2 This file is a part of libcds - Concurrent Data Structures library
4 (C) Copyright Maxim Khizhinsky (libcds.dev@gmail.com) 2006-2017
6 Source code repo: http://github.com/khizmax/libcds/
7 Download: http://sourceforge.net/projects/libcds/files/
9 Redistribution and use in source and binary forms, with or without
10 modification, are permitted provided that the following conditions are met:
12 * Redistributions of source code must retain the above copyright notice, this
13 list of conditions and the following disclaimer.
15 * Redistributions in binary form must reproduce the above copyright notice,
16 this list of conditions and the following disclaimer in the documentation
17 and/or other materials provided with the distribution.
19 THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
20 AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21 IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
22 DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
23 FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24 DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
25 SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
26 CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
27 OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
28 OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
31 #ifndef CDSLIB_URCU_DETAILS_BASE_H
32 #define CDSLIB_URCU_DETAILS_BASE_H
34 #include <cds/algo/atomic.h>
35 #include <cds/gc/details/retired_ptr.h>
36 #include <cds/details/allocator.h>
37 #include <cds/os/thread.h>
38 #include <cds/details/marked_ptr.h>
41 /// User-space Read-Copy Update (URCU) namespace
42 /** @ingroup cds_garbage_collector
45 This namespace contains declarations for different types of Read-Copy Update (%RCU)
46 synchronization primitive and data structures developed for RCU.
47 In <b>libcds</b> %RCU is used as garbage collector.
50 - [2009] M.Desnoyers "Low-Impact Operating System Tracing" PhD Thesis,
51 Chapter 6 "User-Level Implementations of Read-Copy Update"
52 - [2011] M.Desnoyers, P.McKenney, A.Stern, M.Dagenias, J.Walpole "User-Level
53 Implementations of Read-Copy Update"
55 <b>Informal introduction to user-space %RCU</b>
57 [<i>From Desnoyer's papers</i>] %RCU is a synchronization mechanism that was added to the
58 Linux kernel in October of 2002. %RCU achieves scalability improvements by allowing
59 reads to occur concurrently with updates. In contrast to conventional locking
60 primitives that ensure mutual exclusion among concurrent threads regardless of whether
61 they be readers or updaters, or with reader-writer locks that allow concurrent reads
62 but not in the presence of updates, %RCU supports concurrency between multiple updaters
63 and multiple readers. %RCU ensures that data are not freed up until all pre-existing
64 critical sections complete. %RCU defines and uses efficient and scalable mechanisms
65 for deferring reclamation of old data. These mechanisms distribute the work among read and update
66 paths in such a way as to make read paths extremely fast.
68 %RCU readers execute within %RCU read-side critical sections. Each such critical section begins with
69 \p rcu_read_lock(), ends with \p rcu_read_unlock() (in \p libcds these primitives are the methods of
70 GC class and are usually called \p access_lock and \p access_unlock respectively). Read-side
71 critical sections can be nested.
72 The performance benefits of %RCU are due to the fact that \p rcu_read_lock()
73 and \p rcu_read_unlock() are exceedingly fast.
75 When a thread is not in an %RCU read-side critical section, it is in a quiescent state.
76 A quiescent state that persists for a significant time period is an extended quiescent state.
77 Any time period during which every thread has been in at least one quiescent state
78 is a grace period; this implies that every %RCU read-side critical section
79 that starts before a grace period must end before that grace period does.
80 Distinct grace periods may overlap, either partially or completely. Any time period
81 that includes a grace period is by definition itself a grace period.
82 Each grace period is guaranteed to complete as long as all read-side critical sections
83 are finite in duration; thus even a constant flow of such critical sections is unable to
84 extend an %RCU grace period indefinitely.
86 Suppose that readers enclose each of their data-structure traversals in
87 an %RCU read-side critical section. If an updater first removes an element
88 from such a data structure and then waits for a grace period, there can be
89 no more readers accessing that element. The updater can then carry out destructive
90 operations, for example freeing the element, without disturbing any readers.
92 The %RCU update is split into two phases, a removal phase and a reclamation phase.
93 These two phases must be separated by a grace period, for example via the \p synchronize_rcu()
94 primitive, which initiates a grace period and waits for it to finish.
