學習中主要參考 @杏林小軒 的 Jemalloc系列:
jemalloc 3.6.0源碼詳解—[3]Run and bins
jemalloc 3.6.0源碼詳解—[4]Thread caches
這裏借用部分內容,便於複習;本文補充在 Android設備上查看 Jemalloc 的內存佈局.
1.概念
jemalloc對內存劃分按照如下由高到低的順序:
- 內存是由一定數量的arenas進行管理.
- 一個arena被分割成若干chunks, 後者主要負責記錄bookkeeping(記錄信息).
- chunk內部又包含着若干runs, 作爲分配小塊內存的基本單元.
- run由pages組成, 最終被劃分成一定數量的regions,
- 對於small size的分配請求來說, 這些region就相當於user memory.
2.使用shadow查看Jemalloc內存
shadow使用及配置:Shadow2.1 查看 arenas
arena 指針: (arena_t *)對應 struct:struct arena_s {(gdb) jearenas index address bins chunks threads ------------------------------------------------------ 0 0x7623a00140 36 17 1 0x7623a8fc00 36 4
可以看到,總共有兩個 arena,每個 arena有36個 bin,但chunk數量不同,兩個 arena共有 21個chunk;需要注意的是:arena本來是 cpu core數目的 4倍,而 Android 設置了最多有 2 個 arena;實際上,arena 指針是存儲在 je_arenas 數組中,shadow的 jearenas就是從這個數據讀取的 arena:從上面 jearenas的打印,我們知道兩個 arena 的指針分別是:(gdb) p je_arenas $1 = (arena_t **) 0x7623a87c00 (gdb) p *je_arenas@2 $2 = {0x7623a00140, 0x7623a8fc00}
{0x7623a00140, 0x7623a8fc00},對應的類型是 (arena_t *);可以通過如下命令查看 arena具體數據:(gdb) p *((arena_t *)0x7623a00140) $4 = { ind = 0, nthreads = {10, 10}, lock = { lock = { __private = {0, 0, 0, 0, 0, 0, 0, 0, 0, 0} }, witness = { name = 0x0, rank = 0, comp = 0x0, link = { qre_next = 0x0, qre_prev = 0x0 } } }, stats = { .....
2.2 查看 chunks
chunk指針 :(arena_chunk_t*)struct arena_chunk_s { /* * A pointer to the arena that owns the chunk is stored within the node. * This field as a whole is used by chunks_rtree to support both * ivsalloc() and core-based debugging. */ extent_node_t node; /* * True if memory could be backed by transparent huge pages. This is * only directly relevant to Linux, since it is the only supported * platform on which jemalloc interacts with explicit transparent huge * page controls. */ bool hugepage; /* * Map of pages within chunk that keeps track of free/large/small. The * first map_bias entries are omitted, since the chunk header does not * need to be tracked in the map. This omission saves a header page * for common chunk sizes (e.g. 4 MiB). */ arena_chunk_map_bits_t map_bits[1]; /* Dynamically sized. */ };
chunk內存佈局:查看chunks:可以看到共有 21 chunk;與上面 jearenas統計的數據符合;(gdb) jechunks addr arena no_runs ------------------------------------------- 0x7604800000 0x7623a00140 4 0x7604a00000 0x7623a00140 13 0x7605600000 0x7623a00140 58 0x7605800000 0x7623a8fc00 12 0x7605a00000 0x7623a00140 99 0x7606000000 0x7623a00140 42 0x7606200000 0x7623a00140 73 0x7607800000 0x7623a8fc00 103 0x7607a00000 0x7623a00140 69 0x7607c00000 0x7623a8fc00 53 0x7608400000 0x7623a00140 91 0x7613400000 0x7623a00140 94 0x7613600000 0x7623a00140 37 0x7613800000 0x7623a00140 31 0x7613a00000 0x7623a00140 23 0x7613c00000 0x7623a00140 33 0x7613e00000 0x7623a00140 22 0x7614000000 0x7623a00140 7 0x7615e00000 0x7623a00140 171 0x7618a00000 0x7623a8fc00 158 0x7623800000 0x7623a00140 174
