| 1 |
#ifndef Py_INTERNAL_OBMALLOC_H |
| 2 |
#define Py_INTERNAL_OBMALLOC_H |
| 3 |
#ifdef __cplusplus |
| 4 |
extern "C" { |
| 5 |
#endif |
| 6 |
|
| 7 |
#ifndef Py_BUILD_CORE |
| 8 |
# error "this header requires Py_BUILD_CORE define" |
| 9 |
#endif |
| 10 |
|
| 11 |
|
| 12 |
typedef unsigned int pymem_uint; /* assuming >= 16 bits */ |
| 13 |
|
| 14 |
#undef uint |
| 15 |
#define uint pymem_uint |
| 16 |
|
| 17 |
|
| 18 |
/* An object allocator for Python. |
| 19 |
|
| 20 |
Here is an introduction to the layers of the Python memory architecture, |
| 21 |
showing where the object allocator is actually used (layer +2), It is |
| 22 |
called for every object allocation and deallocation (PyObject_New/Del), |
| 23 |
unless the object-specific allocators implement a proprietary allocation |
| 24 |
scheme (ex.: ints use a simple free list). This is also the place where |
| 25 |
the cyclic garbage collector operates selectively on container objects. |
| 26 |
|
| 27 |
|
| 28 |
Object-specific allocators |
| 29 |
_____ ______ ______ ________ |
| 30 |
[ int ] [ dict ] [ list ] ... [ string ] Python core | |
| 31 |
+3 | <----- Object-specific memory -----> | <-- Non-object memory --> | |
| 32 |
_______________________________ | | |
| 33 |
[ Python's object allocator ] | | |
| 34 |
+2 | ####### Object memory ####### | <------ Internal buffers ------> | |
| 35 |
______________________________________________________________ | |
| 36 |
[ Python's raw memory allocator (PyMem_ API) ] | |
| 37 |
+1 | <----- Python memory (under PyMem manager's control) ------> | | |
| 38 |
__________________________________________________________________ |
| 39 |
[ Underlying general-purpose allocator (ex: C library malloc) ] |
| 40 |
0 | <------ Virtual memory allocated for the python process -------> | |
| 41 |
|
| 42 |
========================================================================= |
| 43 |
_______________________________________________________________________ |
| 44 |
[ OS-specific Virtual Memory Manager (VMM) ] |
| 45 |
-1 | <--- Kernel dynamic storage allocation & management (page-based) ---> | |
| 46 |
__________________________________ __________________________________ |
| 47 |
[ ] [ ] |
| 48 |
-2 | <-- Physical memory: ROM/RAM --> | | <-- Secondary storage (swap) --> | |
| 49 |
|
| 50 |
*/ |
| 51 |
/*==========================================================================*/ |
| 52 |
|
| 53 |
/* A fast, special-purpose memory allocator for small blocks, to be used |
| 54 |
on top of a general-purpose malloc -- heavily based on previous art. */ |
| 55 |
|
| 56 |
/* Vladimir Marangozov -- August 2000 */ |
| 57 |
|
| 58 |
/* |
| 59 |
* "Memory management is where the rubber meets the road -- if we do the wrong |
| 60 |
* thing at any level, the results will not be good. And if we don't make the |
| 61 |
* levels work well together, we are in serious trouble." (1) |
| 62 |
* |
| 63 |
* (1) Paul R. Wilson, Mark S. Johnstone, Michael Neely, and David Boles, |
| 64 |
* "Dynamic Storage Allocation: A Survey and Critical Review", |
| 65 |
* in Proc. 1995 Int'l. Workshop on Memory Management, September 1995. |
| 66 |
*/ |
| 67 |
|
| 68 |
/* #undef WITH_MEMORY_LIMITS */ /* disable mem limit checks */ |
| 69 |
|
| 70 |
/*==========================================================================*/ |
| 71 |
|
| 72 |
/* |
| 73 |
* Allocation strategy abstract: |
| 74 |
* |
| 75 |
* For small requests, the allocator sub-allocates <Big> blocks of memory. |
| 76 |
* Requests greater than SMALL_REQUEST_THRESHOLD bytes are routed to the |
| 77 |
* system's allocator. |
| 78 |
* |
| 79 |
* Small requests are grouped in size classes spaced 8 bytes apart, due |
| 80 |
* to the required valid alignment of the returned address. Requests of |
| 81 |
* a particular size are serviced from memory pools of 4K (one VMM page). |
| 82 |
* Pools are fragmented on demand and contain free lists of blocks of one |
| 83 |
* particular size class. In other words, there is a fixed-size allocator |
