1 #include "Python.h" 2 3 #if defined(__has_feature) /* Clang */ 4 #if __has_feature(address_sanitizer) /* is ASAN enabled? */ 5 #define ATTRIBUTE_NO_ADDRESS_SAFETY_ANALYSIS \ 6 __attribute__((no_address_safety_analysis)) \ 7 __attribute__ ((noinline)) 8 #else 9 #define ATTRIBUTE_NO_ADDRESS_SAFETY_ANALYSIS 10 #endif 11 #else 12 #if defined(__SANITIZE_ADDRESS__) /* GCC 4.8.x, is ASAN enabled? */ 13 #define ATTRIBUTE_NO_ADDRESS_SAFETY_ANALYSIS \ 14 __attribute__((no_address_safety_analysis)) \ 15 __attribute__ ((noinline)) 16 #else 17 #define ATTRIBUTE_NO_ADDRESS_SAFETY_ANALYSIS 18 #endif 19 #endif 20 21 #ifdef WITH_PYMALLOC 22 23 #ifdef HAVE_MMAP 24 #include <sys/mman.h> 25 #ifdef MAP_ANONYMOUS 26 #define ARENAS_USE_MMAP 27 #endif 28 #endif 29 30 #ifdef WITH_VALGRIND 31 #include <valgrind/valgrind.h> 32 33 /* If we're using GCC, use __builtin_expect() to reduce overhead of 34 the valgrind checks */ 35 #if defined(__GNUC__) && (__GNUC__ > 2) && defined(__OPTIMIZE__) 36 # define UNLIKELY(value) __builtin_expect((value), 0) 37 #else 38 # define UNLIKELY(value) (value) 39 #endif 40 41 /* -1 indicates that we haven't checked that we're running on valgrind yet. */ 42 static int running_on_valgrind = -1; 43 #endif 44 45 /* An object allocator for Python. 46 47 Here is an introduction to the layers of the Python memory architecture, 48 showing where the object allocator is actually used (layer +2), It is 49 called for every object allocation and deallocation (PyObject_New/Del), 50 unless the object-specific allocators implement a proprietary allocation 51 scheme (ex.: ints use a simple free list). This is also the place where 52 the cyclic garbage collector operates selectively on container objects. 53 54 55 Object-specific allocators 56 _____ ______ ______ ________ 57 [ int ] [ dict ] [ list ] ... [ string ] Python core | 58 +3 | <----- Object-specific memory -----> | <-- Non-object memory --> | 59 _______________________________ | | 60 [ Python's object allocator ] | | 61 +2 | ####### Object memory ####### | <------ Internal buffers ------> | 62 ______________________________________________________________ | 63 [ Python's raw memory allocator (PyMem_ API) ] | 64 +1 | <----- Python memory (under PyMem manager's control) ------> | | 65 __________________________________________________________________ 66 [ Underlying general-purpose allocator (ex: C library malloc) ] 67 0 | <------ Virtual memory allocated for the python process -------> | 68 69 ========================================================================= 70 _______________________________________________________________________ 71 [ OS-specific Virtual Memory Manager (VMM) ] 72 -1 | <--- Kernel dynamic storage allocation & management (page-based) ---> | 73 __________________________________ __________________________________ 74 [ ] [ ] 75 -2 | <-- Physical memory: ROM/RAM --> | | <-- Secondary storage (swap) --> | 76 77 */ 78 /*==========================================================================*/ 79 80 /* A fast, special-purpose memory allocator for small blocks, to be used 81 on top of a general-purpose malloc -- heavily based on previous art. */ 82 83 /* Vladimir Marangozov -- August 2000 */ 84 85 /* 86 * "Memory management is where the rubber meets the road -- if we do the wrong 87 * thing at any level, the results will not be good. And if we don't make the 88 * levels work well together, we are in serious trouble." (1) 89 * 90 * (1) Paul R. Wilson, Mark S. Johnstone, Michael Neely, and David Boles, 91 * "Dynamic Storage Allocation: A Survey and Critical Review", 92 * in Proc. 1995 Int'l. Workshop on Memory Management, September 1995. 93 */ 94 95 /* #undef WITH_MEMORY_LIMITS */ /* disable mem limit checks */ 96 97 /*==========================================================================*/ 98 99 /* 100 * Allocation strategy abstract: 101 * 102 * For small requests, the allocator sub-allocates <Big> blocks of memory. 103 * Requests greater than SMALL_REQUEST_THRESHOLD bytes are routed to the 104 * system's allocator. 105 * 106 * Small requests are grouped in size classes spaced 8 bytes apart, due 107 * to the required valid alignment of the returned address. Requests of 108 * a particular size are serviced from memory pools of 4K (one VMM page). 109 * Pools are fragmented on demand and contain free lists of blocks of one 110 * particular size class. In other words, there is a fixed-size allocator 111 * for each size class. Free pools are shared by the different allocators 112 * thus minimizing the space reserved for a particular size class. 113 * 114 * This allocation strategy is a variant of what is known as "simple 115 * segregated storage based on array of free lists". The main drawback of 116 * simple segregated storage is that we might end up with lot of reserved 117 * memory for the different free lists, which degenerate in time. To avoid 118 * this, we partition each free list in pools and we share dynamically the 119 * reserved space between all free lists. This technique is quite efficient 120 * for memory intensive programs which allocate mainly small-sized blocks. 121 * 122 * For small requests we have the following table: 123 * 124 * Request in bytes Size of allocated block Size class idx 125 * ---------------------------------------------------------------- 126 * 1-8 8 0 127 * 9-16 16 1 128 * 17-24 24 2 129 * 25-32 32 3 130 * 33-40 40 4 131 * 41-48 48 5 132 * 49-56 56 6 133 * 57-64 64 7 134 * 65-72 72 8 135 * ... ... ... 136 * 497-504 504 62 137 * 505-512 512 63 138 * 139 * 0, SMALL_REQUEST_THRESHOLD + 1 and up: routed to the underlying 140 * allocator. 141 */ 142 143 /*==========================================================================*/ 144 145 /* 146 * -- Main tunable settings section -- 147 */ 148 149 /* 150 * Alignment of addresses returned to the user. 8-bytes alignment works 151 * on most current architectures (with 32-bit or 64-bit address busses). 152 * The alignment value is also used for grouping small requests in size 153 * classes spaced ALIGNMENT bytes apart. 154 * 155 * You shouldn't change this unless you know what you are doing. 156 */ 157 #define ALIGNMENT 8 /* must be 2^N */ 158 #define ALIGNMENT_SHIFT 3 159 #define ALIGNMENT_MASK (ALIGNMENT - 1) 160 161 /* Return the number of bytes in size class I, as a uint. */ 162 #define INDEX2SIZE(I) (((uint)(I) + 1) << ALIGNMENT_SHIFT) 163 164 /* 165 * Max size threshold below which malloc requests are considered to be 166 * small enough in order to use preallocated memory pools. You can tune 167 * this value according to your application behaviour and memory needs. 168 * 169 * The following invariants must hold: 170 * 1) ALIGNMENT <= SMALL_REQUEST_THRESHOLD <= 256 171 * 2) SMALL_REQUEST_THRESHOLD is evenly divisible by ALIGNMENT 172 * 173 * Note: a size threshold of 512 guarantees that newly created dictionaries 174 * will be allocated from preallocated memory pools on 64-bit. 175 * 176 * Although not required, for better performance and space efficiency, 177 * it is recommended that SMALL_REQUEST_THRESHOLD is set to a power of 2. 