這個數據除了指明各個 chunk指針 (arena_chunk_t),還統計了每個 chunk中的 runs數量;查看一個chunk:可以看到這個 chunk所屬的 arena是 0x7623a00140, chunk的大小是 2M。(gdb) p *(arena_chunk_t*)0x7604800000 $12 = { node = { en_arena = 0x7623a00140, /* arena */ en_addr = 0x7604800000, en_size = 2097152, /* Total region size. */ en_sn = 16, en_zeroed = true, en_committed = true, en_achunk = true, en_prof_tctx = 0x0, rd = { rd_link = { qre_next = 0x0, qre_prev = 0x0 } }, cc_link = { qre_next = 0x0, qre_prev = 0x0 }, { szsnad_link = { rbn_left = 0x7623800000, rbn_right_red = 0x7604a00000 }, ql_link = { qre_next = 0x7623800000, qre_prev = 0x7604a00000 } }, ad_link = { rbn_left = 0x0, rbn_right_red = 0x0 } }, hugepage = true, map_bits = {{ bits = 2015216 }} }
可以知道的是 chunks是保存在 je_chunks_rtree 數據結構中的。還沒有搞明白這個數據結構如何訪問,搞明白了,再補充。詳細查看單個的 chunk:arena_chunk_t中的 map_bits用來記錄當前chunk偏移 0xd000(je_map_bias) 之後的所有的 page的使用狀態;(gdb) jechunk 0x7623800000 This chunk belongs to the arena at 0x7623a00140. addr info size usage ------------------------------------------------------------ 0x7623800000 headers 0xd000 - 0x762380d000 small run (0x1c00) 0x7000 4/4 0x7623814000 small run (0xc00) 0x3000 4/4 0x7623817000 small run (0x50) 0x5000 256/256 0x762381c000 small run (0x20) 0x1000 128/128 0x762381d000 small run (0x1000) 0x1000 1/1 0x762381e000 small run (0x1000) 0x1000 1/1 0x762381f000 small run (0x1000) 0x1000 1/1 0x7623820000 small run (0x200) 0x1000 8/8 0x7623821000 small run (0x50) 0x5000 256/256 0x7623826000 small run (0x20) 0x1000 128/128 ...
比如第一個 run是 small run,run大小是0x7000, 包含 7個page,這 7 個 page被平分成了 4個 region;每個 run 包含 N 個 region,每個 region又包含 N個 PAGE,所以 run大小是 page整數倍;查看一個run:這個 run是當前chunk的第一個run,包含4個 region,都是使用狀態,每個 region的開始地址也標註了;(gdb) jerun 0x762380d000 * status address preview --------------------------------------------------- 0 used 0x762380d000 0000007623812400 1 used 0x762380ec00 0000000000000000 2 used 0x7623810800 0000007613505000 3 used 0x7623812400 0000007613506c00
我們通過 當前chunk的 map_bits 看下,第一個run的所有page的狀態;通過 struct arena_chunk_map_bits_s,我們知道, 對應 bits的 第 [0] bit,代表是否被 allocated,第[1] bit代表是否 large run,(gdb) p /t ((arena_chunk_t*)0x7623800000)->map_bits[0] $155 = { bits = 1111100001 } (gdb) p /t ((arena_chunk_t*)0x7623800000)->map_bits[1] $156 = { bits = 10001111100001 } (gdb) p /t ((arena_chunk_t*)0x7623800000)->map_bits[2] $157 = { bits = 100001111100001 } (gdb) p /t ((arena_chunk_t*)0x7623800000)->map_bits[3] $158 = { bits = 110001111100001 } (gdb) p /t ((arena_chunk_t*)0x7623800000)->map_bits[4] $159 = { bits = 1000001111100001 } (gdb) p /t ((arena_chunk_t*)0x7623800000)->map_bits[5] $160 = { bits = 1010001111100001 } (gdb) p /t ((arena_chunk_t*)0x7623800000)->map_bits[6] $161 = { bits = 1100001111100001 }
可以看到 第 0 ~ 6 的這7個 page都是 allocated狀態[1],且屬於small run [0];我們可以找一個 large run驗證下:可以看到,0x7623884000 對應的 large run,其第1個 bit是 1;(gdb) jechunk 0x7623800000 This chunk belongs to the arena at 0x7623a00140. addr info size usage ------------------------------------------------------------ 0x7623800000 headers 0xd000 - 0x762380d000 small run (0x1c00) 0x7000 4/4 0x7623814000 small run (0xc00) 0x3000 4/4 ... 0x7623883000 small run (0x40) 0x1000 64/64 0x7623884000 large run 0x7000 - 0x762388b000 small run (0xc00) 0x3000 4/4 (gdb) p (0x7623884000-0x762380d000)/4096 $163 = 119 (gdb) p /t ((arena_chunk_t*)0x7623800000)->map_bits[118] $164 = { bits = 10000001 } (gdb) p /t ((arena_chunk_t*)0x7623800000)->map_bits[119] $165 = { bits = 1111111111100011 }