| 84 |
* for each size class. Free pools are shared by the different allocators |
| 85 |
* thus minimizing the space reserved for a particular size class. |
| 86 |
* |
| 87 |
* This allocation strategy is a variant of what is known as "simple |
| 88 |
* segregated storage based on array of free lists". The main drawback of |
| 89 |
* simple segregated storage is that we might end up with lot of reserved |
| 90 |
* memory for the different free lists, which degenerate in time. To avoid |
| 91 |
* this, we partition each free list in pools and we share dynamically the |
| 92 |
* reserved space between all free lists. This technique is quite efficient |
| 93 |
* for memory intensive programs which allocate mainly small-sized blocks. |
| 94 |
* |
| 95 |
* For small requests we have the following table: |
| 96 |
* |
| 97 |
* Request in bytes Size of allocated block Size class idx |
| 98 |
* ---------------------------------------------------------------- |
| 99 |
* 1-8 8 0 |
| 100 |
* 9-16 16 1 |
| 101 |
* 17-24 24 2 |
| 102 |
* 25-32 32 3 |
| 103 |
* 33-40 40 4 |
| 104 |
* 41-48 48 5 |
| 105 |
* 49-56 56 6 |
| 106 |
* 57-64 64 7 |
| 107 |
* 65-72 72 8 |
| 108 |
* ... ... ... |
| 109 |
* 497-504 504 62 |
| 110 |
* 505-512 512 63 |
| 111 |
* |
| 112 |
* 0, SMALL_REQUEST_THRESHOLD + 1 and up: routed to the underlying |
| 113 |
* allocator. |
| 114 |
*/ |
| 115 |
|
| 116 |
/*==========================================================================*/ |
| 117 |
|
| 118 |
/* |
| 119 |
* -- Main tunable settings section -- |
| 120 |
*/ |
| 121 |
|
| 122 |
/* |
| 123 |
* Alignment of addresses returned to the user. 8-bytes alignment works |
| 124 |
* on most current architectures (with 32-bit or 64-bit address buses). |
| 125 |
* The alignment value is also used for grouping small requests in size |
| 126 |
* classes spaced ALIGNMENT bytes apart. |
| 127 |
* |
| 128 |
* You shouldn't change this unless you know what you are doing. |
| 129 |
*/ |
| 130 |
|
| 131 |
#if SIZEOF_VOID_P > 4 |
| 132 |
#define ALIGNMENT 16 /* must be 2^N */ |
| 133 |
#define ALIGNMENT_SHIFT 4 |
| 134 |
#else |
| 135 |
#define ALIGNMENT 8 /* must be 2^N */ |
| 136 |
#define ALIGNMENT_SHIFT 3 |
| 137 |
#endif |
| 138 |
|
| 139 |
/* Return the number of bytes in size class I, as a uint. */ |
| 140 |
#define INDEX2SIZE(I) (((pymem_uint)(I) + 1) << ALIGNMENT_SHIFT) |
| 141 |
|
| 142 |
/* |
| 143 |
* Max size threshold below which malloc requests are considered to be |
| 144 |
* small enough in order to use preallocated memory pools. You can tune |
| 145 |
* this value according to your application behaviour and memory needs. |
| 146 |
* |
| 147 |
* Note: a size threshold of 512 guarantees that newly created dictionaries |
| 148 |
* will be allocated from preallocated memory pools on 64-bit. |
| 149 |
* |
| 150 |
* The following invariants must hold: |
| 151 |
* 1) ALIGNMENT <= SMALL_REQUEST_THRESHOLD <= 512 |
| 152 |
* 2) SMALL_REQUEST_THRESHOLD is evenly divisible by ALIGNMENT |
| 153 |
* |
| 154 |
* Although not required, for better performance and space efficiency, |
| 155 |
* it is recommended that SMALL_REQUEST_THRESHOLD is set to a power of 2. |
| 156 |
*/ |
| 157 |
#define SMALL_REQUEST_THRESHOLD 512 |
| 158 |
#define NB_SMALL_SIZE_CLASSES (SMALL_REQUEST_THRESHOLD / ALIGNMENT) |
| 159 |
|
| 160 |
/* |
| 161 |
* The system's VMM page size can be obtained on most unices with a |
| 162 |
* getpagesize() call or deduced from various header files. To make |
| 163 |
* things simpler, we assume that it is 4K, which is OK for most systems. |
| 164 |
* It is probably better if this is the native page size, but it doesn't |
| 165 |
* have to be. In theory, if SYSTEM_PAGE_SIZE is larger than the native page |
| 166 |
* size, then `POOL_ADDR(p)->arenaindex' could rarely cause a segmentation |
| 167 |
* violation fault. 4K is apparently OK for all the platforms that python |
| 168 |
* currently targets. |
| 169 |
*/ |
| 170 |
#define SYSTEM_PAGE_SIZE (4 * 1024) |
