| blob | cab6bde4fbab9682471022bca6b098a69af86486 |
1 <HTML>
2 <HEAD>
5 </HEAD>
6 <BODY>
9 <P>
10 This is a description of the algorithms and data structures used in our
11 conservative garbage collector. I expect the level of detail to increase
12 with time. For a survey of GC algorithms, see for example
14 excellent paper</a>. For an overview of the collector interface,
16 <P>
17 This description is targeted primarily at someone trying to understand the
18 source code. It specifically refers to variable and function names.
19 It may also be useful for understanding the algorithms at a higher level.
20 <P>
21 The description here assumes that the collector is used in default mode.
22 In particular, we assume that it used as a garbage collector, and not just
23 a leak detector. We initially assume that it is used in stop-the-world,
24 non-incremental mode, though the presence of the incremental collector
25 will be apparent in the design.
26 We assume the default finalization model, but the code affected by that
27 is very localized.
29 The garbage collector uses a modified mark-sweep algorithm. Conceptually
30 it operates roughly in four phases, which are performed occasionally
31 as part of a memory allocation:
33 <OL>
35 <LI>
37 Clear all mark bits, indicating that all objects
38 are potentially unreachable.
40 <LI>
42 pointers from variables. Often the collector has no real information
43 about the location of pointer variables in the heap, so it
44 views all static data areas, stacks and registers as potentially containing
45 pointers. Any bit patterns that represent addresses inside
46 heap objects managed by the collector are viewed as pointers.
47 Unless the client program has made heap object layout information
48 available to the collector, any heap objects found to be reachable from
49 variables are again scanned similarly.
51 <LI>
53 objects, and returns them to an appropriate free list for reuse. This is
54 not really a separate phase; even in non incremental mode this is operation
55 is usually performed on demand during an allocation that discovers an empty
56 free list. Thus the sweep phase is very unlikely to touch a page that
57 would not have been touched shortly thereafter anyway.
59 <LI>
61 for finalization are enqueued for finalization outside the collector.
63 </ol>
65 <P>
66 The remaining sections describe the memory allocation data structures,
67 and then the last 3 collection phases in more detail. We conclude by
68 outlining some of the additional features implemented in the collector.
71 The collector includes its own memory allocator. The allocator obtains
72 memory from the system in a platform-dependent way. Under UNIX, it
74 <P>
75 Most static data used by the allocator, as well as that needed by the
76 rest of the garbage collector is stored inside the
78 This allows the garbage collector to easily ignore the collectors own
79 data structures when it searches for root pointers. Other allocator
80 and collector internal data structures are allocated dynamically
82 allow for deallocation, and is therefore used only for permanent data
83 structures.
84 <P>
86 Different kinds are handled somewhat differently by certain parts
87 of the garbage collector. Certain kinds are scanned for pointers,
88 others are not. Some may have per-object type descriptors that
89 determine pointer locations. Or a specific kind may correspond
90 to one specific object layout. Two built-in kinds are uncollectable.
92 In spite of that, it is very likely that most C clients of the
93 collector currently
96 heavy use of a kind (allocated with GC_gcj_malloc) that stores
97 type information at a known offset in method tables.
98 <P>
99 The collector uses a two level allocator. A large block is defined to
101 typically on the order of the page size.
102 <P>
103 Large block sizes are rounded up to
106 use an approximate best fit algorithm by keeping free lists for
107 several large block sizes.
108 The actual
110 is significantly complicated by black-listing issues
111 (see below).
112 <P>
114 Each chunk is
115 dedicated to only one object size and kind. The allocator maintains
116 separate free lists for each size and kind of object.
117 <P>
118 Once a large block is split for use in smaller objects, it can only
119 be used for objects of that size, unless the collector discovers a completely
120 empty chunk. Completely empty chunks are restored to the appropriate
121 large block free list.
122 <P>
123 In order to avoid allocating blocks for too many distinct object sizes,
124 the collector normally does not directly allocate objects of every possible
125 request size. Instead request are rounded up to one of a smaller number
126 of allocated sizes, for which free lists are maintained. The exact
127 allocated sizes are computed on demand, but subject to the constraint
128 that they increase roughly in geometric progression. Thus objects
129 requested early in the execution are likely to be allocated with exactly
130 the requested size, subject to alignment constraints.