對於 map_bits的解釋,在 arena.h文件中:所以,對於該 large run :/* Each element of the chunk map corresponds to one page within the chunk. */ struct arena_chunk_map_bits_s { /* * Run address (or size) and various flags are stored together. The bit * layout looks like (assuming 32-bit system): * * ???????? ???????? ???nnnnn nnndumla * * ? : Unallocated: Run address for first/last pages, unset for internal * pages. * Small: Run page offset. * Large: Run page count for first page, unset for trailing pages. * n : binind for small size class, BININD_INVALID for large size class. * d : dirty? * u : unzeroed? * m : decommitted? * l : large? * a : allocated? * * Following are example bit patterns for the three types of runs. * * p : run page offset * s : run size * n : binind for size class; large objects set these to BININD_INVALID * x : don't care * - : 0 * + : 1 * [DUMLA] : bit set * [dumla] : bit unset * * Unallocated (clean): * ssssssss ssssssss sss+++++ +++dum-a * xxxxxxxx xxxxxxxx xxxxxxxx xxx-Uxxx * ssssssss ssssssss sss+++++ +++dUm-a * * Unallocated (dirty): * ssssssss ssssssss sss+++++ +++D-m-a * xxxxxxxx xxxxxxxx xxxxxxxx xxxxxxxx * ssssssss ssssssss sss+++++ +++D-m-a * * Small: * pppppppp pppppppp pppnnnnn nnnd---A * pppppppp pppppppp pppnnnnn nnn----A * pppppppp pppppppp pppnnnnn nnnd---A * * Large: * ssssssss ssssssss sss+++++ +++D--LA * xxxxxxxx xxxxxxxx xxxxxxxx xxxxxxxx * -------- -------- ---+++++ +++D--LA * * Large (sampled, size <= LARGE_MINCLASS): * ssssssss ssssssss sssnnnnn nnnD--LA * xxxxxxxx xxxxxxxx xxxxxxxx xxxxxxxx * -------- -------- ---+++++ +++D--LA * * Large (not sampled, size == LARGE_MINCLASS): * ssssssss ssssssss sss+++++ +++D--LA * xxxxxxxx xxxxxxxx xxxxxxxx xxxxxxxx * -------- -------- ---+++++ +++D--LA */ size_t bits; #define CHUNK_MAP_ALLOCATED ((size_t)0x01U) #define CHUNK_MAP_LARGE ((size_t)0x02U) #define CHUNK_MAP_STATE_MASK ((size_t)0x3U) #define CHUNK_MAP_DECOMMITTED ((size_t)0x04U) #define CHUNK_MAP_UNZEROED ((size_t)0x08U) #define CHUNK_MAP_DIRTY ((size_t)0x10U) #define CHUNK_MAP_FLAGS_MASK ((size_t)0x1cU) #define CHUNK_MAP_BININD_SHIFT 5 #define BININD_INVALID ((size_t)0xffU) #define CHUNK_MAP_BININD_MASK (BININD_INVALID << CHUNK_MAP_BININD_SHIFT) #define CHUNK_MAP_BININD_INVALID CHUNK_MAP_BININD_MASK #define CHUNK_MAP_RUNIND_SHIFT (CHUNK_MAP_BININD_SHIFT + 8) #define CHUNK_MAP_SIZE_SHIFT (CHUNK_MAP_RUNIND_SHIFT - LG_PAGE) #define CHUNK_MAP_SIZE_MASK \ (~(CHUNK_MAP_BININD_MASK | CHUNK_MAP_FLAGS_MASK | CHUNK_MAP_STATE_MASK)) };
0x7623884000 large run 0x7000 -其總共佔了 7個page,其對應的 map_bits的 第 [119] ~ [125] 的7 個page,所以再查看其最後一個 page的狀態看下:(gdb) p /t ((arena_chunk_t*)0x7623800000)->map_bits[125] $169 = { bits = 1111111100011 }
(gdb) p ((arena_chunk_t*)0x7623800000)->map_bits[119]->bits >> 13 $179 = 7