| 171 |
|
| 172 |
/* |
| 173 |
* Maximum amount of memory managed by the allocator for small requests. |
| 174 |
*/ |
| 175 |
#ifdef WITH_MEMORY_LIMITS |
| 176 |
#ifndef SMALL_MEMORY_LIMIT |
| 177 |
#define SMALL_MEMORY_LIMIT (64 * 1024 * 1024) /* 64 MB -- more? */ |
| 178 |
#endif |
| 179 |
#endif |
| 180 |
|
| 181 |
#if !defined(WITH_PYMALLOC_RADIX_TREE) |
| 182 |
/* Use radix-tree to track arena memory regions, for address_in_range(). |
| 183 |
* Enable by default since it allows larger pool sizes. Can be disabled |
| 184 |
* using -DWITH_PYMALLOC_RADIX_TREE=0 */ |
| 185 |
#define WITH_PYMALLOC_RADIX_TREE 1 |
| 186 |
#endif |
| 187 |
|
| 188 |
#if SIZEOF_VOID_P > 4 |
| 189 |
/* on 64-bit platforms use larger pools and arenas if we can */ |
| 190 |
#define USE_LARGE_ARENAS |
| 191 |
#if WITH_PYMALLOC_RADIX_TREE |
| 192 |
/* large pools only supported if radix-tree is enabled */ |
| 193 |
#define USE_LARGE_POOLS |
| 194 |
#endif |
| 195 |
#endif |
| 196 |
|
| 197 |
/* |
| 198 |
* The allocator sub-allocates <Big> blocks of memory (called arenas) aligned |
| 199 |
* on a page boundary. This is a reserved virtual address space for the |
| 200 |
* current process (obtained through a malloc()/mmap() call). In no way this |
| 201 |
* means that the memory arenas will be used entirely. A malloc(<Big>) is |
| 202 |
* usually an address range reservation for <Big> bytes, unless all pages within |
| 203 |
* this space are referenced subsequently. So malloc'ing big blocks and not |
| 204 |
* using them does not mean "wasting memory". It's an addressable range |
| 205 |
* wastage... |
| 206 |
* |
| 207 |
* Arenas are allocated with mmap() on systems supporting anonymous memory |
| 208 |
* mappings to reduce heap fragmentation. |
| 209 |
*/ |
| 210 |
#ifdef USE_LARGE_ARENAS |
| 211 |
#define ARENA_BITS 20 /* 1 MiB */ |
| 212 |
#else |
| 213 |
#define ARENA_BITS 18 /* 256 KiB */ |
| 214 |
#endif |
| 215 |
#define ARENA_SIZE (1 << ARENA_BITS) |
| 216 |
#define ARENA_SIZE_MASK (ARENA_SIZE - 1) |
| 217 |
|
| 218 |
#ifdef WITH_MEMORY_LIMITS |
| 219 |
#define MAX_ARENAS (SMALL_MEMORY_LIMIT / ARENA_SIZE) |
| 220 |
#endif |
| 221 |
|
| 222 |
/* |
| 223 |
* Size of the pools used for small blocks. Must be a power of 2. |
| 224 |
*/ |
| 225 |
#ifdef USE_LARGE_POOLS |
| 226 |
#define POOL_BITS 14 /* 16 KiB */ |
| 227 |
#else |
| 228 |
#define POOL_BITS 12 /* 4 KiB */ |
| 229 |
#endif |
| 230 |
#define POOL_SIZE (1 << POOL_BITS) |
| 231 |
#define POOL_SIZE_MASK (POOL_SIZE - 1) |
| 232 |
|
| 233 |
#if !WITH_PYMALLOC_RADIX_TREE |
| 234 |
#if POOL_SIZE != SYSTEM_PAGE_SIZE |
| 235 |
# error "pool size must be equal to system page size" |
| 236 |
#endif |
| 237 |
#endif |
| 238 |
|
| 239 |
#define MAX_POOLS_IN_ARENA (ARENA_SIZE / POOL_SIZE) |
| 240 |
#if MAX_POOLS_IN_ARENA * POOL_SIZE != ARENA_SIZE |
| 241 |
# error "arena size not an exact multiple of pool size" |
| 242 |
#endif |
| 243 |
|
| 244 |
/* |
| 245 |
* -- End of tunable settings section -- |
| 246 |
*/ |
| 247 |
|
| 248 |
/*==========================================================================*/ |
| 249 |
|
| 250 |
/* When you say memory, my mind reasons in terms of (pointers to) blocks */ |
| 251 |
typedef uint8_t pymem_block; |
| 252 |
|
| 253 |
/* Pool for small blocks. */ |
| 254 |
struct pool_header { |
| 255 |
union { pymem_block *_padding; |
| 256 |
uint count; } ref; /* number of allocated blocks */ |
| 257 |
pymem_block *freeblock; /* pool's free list head */ |
| 258 |
struct pool_header *nextpool; /* see "Pool table" for meaning */ |
| 259 |
struct pool_header *prevpool; /* " */ |
| 260 |
uint arenaindex; /* index into arenas of base adr */ |
| 261 |
uint szidx; /* block size class index */ |
| 262 |
uint nextoffset; /* bytes to virgin block */ |
| 263 |
uint maxnextoffset; /* largest valid nextoffset */ |
| 264 |
}; |
| 265 |
|
| 266 |
typedef struct pool_header *poolp; |
| 267 |
|
| 268 |
/* Record keeping for arenas. */ |
| 269 |
struct arena_object { |
| 270 |