178 */ 179 #define SMALL_REQUEST_THRESHOLD 512 180 #define NB_SMALL_SIZE_CLASSES (SMALL_REQUEST_THRESHOLD / ALIGNMENT) 181 182 /* 183 * The system's VMM page size can be obtained on most unices with a 184 * getpagesize() call or deduced from various header files. To make 185 * things simpler, we assume that it is 4K, which is OK for most systems. 186 * It is probably better if this is the native page size, but it doesn't 187 * have to be. In theory, if SYSTEM_PAGE_SIZE is larger than the native page 188 * size, then `POOL_ADDR(p)->arenaindex' could rarely cause a segmentation 189 * violation fault. 4K is apparently OK for all the platforms that python 190 * currently targets. 191 */ 192 #define SYSTEM_PAGE_SIZE (4 * 1024) 193 #define SYSTEM_PAGE_SIZE_MASK (SYSTEM_PAGE_SIZE - 1) 194 195 /* 196 * Maximum amount of memory managed by the allocator for small requests. 197 */ 198 #ifdef WITH_MEMORY_LIMITS 199 #ifndef SMALL_MEMORY_LIMIT 200 #define SMALL_MEMORY_LIMIT (64 * 1024 * 1024) /* 64 MB -- more? */ 201 #endif 202 #endif 203 204 /* 205 * The allocator sub-allocates <Big> blocks of memory (called arenas) aligned 206 * on a page boundary. This is a reserved virtual address space for the 207 * current process (obtained through a malloc()/mmap() call). In no way this 208 * means that the memory arenas will be used entirely. A malloc(<Big>) is 209 * usually an address range reservation for <Big> bytes, unless all pages within 210 * this space are referenced subsequently. So malloc'ing big blocks and not 211 * using them does not mean "wasting memory". It's an addressable range 212 * wastage... 213 * 214 * Arenas are allocated with mmap() on systems supporting anonymous memory 215 * mappings to reduce heap fragmentation. 216 */ 217 #define ARENA_SIZE (256 << 10) /* 256KB */ 218 219 #ifdef WITH_MEMORY_LIMITS 220 #define MAX_ARENAS (SMALL_MEMORY_LIMIT / ARENA_SIZE) 221 #endif 222 223 /* 224 * Size of the pools used for small blocks. Should be a power of 2, 225 * between 1K and SYSTEM_PAGE_SIZE, that is: 1k, 2k, 4k. 226 */ 227 #define POOL_SIZE SYSTEM_PAGE_SIZE /* must be 2^N */ 228 #define POOL_SIZE_MASK SYSTEM_PAGE_SIZE_MASK 229 230 /* 231 * -- End of tunable settings section -- 232 */ 233 234 /*==========================================================================*/ 235 236 /* 237 * Locking 238 * 239 * To reduce lock contention, it would probably be better to refine the 240 * crude function locking with per size class locking. I'm not positive 241 * however, whether it's worth switching to such locking policy because 242 * of the performance penalty it might introduce. 243 * 244 * The following macros describe the simplest (should also be the fastest) 245 * lock object on a particular platform and the init/fini/lock/unlock 246 * operations on it. The locks defined here are not expected to be recursive 247 * because it is assumed that they will always be called in the order: 248 * INIT, [LOCK, UNLOCK]*, FINI. 249 */ 250 251 /* 252 * Python's threads are serialized, so object malloc locking is disabled. 253 */ 254 #define SIMPLELOCK_DECL(lock) /* simple lock declaration */ 255 #define SIMPLELOCK_INIT(lock) /* allocate (if needed) and initialize */ 256 #define SIMPLELOCK_FINI(lock) /* free/destroy an existing lock */ 257 #define SIMPLELOCK_LOCK(lock) /* acquire released lock */ 258 #define SIMPLELOCK_UNLOCK(lock) /* release acquired lock */ 259 260 /* 261 * Basic types 262 * I don't care if these are defined in <sys/types.h> or elsewhere. Axiom. 263 */ 264 #undef uchar 265 #define uchar unsigned char /* assuming == 8 bits */ 266 267 #undef uint 268 #define uint unsigned int /* assuming >= 16 bits */ 269 270 #undef ulong 271 #define ulong unsigned long /* assuming >= 32 bits */ 272 273 #undef uptr 274 #define uptr Py_uintptr_t 275 276 /* When you say memory, my mind reasons in terms of (pointers to) blocks */ 277 typedef uchar block; 278 279 /* Pool for small blocks. */ 280 struct pool_header { 281 union { block *_padding; 282 uint count; } ref; /* number of allocated blocks */ 283 block *freeblock; /* pool's free list head */ 284 struct pool_header *nextpool; /* next pool of this size class */ 285 struct pool_header *prevpool; /* previous pool "" */ 286 uint arenaindex; /* index into arenas of base adr */ 287 uint szidx; /* block size class index */ 288 uint nextoffset; /* bytes to virgin block */ 289 uint maxnextoffset; /* largest valid nextoffset */ 290 }; 291 292 typedef struct pool_header *poolp; 293 294 /* Record keeping for arenas. */ 295 struct arena_object { 296 /* The address of the arena, as returned by malloc. Note that 0 297 * will never be returned by a successful malloc, and is used 298 * here to mark an arena_object that doesn't correspond to an 299 * allocated arena. 300 */ 301 uptr address; 302 303 /* Pool-aligned pointer to the next pool to be carved off. */ 304 block* pool_address; 305 306 /* The number of available pools in the arena: free pools + never- 307 * allocated pools. 308 */ 309 uint nfreepools; 310 311 /* The total number of pools in the arena, whether or not available. */ 312 uint ntotalpools; 313 314 /* Singly-linked list of available pools. */ 315 struct pool_header* freepools; 316 317 /* Whenever this arena_object is not associated with an allocated 318 * arena, the nextarena member is used to link all unassociated 319 * arena_objects in the singly-linked `unused_arena_objects` list. 320 * The prevarena member is unused in this case. 321 * 322 * When this arena_object is associated with an allocated arena 323 * with at least one available pool, both members are used in the 324 * doubly-linked `usable_arenas` list, which is maintained in 325 * increasing order of `nfreepools` values. 326 * 327 * Else this arena_object is associated with an allocated arena 328 * all of whose pools are in use. `nextarena` and `prevarena` 329 * are both meaningless in this case. 330 */ 331 struct arena_object* nextarena; 332 struct arena_object* prevarena; 333 }; 334 335 #undef ROUNDUP 336 #define ROUNDUP(x) (((x) + ALIGNMENT_MASK) & ~ALIGNMENT_MASK) 337 #define POOL_OVERHEAD ROUNDUP(sizeof(struct pool_header)) 338 339 #define DUMMY_SIZE_IDX 0xffff /* size class of newly cached pools */ 340 341 /* Round pointer P down to the closest pool-aligned address <= P, as a poolp */ 342 #define POOL_ADDR(P) ((poolp)((uptr)(P) & ~(uptr)POOL_SIZE_MASK)) 343 344 /* Return total number of blocks in pool of size index I, as a uint. */ 345 #define NUMBLOCKS(I) ((uint)(POOL_SIZE - POOL_OVERHEAD) / INDEX2SIZE(I)) 346 347 /*==========================================================================*/ 348 349 /* 350 * This malloc lock 351 */ 352 SIMPLELOCK_DECL(_malloc_lock) 353 #define LOCK() SIMPLELOCK_LOCK(_malloc_lock) 354 #define UNLOCK() SIMPLELOCK_UNLOCK(_malloc_lock) 355 #define LOCK_INIT() SIMPLELOCK_INIT(_malloc_lock) 356 #define LOCK_FINI() SIMPLELOCK_FINI(_malloc_lock) 357 358 /* 359 * Pool table -- headed, circular, doubly-linked lists of partially used pools. 