其對應的第一個page的 map_bits中也記錄了該 large run的大小:7 * PAGE_SIZE其內容,確實符合規則;對於chunk中第一個run 0x762380d000,是個 small run:也驗證了,map_bits保存其的狀態 bits的高19位,保存着當前 page 在該run中的 offset;(gdb) p ((arena_chunk_t*)0x7623800000)->map_bits[0]->bits >> 13 $172 = 0 (gdb) p ((arena_chunk_t*)0x7623800000)->map_bits[1]->bits >> 13 $173 = 1 (gdb) p ((arena_chunk_t*)0x7623800000)->map_bits[2]->bits >> 13 $174 = 2 (gdb) p ((arena_chunk_t*)0x7623800000)->map_bits[3]->bits >> 13 $175 = 3 (gdb) p ((arena_chunk_t*)0x7623800000)->map_bits[4]->bits >> 13 $176 = 4 (gdb) p ((arena_chunk_t*)0x7623800000)->map_bits[5]->bits >> 13 $177 = 5 (gdb) p ((arena_chunk_t*)0x7623800000)->map_bits[6]->bits >> 13 $178 = 6
而對於 small run,對應page的 map_bits的第[5~12]的 8個bit保存其對應的 bin的index該run共 7個page,其狀態對應 map_bits的起始 7個元素,當前run第一個 page的狀態中保存有該run對應的bin index:0x762380d000 small run (0x1c00) 0x7000 4/4
可以看到 31號bin對應的 run size確實是 0x1c00;(gdb) p ((arena_chunk_t*)0x7623800000)->map_bits[0]->bits << 19 >> 24 $194 = 31 (gdb) jebins arena @ 0x7623a00140 index addr size runcur --------------------------------------------------- 0 0x7623a00ac0 0x8 0x7623801c28 1 0x7623a00b68 0x10 0x7613408348 2 0x7623a00c10 0x20 0x7615e03b48 3 0x7623a00cb8 0x30 0x76134084c8 4 0x7623a00d60 0x40 0x7623803548 ... 30 0x7623a01e70 0x1800 0x76084044a8 31 0x7623a01f18 0x1c00 0x7608401808 32 0x7623a01fc0 0x2000 0x7608403f68 33 0x7623a02068 0x2800 0x7606207568 34 0x7623a02110 0x3000 0x7607a07f88 35 0x7623a021b8 - -
2.3 查看 runs
run指針:(arena_run_t *)struct arena_run_s { /* Index of bin this run is associated with. */ szind_t binind; /* Number of free regions in run. */ unsigned nfree; /* Per region allocated/deallocated bitmap. */ bitmap_t bitmap[BITMAP_GROUPS_MAX]; };
查看所有 runs:可以看到,對應的 run指針()地址,當前 run 的 size,當前run中的 region數量,以及每個 region的size,和 處於 free狀態的region數量。(gdb) jeruns * run_addr run_size region_size no_regions no_free ---------------------------------------------------------------------------- 1 0x7604902000 0x31000 - - - 2 0x7604933000 0x31000 - - - 3 0x7604964000 0x31000 - - - 4 0x7604995000 0x31000 - - - 5 0x7604a0d000 0x31000 - - - 6 0x7604a4f000 0x5000 0x1400 4 0 7 0x7604a54000 0x3000 0x600 8 0 8 0x7604a57000 0x5000 0x1400 4 0 9 0x7604a5c000 0x5000 0x1400 4 3 10 0x7604a77000 0x1000 0x800 2 0 11 0x7604a78000 0x31000 - - - 12 0x7604b0e000 0xb000 - - - 13 0x7604b1a000 0x3000 0x1800 2 1 14 0x7604b33000 0x3000 0x3000 1 0 15 0x7604b36000 0x3000 0x3000 1 0 16 0x7604b5f000 0xb000 - - - 17 0x7604b6b000 0x1000 0x1000 1 0 18 0x760560d000 0x29000 - - - 19 0x7605636000 0x29000 - - - 20 0x760565f000 0x29000 - - - 21 0x7605688000 0x5000 0x280 32 29 22 0x7605690000 0x8000 - - - 23 0x7605698000 0x5000 0xa00 8 6 24 0x76056b1000 0x3000 0x3000 1 0 25 0x76056b4000 0x3000 0x3000 1 0 26 0x76056b7000 0x3000 0x3000 1 0 ....
