2008-01-10 Vladimir Makarov <vmakarov@redhat.com>
[official-gcc.git] / libstdc++-v3 / docs / html / ext / mt_allocator.html
blobe91760e8b4863af042742f61607b226794268772
1 <?xml version="1.0" encoding="ISO-8859-1"?>
2 <!DOCTYPE html
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">
7 <head>
8 <meta name="AUTHOR" content="Stefan Olsson &lt;stefan@xapa.se&gt;" />
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" />
18 </head>
19 <body>
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>.
27 </em></p>
29 <p><em>
30 To the <a href="http://gcc.gnu.org/libstdc++/">libstdc++-v3 homepage</a>.
31 </em></p>
33 <!-- ####################################################### -->
34 <hr />
35 <h3 class="left">
36 <a name="intro">Introduction</a>
37 </h3>
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, in
44 particular 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
49 __GTHREADS). This allocator is tunable, very flexible, and capable of
50 high-performance.
51 </p>
53 <p>
54 The aim of this document is to describe - from a application point of
55 view - the "inner workings" of the allocator.
56 </p>
58 <h3 class="left">
59 <a name="design">Design Overview</a>
60 </h3>
62 <p> There are three general components to the allocator: a datum
63 describing the characteristics of the memory pool, a policy class
64 containing this pool that links instantiation types to common or
65 individual pools, and a class inheriting from the policy class that is
66 the actual allocator.
67 </p>
69 <p>The datum describing pools characteristics is
70 <pre>
71 template&lt;bool _Thread&gt;
72 class __pool
73 </pre>
74 This class is parametrized on thread support, and is explicitly
75 specialized for both multiple threads (with <code>bool==true</code>)
76 and single threads (via <code>bool==false</code>.) It is possible to
77 use a custom pool datum instead of the default class that is provided.
78 </p>
80 <p> There are two distinct policy classes, each of which can be used
81 with either type of underlying pool datum.
82 </p>
84 <pre>
85 template&lt;bool _Thread&gt;
86 struct __common_pool_policy
88 template&lt;typename _Tp, bool _Thread&gt;
89 struct __per_type_pool_policy
90 </pre>
92 <p> The first policy, <code>__common_pool_policy</code>, implements a
93 common pool. This means that allocators that are instantiated with
94 different types, say <code>char</code> and <code>long</code> will both
95 use the same pool. This is the default policy.
96 </p>
98 <p> The second policy, <code>__per_type_pool_policy</code>, implements
99 a separate pool for each instantiating type. Thus, <code>char</code>
100 and <code>long</code> will use separate pools. This allows per-type
101 tuning, for instance.
102 </p>
104 <p> Putting this all together, the actual allocator class is
105 <pre>
106 template&lt;typename _Tp, typename _Poolp = __default_policy&gt;
107 class __mt_alloc : public __mt_alloc_base&lt;_Tp&gt;, _Poolp
108 </pre>
109 This class has the interface required for standard library allocator
110 classes, namely member functions <code>allocate</code> and
111 <code>deallocate</code>, plus others.