360 361 This is involved. For an index i, usedpools[i+i] is the header for a list of 362 all partially used pools holding small blocks with "size class idx" i. So 363 usedpools[0] corresponds to blocks of size 8, usedpools[2] to blocks of size 364 16, and so on: index 2*i <-> blocks of size (i+1)<<ALIGNMENT_SHIFT. 365 366 Pools are carved off an arena's highwater mark (an arena_object's pool_address 367 member) as needed. Once carved off, a pool is in one of three states forever 368 after: 369 370 used == partially used, neither empty nor full 371 At least one block in the pool is currently allocated, and at least one 372 block in the pool is not currently allocated (note this implies a pool 373 has room for at least two blocks). 374 This is a pool's initial state, as a pool is created only when malloc 375 needs space. 376 The pool holds blocks of a fixed size, and is in the circular list headed 377 at usedpools[i] (see above). It's linked to the other used pools of the 378 same size class via the pool_header's nextpool and prevpool members. 379 If all but one block is currently allocated, a malloc can cause a 380 transition to the full state. If all but one block is not currently 381 allocated, a free can cause a transition to the empty state. 382 383 full == all the pool's blocks are currently allocated 384 On transition to full, a pool is unlinked from its usedpools[] list. 385 It's not linked to from anything then anymore, and its nextpool and 386 prevpool members are meaningless until it transitions back to used. 387 A free of a block in a full pool puts the pool back in the used state. 388 Then it's linked in at the front of the appropriate usedpools[] list, so 389 that the next allocation for its size class will reuse the freed block. 390 391 empty == all the pool's blocks are currently available for allocation 392 On transition to empty, a pool is unlinked from its usedpools[] list, 393 and linked to the front of its arena_object's singly-linked freepools list, 394 via its nextpool member. The prevpool member has no meaning in this case. 395 Empty pools have no inherent size class: the next time a malloc finds 396 an empty list in usedpools[], it takes the first pool off of freepools. 397 If the size class needed happens to be the same as the size class the pool 398 last had, some pool initialization can be skipped. 399 400 401 Block Management 402 403 Blocks within pools are again carved out as needed. pool->freeblock points to 404 the start of a singly-linked list of free blocks within the pool. When a 405 block is freed, it's inserted at the front of its pool's freeblock list. Note 406 that the available blocks in a pool are *not* linked all together when a pool 407 is initialized. Instead only "the first two" (lowest addresses) blocks are 408 set up, returning the first such block, and setting pool->freeblock to a 409 one-block list holding the second such block. This is consistent with that 410 pymalloc strives at all levels (arena, pool, and block) never to touch a piece 411 of memory until it's actually needed. 412 413 So long as a pool is in the used state, we're certain there *is* a block 414 available for allocating, and pool->freeblock is not NULL. If pool->freeblock 415 points to the end of the free list before we've carved the entire pool into 416 blocks, that means we simply haven't yet gotten to one of the higher-address 417 blocks. The offset from the pool_header to the start of "the next" virgin 418 block is stored in the pool_header nextoffset member, and the largest value 419 of nextoffset that makes sense is stored in the maxnextoffset member when a 420 pool is initialized. All the blocks in a pool have been passed out at least 421 once when and only when nextoffset > maxnextoffset. 422 423 424 Major obscurity: While the usedpools vector is declared to have poolp 425 entries, it doesn't really. It really contains two pointers per (conceptual) 426 poolp entry, the nextpool and prevpool members of a pool_header. The 427 excruciating initialization code below fools C so that 428 429 usedpool[i+i] 430 431 "acts like" a genuine poolp, but only so long as you only reference its 432 nextpool and prevpool members. The "- 2*sizeof(block *)" gibberish is 433 compensating for that a pool_header's nextpool and prevpool members 434 immediately follow a pool_header's first two members: 435 436 union { block *_padding; 437 uint count; } ref; 438 block *freeblock; 439 440 each of which consume sizeof(block *) bytes. So what usedpools[i+i] really 441 contains is a fudged-up pointer p such that *if* C believes it's a poolp 442 pointer, then p->nextpool and p->prevpool are both p (meaning that the headed 443 circular list is empty). 444 445 It's unclear why the usedpools setup is so convoluted. It could be to 446 minimize the amount of cache required to hold this heavily-referenced table 447 (which only *needs* the two interpool pointer members of a pool_header). OTOH, 448 referencing code has to remember to "double the index" and doing so isn't 449 free, usedpools[0] isn't a strictly legal pointer, and we're crucially relying 450 on that C doesn't insert any padding anywhere in a pool_header at or before 451 the prevpool member. 452 **************************************************************************** */ 453 454 #define PTA(x) ((poolp )((uchar *)&(usedpools[2*(x)]) - 2*sizeof(block *))) 455 #define PT(x) PTA(x), PTA(x) 456 457 static poolp usedpools[2 * ((NB_SMALL_SIZE_CLASSES + 7) / 8) * 8] = { 458 PT(0), PT(1), PT(2), PT(3), PT(4), PT(5), PT(6), PT(7) 459 #if NB_SMALL_SIZE_CLASSES > 8 460 , PT(8), PT(9), PT(10), PT(11), PT(12), PT(13), PT(14), PT(15) 461 #if NB_SMALL_SIZE_CLASSES > 16 462 , PT(16), PT(17), PT(18), PT(19), PT(20), PT(21), PT(22), PT(23) 463 #if NB_SMALL_SIZE_CLASSES > 24 464 , PT(24), PT(25), PT(26), PT(27), PT(28), PT(29), PT(30), PT(31) 465 #if NB_SMALL_SIZE_CLASSES > 32 466 , PT(32), PT(33), PT(34), PT(35), PT(36), PT(37), PT(38), PT(39) 467 #if NB_SMALL_SIZE_CLASSES > 40 468 , PT(40), PT(41), PT(42), PT(43), PT(44), PT(45), PT(46), PT(47) 469 #if NB_SMALL_SIZE_CLASSES > 48 470 , PT(48), PT(49), PT(50), PT(51), PT(52), PT(53), PT(54), PT(55) 471 #if NB_SMALL_SIZE_CLASSES > 56 472 , PT(56), PT(57), PT(58), PT(59), PT(60), PT(61), PT(62), PT(63) 473 #if NB_SMALL_SIZE_CLASSES > 64 474 #error "NB_SMALL_SIZE_CLASSES should be less than 64" 475 #endif /* NB_SMALL_SIZE_CLASSES > 64 */ 476 #endif /* NB_SMALL_SIZE_CLASSES > 56 */ 477 #endif /* NB_SMALL_SIZE_CLASSES > 48 */ 478 #endif /* NB_SMALL_SIZE_CLASSES > 40 */ 479 #endif /* NB_SMALL_SIZE_CLASSES > 32 */ 480 #endif /* NB_SMALL_SIZE_CLASSES > 24 */ 481 #endif /* NB_SMALL_SIZE_CLASSES > 16 */ 482 #endif /* NB_SMALL_SIZE_CLASSES > 8 */ 483 }; 484 485 /*========================================================================== 486 Arena management. 