runs是連續存儲在 chunks中的:(gdb) p /x 0x1400*4 $18 = 0x5000 (gdb) p /x 0x7604a4f000+0x5000 $20 = 0x7604a54000 (gdb) jechunk 0x7604a54000 This chunk belongs to the arena at 0x7623a00140. addr info size usage ---------------------------------------------------------- 0x7604a00000 headers 0xd000 - 0x7604a0d000 large run 0x31000 - 0x7604a3e000 unused range 0x73000 - 0x7604a4f000 small run (0x1400) 0x5000 4/4 0x7604a54000 small run (0x600) 0x3000 8/8 0x7604a57000 small run (0x1400) 0x5000 4/4 0x7604a5c000 small run (0x1400) 0x5000 1/4 0x7604a61000 unused range 0x20000 - 0x7604a77000 small run (0x800) 0x1000 2/2 0x7604a78000 large run 0x31000 - 0x7604aa9000 unused range 0xc7000 - 0x7604b0e000 large run 0xb000 - 0x7604b19000 unused range 0x17000 - 0x7604b1a000 small run (0x1800) 0x3000 1/2 0x7604b1d000 unused range 0x1c000 - 0x7604b33000 small run (0x3000) 0x3000 1/1 0x7604b36000 small run (0x3000) 0x3000 1/1 0x7604b39000 unused range 0x2c000 - 0x7604b5f000 large run 0xb000 - 0x7604b6a000 unused range 0x17000 - 0x7604b6b000 small run (0x1000) 0x1000 1/1 0x7604b6c000 unused range 0x96000 -
可以確認的是,runs的獲取,是由 chunk獲得的,使用 chunk及其成員 arena_chunk_map_bits_tmap_bits[1];在chunk中,後續的每個run都是page size的整數倍,且每個run是緊鄰的;來獲取的每個 chunk的 runs;其中sizeof (arena_chunk_t) =128,而 header size = 0xd000(13個page = je_map_bias)從上面數據可以看到,緊跟着 header後面就是 chunk中的連續的 runs,而 header 的大小是固定的:je_map_bias * PAGE_SIZE;(gdb) p sizeof(arena_chunk_t) $94 = 128 (gdb) p je_map_bias $95 = 13
je_map_bias是根據chunk的大小 je_chunksize (je_chunk_npages*PAGE_SIZE),以及 arena_chunk_t計算出來的,在 Android8.0的一個手機上:每個chunk最大是 2M,一個 chunk 中有 1 ~ N 個run,所以每個run最大是:(gdb) p /x je_chunksize $117 = 0x200000 (gdb) p je_map_bias $118 = 13
(gdb) p je_chunksize-je_map_bias*4096 $123 = 2043904 (gdb) p je_arena_maxrun $121 = 2043904
chunk與 runs的關係:
查看其中的一個 run 的bitmap:
可以看到,其 bitmap數組只有 8個元素,而它有 32 個 region;(gdb) jeruns * run_addr run_size region_size no_regions no_free ---------------------------------------------------------------------------- 1 0x7604902000 0x31000 - - - ... 32 0x76056c4000 0x1000 0x80 32 5 ... (gdb) p &((arena_run_t*)0x76056c4000)->bitmap $261 = (bitmap_t (*)[8]) 0x76056c4008
(gdb) jerun 0x76056c4000 * status address preview ---------------------------------------------------- 0 used 0x76056c4000 0000003800000000 1 used 0x76056c4080 0000002200000000 2 used 0x76056c4100 0000441400000000 3 used 0x76056c4180 ff917754ff917701 4 used 0x76056c4200 ff917754ff917754 5 used 0x76056c4280 0e00000050221c13 6 used 0x76056c4300 0c00000000000000 7 used 0x76056c4380 ff917754ff917754 8 used 0x76056c4400 ff917754ff917701 9 used 0x76056c4480 0000223300000000 10 used 0x76056c4500 0000261400000000 11 used 0x76056c4580 00001d1600000000 12 used 0x76056c4600 0000457000000000 13 used 0x76056c4680 0020670962405c06 14 used 0x76056c4700 00003a4e00000000 15 used 0x76056c4780 0000002200000000 16 used 0x76056c4800 0000002200000000 17 used 0x76056c4880 0000003400000000 18 used 0x76056c4900 0000003400000000 19 used 0x76056c4980 0000002700000000 20 used 0x76056c4a00 0000003000000000 21 used 0x76056c4a80 00002d8000000000 22 used 0x76056c4b00 000064d900000000 23 used 0x76056c4b80 0000002200000000 24 used 0x76056c4c00 0000441400000000 25 free 0x76056c4c80 0000000006000002 26 free 0x76056c4d00 0000000006000002 27 free 0x76056c4d80 0000000006000002 28 free 0x76056c4e00 0000000006000002 29 free 0x76056c4e80 0000000006000002 30 used 0x76056c4f00 0000000006000002 31 used 0x76056c4f80 0000000006000002
看到 jerun從開始的位置就是屬於第一個 region,header在哪裏 ?
