_USE_POSIX */
309             state = st_starting;
310             if (!my_state.compare_exchange_strong(state, st_normal)) {
311                 // Do shutdown during startup. my_handle can't be released
312                 // by start_shutdown, because my_handle value might be not set yet
313                 // at time of transition from st_starting to st_quit.
314                 __TBB_ASSERT(state == st_quit, nullptr);
315                 release_handle(my_handle, governor::does_client_join_workers(my_client));
316             }
317         }
318         break;
319     default:
320         __TBB_ASSERT(state == st_quit, nullptr);
321     }
322 }
323 
324 //------------------------------------------------------------------------
325 // Methods of private_server
326 //------------------------------------------------------------------------
327 private_server::private_server( tbb_client& client ) :
328     my_client(client),
329     my_n_thread(client.max_job_count()),
330     my_stack_size(client.min_stack_size()),
331     my_slack(0),
332     my_ref_count(my_n_thread+1),
333     my_thread_array(nullptr),
334     my_asleep_list_root(nullptr)
335 #if TBB_USE_ASSERT
336     , my_net_slack_requests(0)
337 #endif /* TBB_USE_ASSERT */
338 {
339     my_thread_array = tbb::cache_aligned_allocator<padded_private_worker>().allocate( my_n_thread );
340     for( std::size_t i=0; i<my_n_thread; ++i ) {
341         private_worker* t = new( &my_thread_array[i] ) padded_private_worker( *this, client, i );
342         t->my_next = my_asleep_list_root.load(std::memory_order_relaxed);
343         my_asleep_list_root.store(t, std::memory_order_relaxed);
344     }
345 }
346 
347 private_server::~private_server() {
348     __TBB_ASSERT( my_net_slack_requests==0, nullptr);
349     for( std::size_t i=my_n_thread; i--; )
350         my_thread_array[i].~padded_private_worker();
351     tbb::cache_aligned_allocator<padded_private_worker>().deallocate( my_thread_array, my_n_thread );
352     tbb::detail::poison_pointer( my_thread_array );
353 }
354 
355 inline bool private_server::try_insert_in_asleep_list( private_worker& t ) {
356     asleep_list_mutex_type::scoped_lock lock;
357     if( !lock.try_acquire(my_asleep_list_mutex) )
358         return false;
359     // Contribute to slack under lock so that if another takes that unit of slack,
360     // it sees us sleeping on the list and wakes us up.
361     auto expected = my_slack.load(std::memory_order_relaxed);
362     while (expected < 0) {
363         if (my_slack.compare_exchange_strong(expected, expected + 1)) {
364             t.my_next = my_asleep_list_root.load(std::memory_order_relaxed);
365             my_asleep_list_root.store(&t, std::memory_order_relaxed);
366             return true;
367         }
368     }
369 
370     return false;
371 }
372 
373 void private_server::wake_some( int additional_slack ) {
374     __TBB_ASSERT( additional_slack>=0, nullptr );
375     private_worker* wakee[2];
376     private_worker**w = wakee;
377 
378     if (additional_slack) {
379         // Contribute our unused slack to my_slack.
380         my_slack += additional_slack;
381     }
382 
383     int allotted_slack = 0;
384     while (allotted_slack < 2) {
385         // Chain reaction; Try to claim unit of slack
386         int old = my_slack.load(std::memory_order_relaxed);
387         do {
388             if (old <= 0) goto done;
389         } while (!my_slack.compare_exchange_strong(old, old - 1));
390         ++allotted_slack;
391     }
392 done:
393 
394     if (allotted_slack) {
395         asleep_list_mutex_type::scoped_lock lock(my_asleep_list_mutex);
396         auto root = my_asleep_list_root.load(std::memory_order_relaxed);
397         while( root && w<wakee+2 && allotted_slack) {
398             --allotted_slack;
399             // Pop sleeping worker to combine with claimed unit of slack
400             *w++ = root;
401             root = root->my_next;
402         }
403         my_asleep_list_root.store(root, std::memory_order_relaxed);
404         if(allotted_slack) {
405             // Contribute our unused slack to my_slack.
406             my_slack += allotted_slack;
407         }
408     }
409     while( w>wakee ) {
410         private_worker* ww = *--w;
411         ww->my_next = nullptr;
412         ww->wake_or_launch();
413     }
414 }
415 
416 void private_server::adjust_job_count_estimate( int delta ) {
417 #if TBB_USE_ASSERT
418     my_net_slack_requests+=delta;
419 #endif /* TBB_USE_ASSERT */
420     if( delta<0 ) {
421         my_slack+=delta;
422     } else if( delta>0 ) {
423         wake_some( delta );
424     }
425 }
426 
427 //! Factory method called from task.cpp to create a private_server.
428 tbb_server* make_private_server( tbb_client& client ) {
429     return new( tbb::cache_aligned_allocator<private_server>().allocate(1) ) private_server(client);
430 }
431 
432 } // namespace rml
433 } // namespace r1
434 } // namespace detail
435 } // namespace tbb
436 
437