95 During the removal phase, the %RCU update removes elements from a shared data structure.
96 The removed data elements will only be accessible to read-side critical sections
97 that ran concurrently with the removal phase, which are guaranteed to complete before the
98 grace period ends. Therefore the reclamation phase can safely free the data elements
99 removed by the removal phase.
101 Desnoyers describes several classes of user-space %RCU implementations:
102 - The Quiescent-State-Based Reclamation (QSBR) %RCU implementation offers
103 the best possible read-side performance, but requires that each thread periodically
104 calls a function to announce that it is in a quiescent state, thus strongly
105 constraining the application design. This type of %RCU is not implemented in \p libcds.
106 - The general-purpose %RCU implementation places almost no constraints on the application's
107 design, thus being appropriate for use within a general-purpose library, but it has
108 relatively higher read-side overhead. The \p libcds contains several implementations of general-purpose
109 %RCU: \ref general_instant, \ref general_buffered, \ref general_threaded.
110 - \ref signal_buffered: the signal-handling %RCU presents an implementation having low read-side overhead and
111 requiring only that the application give up one POSIX signal to %RCU update processing.
113 @note The signal-handled %RCU is defined only for UNIX-like systems, not for Windows.
115 @anchor cds_urcu_type
116 <b>RCU implementation type</b>
118 There are several internal implementation of RCU (all declared in \p %cds::urcu namespace):
119 - \ref general_instant - general purpose RCU with immediate reclamation
120 - \ref general_buffered - general purpose RCU with deferred (buffered) reclamation
121 - \ref general_threaded - general purpose RCU with special reclamation thread
122 - \ref signal_buffered - signal-handling RCU with deferred (buffered) reclamation
124 You cannot create an object of any of those classes directly.
125 Instead, you should use wrapper classes.
126 The wrapper simplifies creation and usage of RCU singleton objects
127 and has the reacher interface that combines interfaces of wrapped class i.e. RCU global part like
128 \p synchronize, and corresponding RCU thread-specific interface like \p access_lock, \p access_unlock and \p retire_ptr.
131 There are several wrapper classes (all declared in \p %cds::urcu namespace)
132 - \ref cds_urcu_general_instant_gc "gc<general_instant>" - general purpose RCU with immediate reclamation,
133 include file <tt><cds/urcu/general_instant.h></tt>
134 - \ref cds_urcu_general_buffered_gc "gc<general_buffered>" - general purpose RCU with deferred (buffered) reclamation,
135 include file <tt><cds/urcu/general_buffered.h></tt>
136 - \ref cds_urcu_general_threaded_gc "gc<general_threaded>" - general purpose RCU with special reclamation thread
137 include file <tt><cds/urcu/general_threaded.h></tt>
138 - \ref cds_urcu_signal_buffered_gc "gc<signal_buffered>" - signal-handling RCU with deferred (buffered) reclamation
139 include file <tt><cds/urcu/signal_buffered.h></tt>
141 Any RCU-related container in \p libcds expects that its \p RCU template parameter is one of those wrapper.
143 @anchor cds_urcu_tags
144 For simplicity, in some algorithms instead of using RCU implementation type
145 you should specify corresponding RCU tags (all declared in \p %cds::urcu namespace):
146 - \ref general_instant_tag - for \ref general_instant
147 - \ref general_buffered_tag - for \ref general_buffered
148 - \ref general_threaded_tag - for \ref general_threaded
149 - \ref signal_buffered_tag - for \ref signal_buffered
151 @anchor cds_urcu_performance
154 As a result of our experiments we can range above %RCU implementation in such order,
155 from high to low performance:
156 - <tt>gc<general_buffered></tt> - high
157 - <tt>gc<general_threaded></tt>
158 - <tt>gc<signal_buffered></tt>
159 - <tt>gc<general_instant></tt> - low
161 This estimation is very rough and depends on many factors:
162 type of payload - mostly read-only (seeking) or read-write (inserting and deleting), -
163 a hardware, your application, and so on.