132 <P>
133 The actual size rounding operation during small object allocation is
135 <P>
136 Both collector initialization and computation of allocated sizes are
137 handled carefully so that they do not slow down the small object fast
138 allocation path. An attempt to allocate before the collector is initialized,
140 will take the same path as an allocation attempt with an empty free list.
142 which performs the appropriate initialization checks.
143 <P>
144 In non-incremental mode, we make a decision about whether to garbage collect
145 whenever an allocation would otherwise have failed with the current heap size.
146 If the total amount of allocation since the last collection is less than
148 expand the heap. Otherwise, we initiate a garbage collection. This ensures
149 that the amount of garbage collection work per allocated byte remains
150 constant.
151 <P>
152 The above is in fact an oversimplification of the real heap expansion
153 and GC triggering heuristic, which adjusts slightly for root size
154 and certain kinds of
155 fragmentation. In particular:
156 <UL>
157 <LI> Programs with a large root set size and
158 little live heap memory will expand the heap to amortize the cost of
159 scanning the roots.
161 nonincremental mode. The large block allocator usually refuses to split
162 large heap blocks once the garbage collection threshold is
163 reached. This often has the effect of collecting well before the
164 heap fills up, thus reducing fragmentation and working set size at the
165 expense of GC time. Versions 6.x choose an intermediate strategy depending
166 on how much large object allocation has taken place in the past.
167 (If the collector is configured to unmap unused pages, versions 6.x
168 use the 5.x strategy.)
169 <LI> In calculating the amount of allocation since the last collection we
170 give partial credit for objects we expect to be explicitly deallocated.
171 Even if all objects are explicitly managed, it is often desirable to collect
172 on rare occasion, since that is our only mechanism for coalescing completely
173 empty chunks.
174 </ul>
175 <P>
176 It has been suggested that this should be adjusted so that we favor
177 expansion if the resulting heap still fits into physical memory.
178 In many cases, that would no doubt help. But it is tricky to do this
179 in a way that remains robust if multiple application are contending
180 for a single pool of physical memory.
184 At each collection, the collector marks all objects that are
185 possibly reachable from pointer variables. Since it cannot generally
186 tell where pointer variables are located, it scans the following
188 <UL>
189 <LI>The registers. Depending on the architecture, this may be done using
191 register contents on the stack.
192 <LI>The stack(s). In the case of a single-threaded application,
193 on most platforms this
194 is done by scanning the memory between (an approximation of) the current
197 a highly platform-specific way depending on the appropriate configuration
199 stack needs to be scanned carefully, since callee-save registers of
200 client code may appear inside collector stack frames, which may
201 change during the mark process. This is addressed by scanning
202 some sections of the stack "eagerly", effectively capturing a snapshot
203 at one point in time.
204 <LI>Static data region(s). In the simplest case, this is the region
207 static data regions associated with dynamic libraries. These are
209 </ul>
210 The marker maintains an explicit stack of memory regions that are known
211 to be accessible, but that have not yet been searched for contained pointers.
212 Each stack entry contains the starting address of the block to be scanned,
213 as well as a descriptor of the block. If no layout information is
214 available for the block, then the descriptor is simply a length.
216 <P>
217 At the beginning of the mark phase, all root segments
218 (as described above) are pushed on the
220 stack sections are processed by pushing the referenced objects instead
222 defined, then stack roots require special treatment. In this case, the
224 explicitly checks for interior pointers and pushes descriptors for target
225 objects.
226 <P>
227 The marker is structured to allow incremental marking.
229 work towards marking the heap.
230 It maintains
232 identifies a particular sub-phase. Some other pieces of state, most
233 notably the mark stack, identify how much work remains to be done
234 in each sub-phase. The normal progression of mark states for
235 a stop-the-world collection is:
236 <OL>
239 be false, so the mark state is advanced to
241 uncollectable objects, roots, and then mark everything reachable from them.
