1 <?xml version="1.0" encoding="ISO-8859-1"?>
3 PUBLIC "-//W3C//DTD XHTML 1.0 Transitional//EN"
4 "http://www.w3.org/TR/xhtml1/DTD/xhtml1-transitional.dtd">
6 <html xmlns="http://www.w3.org/1999/xhtml" xml:lang="en" lang="en">
8 <meta name="AUTHOR" content="Stefan Olsson <stefan@xapa.se>" />
9 <meta name="KEYWORDS" content="c++, libstdc++, g++, allocator, memory" />
10 <meta name="DESCRIPTION" content="Allocators and allocation" />
11 <meta name="GENERATOR" content="emacs and ten fingers" />
12 <title>A fixed-size, multi-thread optimized allocator</title>
13 <link rel="StyleSheet" href="../lib3styles.css" type="text/css" />
14 <link rel="Start" href="../documentation.html" type="text/html"
15 title="GNU C++ Standard Library" />
16 <link rel="Bookmark" href="howto.html" type="text/html" title="Extensions" />
17 <link rel="Copyright" href="../17_intro/license.html" type="text/html" />
21 <h1 class="centered"><a name="top">A fixed-size, multi-thread optimized allocator</a></h1>
23 <p class="fineprint"><em>
24 The latest version of this document is always available at
25 <a href="http://gcc.gnu.org/onlinedocs/libstdc++/ext/mt_allocator.html">
26 http://gcc.gnu.org/onlinedocs/libstdc++/ext/mt_allocator.html</a>.
30 To the <a href="http://gcc.gnu.org/libstdc++/">libstdc++-v3 homepage</a>.
33 <!-- ####################################################### -->
36 <a name="intro">Introduction</a>
39 <p> The mt allocator [hereinafter referred to simply as "the
40 allocator"] is a fixed size (power of two) allocator that was
41 initially developed specifically to suit the needs of multi threaded
42 applications [hereinafter referred to as an MT application]. Over time
43 the allocator has evolved and been improved in many ways, one of the
44 being that it now also does a good job in single threaded applications
45 [hereinafter referred to as a ST application]. (Note: In this
46 document, when referring to single threaded applications this also
47 includes applications that are compiled with gcc without thread
48 support enabled. This is accomplished using ifdef's on __GTHREADS)
52 The aim of this document is to describe - from a application point of
53 view - the "inner workings" of the allocator.
57 <a name="init">Tunable parameters</a>
60 <p>Certain allocation parameters can be modified on a per-type
61 basis. There exists a nested <pre>struct _Tune</pre> that contains all
62 these parameters, which include settings for
66 <li>Maximum bytes before calling <code>::operator new</code> directly</li>
67 <li>Minimum bytes</li>
68 <li>Size of underlying global allocations</li>
69 <li>Maximum number of supported threads</li>
70 <li>Migration of deallocations to the global free list</li>
71 <li>Shunt for global <code>new</code> and <code>delete</code></li>
73 <p>Adjusting parameters for a given instance of an allocator can only
74 happen before any allocations take place, when the allocator itself is
75 initialized. For instance:
78 #include <ext/mt_allocator.h>
88 typedef pod value_type;
89 typedef __gnu_cxx::__mt_alloc<value_type> allocator_type;
90 typedef allocator_type::_Tune tune_type;
93 tune_type t_opt(16, 5120, 32, 5120, 20, 10, false);
94 tune_type t_single(16, 5120, 32, 5120, 1, 10, false);
97 t = allocator_type::_S_get_options();
98 allocator_type::_S_set_options(t_opt);
99 t = allocator_type::_S_get_options();
102 allocator_type::pointer p1 = a.allocate(128);
103 allocator_type::pointer p2 = a.allocate(5128);
105 a.deallocate(p1, 128);
106 a.deallocate(p2, 5128);
113 <a name="init">Initialization</a>
117 The static variables (pointers to freelists, tuning parameters etc)
118 are initialized as above, or are set to the global defaults.