/* The address of the arena, as returned by malloc. Note that 0 |
| 271 |
* will never be returned by a successful malloc, and is used |
| 272 |
* here to mark an arena_object that doesn't correspond to an |
| 273 |
* allocated arena. |
| 274 |
*/ |
| 275 |
uintptr_t address; |
| 276 |
|
| 277 |
/* Pool-aligned pointer to the next pool to be carved off. */ |
| 278 |
pymem_block* pool_address; |
| 279 |
|
| 280 |
/* The number of available pools in the arena: free pools + never- |
| 281 |
* allocated pools. |
| 282 |
*/ |
| 283 |
uint nfreepools; |
| 284 |
|
| 285 |
/* The total number of pools in the arena, whether or not available. */ |
| 286 |
uint ntotalpools; |
| 287 |
|
| 288 |
/* Singly-linked list of available pools. */ |
| 289 |
struct pool_header* freepools; |
| 290 |
|
| 291 |
/* Whenever this arena_object is not associated with an allocated |
| 292 |
* arena, the nextarena member is used to link all unassociated |
| 293 |
* arena_objects in the singly-linked `unused_arena_objects` list. |
| 294 |
* The prevarena member is unused in this case. |
| 295 |
* |
| 296 |
* When this arena_object is associated with an allocated arena |
| 297 |
* with at least one available pool, both members are used in the |
| 298 |
* doubly-linked `usable_arenas` list, which is maintained in |
| 299 |
* increasing order of `nfreepools` values. |
| 300 |
* |
| 301 |
* Else this arena_object is associated with an allocated arena |
| 302 |
* all of whose pools are in use. `nextarena` and `prevarena` |
| 303 |
* are both meaningless in this case. |
| 304 |
*/ |
| 305 |
struct arena_object* nextarena; |
| 306 |
struct arena_object* prevarena; |
| 307 |
}; |
| 308 |
|
| 309 |
#define POOL_OVERHEAD _Py_SIZE_ROUND_UP(sizeof(struct pool_header), ALIGNMENT) |
| 310 |
|
| 311 |
#define DUMMY_SIZE_IDX 0xffff /* size class of newly cached pools */ |
| 312 |
|
| 313 |
/* Round pointer P down to the closest pool-aligned address <= P, as a poolp */ |
| 314 |
#define POOL_ADDR(P) ((poolp)_Py_ALIGN_DOWN((P), POOL_SIZE)) |
| 315 |
|
| 316 |
/* Return total number of blocks in pool of size index I, as a uint. */ |
| 317 |
#define NUMBLOCKS(I) ((pymem_uint)(POOL_SIZE - POOL_OVERHEAD) / INDEX2SIZE(I)) |
| 318 |
|
| 319 |
/*==========================================================================*/ |
| 320 |
|
| 321 |
/* |
| 322 |
* Pool table -- headed, circular, doubly-linked lists of partially used pools. |
| 323 |
|
| 324 |
This is involved. For an index i, usedpools[i+i] is the header for a list of |
| 325 |
all partially used pools holding small blocks with "size class idx" i. So |
| 326 |
usedpools[0] corresponds to blocks of size 8, usedpools[2] to blocks of size |
| 327 |
16, and so on: index 2*i <-> blocks of size (i+1)<<ALIGNMENT_SHIFT. |
| 328 |
|
| 329 |
Pools are carved off an arena's highwater mark (an arena_object's pool_address |
| 330 |
member) as needed. Once carved off, a pool is in one of three states forever |
| 331 |
after: |
| 332 |
|
| 333 |
used == partially used, neither empty nor full |
| 334 |
At least one block in the pool is currently allocated, and at least one |
| 335 |
block in the pool is not currently allocated (note this implies a pool |
| 336 |
has room for at least two blocks). |
| 337 |
This is a pool's initial state, as a pool is created only when malloc |
| 338 |
needs space. |
| 339 |
The pool holds blocks of a fixed size, and is in the circular list headed |
| 340 |
at usedpools[i] (see above). It's linked to the other used pools of the |
| 341 |
same size class via the pool_header's nextpool and prevpool members. |
| 342 |
If all but one block is currently allocated, a malloc can cause a |
| 343 |
transition to the full state. If all but one block is not currently |
| 344 |
allocated, a free can cause a transition to the empty state. |
| 345 |
|
| 346 |
full == all the pool's blocks are currently allocated |
| 347 |
On transition to full, a pool is unlinked from its usedpools[] list. |
| 348 |
It's not linked to from anything then anymore, and its nextpool and |