243 objects are pushed, and objects reachable from them are marked.
246 mark state to
248 empty, all reachable objects are marked. Once in this state, we work
249 only on emptying the mark stack. Once this is completed, the state
250 changes to
252 </ol>
255 repeatedly by several of the sub-phases when the mark stack starts to fill
257 to empty the mark stack.
258 The routine is designed to only perform a limited amount of marking at
259 each call, so that it can also be used by the incremental collector.
260 It is fairly carefully tuned, since it usually consumes a large majority
261 of the garbage collection time.
262 <P>
263 The fact that it perform a only a small amount of work per call also
264 allows it to be used as the core routine of the parallel marker. In that
265 case it is normally invoked on thread-private mark stacks instead of the
266 global mark stack. More details can be found in
268 <P>
269 The marker correctly handles mark stack overflows. Whenever the mark stack
271 Since there are already marked objects in the heap,
272 this eventually forces a complete
273 scan of the heap, searching for pointers, during which any unmarked objects
274 referenced by marked objects are again pushed on the mark stack. This
275 process is repeated until the mark phase completes without a stack overflow.
276 Each time the stack overflows, an attempt is made to grow the mark stack.
277 All pieces of the collector that push regions onto the mark stack have to be
278 careful to ensure forward progress, even in case of repeated mark stack
279 overflows. Every mark attempt results in additional marked objects.
280 <P>
281 Each mark stack entry is processed by examining all candidate pointers
282 in the range described by the entry. If the region has no associated
283 type information, then this typically requires that each 4-byte aligned
285 pointer.
286 <P>
287 We determine whether a candidate pointer is actually the address of
288 a heap block. This is done in the following steps:
289 <NL>
290 <LI> The candidate pointer is checked against rough heap bounds.
291 These heap bounds are maintained such that all actual heap objects
292 fall between them. In order to facilitate black-listing (see below)
293 we also include address regions that the heap is likely to expand into.
294 Most non-pointers fail this initial test.
295 <LI> The candidate pointer is divided into two pieces; the most significant
297 the least significant bits specify an offset within that page.
298 (A hardware page may actually consist of multiple such pages.
299 HBLKSIZE is usually the page size divided by a small power of two.)
300 <LI>
301 The page address part of the candidate pointer is looked up in a
303 Each table entry contains either 0, indicating that the page is not part
306 back, or a pointer to a descriptor for the page. In the first case,
307 the candidate pointer i not a true pointer and can be safely ignored.
308 In the last two cases, we can obtain a descriptor for the page containing
309 the beginning of the object.
310 <LI>
311 The starting address of the referenced object is computed.
312 The page descriptor contains the size of the object(s)
313 in that page, the object kind, and the necessary mark bits for those
314 objects. The size information can be used to map the candidate pointer
315 to the object starting address. To accelerate this process, the page header
316 also contains a pointer to a precomputed map of page offsets to displacements
317 from the beginning of an object. The use of this map avoids a
318 potentially slow integer remainder operation in computing the object
319 start address.
320 <LI>
321 The mark bit for the target object is checked and set. If the object
322 was previously unmarked, the object is pushed on the mark stack.
323 The descriptor is read from the page descriptor. (This is computed
325 </nl>
326 <P>
327 At the end of the mark phase, mark bits for left-over free lists are cleared,
328 in case a free list was accidentally marked due to a stray pointer.
332 At the end of the mark phase, all blocks in the heap are examined.
333 Unmarked large objects are immediately returned to the large object free list.
334 Each small object page is checked to see if all mark bits are clear.
335 If so, the entire page is returned to the large object free list.
336 Small object pages containing some reachable object are queued for later
337 sweeping, unless we determine that the page contains very little free
338 space, in which case it is not examined further.
339 <P>
340 This initial sweep pass touches only block headers, not
341 the blocks themselves. Thus it does not require significant paging, even
342 if large sections of the heap are not in physical memory.
343 <P>
344 Nonempty small object pages are swept when an allocation attempt
345 encounters an empty free list for that object size and kind.