122 The very first allocate() call will always call the _S_init() function.
123 In order to make sure that this function is called exactly once we make use
124 of a __gthread_once (with _S_once_mt and _S_init as arguments) call in MT
125 applications and check a static bool (_S_initialized) in ST applications.
129 The _S_init() function:
130 - If the GLIBCXX_FORCE_NEW environment variable is set, it sets the bool
131 _S_force_new to true and then returns. This will cause subsequent calls to
132 allocate() to return memory directly from a new() call, and deallocate will
133 only do a delete() call.
137 - If the GLIBCXX_FORCE_NEW environment variable is not set, both ST and MT
139 - Calculate the number of bins needed. A bin is a specific power of two size
140 of bytes. I.e., by default the allocator will deal with requests of up to
141 128 bytes (or whatever the value of _S_max_bytes is when _S_init() is
142 called). This means that there will be bins of the following sizes
143 (in bytes): 1, 2, 4, 8, 16, 32, 64, 128.
145 - Create the _S_binmap array. All requests are rounded up to the next
146 "large enough" bin. I.e., a request for 29 bytes will cause a block from
147 the "32 byte bin" to be returned to the application. The purpose of
148 _S_binmap is to speed up the process of finding out which bin to use.
149 I.e., the value of _S_binmap[ 29 ] is initialized to 5 (bin 5 = 32 bytes).
152 - Create the _S_bin array. This array consists of bin_records. There will be
153 as many bin_records in this array as the number of bins that we calculated
154 earlier. I.e., if _S_max_bytes = 128 there will be 8 entries.
155 Each bin_record is then initialized:
156 - bin_record->first = An array of pointers to block_records. There will be
157 as many block_records pointers as there are maximum number of threads
158 (in a ST application there is only 1 thread, in a MT application there
160 This holds the pointer to the first free block for each thread in this
161 bin. I.e., if we would like to know where the first free block of size 32
162 for thread number 3 is we would look this up by: _S_bin[ 5 ].first[ 3 ]
164 The above created block_record pointers members are now initialized to
165 their initial values. I.e. _S_bin[ n ].first[ n ] = NULL;
169 - Additionally a MT application will:
170 - Create a list of free thread id's. The pointer to the first entry
171 is stored in _S_thread_freelist_first. The reason for this approach is
172 that the __gthread_self() call will not return a value that corresponds to
173 the maximum number of threads allowed but rather a process id number or
174 something else. So what we do is that we create a list of thread_records.
175 This list is _S_max_threads long and each entry holds a size_t thread_id
176 which is initialized to 1, 2, 3, 4, 5 and so on up to _S_max_threads.
177 Each time a thread calls allocate() or deallocate() we call
178 _S_get_thread_id() which looks at the value of _S_thread_key which is a
179 thread local storage pointer. If this is NULL we know that this is a newly
180 created thread and we pop the first entry from this list and saves the
181 pointer to this record in the _S_thread_key variable. The next time
182 we will get the pointer to the thread_record back and we use the
183 thread_record->thread_id as identification. I.e., the first thread that
184 calls allocate will get the first record in this list and thus be thread
185 number 1 and will then find the pointer to its first free 32 byte block
186 in _S_bin[ 5 ].first[ 1 ]
187 When we create the _S_thread_key we also define a destructor
188 (_S_thread_key_destr) which means that when the thread dies, this
189 thread_record is returned to the front of this list and the thread id
190 can then be reused if a new thread is created.
191 This list is protected by a mutex (_S_thread_freelist_mutex) which is only
192 locked when records are removed/added to the list.
195 - Initialize the free and used counters of each bin_record:
196 - bin_record->free = An array of size_t. This keeps track of the number
197 of blocks on a specific thread's freelist in each bin. I.e., if a thread
198 has 12 32-byte blocks on it's freelists and allocates one of these, this
199 counter would be decreased to 11.
201 - bin_record->used = An array of size_t. This keeps track of the number
202 of blocks currently in use of this size by this thread. I.e., if a thread
203 has made 678 requests (and no deallocations...) of 32-byte blocks this
204 counter will read 678.