| 349 |
prevpool members are meaningless until it transitions back to used. |
| 350 |
A free of a block in a full pool puts the pool back in the used state. |
| 351 |
Then it's linked in at the front of the appropriate usedpools[] list, so |
| 352 |
that the next allocation for its size class will reuse the freed block. |
| 353 |
|
| 354 |
empty == all the pool's blocks are currently available for allocation |
| 355 |
On transition to empty, a pool is unlinked from its usedpools[] list, |
| 356 |
and linked to the front of its arena_object's singly-linked freepools list, |
| 357 |
via its nextpool member. The prevpool member has no meaning in this case. |
| 358 |
Empty pools have no inherent size class: the next time a malloc finds |
| 359 |
an empty list in usedpools[], it takes the first pool off of freepools. |
| 360 |
If the size class needed happens to be the same as the size class the pool |
| 361 |
last had, some pool initialization can be skipped. |
| 362 |
|
| 363 |
|
| 364 |
Block Management |
| 365 |
|
| 366 |
Blocks within pools are again carved out as needed. pool->freeblock points to |
| 367 |
the start of a singly-linked list of free blocks within the pool. When a |
| 368 |
block is freed, it's inserted at the front of its pool's freeblock list. Note |
| 369 |
that the available blocks in a pool are *not* linked all together when a pool |
| 370 |
is initialized. Instead only "the first two" (lowest addresses) blocks are |
| 371 |
set up, returning the first such block, and setting pool->freeblock to a |
| 372 |
one-block list holding the second such block. This is consistent with that |
| 373 |
pymalloc strives at all levels (arena, pool, and block) never to touch a piece |
| 374 |
of memory until it's actually needed. |
| 375 |
|
| 376 |
So long as a pool is in the used state, we're certain there *is* a block |
| 377 |
available for allocating, and pool->freeblock is not NULL. If pool->freeblock |
| 378 |
points to the end of the free list before we've carved the entire pool into |
| 379 |
blocks, that means we simply haven't yet gotten to one of the higher-address |
| 380 |
blocks. The offset from the pool_header to the start of "the next" virgin |
| 381 |
block is stored in the pool_header nextoffset member, and the largest value |
| 382 |
of nextoffset that makes sense is stored in the maxnextoffset member when a |
| 383 |
pool is initialized. All the blocks in a pool have been passed out at least |
| 384 |
once when and only when nextoffset > maxnextoffset. |
| 385 |
|
| 386 |
|
| 387 |
Major obscurity: While the usedpools vector is declared to have poolp |
| 388 |
entries, it doesn't really. It really contains two pointers per (conceptual) |
| 389 |
poolp entry, the nextpool and prevpool members of a pool_header. The |
| 390 |
excruciating initialization code below fools C so that |
| 391 |
|
| 392 |
usedpool[i+i] |
| 393 |
|
| 394 |
"acts like" a genuine poolp, but only so long as you only reference its |
| 395 |
nextpool and prevpool members. The "- 2*sizeof(pymem_block *)" gibberish is |
| 396 |
compensating for that a pool_header's nextpool and prevpool members |
| 397 |
immediately follow a pool_header's first two members: |
| 398 |
|
| 399 |
union { pymem_block *_padding; |
| 400 |
uint count; } ref; |
| 401 |
pymem_block *freeblock; |
| 402 |
|
| 403 |
each of which consume sizeof(pymem_block *) bytes. So what usedpools[i+i] really |
| 404 |
contains is a fudged-up pointer p such that *if* C believes it's a poolp |
| 405 |
pointer, then p->nextpool and p->prevpool are both p (meaning that the headed |
| 406 |
circular list is empty). |
| 407 |
|
| 408 |
It's unclear why the usedpools setup is so convoluted. It could be to |
| 409 |
minimize the amount of cache required to hold this heavily-referenced table |
| 410 |
(which only *needs* the two interpool pointer members of a pool_header). OTOH, |
| 411 |
referencing code has to remember to "double the index" and doing so isn't |