112 </p>
114 <h3 class="left">
115 <a name="init">Tunable parameters</a>
116 </h3>
118 <p>Certain allocation parameters can be modified, or tuned. There
119 exists a nested <pre>struct __pool_base::_Tune</pre> that contains all
120 these parameters, which include settings for
121 </p>
122 <ul>
123 <li>Alignment </li>
124 <li>Maximum bytes before calling <code>::operator new</code> directly</li>
125 <li>Minimum bytes</li>
126 <li>Size of underlying global allocations</li>
127 <li>Maximum number of supported threads</li>
128 <li>Migration of deallocations to the global free list</li>
129 <li>Shunt for global <code>new</code> and <code>delete</code></li>
130 </ul>
131 <p>Adjusting parameters for a given instance of an allocator can only
132 happen before any allocations take place, when the allocator itself is
133 initialized. For instance:
134 </p>
135 <pre>
136 #include &lt;ext/mt_allocator.h&gt;
138 struct pod
140 int i;
141 int j;
144 int main()
146 typedef pod value_type;
147 typedef __gnu_cxx::__mt_alloc&lt;value_type&gt; allocator_type;
148 typedef __gnu_cxx::__pool_base::_Tune tune_type;
150 tune_type t_default;
151 tune_type t_opt(16, 5120, 32, 5120, 20, 10, false);
152 tune_type t_single(16, 5120, 32, 5120, 1, 10, false);
154 tune_type t;
155 t = allocator_type::_M_get_options();
156 allocator_type::_M_set_options(t_opt);
157 t = allocator_type::_M_get_options();
159 allocator_type a;
160 allocator_type::pointer p1 = a.allocate(128);
161 allocator_type::pointer p2 = a.allocate(5128);
163 a.deallocate(p1, 128);
164 a.deallocate(p2, 5128);
166 return 0;
168 </pre>
170 <h3 class="left">
171 <a name="init">Initialization</a>
172 </h3>
175 The static variables (pointers to freelists, tuning parameters etc)
176 are initialized as above, or are set to the global defaults.
177 </p>
180 The very first allocate() call will always call the
181 _S_initialize_once() function. In order to make sure that this
182 function is called exactly once we make use of a __gthread_once call
183 in MT applications and check a static bool (_S_init) in ST
184 applications.
185 </p>
188 The _S_initialize() function:
189 - If the GLIBCXX_FORCE_NEW environment variable is set, it sets the bool
190 _S_force_new to true and then returns. This will cause subsequent calls to
191 allocate() to return memory directly from a new() call, and deallocate will
192 only do a delete() call.
193 </p>
196 - If the GLIBCXX_FORCE_NEW environment variable is not set, both ST and MT
197 applications will:
198 - Calculate the number of bins needed. A bin is a specific power of two size
199 of bytes. I.e., by default the allocator will deal with requests of up to
200 128 bytes (or whatever the value of _S_max_bytes is when _S_init() is
201 called). This means that there will be bins of the following sizes
202 (in bytes): 1, 2, 4, 8, 16, 32, 64, 128.
204 - Create the _S_binmap array. All requests are rounded up to the next
205 "large enough" bin. I.e., a request for 29 bytes will cause a block from
206 the "32 byte bin" to be returned to the application. The purpose of
207 _S_binmap is to speed up the process of finding out which bin to use.
208 I.e., the value of _S_binmap[ 29 ] is initialized to 5 (bin 5 = 32 bytes).
209 </p>
211 - Create the _S_bin array. This array consists of bin_records. There will be
212 as many bin_records in this array as the number of bins that we calculated
213 earlier. I.e., if _S_max_bytes = 128 there will be 8 entries.
214 Each bin_record is then initialized:
215 - bin_record-&gt;first = An array of pointers to block_records. There will be
216 as many block_records pointers as there are maximum number of threads
217 (in a ST application there is only 1 thread, in a MT application there
218 are _S_max_threads).
219 This holds the pointer to the first free block for each thread in this
220 bin. I.e., if we would like to know where the first free block of size 32
221 for thread number 3 is we would look this up by: _S_bin[ 5 ].first[ 3 ]
223 The above created block_record pointers members are now initialized to
224 their initial values. I.e. _S_bin[ n ].first[ n ] = NULL;
225 </p>
228 - Additionally a MT application will:
229 - Create a list of free thread id's. The pointer to the first entry
230 is stored in _S_thread_freelist_first. The reason for this approach is
231 that the __gthread_self() call will not return a value that corresponds to
232 the maximum number of threads allowed but rather a process id number or
233 something else. So what we do is that we create a list of thread_records.
234 This list is _S_max_threads long and each entry holds a size_t thread_id
235 which is initialized to 1, 2, 3, 4, 5 and so on up to _S_max_threads.