487 488 `arenas` is a vector of arena_objects. It contains maxarenas entries, some of 489 which may not be currently used (== they're arena_objects that aren't 490 currently associated with an allocated arena). Note that arenas proper are 491 separately malloc'ed. 492 493 Prior to Python 2.5, arenas were never free()'ed. Starting with Python 2.5, 494 we do try to free() arenas, and use some mild heuristic strategies to increase 495 the likelihood that arenas eventually can be freed. 496 497 unused_arena_objects 498 499 This is a singly-linked list of the arena_objects that are currently not 500 being used (no arena is associated with them). Objects are taken off the 501 head of the list in new_arena(), and are pushed on the head of the list in 502 PyObject_Free() when the arena is empty. Key invariant: an arena_object 503 is on this list if and only if its .address member is 0. 504 505 usable_arenas 506 507 This is a doubly-linked list of the arena_objects associated with arenas 508 that have pools available. These pools are either waiting to be reused, 509 or have not been used before. The list is sorted to have the most- 510 allocated arenas first (ascending order based on the nfreepools member). 511 This means that the next allocation will come from a heavily used arena, 512 which gives the nearly empty arenas a chance to be returned to the system. 513 In my unscientific tests this dramatically improved the number of arenas 514 that could be freed. 515 516 Note that an arena_object associated with an arena all of whose pools are 517 currently in use isn't on either list. 518 */ 519 520 /* Array of objects used to track chunks of memory (arenas). */ 521 static struct arena_object* arenas = NULL; 522 /* Number of slots currently allocated in the `arenas` vector. */ 523 static uint maxarenas = 0; 524 525 /* The head of the singly-linked, NULL-terminated list of available 526 * arena_objects. 527 */ 528 static struct arena_object* unused_arena_objects = NULL; 529 530 /* The head of the doubly-linked, NULL-terminated at each end, list of 531 * arena_objects associated with arenas that have pools available. 532 */ 533 static struct arena_object* usable_arenas = NULL; 534 535 /* How many arena_objects do we initially allocate? 536 * 16 = can allocate 16 arenas = 16 * ARENA_SIZE = 4MB before growing the 537 * `arenas` vector. 538 */ 539 #define INITIAL_ARENA_OBJECTS 16 540 541 /* Number of arenas allocated that haven't been free()'d. */ 542 static size_t narenas_currently_allocated = 0; 543 544 #ifdef PYMALLOC_DEBUG 545 /* Total number of times malloc() called to allocate an arena. */ 546 static size_t ntimes_arena_allocated = 0; 547 /* High water mark (max value ever seen) for narenas_currently_allocated. */ 548 static size_t narenas_highwater = 0; 549 #endif 550 551 /* Allocate a new arena. If we run out of memory, return NULL. Else 552 * allocate a new arena, and return the address of an arena_object 553 * describing the new arena. It's expected that the caller will set 554 * `usable_arenas` to the return value. 555 */ 556 static struct arena_object* 557 new_arena(void) 558 { 559 struct arena_object* arenaobj; 560 uint excess; /* number of bytes above pool alignment */ 561 void *address; 562 int err; 563 564 #ifdef PYMALLOC_DEBUG 565 if (Py_GETENV("PYTHONMALLOCSTATS")) 566 _PyObject_DebugMallocStats(); 567 #endif 568 if (unused_arena_objects == NULL) { 569 uint i; 570 uint numarenas; 571 size_t nbytes; 572 573 /* Double the number of arena objects on each allocation. 574 * Note that it's possible for `numarenas` to overflow. 575 */ 576 numarenas = maxarenas ? maxarenas << 1 : INITIAL_ARENA_OBJECTS; 577 if (numarenas <= maxarenas) 578 return NULL; /* overflow */ 579 #if SIZEOF_SIZE_T <= SIZEOF_INT 580 if (numarenas > PY_SIZE_MAX / sizeof(*arenas)) 581 return NULL; /* overflow */ 582 #endif 583 nbytes = numarenas * sizeof(*arenas); 584 arenaobj = (struct arena_object *)realloc(arenas, nbytes); 585 if (arenaobj == NULL) 586 return NULL; 587 arenas = arenaobj; 588 589 /* We might need to fix pointers that were copied. However, 590 * new_arena only gets called when all the pages in the 591 * previous arenas are full. Thus, there are *no* pointers 592 * into the old array. Thus, we don't have to worry about 593 * invalid pointers. Just to be sure, some asserts: 594 */ 595 assert(usable_arenas == NULL); 596 assert(unused_arena_objects == NULL); 597 598 /* Put the new arenas on the unused_arena_objects list. */ 599 for (i = maxarenas; i < numarenas; ++i) { 600 arenas[i].address = 0; /* mark as unassociated */ 601 arenas[i].nextarena = i < numarenas - 1 ? 602 &arenas[i+1] : NULL; 603 } 604 605 /* Update globals. */ 606 unused_arena_objects = &arenas[maxarenas]; 607 maxarenas = numarenas; 608 } 609 610 /* Take the next available arena object off the head of the list. */ 611 assert(unused_arena_objects != NULL); 612 arenaobj = unused_arena_objects; 613 unused_arena_objects = arenaobj->nextarena; 614 assert(arenaobj->address == 0); 615 #ifdef ARENAS_USE_MMAP 616 address = mmap(NULL, ARENA_SIZE, PROT_READ|PROT_WRITE, 617 MAP_PRIVATE|MAP_ANONYMOUS, -1, 0); 618 err = (address == MAP_FAILED); 619 #else 620 address = malloc(ARENA_SIZE); 621 err = (address == 0); 622 #endif 623 if (err) { 624 /* The allocation failed: return NULL after putting the 625 * arenaobj back. 626 */ 627 arenaobj->nextarena = unused_arena_objects; 628 unused_arena_objects = arenaobj; 629 return NULL; 630 } 631 arenaobj->address = (uptr)address; 632 633 ++narenas_currently_allocated; 634 #ifdef PYMALLOC_DEBUG 635 ++ntimes_arena_allocated; 636 if (narenas_currently_allocated > narenas_highwater) 637 narenas_highwater = narenas_currently_allocated; 638 #endif 639 arenaobj->freepools = NULL; 640 /* pool_address <- first pool-aligned address in the arena 641 nfreepools <- number of whole pools that fit after alignment */ 642 arenaobj->pool_address = (block*)arenaobj->address; 643 arenaobj->nfreepools = ARENA_SIZE / POOL_SIZE; 644 assert(POOL_SIZE * arenaobj->nfreepools == ARENA_SIZE); 645 excess = (uint)(arenaobj->address & POOL_SIZE_MASK); 646 if (excess != 0) { 647 --arenaobj->nfreepools; 648 arenaobj->pool_address += POOL_SIZE - excess; 649 } 650 arenaobj->ntotalpools = arenaobj->nfreepools; 651 652 return arenaobj; 653 } 654 655 /* 656 Py_ADDRESS_IN_RANGE(P, POOL) 657 658 Return true if and only if P is an address that was allocated by pymalloc. 