206 The above created arrays are now initialized with their initial values.
207 I.e. _S_bin[ n ].free[ n ] = 0;
210 - Initialize the mutex of each bin_record:
211 The bin_record->mutex is used to protect the global freelist. This concept
212 of a global freelist is explained in more detail in the section
213 "A multi threaded example", but basically this mutex is locked whenever
214 a block of memory is retrieved or returned to the global freelist for this
215 specific bin. This only occurs when a number of blocks are grabbed from the
216 global list to a thread specific list or when a thread decides to return
217 some blocks to the global freelist.
221 <a name="st_example">A single threaded example (and a primer for the multi threaded example!)</a>
225 Let's start by describing how the data on a freelist is laid out in memory.
226 This is the first two blocks in freelist for thread id 3 in bin 3 (8 bytes):
230 | next* ---------|--+ (_S_bin[ 3 ].first[ 3 ] points here)
240 | DATA | | (A pointer to here is what is returned to the
241 | | | the application when needed)
250 | next* |<-+ (If next == NULL it's the last one on the list)
272 With this in mind we simplify things a bit for a while and say that there is
273 only one thread (a ST application). In this case all operations are made to
274 what is referred to as the global pool - thread id 0 (No thread may be
275 assigned this id since they span from 1 to _S_max_threads in a MT application).
278 When the application requests memory (calling allocate()) we first look at the
279 requested size and if this is > _S_max_bytes we call new() directly and return.
282 If the requested size is within limits we start by finding out from which
283 bin we should serve this request by looking in _S_binmap.
286 A quick look at _S_bin[ bin ].first[ 0 ] tells us if there are any blocks of
287 this size on the freelist (0). If this is not NULL - fine, just remove the
288 block that _S_bin[ bin ].first[ 0 ] points to from the list,
289 update _S_bin[ bin ].first[ 0 ] and return a pointer to that blocks data.
292 If the freelist is empty (the pointer is NULL) we must get memory from the
293 system and build us a freelist within this memory. All requests for new memory
294 is made in chunks of _S_chunk_size. Knowing the size of a block_record and
295 the bytes that this bin stores we then calculate how many blocks we can create
296 within this chunk, build the list, remove the first block, update the pointer
297 (_S_bin[ bin ].first[ 0 ]) and return a pointer to that blocks data.
301 Deallocation is equally simple; the pointer is casted back to a block_record
302 pointer, lookup which bin to use based on the size, add the block to the front
303 of the global freelist and update the pointer as needed
304 (_S_bin[ bin ].first[ 0 ]).
308 The decision to add deallocated blocks to the front of the freelist was made
309 after a set of performance measurements that showed that this is roughly 10%
310 faster than maintaining a set of "last pointers" as well.
314 <a name="mt_example">A multi threaded example</a>
318 In the ST example we never used the thread_id variable present in each block.
319 Let's start by explaining the purpose of this in a MT application.
323 The concept of "ownership" was introduced since many MT applications
324 allocate and deallocate memory to shared containers from different
325 threads (such as a cache shared amongst all threads). This introduces
326 a problem if the allocator only returns memory to the current threads
327 freelist (I.e., there might be one thread doing all the allocation and
328 thus obtaining ever more memory from the system and another thread
329 that is getting a longer and longer freelist - this will in the end
330 consume all available memory).
334 Each time a block is moved from the global list (where ownership is
335 irrelevant), to a threads freelist (or when a new freelist is built
336 from a chunk directly onto a threads freelist or when a deallocation
337 occurs on a block which was not allocated by the same thread id as the
338 one doing the deallocation) the thread id is set to the current one.
342 What's the use? Well, when a deallocation occurs we can now look at
343 the thread id and find out if it was allocated by another thread id
344 and decrease the used counter of that thread instead, thus keeping the
345 free and used counters correct. And keeping the free and used counters
346 corrects is very important since the relationship between these two
347 variables decides if memory should be returned to the global pool or
348 not when a deallocation occurs.
