1 Title: Understanding the current GC, the end
2 Tags: en, d, dgc, understanding the current gc, druntime, gc, mark-sweep, mark, sweep
4 In this post I will take a closer look at the ``Gcx.mark()`` and
5 ``Gcx.fullcollect()`` functions.
7 This is a simplified version of the **mark** algorithm::
13 offset = from - pool.baseAddr
14 page_index = offset / PAGESIZE
15 bin_size = pool.pagetable[page_index]
16 bit_index = find_bit_index(bin_size, pool, offset)
17 if not pool.mark.test(bit_index)
18 pool.mark.set(bit_index)
19 if not pool.noscan.test(bit_index)
20 pool.scan.set(bit_index)
23 anychanges |= changes // anychanges is global
25 In the original version, there are some optimizations_ and the
26 ``find_bit_index()`` function doesn't exist (it does some bit masking to find
27 the right bit index for the bit set). But everything else is pretty much the
30 So far, is evident that the algorithm don't mark the whole heap in one step,
31 because it doesn't follow pointers. It just marks a consecutive chunk of
32 memory, assuming that pointers can be at any place in that memory, as long as
33 they are aligned (from increments in word-sized steps).
35 ``fullcollect()`` is the one in charge of following pointers, and marking
36 chunks of memory. It does it in an iterative way (that's why ``mark()`` informs
37 about ``anychanges`` (when new pointer should be followed to mark them, or,
38 speaking in the `tri-colour abstraction`_, when grey cells are found).
40 ``fullcollect()`` is huge, so I'll split it up in smaller pieces for the sake
41 of clarity. Let's see what are the basic blocks (see the `second part of this
48 rt_scanStaticData(mark)
49 thread_scanAll(mark, stackTop)
55 Generaly speaking, all the functions that have some *CamelCasing* are real
56 functions and the ones that are *all_lowercase* and made up by me.
58 Let's see each function.
60 ``thread_suspendAll()``
61 This is part of the threads runtime (found in
62 ``src/common/core/thread.d``). A simple peak at it shows it uses
63 ``SIGUSR1`` to stop the thread. When the signal is caught it pushes all the
64 registers into the stack to be sure any pointers there are scanned in the
65 future. The threads waits for ``SIGUSR2`` to resume.
70 foreach pool in pooltable
78 foreach n in B_16 .. B_PAGE
79 foreach node in bucket
81 pool.freebits.set(find_bit_index(pool, node))
82 pool.mark.set(find_bit_index(pool, node))
84 ``rt_scanStaticData(mark)``
85 This function, as the name suggests, uses the provided mark function
86 callback to scan the program's static data.
88 ``thread_scanAll(mark, stackTop)``
89 This is another threads runtime function, used to mark the suspended
90 threads stacks. I does some calculation about the stack bottom and top, and
91 calls ``mark(bottom, top)``, so at this point we have marked all reachable
92 memory from the stack(s).
97 mark(roots, roots + nroots)
98 foreach range in ranges
99 mark(range.pbot, range.ptop)
102 This is where most of the marking work is done. The code is **really**
103 ugly, very hard to read (mainly because of bad variable names) but what it
104 does it's relatively simple, here is the simplified algorithm::
106 // anychanges is global and was set by the mark()ing of the
107 // stacks and root set
110 foreach pool in pooltable
111 foreach bit_pos in pool.scan
112 if not pool.scan.test(bit_pos)
114 pool.scan.clear(bit_pos) // mark as already scanned
115 bin_size = find_bin_for_bit(pool, bit_pos)
116 bin_base_addr = find_base_addr_for_bit(pool, bit_pos)
117 if bin_size < B_PAGE // small object
118 bin_top_addr = bin_base_addr + bin_size
119 else if bin_size in [B_PAGE, B_PAGEPLUS] // big object
120 page_num = (bin_base_addr - pool.baseAddr) / PAGESIZE
121 if bin == B_PAGEPLUS // search for the base page
122 while pool.pagetable[page_num - 1] != B_PAGE
125 while page_num + n_pages < pool.ncommitted
126 and pool.pagetable[page_num + n_pages] == B_PAGEPLUS
128 bin_top_addr = bin_base_addr + n_pages * PAGESIZE
129 mark(bin_base_addr, bin_top_addr)
131 The original algorithm has some optimizations for proccessing bits in
132 clusters (skips groups of bins without the scan bit) and some kind-of
135 Again, the functions in ``all_lower_case`` don't really exist, some pointer
136 arithmetics are done in place for finding those values.
138 Note that the pools are iterated over and over again until there are no
139 unvisited bins. I guess this is a fair price to pay for not having a mark
140 stack (but I'm not really sure =).
142 ``thread_resumeAll()``
143 This is, again, part of the threads runtime and resume all the paused
144 threads by signaling a ``SIGUSR2`` to them.
152 ``mark_unmarked_free()``
153 This (invented) function looks for unmarked bins and set the ``freebits``
154 bit on them if they are small objects (bin size smaller than ``B_PAGE``) or
155 mark the entire page as free (``B_FREE``) in case of large objects.
157 This step is in charge of executing destructors too (through
158 ``rt_finalize()`` the runtime function).
160 ``rebuild_free_list()``
161 This (also invented) function first clear the free list (``bucket``) and
162 then rebuild it using the information collected in the previous step.
164 As usual, only bins with size smaller than ``B_PAGE`` are linked to the
165 free list, except if the pages they belong to have all the bins freed, in
166 which case the page is marked with the special ``B_FREE`` bin size. The
167 same goes for big objects freed in the previous step.
169 I think rebuilding the whole free list is not necessary, the new free bins
170 could be just linked to the existing free list. I guess this step exists to
171 help reducing fragmentation, since the rebuilt free list group bins
172 belonging to the same page together.
175 .. _optimizations: /blog/blog/post/7bdad55d
176 .. _`tri-colour abstraction`: http://en.wikipedia.org/wiki/Garbage_collection_(computer_science)#Tri-colour_marking
177 .. _`second part of this series`: /blog/blog/post/526122be
178 .. _bugs: http://www.dsource.org/projects/druntime/ticket/20
180 .. vim: set et sw=4 sts=4 :