| 412 |
free, usedpools[0] isn't a strictly legal pointer, and we're crucially relying |
| 413 |
on that C doesn't insert any padding anywhere in a pool_header at or before |
| 414 |
the prevpool member. |
| 415 |
**************************************************************************** */ |
| 416 |
|
| 417 |
#define OBMALLOC_USED_POOLS_SIZE (2 * ((NB_SMALL_SIZE_CLASSES + 7) / 8) * 8) |
| 418 |
|
| 419 |
struct _obmalloc_pools { |
| 420 |
poolp used[OBMALLOC_USED_POOLS_SIZE]; |
| 421 |
}; |
| 422 |
|
| 423 |
|
| 424 |
/*========================================================================== |
| 425 |
Arena management. |
| 426 |
|
| 427 |
`arenas` is a vector of arena_objects. It contains maxarenas entries, some of |
| 428 |
which may not be currently used (== they're arena_objects that aren't |
| 429 |
currently associated with an allocated arena). Note that arenas proper are |
| 430 |
separately malloc'ed. |
| 431 |
|
| 432 |
Prior to Python 2.5, arenas were never free()'ed. Starting with Python 2.5, |
| 433 |
we do try to free() arenas, and use some mild heuristic strategies to increase |
| 434 |
the likelihood that arenas eventually can be freed. |
| 435 |
|
| 436 |
unused_arena_objects |
| 437 |
|
| 438 |
This is a singly-linked list of the arena_objects that are currently not |
| 439 |
being used (no arena is associated with them). Objects are taken off the |
| 440 |
head of the list in new_arena(), and are pushed on the head of the list in |
| 441 |
PyObject_Free() when the arena is empty. Key invariant: an arena_object |
| 442 |
is on this list if and only if its .address member is 0. |
| 443 |
|
| 444 |
usable_arenas |
| 445 |
|
| 446 |
This is a doubly-linked list of the arena_objects associated with arenas |
| 447 |
that have pools available. These pools are either waiting to be reused, |
| 448 |
or have not been used before. The list is sorted to have the most- |
| 449 |
allocated arenas first (ascending order based on the nfreepools member). |
| 450 |
This means that the next allocation will come from a heavily used arena, |
| 451 |
which gives the nearly empty arenas a chance to be returned to the system. |
| 452 |
In my unscientific tests this dramatically improved the number of arenas |
| 453 |
that could be freed. |
| 454 |
|
| 455 |
Note that an arena_object associated with an arena all of whose pools are |
| 456 |
currently in use isn't on either list. |
| 457 |
|
| 458 |
Changed in Python 3.8: keeping usable_arenas sorted by number of free pools |
| 459 |
used to be done by one-at-a-time linear search when an arena's number of |
| 460 |
free pools changed. That could, overall, consume time quadratic in the |
| 461 |
number of arenas. That didn't really matter when there were only a few |
| 462 |
hundred arenas (typical!), but could be a timing disaster when there were |
| 463 |
hundreds of thousands. See bpo-37029. |
| 464 |
|
| 465 |
Now we have a vector of "search fingers" to eliminate the need to search: |
| 466 |
nfp2lasta[nfp] returns the last ("rightmost") arena in usable_arenas |
| 467 |
with nfp free pools. This is NULL if and only if there is no arena with |
| 468 |
nfp free pools in usable_arenas. |
| 469 |
*/ |
| 470 |
|
| 471 |
/* How many arena_objects do we initially allocate? |
| 472 |
* 16 = can allocate 16 arenas = 16 * ARENA_SIZE = 4MB before growing the |
| 473 |
* `arenas` vector. |
| 474 |
*/ |
| 475 |
#define INITIAL_ARENA_OBJECTS 16 |
| 476 |
|
| 477 |
struct _obmalloc_mgmt { |
| 478 |
/* Array of objects used to track chunks of memory (arenas). */ |
| 479 |
struct arena_object* arenas; |
| 480 |
/* Number of slots currently allocated in the `arenas` vector. */ |
| 481 |
uint maxarenas; |
| 482 |
|
| 483 |
/* The head of the singly-linked, NULL-terminated list of available |
| 484 |
* arena_objects. |
| 485 |
*/ |
| 486 |
struct arena_object* unused_arena_objects; |
| 487 |
|
| 488 |
/* The head of the doubly-linked, NULL-terminated at each end, list of |
| 489 |