且查看 8.0 代碼中runlayout,與上圖有所不同:
2.3 查看 regions
(gdb) jeruns * run_addr run_size region_size no_regions no_free ---------------------------------------------------------------------------- 1 0x7604902000 0x31000 - - - 2 0x7604933000 0x31000 - - - 3 0x7604964000 0x31000 - - - 4 0x7604995000 0x31000 - - - 5 0x7604a0d000 0x31000 - - - 6 0x7604a4f000 0x5000 0x1400 4 0 ... (gdb) jerun 0x7604a4f000 * status address preview --------------------------------------------------- 0 used 0x7604a4f000 00000076268e01b8 1 used 0x7604a50400 00000076268e01b8 2 used 0x7604a51800 00000076268e01b8 3 used 0x7604a52c00 00000076268e01b8
可以看到,當前 run總共有 4個region,且都是使用狀態,每個 region的 size 是 0x1400。region大小爲 0x1400的 run共有 21個,第一個 run就是我們剛剛查看的 那個run;(gdb) jeregions 0x1400 * run_addr reg_size run_size usage ------------------------------------------------------- 1 0x7604a4f000 5120 0x5000 4/4 2 0x7604a57000 5120 0x5000 4/4 3 0x7604a5c000 5120 0x5000 1/4 4 0x76057bd000 5120 0x5000 2/4 5 0x7605877000 5120 0x5000 1/4 6 0x7605b2a000 5120 0x5000 2/4 7 0x7605b34000 5120 0x5000 4/4 8 0x7605b39000 5120 0x5000 3/4 9 0x7605be7000 5120 0x5000 2/4 10 0x7605bec000 5120 0x5000 4/4 11 0x7606187000 5120 0x5000 4/4 12 0x76062fd000 5120 0x5000 4/4 13 0x7606302000 5120 0x5000 4/4 14 0x76078bd000 5120 0x5000 4/4 15 0x76079da000 5120 0x5000 4/4 16 0x7607b60000 5120 0x5000 1/4 17 0x7607dea000 5120 0x5000 3/4 18 0x7613c84000 5120 0x5000 4/4 19 0x7613fb3000 5120 0x5000 4/4 20 0x7623973000 5120 0x5000 4/4 (gdb) p /x 5120 $213 = 0x1400
2.4 查看 bins
run是分配的執行者, 而分配的調度者是bin. 這個概念同dlmalloc中的bin是類似的, 但jemalloc中bin要更復雜一些. 直白地說, 可以把bin看作non-full run的倉庫, bin負責記錄當前arena中某一個size class範圍內所有non-full run的使用情況. 當有分配請求時, arena查找相應size class的bin, 找出可用於分配的run, 再由run分配region. 當然, 因爲只有small region分配需要run, 所以bin也只對應small size class.struct arena_bin_s { malloc_mutex_t lock; arena_run_t *runcur; arena_run_tree_t runs; malloc_bin_stats_t stats; };
這段粗體,摘自 @杏林小軒 的博客;
lock: 該lock同arena內部的lock不同, 主要負責保護current run. 而對於run本身的分配和釋放還是需要依賴arena lock. 通常情況下, 獲得bin lock的前提是獲得arena lock, 但反之不成立.
runcur: 當前可用於分配的run, 一般情況下指向地址最低的non-full run, 同一時間一個bin只有一個current run用於分配.
runs: rb tree, 記錄當前arena中該bin對應size class的所有non-full runs. 因爲分配是通過current run完成的, 所以也相當於current run的倉庫.
stats: 統計信息.
查看bins:可以看到,每個 arena都對應 36個 bin,對應 36種大小的 mem class;(gdb) jebins arena @ 0x7623a00140 index addr size runcur --------------------------------------------------- 0 0x7623a00ac0 0x8 0x7623801c28 1 0x7623a00b68 0x10 0x7613408348 2 0x7623a00c10 0x20 0x7615e03b48 3 0x7623a00cb8 0x30 0x76134084c8 4 0x7623a00d60 0x40 0x7623803548 5 0x7623a00e08 0x50 0x7608408168 6 0x7623a00eb0 0x60 0x7607a088e8 7 0x7623a00f58 0x70 0x7607a071a8 8 0x7623a01000 0x80 0x7613c04328 9 0x7623a010a8 0xa0 0x762380b2e8 10 0x7623a01150 0xc0 0x7608405348 11 0x7623a011f8 0xe0 0x762380c368 12 0x7623a012a0 0x100 0x7608408888 13 0x7623a01348 0x140 0x761340b0a8 14 0x7623a013f0 0x180 0x7623802468 15 0x7623a01498 0x1c0 0x7623804b08 16 0x7623a01540 0x200 0x76084017a8 