236 Each time a thread calls allocate() or deallocate() we call
237 _S_get_thread_id() which looks at the value of _S_thread_key which is a
238 thread local storage pointer. If this is NULL we know that this is a newly
239 created thread and we pop the first entry from this list and saves the
240 pointer to this record in the _S_thread_key variable. The next time
241 we will get the pointer to the thread_record back and we use the
242 thread_record-&gt;thread_id as identification. I.e., the first thread that
243 calls allocate will get the first record in this list and thus be thread
244 number 1 and will then find the pointer to its first free 32 byte block
245 in _S_bin[ 5 ].first[ 1 ]
246 When we create the _S_thread_key we also define a destructor
247 (_S_thread_key_destr) which means that when the thread dies, this
248 thread_record is returned to the front of this list and the thread id
249 can then be reused if a new thread is created.
250 This list is protected by a mutex (_S_thread_freelist_mutex) which is only
251 locked when records are removed/added to the list.
252 </p>
254 - Initialize the free and used counters of each bin_record:
255 - bin_record-&gt;free = An array of size_t. This keeps track of the number
256 of blocks on a specific thread's freelist in each bin. I.e., if a thread
257 has 12 32-byte blocks on it's freelists and allocates one of these, this
258 counter would be decreased to 11.
260 - bin_record-&gt;used = An array of size_t. This keeps track of the number
261 of blocks currently in use of this size by this thread. I.e., if a thread
262 has made 678 requests (and no deallocations...) of 32-byte blocks this
263 counter will read 678.
265 The above created arrays are now initialized with their initial values.
266 I.e. _S_bin[ n ].free[ n ] = 0;
267 </p>
269 - Initialize the mutex of each bin_record: The bin_record-&gt;mutex
270 is used to protect the global freelist. This concept of a global
271 freelist is explained in more detail in the section "A multi
272 threaded example", but basically this mutex is locked whenever a
273 block of memory is retrieved or returned to the global freelist
274 for this specific bin. This only occurs when a number of blocks
275 are grabbed from the global list to a thread specific list or when
276 a thread decides to return some blocks to the global freelist.
277 </p>
279 <p> Notes about deallocation. This allocator does not explicitly
280 release memory. Because of this, memory debugging programs like
281 valgrind or purify may notice leaks: sorry about this
282 inconvenience. Operating systems will reclaim allocated memory at
283 program termination anyway. If sidestepping this kind of noise is
284 desired, there are three options: use an allocator, like
285 <code>new_allocator</code> that releases memory while debugging, use
286 GLIBCXX_FORCE_NEW to bypass the allocator's internal pools, or use a
287 custom pool datum that releases resources on destruction.</p>
289 <p>On systems with the function <code>__cxa_atexit</code>, the
290 allocator can be forced to free all memory allocated before program
291 termination with the member function
292 <code>__pool_type::_M_destroy</code>. However, because this member
293 function relies on the precise and exactly-conforming ordering of
294 static destructors, including those of a static local
295 <code>__pool</code> object, it should not be used, ever, on systems
296 that don't have the necessary underlying support. In addition, in
297 practice, forcing deallocation can be tricky, as it requires the
298 <code>__pool</code> object to be fully-constructed before the object
299 that uses it is fully constructed. For most (but not all) STL
300 containers, this works, as an instance of the allocator is constructed
301 as part of a container's constructor. However, this assumption is
302 implementation-specific, and subject to change. For an example of a
303 pool that frees memory, see the following
304 <a href="http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/libstdc%2b%2b-v3/testsuite/ext/mt_allocator/deallocate_local-6.cc">
305 example.</a>
306 </p>
308 <h3 class="left">
309 <a name="st_example">A single threaded example (and a primer for the multi threaded example!)</a>