* arena_objects associated with arenas that have pools available. |
| 490 |
*/ |
| 491 |
struct arena_object* usable_arenas; |
| 492 |
|
| 493 |
/* nfp2lasta[nfp] is the last arena in usable_arenas with nfp free pools */ |
| 494 |
struct arena_object* nfp2lasta[MAX_POOLS_IN_ARENA + 1]; |
| 495 |
|
| 496 |
/* Number of arenas allocated that haven't been free()'d. */ |
| 497 |
size_t narenas_currently_allocated; |
| 498 |
|
| 499 |
/* Total number of times malloc() called to allocate an arena. */ |
| 500 |
size_t ntimes_arena_allocated; |
| 501 |
/* High water mark (max value ever seen) for narenas_currently_allocated. */ |
| 502 |
size_t narenas_highwater; |
| 503 |
|
| 504 |
Py_ssize_t raw_allocated_blocks; |
| 505 |
}; |
| 506 |
|
| 507 |
|
| 508 |
#if WITH_PYMALLOC_RADIX_TREE |
| 509 |
/*==========================================================================*/ |
| 510 |
/* radix tree for tracking arena usage. If enabled, used to implement |
| 511 |
address_in_range(). |
| 512 |
|
| 513 |
memory address bit allocation for keys |
| 514 |
|
| 515 |
64-bit pointers, IGNORE_BITS=0 and 2^20 arena size: |
| 516 |
15 -> MAP_TOP_BITS |
| 517 |
15 -> MAP_MID_BITS |
| 518 |
14 -> MAP_BOT_BITS |
| 519 |
20 -> ideal aligned arena |
| 520 |
---- |
| 521 |
64 |
| 522 |
|
| 523 |
64-bit pointers, IGNORE_BITS=16, and 2^20 arena size: |
| 524 |
16 -> IGNORE_BITS |
| 525 |
10 -> MAP_TOP_BITS |
| 526 |
10 -> MAP_MID_BITS |
| 527 |
8 -> MAP_BOT_BITS |
| 528 |
20 -> ideal aligned arena |
| 529 |
---- |
| 530 |
64 |
| 531 |
|
| 532 |
32-bit pointers and 2^18 arena size: |
| 533 |
14 -> MAP_BOT_BITS |
| 534 |
18 -> ideal aligned arena |
| 535 |
---- |
| 536 |
32 |
| 537 |
|
| 538 |
*/ |
| 539 |
|
| 540 |
#if SIZEOF_VOID_P == 8 |
| 541 |
|
| 542 |
/* number of bits in a pointer */ |
| 543 |
#define POINTER_BITS 64 |
| 544 |
|
| 545 |
/* High bits of memory addresses that will be ignored when indexing into the |
| 546 |
* radix tree. Setting this to zero is the safe default. For most 64-bit |
| 547 |
* machines, setting this to 16 would be safe. The kernel would not give |
| 548 |
* user-space virtual memory addresses that have significant information in |
| 549 |
* those high bits. The main advantage to setting IGNORE_BITS > 0 is that less |
| 550 |
* virtual memory will be used for the top and middle radix tree arrays. Those |
| 551 |
* arrays are allocated in the BSS segment and so will typically consume real |
| 552 |
* memory only if actually accessed. |
| 553 |
*/ |
| 554 |
#define IGNORE_BITS 0 |
| 555 |
|
| 556 |
/* use the top and mid layers of the radix tree */ |
| 557 |
#define USE_INTERIOR_NODES |
| 558 |
|
| 559 |
#elif SIZEOF_VOID_P == 4 |
| 560 |
|
| 561 |
#define POINTER_BITS 32 |
| 562 |
#define IGNORE_BITS 0 |
| 563 |
|
| 564 |
#else |
| 565 |
|
| 566 |
/* Currently this code works for 64-bit or 32-bit pointers only. */ |
| 567 |
#error "obmalloc radix tree requires 64-bit or 32-bit pointers." |
| 568 |
|
| 569 |
#endif /* SIZEOF_VOID_P */ |
| 570 |
|
| 571 |
/* arena_coverage_t members require this to be true */ |
| 572 |
#if ARENA_BITS >= 32 |
| 573 |
# error "arena size must be < 2^32" |
| 574 |
#endif |
| 575 |
|
| 576 |
/* the lower bits of the address that are not ignored */ |
| 577 |
#define ADDRESS_BITS (POINTER_BITS - IGNORE_BITS) |
| 578 |
|
| 579 |
#ifdef USE_INTERIOR_NODES |
| 580 |
/* number of bits used for MAP_TOP and MAP_MID nodes */ |
| 581 |
#define INTERIOR_BITS ((ADDRESS_BITS - ARENA_BITS + 2) / 3) |
| 582 |
#else |
| 583 |
#define INTERIOR_BITS 0 |
| 584 |
#endif |
| 585 |
|
| 586 |
#define MAP_TOP_BITS INTERIOR_BITS |
| 587 |
#define MAP_TOP_LENGTH (1 << MAP_TOP_BITS) |
| 588 |
#define MAP_TOP_MASK (MAP_TOP_LENGTH - 1) |
| 589 |
|
| 590 |
#define MAP_MID_BITS INTERIOR_BITS |
| 591 |
#define MAP_MID_LENGTH (1 << MAP_MID_BITS) |
| 592 |
#define MAP_MID_MASK (MAP_MID_LENGTH - 1) |
| 593 |
|
| 594 |
#define MAP_BOT_BITS (ADDRESS_BITS - ARENA_BITS - 2*INTERIOR_BITS) |
| 595 |
#define MAP_BOT_LENGTH (1 << MAP_BOT_BITS) |