659 POOL must be the pool address associated with P, i.e., POOL = POOL_ADDR(P) 660 (the caller is asked to compute this because the macro expands POOL more than 661 once, and for efficiency it's best for the caller to assign POOL_ADDR(P) to a 662 variable and pass the latter to the macro; because Py_ADDRESS_IN_RANGE is 663 called on every alloc/realloc/free, micro-efficiency is important here). 664 665 Tricky: Let B be the arena base address associated with the pool, B = 666 arenas[(POOL)->arenaindex].address. Then P belongs to the arena if and only if 667 668 B <= P < B + ARENA_SIZE 669 670 Subtracting B throughout, this is true iff 671 672 0 <= P-B < ARENA_SIZE 673 674 By using unsigned arithmetic, the "0 <=" half of the test can be skipped. 675 676 Obscure: A PyMem "free memory" function can call the pymalloc free or realloc 677 before the first arena has been allocated. `arenas` is still NULL in that 678 case. We're relying on that maxarenas is also 0 in that case, so that 679 (POOL)->arenaindex < maxarenas must be false, saving us from trying to index 680 into a NULL arenas. 681 682 Details: given P and POOL, the arena_object corresponding to P is AO = 683 arenas[(POOL)->arenaindex]. Suppose obmalloc controls P. Then (barring wild 684 stores, etc), POOL is the correct address of P's pool, AO.address is the 685 correct base address of the pool's arena, and P must be within ARENA_SIZE of 686 AO.address. In addition, AO.address is not 0 (no arena can start at address 0 687 (NULL)). Therefore Py_ADDRESS_IN_RANGE correctly reports that obmalloc 688 controls P. 689 690 Now suppose obmalloc does not control P (e.g., P was obtained via a direct 691 call to the system malloc() or realloc()). (POOL)->arenaindex may be anything 692 in this case -- it may even be uninitialized trash. If the trash arenaindex 693 is >= maxarenas, the macro correctly concludes at once that obmalloc doesn't 694 control P. 695 696 Else arenaindex is < maxarena, and AO is read up. If AO corresponds to an 697 allocated arena, obmalloc controls all the memory in slice AO.address : 698 AO.address+ARENA_SIZE. By case assumption, P is not controlled by obmalloc, 699 so P doesn't lie in that slice, so the macro correctly reports that P is not 700 controlled by obmalloc. 701 702 Finally, if P is not controlled by obmalloc and AO corresponds to an unused 703 arena_object (one not currently associated with an allocated arena), 704 AO.address is 0, and the second test in the macro reduces to: 705 706 P < ARENA_SIZE 707 708 If P >= ARENA_SIZE (extremely likely), the macro again correctly concludes 709 that P is not controlled by obmalloc. However, if P < ARENA_SIZE, this part 710 of the test still passes, and the third clause (AO.address != 0) is necessary 711 to get the correct result: AO.address is 0 in this case, so the macro 712 correctly reports that P is not controlled by obmalloc (despite that P lies in 713 slice AO.address : AO.address + ARENA_SIZE). 714 715 Note: The third (AO.address != 0) clause was added in Python 2.5. Before 716 2.5, arenas were never free()'ed, and an arenaindex < maxarena always 717 corresponded to a currently-allocated arena, so the "P is not controlled by 718 obmalloc, AO corresponds to an unused arena_object, and P < ARENA_SIZE" case 719 was impossible. 720 721 Note that the logic is excruciating, and reading up possibly uninitialized 722 memory when P is not controlled by obmalloc (to get at (POOL)->arenaindex) 723 creates problems for some memory debuggers. The overwhelming advantage is 724 that this test determines whether an arbitrary address is controlled by 725 obmalloc in a small constant time, independent of the number of arenas 726 obmalloc controls. Since this test is needed at every entry point, it's 727 extremely desirable that it be this fast. 728 729 Since Py_ADDRESS_IN_RANGE may be reading from memory which was not allocated 730 by Python, it is important that (POOL)->arenaindex is read only once, as 731 another thread may be concurrently modifying the value without holding the 732 GIL. To accomplish this, the arenaindex_temp variable is used to store 733 (POOL)->arenaindex for the duration of the Py_ADDRESS_IN_RANGE macro's 734 execution. The caller of the macro is responsible for declaring this 735 variable. 736 */ 737 #define Py_ADDRESS_IN_RANGE(P, POOL) \ 738 ((arenaindex_temp = (POOL)->arenaindex) < maxarenas && \ 739 (uptr)(P) - arenas[arenaindex_temp].address < (uptr)ARENA_SIZE && \ 740 arenas[arenaindex_temp].address != 0) 741 742 743 /* This is only useful when running memory debuggers such as 744 * Purify or Valgrind. Uncomment to use. 745 * 746 #define Py_USING_MEMORY_DEBUGGER 747 */ 748 749 #ifdef Py_USING_MEMORY_DEBUGGER 750 751 /* Py_ADDRESS_IN_RANGE may access uninitialized memory by design 752 * This leads to thousands of spurious warnings when using 753 * Purify or Valgrind. By making a function, we can easily 754 * suppress the uninitialized memory reads in this one function. 755 * So we won't ignore real errors elsewhere. 756 * 757 * Disable the macro and use a function. 758 */ 759 760 #undef Py_ADDRESS_IN_RANGE 761 762 #if defined(__GNUC__) && ((__GNUC__ == 3) && (__GNUC_MINOR__ >= 1) || \ 763 (__GNUC__ >= 4)) 764 #define Py_NO_INLINE __attribute__((__noinline__)) 765 #else 766 #define Py_NO_INLINE 767 #endif 768 769 /* Don't make static, to try to ensure this isn't inlined. */ 770 int Py_ADDRESS_IN_RANGE(void *P, poolp pool) Py_NO_INLINE; 771 #undef Py_NO_INLINE 772 #endif 773 774 /*==========================================================================*/ 775 776 /* malloc. Note that nbytes==0 tries to return a non-NULL pointer, distinct 777 * from all other currently live pointers. This may not be possible. 778 */ 779 780 /* 781 * The basic blocks are ordered by decreasing execution frequency, 782 * which minimizes the number of jumps in the most common cases, 783 * improves branching prediction and instruction scheduling (small 784 * block allocations typically result in a couple of instructions). 785 * Unless the optimizer reorders everything, being too smart... 786 */ 787 788 #undef PyObject_Malloc 789 void * 790 PyObject_Malloc(size_t nbytes) 791 { 792 block *bp; 793 poolp pool; 794 poolp next; 795 uint size; 796 797 #ifdef WITH_VALGRIND 798 if (UNLIKELY(running_on_valgrind == -1)) 799 running_on_valgrind = RUNNING_ON_VALGRIND; 800 if (UNLIKELY(running_on_valgrind)) 801 goto redirect; 802 #endif 803 804 /* 805 * Limit ourselves to PY_SSIZE_T_MAX bytes to prevent security holes. 806 * Most python internals blindly use a signed Py_ssize_t to track 807 * things without checking for overflows or negatives. 808 * As size_t is unsigned, checking for nbytes < 0 is not required. 809 */ 810 if (nbytes > PY_SSIZE_T_MAX) 811 return NULL; 812 813 /* 814 * This implicitly redirects malloc(0). 815 */ 816 if ((nbytes - 1) < SMALL_REQUEST_THRESHOLD) { 817 LOCK(); 818 /* 819 * Most frequent paths first 820 */ 821 size = (uint)(nbytes - 1) >> ALIGNMENT_SHIFT; 822 pool = usedpools[size + size]; 823 if (pool != pool->nextpool) { 824 /* 825 * There is a used pool for this size class. 