310 </h3>
313 Let's start by describing how the data on a freelist is laid out in memory.
314 This is the first two blocks in freelist for thread id 3 in bin 3 (8 bytes):
315 </p>
316 <pre>
317 +----------------+
318 | next* ---------|--+ (_S_bin[ 3 ].first[ 3 ] points here)
319 | | |
320 | | |
321 | | |
322 +----------------+ |
323 | thread_id = 3 | |
324 | | |
325 | | |
326 | | |
327 +----------------+ |
328 | DATA | | (A pointer to here is what is returned to the
329 | | | the application when needed)
330 | | |
331 | | |
332 | | |
333 | | |
334 | | |
335 | | |
336 +----------------+ |
337 +----------------+ |
338 | next* |&lt;-+ (If next == NULL it's the last one on the list)
342 +----------------+
343 | thread_id = 3 |
347 +----------------+
348 | DATA |
356 +----------------+
357 </pre>
360 With this in mind we simplify things a bit for a while and say that there is
361 only one thread (a ST application). In this case all operations are made to
362 what is referred to as the global pool - thread id 0 (No thread may be
363 assigned this id since they span from 1 to _S_max_threads in a MT application).
364 </p>
366 When the application requests memory (calling allocate()) we first look at the
367 requested size and if this is &gt; _S_max_bytes we call new() directly and return.
368 </p>
370 If the requested size is within limits we start by finding out from which
371 bin we should serve this request by looking in _S_binmap.
372 </p>
374 A quick look at _S_bin[ bin ].first[ 0 ] tells us if there are any blocks of
375 this size on the freelist (0). If this is not NULL - fine, just remove the
376 block that _S_bin[ bin ].first[ 0 ] points to from the list,
377 update _S_bin[ bin ].first[ 0 ] and return a pointer to that blocks data.
378 </p>
380 If the freelist is empty (the pointer is NULL) we must get memory from the
381 system and build us a freelist within this memory. All requests for new memory
382 is made in chunks of _S_chunk_size. Knowing the size of a block_record and
383 the bytes that this bin stores we then calculate how many blocks we can create
384 within this chunk, build the list, remove the first block, update the pointer
385 (_S_bin[ bin ].first[ 0 ]) and return a pointer to that blocks data.
386 </p>
389 Deallocation is equally simple; the pointer is casted back to a block_record
390 pointer, lookup which bin to use based on the size, add the block to the front
391 of the global freelist and update the pointer as needed
392 (_S_bin[ bin ].first[ 0 ]).
393 </p>
396 The decision to add deallocated blocks to the front of the freelist was made
397 after a set of performance measurements that showed that this is roughly 10%
398 faster than maintaining a set of "last pointers" as well.
399 </p>
401 <h3 class="left">
402 <a name="mt_example">A multi threaded example</a>
403 </h3>
406 In the ST example we never used the thread_id variable present in each block.
407 Let's start by explaining the purpose of this in a MT application.
408 </p>
411 The concept of "ownership" was introduced since many MT applications
412 allocate and deallocate memory to shared containers from different
413 threads (such as a cache shared amongst all threads). This introduces
414 a problem if the allocator only returns memory to the current threads
415 freelist (I.e., there might be one thread doing all the allocation and
416 thus obtaining ever more memory from the system and another thread
417 that is getting a longer and longer freelist - this will in the end
418 consume all available memory).
419 </p>
422 Each time a block is moved from the global list (where ownership is
423 irrelevant), to a threads freelist (or when a new freelist is built
424 from a chunk directly onto a threads freelist or when a deallocation
425 occurs on a block which was not allocated by the same thread id as the
426 one doing the deallocation) the thread id is set to the current one.
427 </p>
430 What's the use? Well, when a deallocation occurs we can now look at
431 the thread id and find out if it was allocated by another thread id
432 and decrease the used counter of that thread instead, thus keeping the
433 free and used counters correct. And keeping the free and used counters
434 corrects is very important since the relationship between these two
435 variables decides if memory should be returned to the global pool or
436 not when a deallocation occurs.
437 </p>
440 When the application requests memory (calling allocate()) we first
441 look at the requested size and if this is &gt; _S_max_bytes we call new()
442 directly and return.
443 </p>
446 If the requested size is within limits we start by finding out from which
447 bin we should serve this request by looking in _S_binmap.