17 0x7623a015e8 0x280 0x7607a09368 18 0x7623a01690 0x300 0x7615e06428 19 0x7623a01738 0x380 0x7615e03de8 20 0x7623a017e0 - - 21 0x7623a01888 0x500 0x7623806cc8 22 0x7623a01930 - - 23 0x7623a019d8 0x700 0x7623802888 24 0x7623a01a80 0x800 0x7613409de8 25 0x7623a01b28 0xa00 0x7623807388 26 0x7623a01bd0 0xc00 0x7605a03248 27 0x7623a01c78 0xe00 0x760560b588 28 0x7623a01d20 0x1000 0x7613407028 29 0x7623a01dc8 0x1400 0x7607a08f48 30 0x7623a01e70 0x1800 0x76084044a8 31 0x7623a01f18 0x1c00 0x7608401808 32 0x7623a01fc0 0x2000 0x7608403f68 33 0x7623a02068 0x2800 0x7606207568 34 0x7623a02110 0x3000 0x7607a07f88 35 0x7623a021b8 - - arena @ 0x7623a8fc00 index addr size runcur --------------------------------------------------- 0 0x7623a90580 0x8 0x7618a01868 1 0x7623a90628 0x10 0x7618a09008 2 0x7623a906d0 0x20 0x7618a04e08 3 0x7623a90778 0x30 0x7607804268 4 0x7623a90820 0x40 0x7607c0c8a8 5 0x7623a908c8 0x50 0x760780ac28 6 0x7623a90970 0x60 0x7618a021c8 7 0x7623a90a18 0x70 0x7618a022e8 8 0x7623a90ac0 0x80 0x7607807b08 9 0x7623a90b68 0xa0 0x7618a02588 10 0x7623a90c10 0xc0 0x7607c0c4e8 11 0x7623a90cb8 0xe0 0x7618a015c8 12 0x7623a90d60 0x100 0x7618a02048 13 0x7623a90e08 0x140 0x7607807568 14 0x7623a90eb0 0x180 0x7607802528 15 0x7623a90f58 0x1c0 0x7618a0a868 16 0x7623a91000 0x200 0x7618a03008 17 0x7623a910a8 0x280 0x7618a035a8 18 0x7623a91150 0x300 0x7607c0c068 19 0x7623a911f8 0x380 0x7607c0c608 20 0x7623a912a0 0x400 0x7607c0aa48 21 0x7623a91348 0x500 0x7618a05528 22 0x7623a913f0 - - 23 0x7623a91498 0x700 0x7607805b88 24 0x7623a91540 0x800 0x76078035a8 25 0x7623a915e8 0xa00 0x76058015c8 26 0x7623a91690 - - 27 0x7623a91738 0xe00 0x7618a09068 28 0x7623a917e0 - - 29 0x7623a91888 0x1400 0x7607c0c308 30 0x7623a91930 0x1800 0x7607c0a808 31 0x7623a919d8 0x1c00 0x76078063c8 32 0x7623a91a80 - - 33 0x7623a91b28 0x2800 0x7607804628 34 0x7623a91bd0 - - 35 0x7623a91c78 - -
而實際上,每個 bin都關聯着多個 run,其中有的 run已經滿了,有的 run還沒滿,有的 run還在使用中;每個 run中的 region size是固定的,其實就是與 bin 的 mem class對應;比如,bin[31] 對應的 mem class是 0x1c00 大小的內存,那麼它關聯的所有 run的 region size都是 0x1c00;另外,有個 table:je_size2index_tab 是用來根據 mem size來查找其對應在 bins中的index的:由於bins是被分成 Group的,除了0號bin之外, 相鄰的4個bin屬於同一group, 相鄰的兩個group,各自group內的相鄰 bin 的差額是2倍;比如說:(gdb) p /d je_size2index_tab $286 = {0, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 9, 9, 10, 10, 10, 10, 11, 11, 11, 11, 12, 12, 12, 12, 13, 13, 13, 13, 13, 13, 13, 13, 14, 14, 14, 14, 14, 14, 14, 14, 15, 15, 15, 15, 15, 15, 15, 15, 16, 16, 16, 16, 16, 16, 16, 16, 17 <repeats 16 times>, 18 <repeats 16 times>, 19 <repeats 16 times>, 20 <repeats 16 times>, 21 <repeats 32 times>, 22 <repeats 32 times>, 23 <repeats 32 times>, 24 <repeats 32 times>, 25 <repeats 64 times>, 26 <repeats 64 times>, 27 <repeats 64 times>, 28 <repeats 64 times>}
已知 group N內有四個bin: bin1,bin2,bin3,bin4,這4個bin的size差距都是 deltaAgroup N+1內有4個bin:bin5,bin6,bin7,bin8,那麼這4個 bin的size差距都是 deltaA * 2group劃分如下:{0}, {1, 2, 3, 4}, {5, 6, 7, 8}, {9, 10, 11, 12}, {13, 14, 15, 16}, ...