826 * Pick up the head block of its free list. 827 */ 828 ++pool->ref.count; 829 bp = pool->freeblock; 830 assert(bp != NULL); 831 if ((pool->freeblock = *(block **)bp) != NULL) { 832 UNLOCK(); 833 return (void *)bp; 834 } 835 /* 836 * Reached the end of the free list, try to extend it. 837 */ 838 if (pool->nextoffset <= pool->maxnextoffset) { 839 /* There is room for another block. */ 840 pool->freeblock = (block*)pool + 841 pool->nextoffset; 842 pool->nextoffset += INDEX2SIZE(size); 843 *(block **)(pool->freeblock) = NULL; 844 UNLOCK(); 845 return (void *)bp; 846 } 847 /* Pool is full, unlink from used pools. */ 848 next = pool->nextpool; 849 pool = pool->prevpool; 850 next->prevpool = pool; 851 pool->nextpool = next; 852 UNLOCK(); 853 return (void *)bp; 854 } 855 856 /* There isn't a pool of the right size class immediately 857 * available: use a free pool. 858 */ 859 if (usable_arenas == NULL) { 860 /* No arena has a free pool: allocate a new arena. */ 861 #ifdef WITH_MEMORY_LIMITS 862 if (narenas_currently_allocated >= MAX_ARENAS) { 863 UNLOCK(); 864 goto redirect; 865 } 866 #endif 867 usable_arenas = new_arena(); 868 if (usable_arenas == NULL) { 869 UNLOCK(); 870 goto redirect; 871 } 872 usable_arenas->nextarena = 873 usable_arenas->prevarena = NULL; 874 } 875 assert(usable_arenas->address != 0); 876 877 /* Try to get a cached free pool. */ 878 pool = usable_arenas->freepools; 879 if (pool != NULL) { 880 /* Unlink from cached pools. */ 881 usable_arenas->freepools = pool->nextpool; 882 883 /* This arena already had the smallest nfreepools 884 * value, so decreasing nfreepools doesn't change 885 * that, and we don't need to rearrange the 886 * usable_arenas list. However, if the arena has 887 * become wholly allocated, we need to remove its 888 * arena_object from usable_arenas. 889 */ 890 --usable_arenas->nfreepools; 891 if (usable_arenas->nfreepools == 0) { 892 /* Wholly allocated: remove. */ 893 assert(usable_arenas->freepools == NULL); 894 assert(usable_arenas->nextarena == NULL || 895 usable_arenas->nextarena->prevarena == 896 usable_arenas); 897 898 usable_arenas = usable_arenas->nextarena; 899 if (usable_arenas != NULL) { 900 usable_arenas->prevarena = NULL; 901 assert(usable_arenas->address != 0); 902 } 903 } 904 else { 905 /* nfreepools > 0: it must be that freepools 906 * isn't NULL, or that we haven't yet carved 907 * off all the arena's pools for the first 908 * time. 909 */ 910 assert(usable_arenas->freepools != NULL || 911 usable_arenas->pool_address <= 912 (block*)usable_arenas->address + 913 ARENA_SIZE - POOL_SIZE); 914 } 915 init_pool: 916 /* Frontlink to used pools. */ 917 next = usedpools[size + size]; /* == prev */ 918 pool->nextpool = next; 919 pool->prevpool = next; 920 next->nextpool = pool; 921 next->prevpool = pool; 922 pool->ref.count = 1; 923 if (pool->szidx == size) { 924 /* Luckily, this pool last contained blocks 925 * of the same size class, so its header 926 * and free list are already initialized. 927 */ 928 bp = pool->freeblock; 929 pool->freeblock = *(block **)bp; 930 UNLOCK(); 931 return (void *)bp; 932 } 933 /* 934 * Initialize the pool header, set up the free list to 935 * contain just the second block, and return the first 936 * block. 937 */ 938 pool->szidx = size; 939 size = INDEX2SIZE(size); 940 bp = (block *)pool + POOL_OVERHEAD; 941 pool->nextoffset = POOL_OVERHEAD + (size << 1); 942 pool->maxnextoffset = POOL_SIZE - size; 943 pool->freeblock = bp + size; 944 *(block **)(pool->freeblock) = NULL; 945 UNLOCK(); 946 return (void *)bp; 947 } 948 949 /* Carve off a new pool. */ 950 assert(usable_arenas->nfreepools > 0); 951 assert(usable_arenas->freepools == NULL); 952 pool = (poolp)usable_arenas->pool_address; 953 assert((block*)pool <= (block*)usable_arenas->address + 954 ARENA_SIZE - POOL_SIZE); 955 pool->arenaindex = usable_arenas - arenas; 956 assert(&arenas[pool->arenaindex] == usable_arenas); 957 pool->szidx = DUMMY_SIZE_IDX; 958 usable_arenas->pool_address += POOL_SIZE; 959 --usable_arenas->nfreepools; 960 961 if (usable_arenas->nfreepools == 0) { 962 assert(usable_arenas->nextarena == NULL || 963 usable_arenas->nextarena->prevarena == 964 usable_arenas); 965 /* Unlink the arena: it is completely allocated. */ 966 usable_arenas = usable_arenas->nextarena; 967 if (usable_arenas != NULL) { 968 usable_arenas->prevarena = NULL; 969 assert(usable_arenas->address != 0); 970 } 971 } 972 973 goto init_pool; 974 } 975 976 /* The small block allocator ends here. */ 977 978 redirect: 979 /* Redirect the original request to the underlying (libc) allocator. 980 * We jump here on bigger requests, on error in the code above (as a 981 * last chance to serve the request) or when the max memory limit 982 * has been reached. 983 */ 984 if (nbytes == 0) 985 nbytes = 1; 986 return (void *)malloc(nbytes); 987 } 988 989 /* free */ 990 991 #undef PyObject_Free 992 ATTRIBUTE_NO_ADDRESS_SAFETY_ANALYSIS 993 void 994 PyObject_Free(void *p) 995 { 996 poolp pool; 997 block *lastfree; 998 poolp next, prev; 999 uint size; 1000 #ifndef Py_USING_MEMORY_DEBUGGER 1001 uint arenaindex_temp; 1002 #endif 1003 1004 if (p == NULL) /* free(NULL) has no effect */ 1005 return; 1006 1007 #ifdef WITH_VALGRIND 1008 if (UNLIKELY(running_on_valgrind > 0)) 1009 goto redirect; 1010 #endif 1011 1012 pool = POOL_ADDR(p); 1013 if (Py_ADDRESS_IN_RANGE(p, pool)) { 1014 /* We allocated this address. */ 1015 LOCK(); 1016 /* Link p to the start of the pool's freeblock list. Since 1017 * the pool had at least the p block outstanding, the pool 1018 * wasn't empty (so it's already in a usedpools[] list, or 1019 * was full and is in no list -- it's not in the freeblocks 1020 * list in any case). 1021 */ 1022 assert(pool->ref.count > 0); /* else it was empty */ 1023 *(block **)p = lastfree = pool->freeblock; 1024 pool->freeblock = (block *)p; 1025 if (lastfree) { 1026 struct arena_object* ao; 1027 uint nf; /* ao->nfreepools */ 1028 1029 /* freeblock wasn't NULL, so the pool wasn't full, 1030 * and the pool is in a usedpools[] list. 1031 */ 1032 if (--pool->ref.count != 0) { 1033 /* pool isn't empty: leave it in usedpools */ 1034 UNLOCK(); 1035 return; 1036 } 1037 /* Pool is now empty: unlink from usedpools, and 1038 * link to the front of freepools. This ensures that 1039 * previously freed pools will be allocated later 1040 * (being not referenced, they are perhaps paged out). 1041 */ 1042 next = pool->nextpool; 1043 prev = pool->prevpool; 1044 next->prevpool = prev; 1045 prev->nextpool = next; 1046 1047 /* Link the pool to freepools. This is a singly-linked 1048 * list, and pool->prevpool isn't used there. 