346 Pages for the correct size and kind are repeatedly swept until at
347 least one empty block is found. Sweeping such a page involves
348 scanning the mark bit array in the page header, and building a free
349 list linked through the first words in the objects themselves.
350 This does involve touching the appropriate data page, but in most cases
351 it will be touched only just before it is used for allocation.
352 Hence any paging is essentially unavoidable.
353 <P>
354 Except in the case of pointer-free objects, we maintain the invariant
355 that any object in a small object free list is cleared (except possibly
356 for the link field). Thus it becomes the burden of the small object
357 sweep routine to clear objects. This has the advantage that we can
358 easily recover from accidentally marking a free list, though that could
359 also be handled by other means. The collector currently spends a fair
360 amount of time clearing objects, and this approach should probably be
361 revisited.
362 <P>
363 In most configurations, we use specialized sweep routines to handle common
364 small object sizes. Since we allocate one mark bit per word, it becomes
365 easier to examine the relevant mark bits if the object size divides
366 the word length evenly. We also suitably unroll the inner sweep loop
367 in each case. (It is conceivable that profile-based procedure cloning
368 in the compiler could make this unnecessary and counterproductive. I
369 know of no existing compiler to which this applies.)
370 <P>
371 The sweeping of small object pages could be avoided completely at the expense
372 of examining mark bits directly in the allocator. This would probably
373 be more expensive, since each allocation call would have to reload
374 a large amount of state (e.g. next object address to be swept, position
375 in mark bit table) before it could do its work. The current scheme
376 keeps the allocator simple and allows useful optimizations in the sweeper.
381 table. The hash table is allocated out of collected memory, but
382 the reference to the finalizable object is hidden from the collector.
383 Currently finalization requests are processed non-incrementally at the
384 end of a mark cycle.
385 <P>
386 The collector makes an initial pass over the table of finalizable objects,
387 pushing the contents of unmarked objects onto the mark stack.
388 After pushing each object, the marker is invoked to mark all objects
389 reachable from it. The object itself is not explicitly marked.
390 This assures that objects on which a finalizer depends are neither
391 collected nor finalized.
392 <P>
393 If in the process of marking from an object the
394 object itself becomes marked, we have uncovered
395 a cycle involving the object. This usually results in a warning from the
396 collector. Such objects are not finalized, since it may be
397 unsafe to do so. See the more detailed
398 <A HREF="http://www.hpl.hp.com/personal/Hans_Boehm/gc/finalization.html"> discussion of finalization semantics</a>.
399 <P>
400 Any objects remaining unmarked at the end of this process are added to
401 a queue of objects whose finalizers can be run. Depending on collector
402 configuration, finalizers are dequeued and run either implicitly during
403 allocation calls, or explicitly in response to a user request.
404 (Note that the former is unfortunately both the default and not generally safe.
405 If finalizers perform synchronization, it may result in deadlocks.
406 Nontrivial finalizers generally need to perform synchronization, and
407 thus require a different collector configuration.)
408 <P>
409 The collector provides a mechanism for replacing the procedure that is
410 used to mark through objects. This is used both to provide support for
411 Java-style unordered finalization, and to ignore certain kinds of cycles,
415 We basically use the concurrent and generational GC algorithm described in
416 <A HREF="http://www.hpl.hp.com/personal/Hans_Boehm/gc/papers/pldi91.ps.Z">"Mostly Parallel Garbage Collection"</a>,
417 by Boehm, Demers, and Shenker.
418 <P>
419 The most significant modification is that
420 the collector always starts running in the allocating thread.
421 There is no separate garbage collector thread. (If parallel GC is
422 enabled, helper threads may also be woken up.)
423 If an allocation attempt either requests a large object, or encounters
424 an empty small object free list, and notices that there is a collection
425 in progress, it immediately performs a small amount of marking work
426 as described above.
427 <P>
428 This change was made both because we wanted to easily accommodate
429 single-threaded environments, and because a separate GC thread requires
430 very careful control over the scheduler to prevent the mutator from
431 out-running the collector, and hence provoking unneeded heap growth.
432 <P>
433 In incremental mode, the heap is always expanded when we encounter
434 insufficient space for an allocation. Garbage collection is triggered
435 whenever we notice that more than
437 bytes of allocation have taken place.