165 @anchor cds_urcu_howto
168 Usually, in your application you use only one \ref cds_urcu_gc "type of RCU" that is the best for your needs.
169 However, the library allows to apply several RCU singleton in one application.
170 The only limitation is that only one object of each RCU type can be instantiated.
171 Since each RCU type is a template class the creation of two object of one RCU type class
172 with different template arguments is an error and is not supported.
173 However, it is correct when your RCU objects relates to different RCU types.
175 In \p libcds, many GC-based ordered list, set and map template classes have %RCU-related specializations
176 that hide the %RCU specific details.
178 RCU GC is initialized in usual way: you should declare an object of type \p cds::urcu::gc< RCU_type >,
181 #include <cds/urcu/general_buffered.h>
183 typedef cds::urcu::gc< cds::urcu::general_buffered<> > rcu_gpb;
189 // Initialize general_buffered RCU
192 // If main thread uses lock-free containers
193 // the main thread should be attached to libcds infrastructure
194 cds::threading::Manager::attachThread();
196 // Now you can use RCU-based containers in the main thread
204 Each thread that deals with RCU-based container should be initialized first:
206 #include <cds/urcu/general_buffered.h>
207 int myThreadEntryPoint(void *)
209 // Attach the thread to libcds infrastructure
210 cds::threading::Manager::attachThread();
212 // Now you can use RCU-based containers in the thread
215 // Detach thread when terminating
216 cds::threading::Manager::detachThread();
222 # if (CDS_OS_INTERFACE == CDS_OSI_UNIX || defined(CDS_DOXYGEN_INVOKED)) && !defined(CDS_THREAD_SANITIZER_ENABLED)
223 # define CDS_URCU_SIGNAL_HANDLING_ENABLED 1
226 /// General-purpose URCU type
227 struct general_purpose_rcu {
229 static uint32_t const c_nControlBit = 0x80000000;
230 static uint32_t const c_nNestMask = c_nControlBit - 1;
234 # ifdef CDS_URCU_SIGNAL_HANDLING_ENABLED
235 /// Signal-handling URCU type
236 struct signal_handling_rcu {
238 static uint32_t const c_nControlBit = 0x80000000;
239 static uint32_t const c_nNestMask = c_nControlBit - 1;
244 /// Tag for general_instant URCU
245 struct general_instant_tag: public general_purpose_rcu {
246 typedef general_purpose_rcu rcu_class ; ///< The URCU type
249 /// Tag for general_buffered URCU
250 struct general_buffered_tag: public general_purpose_rcu
252 typedef general_purpose_rcu rcu_class ; ///< The URCU type
255 /// Tag for general_threaded URCU
256 struct general_threaded_tag: public general_purpose_rcu {
257 typedef general_purpose_rcu rcu_class ; ///< The URCU type
260 # ifdef CDS_URCU_SIGNAL_HANDLING_ENABLED
261 /// Tag for signal_buffered URCU
262 struct signal_buffered_tag: public signal_handling_rcu {
263 typedef signal_handling_rcu rcu_class ; ///< The URCU type
267 ///@anchor cds_urcu_retired_ptr Retired pointer, i.e. pointer that ready for reclamation
268 typedef cds::gc::details::retired_ptr retired_ptr;
269 using cds::gc::make_retired_ptr;
271 /// Pointer to function to free (destruct and deallocate) retired pointer of specific type
272 typedef cds::gc::details::free_retired_ptr_func free_retired_ptr_func;
275 /// Implementation-specific URCU details
277 /// forward declarations
278 template <typename RCUtag>
281 template <typename RCUtag>
284 template <typename RCUtag >
288 class singleton_vtbl {
290 virtual ~singleton_vtbl()
293 virtual void retire_ptr( retired_ptr& p ) = 0;
299 template <typename MarkedPtr> using atomic_marked_ptr = atomics::atomic<MarkedPtr>;
304 template <typename ThreadData>