1049 */ 1050 ao = &arenas[pool->arenaindex]; 1051 pool->nextpool = ao->freepools; 1052 ao->freepools = pool; 1053 nf = ++ao->nfreepools; 1054 1055 /* All the rest is arena management. We just freed 1056 * a pool, and there are 4 cases for arena mgmt: 1057 * 1. If all the pools are free, return the arena to 1058 * the system free(). 1059 * 2. If this is the only free pool in the arena, 1060 * add the arena back to the `usable_arenas` list. 1061 * 3. If the "next" arena has a smaller count of free 1062 * pools, we have to "slide this arena right" to 1063 * restore that usable_arenas is sorted in order of 1064 * nfreepools. 1065 * 4. Else there's nothing more to do. 1066 */ 1067 if (nf == ao->ntotalpools) { 1068 /* Case 1. First unlink ao from usable_arenas. 1069 */ 1070 assert(ao->prevarena == NULL || 1071 ao->prevarena->address != 0); 1072 assert(ao ->nextarena == NULL || 1073 ao->nextarena->address != 0); 1074 1075 /* Fix the pointer in the prevarena, or the 1076 * usable_arenas pointer. 1077 */ 1078 if (ao->prevarena == NULL) { 1079 usable_arenas = ao->nextarena; 1080 assert(usable_arenas == NULL || 1081 usable_arenas->address != 0); 1082 } 1083 else { 1084 assert(ao->prevarena->nextarena == ao); 1085 ao->prevarena->nextarena = 1086 ao->nextarena; 1087 } 1088 /* Fix the pointer in the nextarena. */ 1089 if (ao->nextarena != NULL) { 1090 assert(ao->nextarena->prevarena == ao); 1091 ao->nextarena->prevarena = 1092 ao->prevarena; 1093 } 1094 /* Record that this arena_object slot is 1095 * available to be reused. 1096 */ 1097 ao->nextarena = unused_arena_objects; 1098 unused_arena_objects = ao; 1099 1100 /* Free the entire arena. */ 1101 #ifdef ARENAS_USE_MMAP 1102 munmap((void *)ao->address, ARENA_SIZE); 1103 #else 1104 free((void *)ao->address); 1105 #endif 1106 ao->address = 0; /* mark unassociated */ 1107 --narenas_currently_allocated; 1108 1109 UNLOCK(); 1110 return; 1111 } 1112 if (nf == 1) { 1113 /* Case 2. Put ao at the head of 1114 * usable_arenas. Note that because 1115 * ao->nfreepools was 0 before, ao isn't 1116 * currently on the usable_arenas list. 1117 */ 1118 ao->nextarena = usable_arenas; 1119 ao->prevarena = NULL; 1120 if (usable_arenas) 1121 usable_arenas->prevarena = ao; 1122 usable_arenas = ao; 1123 assert(usable_arenas->address != 0); 1124 1125 UNLOCK(); 1126 return; 1127 } 1128 /* If this arena is now out of order, we need to keep 1129 * the list sorted. The list is kept sorted so that 1130 * the "most full" arenas are used first, which allows 1131 * the nearly empty arenas to be completely freed. In 1132 * a few un-scientific tests, it seems like this 1133 * approach allowed a lot more memory to be freed. 1134 */ 1135 if (ao->nextarena == NULL || 1136 nf <= ao->nextarena->nfreepools) { 1137 /* Case 4. Nothing to do. */ 1138 UNLOCK(); 1139 return; 1140 } 1141 /* Case 3: We have to move the arena towards the end 1142 * of the list, because it has more free pools than 1143 * the arena to its right. 1144 * First unlink ao from usable_arenas. 1145 */ 1146 if (ao->prevarena != NULL) { 1147 /* ao isn't at the head of the list */ 1148 assert(ao->prevarena->nextarena == ao); 1149 ao->prevarena->nextarena = ao->nextarena; 1150 } 1151 else { 1152 /* ao is at the head of the list */ 1153 assert(usable_arenas == ao); 1154 usable_arenas = ao->nextarena; 1155 } 1156 ao->nextarena->prevarena = ao->prevarena; 1157 1158 /* Locate the new insertion point by iterating over 1159 * the list, using our nextarena pointer. 1160 */ 1161 while (ao->nextarena != NULL && 1162 nf > ao->nextarena->nfreepools) { 1163 ao->prevarena = ao->nextarena; 1164 ao->nextarena = ao->nextarena->nextarena; 1165 } 1166 1167 /* Insert ao at this point. */ 1168 assert(ao->nextarena == NULL || 1169 ao->prevarena == ao->nextarena->prevarena); 1170 assert(ao->prevarena->nextarena == ao->nextarena); 1171 1172 ao->prevarena->nextarena = ao; 1173 if (ao->nextarena != NULL) 1174 ao->nextarena->prevarena = ao; 1175 1176 /* Verify that the swaps worked. */ 1177 assert(ao->nextarena == NULL || 1178 nf <= ao->nextarena->nfreepools); 1179 assert(ao->prevarena == NULL || 1180 nf > ao->prevarena->nfreepools); 1181 assert(ao->nextarena == NULL || 1182 ao->nextarena->prevarena == ao); 1183 assert((usable_arenas == ao && 1184 ao->prevarena == NULL) || 1185 ao->prevarena->nextarena == ao); 1186 1187 UNLOCK(); 1188 return; 1189 } 1190 /* Pool was full, so doesn't currently live in any list: 1191 * link it to the front of the appropriate usedpools[] list. 1192 * This mimics LRU pool usage for new allocations and 1193 * targets optimal filling when several pools contain 1194 * blocks of the same size class. 1195 */ 1196 --pool->ref.count; 1197 assert(pool->ref.count > 0); /* else the pool is empty */ 1198 size = pool->szidx; 1199 next = usedpools[size + size]; 1200 prev = next->prevpool; 1201 /* insert pool before next: prev <-> pool <-> next */ 1202 pool->nextpool = next; 1203 pool->prevpool = prev; 1204 next->prevpool = pool; 1205 prev->nextpool = pool; 1206 UNLOCK(); 1207 return; 1208 } 1209 1210 #ifdef WITH_VALGRIND 1211 redirect: 1212 #endif 1213 /* We didn't allocate this address. */ 1214 free(p); 1215 } 1216 1217 /* realloc. If p is NULL, this acts like malloc(nbytes). Else if nbytes==0, 1218 * then as the Python docs promise, we do not treat this like free(p), and 1219 * return a non-NULL result. 1220 */ 1221 1222 #undef PyObject_Realloc 1223 ATTRIBUTE_NO_ADDRESS_SAFETY_ANALYSIS 1224 void * 1225 PyObject_Realloc(void *p, size_t nbytes) 1226 { 1227 void *bp; 1228 poolp pool; 1229 size_t size; 1230 #ifndef Py_USING_MEMORY_DEBUGGER 1231 uint arenaindex_temp; 1232 #endif 1233 1234 if (p == NULL) 1235 return PyObject_Malloc(nbytes); 1236 1237 /* 1238 * Limit ourselves to PY_SSIZE_T_MAX bytes to prevent security holes. 1239 * Most python internals blindly use a signed Py_ssize_t to track 1240 * things without checking for overflows or negatives. 1241 * As size_t is unsigned, checking for nbytes < 0 is not required. 1242 */ 1243 if (nbytes > PY_SSIZE_T_MAX) 1244 return NULL; 1245 1246 #ifdef WITH_VALGRIND 1247 /* Treat running_on_valgrind == -1 the same as 0 */ 1248 if (UNLIKELY(running_on_valgrind > 0)) 1249 goto redirect; 1250 #endif 1251 1252 pool = POOL_ADDR(p); 1253 if (Py_ADDRESS_IN_RANGE(p, pool)) { 1254 /* We're in charge of this block */ 1255 size = INDEX2SIZE(pool->szidx); 1256 if (nbytes <= size) { 1257 /* The block is staying the same or shrinking. If 1258 * it's shrinking, there's a tradeoff: it costs 1259 * cycles to copy the block to a smaller size class, 1260 * but it wastes memory not to copy it. The 1261 * compromise here is to copy on shrink only if at 1262 * least 25% of size can be shaved off. 1263 */ 1264 if (4 * nbytes > 3 * size) { 1265 /* It's the same, 1266 * or shrinking and new/old > 3/4. 