439 is started.
440 <P>
441 All collections initially run interrupted until a predetermined
442 amount of time (50 msecs by default) has expired. If this allows
443 the collection to complete entirely, we can avoid correcting
444 for data structure modifications during the collection. If it does
445 not complete, we return control to the mutator, and perform small
446 amounts of additional GC work during those later allocations that
447 cannot be satisfied from small object free lists. When marking completes,
448 the set of modified pages is retrieved, and we mark once again from
449 marked objects on those pages, this time with the mutator stopped.
450 <P>
451 We keep track of modified pages using one of several distinct mechanisms:
452 <OL>
453 <LI>
454 Through explicit mutator cooperation. Currently this requires
456 <LI>
458 catching write faults. This is
459 implemented for many Unix-like systems and for win32. It is not possible
460 in a few environments.
461 <LI>
463 (Currently only Sun's
464 Solaris supports this. Though this is considerably cleaner, performance
465 may actually be better with mprotect and signals.)
466 <LI>
468 case the one in Xerox PCR.
469 <LI>
471 none of the other techniques is known to be usable, and
474 </ol>
479 in
483 <P>
484 During the mark phase, the collector tracks ``near misses'', i.e. attempts
485 to follow a ``pointer'' to just outside the garbage-collected heap, or
486 to a currently unallocated page inside the heap. Pages that have been
487 the targets of such near misses are likely to be the targets of
488 misidentified ``pointers'' in the future. To minimize the future
489 damage caused by such misidentifications they will be allocated only to
490 small pointerfree objects.
491 <P>
492 The collector understands two different kinds of black-listing. A
493 page may be black listed for interior pointer references
495 miss from a location that requires interior pointer recognition,
497 is set. In this case, we also avoid allocating large blocks that include
498 this page.
499 <P>
500 If the near miss came from a source that did not require interior
501 pointer recognition, it is black-listed with
503 A page black-listed in this way may appear inside a large object,
504 so long as it is not the first page of a large object.
505 <P>
507 a block to a particular object kind and size. It occasionally
508 drops (i.e. allocates and forgets) blocks that are completely black-listed
509 in order to avoid excessively long large block free lists containing
510 only unusable blocks. This would otherwise become an issue
511 if there is low demand for small pointerfree objects.
514 We support several different threading models. Unfortunately Pthreads,
515 the only reasonably well standardized thread model, supports too narrow
516 an interface for conservative garbage collection. There appears to be
517 no completely portable way to allow the collector
518 to coexist with various Pthreads
519 implementations. Hence we currently support only the more
520 common Pthreads implementations.
521 <P>
522 In particular, it is very difficult for the collector to stop all other
523 threads in the system and examine the register contents. This is currently
524 accomplished with very different mechanisms for some Pthreads
525 implementations. The Solaris implementation temporarily disables much
526 of the user-level threads implementation by stopping kernel-level threads
527 ("lwp"s). The Linux/HPUX/OSF1 and Irix implementations sends signals to
528 individual Pthreads and has them wait in the signal handler.
529 <P>
530 The Linux and Irix implementations use
531 only documented Pthreads calls, but rely on extensions to their semantics.
533 mild extensions to the pthreads semantics, and already supports a large number
534 of other Unix-like pthreads implementations. Our goal is to make this the
535 only pthread support in the collector.
536 <P>
537 (The Irix implementation is separate only for historical reasons and should
538 clearly be merged. The current Solaris implementation probably performs
539 better in the uniprocessor case, but does not support thread operations in the
540 collector. Hence it cannot support the parallel marker.)
541 <P>
542 All implementations must
543 intercept thread creation and a few other thread-specific calls to allow
544 enumeration of threads and location of thread stacks. This is current
547 by using ld's function call wrapping mechanism under Linux.
548 <P>
549 Recent versions of the collector support several facilites to enhance
550 the processor-scalability and thread performance of the collector.
552 <P>
553 Comments are appreciated. Please send mail to
556 <P>
557 This is a modified copy of a page written while the author was at SGI.
559 </body>
560 </html>