448 </p>
451 A call to _S_get_thread_id() returns the thread id for the calling thread
452 (and if no value has been set in _S_thread_key, a new id is assigned and
453 returned).
454 </p>
457 A quick look at _S_bin[ bin ].first[ thread_id ] tells us if there are
458 any blocks of this size on the current threads freelist. If this is
459 not NULL - fine, just remove the block that _S_bin[ bin ].first[
460 thread_id ] points to from the list, update _S_bin[ bin ].first[
461 thread_id ], update the free and used counters and return a pointer to
462 that blocks data.
463 </p>
466 If the freelist is empty (the pointer is NULL) we start by looking at
467 the global freelist (0). If there are blocks available on the global
468 freelist we lock this bins mutex and move up to block_count (the
469 number of blocks of this bins size that will fit into a _S_chunk_size)
470 or until end of list - whatever comes first - to the current threads
471 freelist and at the same time change the thread_id ownership and
472 update the counters and pointers. When the bins mutex has been
473 unlocked, we remove the block that _S_bin[ bin ].first[ thread_id ]
474 points to from the list, update _S_bin[ bin ].first[ thread_id ],
475 update the free and used counters, and return a pointer to that blocks
476 data.
477 </p>
480 The reason that the number of blocks moved to the current threads
481 freelist is limited to block_count is to minimize the chance that a
482 subsequent deallocate() call will return the excess blocks to the
483 global freelist (based on the _S_freelist_headroom calculation, see
484 below).
485 </p>
488 However if there isn't any memory on the global pool we need to get
489 memory from the system - this is done in exactly the same way as in a
490 single threaded application with one major difference; the list built
491 in the newly allocated memory (of _S_chunk_size size) is added to the
492 current threads freelist instead of to the global.
493 </p>
496 The basic process of a deallocation call is simple: always add the
497 block to the front of the current threads freelist and update the
498 counters and pointers (as described earlier with the specific check of
499 ownership that causes the used counter of the thread that originally
500 allocated the block to be decreased instead of the current threads
501 counter).
502 </p>
505 And here comes the free and used counters to service. Each time a
506 deallocation() call is made, the length of the current threads
507 freelist is compared to the amount memory in use by this thread.
508 </p>
511 Let's go back to the example of an application that has one thread
512 that does all the allocations and one that deallocates. Both these
513 threads use say 516 32-byte blocks that was allocated during thread
514 creation for example. Their used counters will both say 516 at this
515 point. The allocation thread now grabs 1000 32-byte blocks and puts
516 them in a shared container. The used counter for this thread is now
517 1516.
518 </p>
521 The deallocation thread now deallocates 500 of these blocks. For each
522 deallocation made the used counter of the allocating thread is
523 decreased and the freelist of the deallocation thread gets longer and
524 longer. But the calculation made in deallocate() will limit the length
525 of the freelist in the deallocation thread to _S_freelist_headroom %
526 of it's used counter. In this case, when the freelist (given that the
527 _S_freelist_headroom is at it's default value of 10%) exceeds 52
528 (516/10) blocks will be returned to the global pool where the
529 allocating thread may pick them up and reuse them.
530 </p>
533 In order to reduce lock contention (since this requires this bins
534 mutex to be locked) this operation is also made in chunks of blocks
535 (just like when chunks of blocks are moved from the global freelist to
536 a threads freelist mentioned above). The "formula" used can probably
537 be improved to further reduce the risk of blocks being "bounced back
538 and forth" between freelists.
539 </p>
541 <hr />
542 <p>Return <a href="#top">to the top of the page</a> or
543 <a href="http://gcc.gnu.org/libstdc++/">to the libstdc++ homepage</a>.
544 </p>
547 <!-- ####################################################### -->
549 <hr />
550 <p class="fineprint"><em>
551 See <a href="../17_intro/license.html">license.html</a> for copying conditions.
552 Comments and suggestions are welcome, and may be sent to
553 <a href="mailto:libstdc++@gcc.gnu.org">the libstdc++ mailing list</a>.
554 </em></p>
557 </body>
558 </html>