352 When the application requests memory (calling allocate()) we first
353 look at the requested size and if this is > _S_max_bytes we call new()
358 If the requested size is within limits we start by finding out from which
359 bin we should serve this request by looking in _S_binmap.
363 A call to _S_get_thread_id() returns the thread id for the calling thread
364 (and if no value has been set in _S_thread_key, a new id is assigned and
369 A quick look at _S_bin[ bin ].first[ thread_id ] tells us if there are
370 any blocks of this size on the current threads freelist. If this is
371 not NULL - fine, just remove the block that _S_bin[ bin ].first[
372 thread_id ] points to from the list, update _S_bin[ bin ].first[
373 thread_id ], update the free and used counters and return a pointer to
378 If the freelist is empty (the pointer is NULL) we start by looking at
379 the global freelist (0). If there are blocks available on the global
380 freelist we lock this bins mutex and move up to block_count (the
381 number of blocks of this bins size that will fit into a _S_chunk_size)
382 or until end of list - whatever comes first - to the current threads
383 freelist and at the same time change the thread_id ownership and
384 update the counters and pointers. When the bins mutex has been
385 unlocked, we remove the block that _S_bin[ bin ].first[ thread_id ]
386 points to from the list, update _S_bin[ bin ].first[ thread_id ],
387 update the free and used counters, and return a pointer to that blocks
392 The reason that the number of blocks moved to the current threads
393 freelist is limited to block_count is to minimize the chance that a
394 subsequent deallocate() call will return the excess blocks to the
395 global freelist (based on the _S_freelist_headroom calculation, see
400 However if there isn't any memory on the global pool we need to get
401 memory from the system - this is done in exactly the same way as in a
402 single threaded application with one major difference; the list built
403 in the newly allocated memory (of _S_chunk_size size) is added to the
404 current threads freelist instead of to the global.
408 The basic process of a deallocation call is simple: always add the
409 block to the front of the current threads freelist and update the
410 counters and pointers (as described earlier with the specific check of
411 ownership that causes the used counter of the thread that originally
412 allocated the block to be decreased instead of the current threads
417 And here comes the free and used counters to service. Each time a
418 deallocation() call is made, the length of the current threads
419 freelist is compared to the amount memory in use by this thread.
423 Let's go back to the example of an application that has one thread
424 that does all the allocations and one that deallocates. Both these
425 threads use say 516 32-byte blocks that was allocated during thread
426 creation for example. Their used counters will both say 516 at this
427 point. The allocation thread now grabs 1000 32-byte blocks and puts
428 them in a shared container. The used counter for this thread is now
433 The deallocation thread now deallocates 500 of these blocks. For each
434 deallocation made the used counter of the allocating thread is
435 decreased and the freelist of the deallocation thread gets longer and
436 longer. But the calculation made in deallocate() will limit the length
437 of the freelist in the deallocation thread to _S_freelist_headroom %
438 of it's used counter. In this case, when the freelist (given that the
439 _S_freelist_headroom is at it's default value of 10%) exceeds 52
440 (516/10) blocks will be returned to the global pool where the
441 allocating thread may pick them up and reuse them.
445 In order to reduce lock contention (since this requires this bins
446 mutex to be locked) this operation is also made in chunks of blocks
447 (just like when chunks of blocks are moved from the global freelist to
448 a threads freelist mentioned above). The "formula" used can probably
449 be improved to further reduce the risk of blocks being "bounced back
450 and forth" between freelists.
454 <p>Return <a href="#top">to the top of the page</a> or
455 <a href="http://gcc.gnu.org/libstdc++/">to the libstdc++ homepage</a>.
459 <!-- ####################################################### -->
462 <p class="fineprint"><em>
463 See <a href="../17_intro/license.html">license.html</a> for copying conditions.
464 Comments and suggestions are welcome, and may be sent to
465 <a href="mailto:libstdc++@gcc.gnu.org">the libstdc++ mailing list</a>.
OSDN Git Service