由於bin的最小 size是 8 byte,所以得到一個 mem地址,獲取其 bin index可以這麼計算:bin_idx = (addr -1 ) >> 3由於每個 bin 的取值範圍不同,所以其對應的 index需要在 je_size2index_tab 中重複 (bin_range_size/8 ) 次,這樣才能得到正確的 bin index,比如:(gdb) jebins arena @ 0x7623a00140 index addr size runcur --------------------------------------------------- 0 0x7623a00ac0 0x8 0x7623801c28 1 0x7623a00b68 0x10 0x7613408348 2 0x7623a00c10 0x20 0x7615e03b48 3 0x7623a00cb8 0x30 0x76134084c8 4 0x7623a00d60 0x40 0x7623803548 5 0x7623a00e08 0x50 0x7608408168 6 0x7623a00eb0 0x60 0x7607a088e8 7 0x7623a00f58 0x70 0x7607a071a8 8 0x7623a01000 0x80 0x7613c04328 9 0x7623a010a8 0xa0 0x762380b2e8 10 0x7623a01150 0xc0 0x7608405348 11 0x7623a011f8 0xe0 0x762380c368 12 0x7623a012a0 0x100 0x7608408888 13 0x7623a01348 0x140 0x761340b0a8 14 0x7623a013f0 0x180 0x7623802468 15 0x7623a01498 0x1c0 0x7623804b08
第1組中,index1的bin,對應的取值範圍是 0x8 ~ 0x10,它的範圍只有8 byte,所以 index1 在 je_size2index_tab 中出現1次即可;第1組中,index2的bin,對應的取值範圍是 0x10 ~ 0x20,它的範圍有16 byte,所以 index1 在 je_size2index_tab 中出現 16/8 = 2次;第3組中,index9的bin,對應的取值範圍是 0xa0 ~ 0x80,它的範圍有32 byte,所以 index9 在 je_size2index_tab 中出現 32/8 = 4次;依次類推即可;這樣,我們就可以很快的爲一個請求的 size找到其合適的 bin 去分配內存了:比如 arenas[idx]->bins[ (size-1) >> 3] 即是 匹配 size的 bin,在該 bin上掛的 runs中給其分配 region 即可;2.5 tcache
TLS/TSD是另一種針對多線程優化使用的分配技術, jemalloc中稱爲tcache. tcache解決的是同一cpu core下不同線程對heap的競爭. 通過爲每個線程指定專屬分配區域,來減小線程間的干擾. 但顯然這種方法會增大整體內存消耗量. 爲了減小副作用,jemalloc將tcache設計成一個bookkeeping結構, 在tcache中保存的僅僅是指向外部region的指針, region對象仍然位於各個run當中. 換句話說, 如果一個region被tcache記錄了, 那麼從run的角度看, 它就已經被分配了.
tcache的內容如下,
struct tcache_s { ql_elm(tcache_t) link; uint64_t prof_accumbytes; arena_t *arena; unsigned ev_cnt; unsigned next_gc_bin; tcache_bin_t tbins[1]; };
link: 鏈接節點, 用於將同一個arena下的所有tcache鏈接起來.
prof_accumbytes: memory profile相關.
arena: 該tcache所屬的arena指針.
ev_cnt: 是tcache內部的一個週期計數器. 每當tcache執行一次分配或釋放時, ev_cnt會記錄一次. 直到週期到來, jemalloc會執行一次incremental gc.這裏的gc會清理tcache中多餘的region, 將它們釋放掉. 儘管這不意味着系統內存會獲得釋放, 但可以解放更多的region交給其他更飢餓的線程以分配.
next_gc_bin: 指向下一次gc的binidx. tcache gc按照一週期清理一個bin執行.
tbins: tcache bin數組. 同樣外掛在tcache後面.
同arena bin類似, tcache同樣有tcache_bin_t和tcache_bin_info_t.tcache_bin_t作用類似於arena bin, 但其結構要比後者更簡單. 準確的說, tcache bin並沒有分配調度的功能, 而僅起到記錄作用. 其內部通過一個stack記錄指向外部arena run中的region指針. 而一旦region被cache到tbins內, 就不能再被其他任何線程所使用, 儘管它可能甚至與其他線程tcache中記錄的region位於同一個arena run中.
tcache bin結構如下,
struct tcache_bin_s { tcache_bin_stats_t tstats; int low_water; unsigned lg_fill_div; unsigned ncached; void **avail; }
tstats: tcache bin內部統計.
low_water: 記錄兩次gc間tcache內部使用的最低水線. 該數值與下一次gc時嘗試釋放的region數量有關. 釋放量相當於low water數值的3/4.
lg_fill_div: 用作tcache refill時作爲除數. 當tcache耗盡時, 會請求arena run進行refill. 但refill不會一次性灌滿tcache, 而是依照其最大容量縮小2^lg_fill_div的倍數. 該數值同low_water一樣是動態的, 兩者互相配合確保tcache處於一個合理的充滿度.
ncached: 指當前緩存的region數量, 同時也代表棧頂index.
avail: 保存region指針的stack, 稱爲avail-stack.
tcache_bin_info_t保存tcache bin的靜態信息. 其本身只保存了tcache max容量. 該數值是在tcache boot時根據相對應的arena bin的nregs決定的. 通常等於nregs的二倍, 但不得超過TCACHE_NSLOTS_SMALL_MAX. 該數值默認爲200, 但在android中大大提升了該限制, small bins不得超過8, large bins則爲16.
tcache layout如下,struct tcache_bin_info_s { unsigned ncached_max; };
tcache的調試目前還沒有搞定,還無法從 gdb中查看 tcache,待調查;
3.Jemalloc分配與釋放
Jemalloc的分配與釋放待學習,分配路徑,參考下面 Jemalloc的框架圖,可以推測一二:1.優先 tcache,根據 tcache_bin查找合適的run,找到空閒的 region進行分配
2.繼而選擇一個 arena,然後根據bins,選擇對應 bins上掛着的run,再從run中選取合適的 region分配
當然遇到分配過程中的某一個失敗時,可以選取一個 arena,進行分配新的 chunk,run,region;