305 struct thread_list_record {
306 atomics::atomic<ThreadData*> m_pNext; ///< Next item in thread list
307 atomics::atomic<OS::ThreadId> m_idOwner; ///< Owner thread id; 0 - the record is free (not owned)
311 , m_idOwner( cds::OS::c_NullThreadId )
314 explicit thread_list_record( OS::ThreadId owner )
319 ~thread_list_record()
325 template <typename RCUtag, class Alloc = CDS_DEFAULT_ALLOCATOR >
328 typedef thread_data<RCUtag> thread_record;
329 typedef cds::details::Allocator< thread_record, Alloc > allocator_type;
332 atomics::atomic<thread_record *> m_pHead;
344 thread_record * alloc()
346 thread_record * pRec;
347 cds::OS::ThreadId const nullThreadId = cds::OS::c_NullThreadId;
348 cds::OS::ThreadId const curThreadId = cds::OS::get_current_thread_id();
350 // First, try to reuse a retired (non-active) HP record
351 for ( pRec = m_pHead.load( atomics::memory_order_acquire ); pRec; pRec = pRec->m_list.m_pNext.load( atomics::memory_order_relaxed )) {
352 cds::OS::ThreadId thId = nullThreadId;
353 if ( !pRec->m_list.m_idOwner.compare_exchange_strong( thId, curThreadId, atomics::memory_order_seq_cst, atomics::memory_order_relaxed ))
358 // No records available for reuse
359 // Allocate and push a new record
360 pRec = allocator_type().New( curThreadId );
362 thread_record * pOldHead = m_pHead.load( atomics::memory_order_acquire );
364 // Compiler barriers: assignment MUST BE inside the loop
365 CDS_COMPILER_RW_BARRIER;
366 pRec->m_list.m_pNext.store( pOldHead, atomics::memory_order_relaxed );
367 CDS_COMPILER_RW_BARRIER;
368 } while ( !m_pHead.compare_exchange_weak( pOldHead, pRec, atomics::memory_order_acq_rel, atomics::memory_order_acquire ));
373 void retire( thread_record * pRec )
375 assert( pRec != nullptr );
376 pRec->m_list.m_idOwner.store( cds::OS::c_NullThreadId, atomics::memory_order_release );
381 thread_record * pNext = nullptr;
382 cds::OS::ThreadId const nullThreadId = cds::OS::c_NullThreadId;
384 for ( thread_record * pRec = m_pHead.load( atomics::memory_order_acquire ); pRec; pRec = pNext ) {
385 pNext = pRec->m_list.m_pNext.load( atomics::memory_order_relaxed );
386 if ( pRec->m_list.m_idOwner.load( atomics::memory_order_relaxed ) != nullThreadId ) {
392 thread_record * head( atomics::memory_order mo ) const
394 return m_pHead.load( mo );
401 CDS_DEBUG_ONLY( cds::OS::ThreadId const nullThreadId = cds::OS::c_NullThreadId; )
402 CDS_DEBUG_ONLY( cds::OS::ThreadId const mainThreadId = cds::OS::get_current_thread_id() ;)
404 thread_record * p = m_pHead.exchange( nullptr, atomics::memory_order_acquire );
406 thread_record * pNext = p->m_list.m_pNext.load( atomics::memory_order_relaxed );
408 assert( p->m_list.m_idOwner.load( atomics::memory_order_relaxed ) == nullThreadId
409 || p->m_list.m_idOwner.load( atomics::memory_order_relaxed ) == mainThreadId
420 template <class ThreadGC>
423 typedef ThreadGC thread_gc;
424 typedef typename thread_gc::rcu_tag rcu_tag;
429 thread_gc::access_lock();
434 thread_gc::access_unlock();
438 } // namespace details
443 template <typename RCUimpl> class gc;
446 /// Epoch-based retired ptr
448 Retired pointer with additional epoch field that prevents early reclamation.
449 This type of retired pointer is used in buffered RCU implementations.
451 struct epoch_retired_ptr: public retired_ptr
453 uint64_t m_nEpoch; ///< The epoch when the object has been retired
460 /// Constructor creates a copy of \p rp retired pointer and saves \p nEpoch reclamation epoch.
461 epoch_retired_ptr( retired_ptr const& rp, uint64_t nEpoch )
470 #endif // #ifndef CDSLIB_URCU_DETAILS_BASE_H