| 596 |
#define MAP_BOT_MASK (MAP_BOT_LENGTH - 1) |
| 597 |
|
| 598 |
#define MAP_BOT_SHIFT ARENA_BITS |
| 599 |
#define MAP_MID_SHIFT (MAP_BOT_BITS + MAP_BOT_SHIFT) |
| 600 |
#define MAP_TOP_SHIFT (MAP_MID_BITS + MAP_MID_SHIFT) |
| 601 |
|
| 602 |
#define AS_UINT(p) ((uintptr_t)(p)) |
| 603 |
#define MAP_BOT_INDEX(p) ((AS_UINT(p) >> MAP_BOT_SHIFT) & MAP_BOT_MASK) |
| 604 |
#define MAP_MID_INDEX(p) ((AS_UINT(p) >> MAP_MID_SHIFT) & MAP_MID_MASK) |
| 605 |
#define MAP_TOP_INDEX(p) ((AS_UINT(p) >> MAP_TOP_SHIFT) & MAP_TOP_MASK) |
| 606 |
|
| 607 |
#if IGNORE_BITS > 0 |
| 608 |
/* Return the ignored part of the pointer address. Those bits should be same |
| 609 |
* for all valid pointers if IGNORE_BITS is set correctly. |
| 610 |
*/ |
| 611 |
#define HIGH_BITS(p) (AS_UINT(p) >> ADDRESS_BITS) |
| 612 |
#else |
| 613 |
#define HIGH_BITS(p) 0 |
| 614 |
#endif |
| 615 |
|
| 616 |
|
| 617 |
/* This is the leaf of the radix tree. See arena_map_mark_used() for the |
| 618 |
* meaning of these members. */ |
| 619 |
typedef struct { |
| 620 |
int32_t tail_hi; |
| 621 |
int32_t tail_lo; |
| 622 |
} arena_coverage_t; |
| 623 |
|
| 624 |
typedef struct arena_map_bot { |
| 625 |
/* The members tail_hi and tail_lo are accessed together. So, it |
| 626 |
* better to have them as an array of structs, rather than two |
| 627 |
* arrays. |
| 628 |
*/ |
| 629 |
arena_coverage_t arenas[MAP_BOT_LENGTH]; |
| 630 |
} arena_map_bot_t; |
| 631 |
|
| 632 |
#ifdef USE_INTERIOR_NODES |
| 633 |
typedef struct arena_map_mid { |
| 634 |
struct arena_map_bot *ptrs[MAP_MID_LENGTH]; |
| 635 |
} arena_map_mid_t; |
| 636 |
|
| 637 |
typedef struct arena_map_top { |
| 638 |
struct arena_map_mid *ptrs[MAP_TOP_LENGTH]; |
| 639 |
} arena_map_top_t; |
| 640 |
#endif |
| 641 |
|
| 642 |
struct _obmalloc_usage { |
| 643 |
/* The root of radix tree. Note that by initializing like this, the memory |
| 644 |
* should be in the BSS. The OS will only memory map pages as the MAP_MID |
| 645 |
* nodes get used (OS pages are demand loaded as needed). |
| 646 |
*/ |
| 647 |
#ifdef USE_INTERIOR_NODES |
| 648 |
arena_map_top_t arena_map_root; |
| 649 |
/* accounting for number of used interior nodes */ |
| 650 |
int arena_map_mid_count; |
| 651 |
int arena_map_bot_count; |
| 652 |
#else |
| 653 |
arena_map_bot_t arena_map_root; |
| 654 |
#endif |
| 655 |
}; |
| 656 |
|
| 657 |
#endif /* WITH_PYMALLOC_RADIX_TREE */ |
| 658 |
|
| 659 |
|
| 660 |
struct _obmalloc_global_state { |
| 661 |
int dump_debug_stats; |
| 662 |
Py_ssize_t interpreter_leaks; |
| 663 |
}; |
| 664 |
|
| 665 |
struct _obmalloc_state { |
| 666 |
struct _obmalloc_pools pools; |
| 667 |
struct _obmalloc_mgmt mgmt; |
| 668 |
#if WITH_PYMALLOC_RADIX_TREE |
| 669 |
struct _obmalloc_usage usage; |
| 670 |
#endif |
| 671 |
}; |
| 672 |
|
| 673 |
|
| 674 |
#undef uint |
| 675 |
|
| 676 |
|
| 677 |
/* Allocate memory directly from the O/S virtual memory system, |
| 678 |
* where supported. Otherwise fallback on malloc */ |
| 679 |
void *_PyObject_VirtualAlloc(size_t size); |
| 680 |
void _PyObject_VirtualFree(void *, size_t size); |
| 681 |
|
| 682 |
|
| 683 |
/* This function returns the number of allocated memory blocks, regardless of size */ |
| 684 |
extern Py_ssize_t _Py_GetGlobalAllocatedBlocks(void); |
| 685 |
#define _Py_GetAllocatedBlocks() \ |
| 686 |
_Py_GetGlobalAllocatedBlocks() |
| 687 |
extern Py_ssize_t _PyInterpreterState_GetAllocatedBlocks(PyInterpreterState *); |
| 688 |
extern void _PyInterpreterState_FinalizeAllocatedBlocks(PyInterpreterState *); |
| 689 |
extern int _PyMem_init_obmalloc(PyInterpreterState *interp); |
| 690 |
extern bool _PyMem_obmalloc_state_on_heap(PyInterpreterState *interp); |
| 691 |
|
| 692 |
|
| 693 |
#ifdef WITH_PYMALLOC |
| 694 |
// Export the symbol for the 3rd party 'guppy3' project |
| 695 |
PyAPI_FUNC(int) _PyObject_DebugMallocStats(FILE *out); |
| 696 |
#endif |
| 697 |
|
| 698 |
|
| 699 |
#ifdef __cplusplus |
| 700 |
} |
| 701 |
#endif |
| 702 |
#endif // !Py_INTERNAL_OBMALLOC_H |