1267 */ 1268 return p; 1269 } 1270 size = nbytes; 1271 } 1272 bp = PyObject_Malloc(nbytes); 1273 if (bp != NULL) { 1274 memcpy(bp, p, size); 1275 PyObject_Free(p); 1276 } 1277 return bp; 1278 } 1279 #ifdef WITH_VALGRIND 1280 redirect: 1281 #endif 1282 /* We're not managing this block. If nbytes <= 1283 * SMALL_REQUEST_THRESHOLD, it's tempting to try to take over this 1284 * block. However, if we do, we need to copy the valid data from 1285 * the C-managed block to one of our blocks, and there's no portable 1286 * way to know how much of the memory space starting at p is valid. 1287 * As bug 1185883 pointed out the hard way, it's possible that the 1288 * C-managed block is "at the end" of allocated VM space, so that 1289 * a memory fault can occur if we try to copy nbytes bytes starting 1290 * at p. Instead we punt: let C continue to manage this block. 1291 */ 1292 if (nbytes) 1293 return realloc(p, nbytes); 1294 /* C doesn't define the result of realloc(p, 0) (it may or may not 1295 * return NULL then), but Python's docs promise that nbytes==0 never 1296 * returns NULL. We don't pass 0 to realloc(), to avoid that endcase 1297 * to begin with. Even then, we can't be sure that realloc() won't 1298 * return NULL. 1299 */ 1300 bp = realloc(p, 1); 1301 return bp ? bp : p; 1302 } 1303 1304 #else /* ! WITH_PYMALLOC */ 1305 1306 /*==========================================================================*/ 1307 /* pymalloc not enabled: Redirect the entry points to malloc. These will 1308 * only be used by extensions that are compiled with pymalloc enabled. */ 1309 1310 void * 1311 PyObject_Malloc(size_t n) 1312 { 1313 return PyMem_MALLOC(n); 1314 } 1315 1316 void * 1317 PyObject_Realloc(void *p, size_t n) 1318 { 1319 return PyMem_REALLOC(p, n); 1320 } 1321 1322 void 1323 PyObject_Free(void *p) 1324 { 1325 PyMem_FREE(p); 1326 } 1327 #endif /* WITH_PYMALLOC */ 1328 1329 #ifdef PYMALLOC_DEBUG 1330 /*==========================================================================*/ 1331 /* A x-platform debugging allocator. This doesn't manage memory directly, 1332 * it wraps a real allocator, adding extra debugging info to the memory blocks. 1333 */ 1334 1335 /* Special bytes broadcast into debug memory blocks at appropriate times. 1336 * Strings of these are unlikely to be valid addresses, floats, ints or 1337 * 7-bit ASCII. 1338 */ 1339 #undef CLEANBYTE 1340 #undef DEADBYTE 1341 #undef FORBIDDENBYTE 1342 #define CLEANBYTE 0xCB /* clean (newly allocated) memory */ 1343 #define DEADBYTE 0xDB /* dead (newly freed) memory */ 1344 #define FORBIDDENBYTE 0xFB /* untouchable bytes at each end of a block */ 1345 1346 /* We tag each block with an API ID in order to tag API violations */ 1347 #define _PYMALLOC_MEM_ID 'm' /* the PyMem_Malloc() API */ 1348 #define _PYMALLOC_OBJ_ID 'o' /* The PyObject_Malloc() API */ 1349 1350 static size_t serialno = 0; /* incremented on each debug {m,re}alloc */ 1351 1352 /* serialno is always incremented via calling this routine. The point is 1353 * to supply a single place to set a breakpoint. 1354 */ 1355 static void 1356 bumpserialno(void) 1357 { 1358 ++serialno; 1359 } 1360 1361 #define SST SIZEOF_SIZE_T 1362 1363 /* Read sizeof(size_t) bytes at p as a big-endian size_t. */ 1364 static size_t 1365 read_size_t(const void *p) 1366 { 1367 const uchar *q = (const uchar *)p; 1368 size_t result = *q++; 1369 int i; 1370 1371 for (i = SST; --i > 0; ++q) 1372 result = (result << 8) | *q; 1373 return result; 1374 } 1375 1376 /* Write n as a big-endian size_t, MSB at address p, LSB at 1377 * p + sizeof(size_t) - 1. 1378 */ 1379 static void 1380 write_size_t(void *p, size_t n) 1381 { 1382 uchar *q = (uchar *)p + SST - 1; 1383 int i; 1384 1385 for (i = SST; --i >= 0; --q) { 1386 *q = (uchar)(n & 0xff); 1387 n >>= 8; 1388 } 1389 } 1390 1391 #ifdef Py_DEBUG 1392 /* Is target in the list? The list is traversed via the nextpool pointers. 1393 * The list may be NULL-terminated, or circular. Return 1 if target is in 1394 * list, else 0. 1395 */ 1396 static int 1397 pool_is_in_list(const poolp target, poolp list) 1398 { 1399 poolp origlist = list; 1400 assert(target != NULL); 1401 if (list == NULL) 1402 return 0; 1403 do { 1404 if (target == list) 1405 return 1; 1406 list = list->nextpool; 1407 } while (list != NULL && list != origlist); 1408 return 0; 1409 } 1410 1411 #else 1412 #define pool_is_in_list(X, Y) 1 1413 1414 #endif /* Py_DEBUG */ 1415 1416 /* Let S = sizeof(size_t). The debug malloc asks for 4*S extra bytes and 1417 fills them with useful stuff, here calling the underlying malloc's result p: 1418 1419 p[0: S] 1420 Number of bytes originally asked for. This is a size_t, big-endian (easier 1421 to read in a memory dump). 1422 p[S: 2*S] 1423 Copies of FORBIDDENBYTE. Used to catch under- writes and reads. 1424 p[2*S: 2*S+n] 1425 The requested memory, filled with copies of CLEANBYTE. 1426 Used to catch reference to uninitialized memory. 1427 &p[2*S] is returned. Note that this is 8-byte aligned if pymalloc 1428 handled the request itself. 1429 p[2*S+n: 2*S+n+S] 1430 Copies of FORBIDDENBYTE. Used to catch over- writes and reads. 1431 p[2*S+n+S: 2*S+n+2*S] 1432 A serial number, incremented by 1 on each call to _PyObject_DebugMalloc 1433 and _PyObject_DebugRealloc. 1434 This is a big-endian size_t. 1435 If "bad memory" is detected later, the serial number gives an 1436 excellent way to set a breakpoint on the next run, to capture the 1437 instant at which this block was passed out. 1438 */ 1439 1440 /* debug replacements for the PyMem_* memory API */ 1441 void * 1442 _PyMem_DebugMalloc(size_t nbytes) 1443 { 1444 return _PyObject_DebugMallocApi(_PYMALLOC_MEM_ID, nbytes); 1445 } 1446 void * 1447 _PyMem_DebugRealloc(void *p, size_t nbytes) 1448 { 1449 return _PyObject_DebugReallocApi(_PYMALLOC_MEM_ID, p, nbytes); 1450 } 1451 void 1452 _PyMem_DebugFree(void *p) 1453 { 1454 _PyObject_DebugFreeApi(_PYMALLOC_MEM_ID, p); 1455 } 1456 1457 /* debug replacements for the PyObject_* memory API */ 1458 void * 1459 _PyObject_DebugMalloc(size_t nbytes) 1460 { 1461 return _PyObject_DebugMallocApi(_PYMALLOC_OBJ_ID, nbytes); 1462 } 1463 void * 1464 _PyObject_DebugRealloc(void *p, size_t nbytes) 1465 { 1466 return _PyObject_DebugReallocApi(_PYMALLOC_OBJ_ID, p, nbytes); 1467 } 1468 void 1469 _PyObject_DebugFree(void *p) 1470 { 1471 _PyObject_DebugFreeApi(_PYMALLOC_OBJ_ID, p); 1472 } 1473 void 1474 _PyObject_DebugCheckAddress(const void *p) 1475 { 1476 _PyObject_DebugCheckAddressApi(_PYMALLOC_OBJ_ID, p); 1477 } 1478 1479 1480 /* generic debug memory api, with an "id" to identify the API in use */ 1481 void * 1482 _PyObject_DebugMallocApi(char id, size_t nbytes) 1483 { 1484 uchar *p; /* base address of malloc'ed block */ 1485 uchar *tail; /* p + 2*SST + nbytes == pointer to tail pad bytes */ 1486 size_t total; /* nbytes + 4*SST */ 1487 1488 bumpserialno(); 1489 total = nbytes + 4*SST; 1490 if (total < nbytes) 1491 /* overflow: can't represent total as a size_t */ 1492 return NULL; 1493 1494 p = (uchar *)PyObject_Malloc(total); 1495 if (p == NULL) 1496 return NULL; 1497 1498 /* at p, write size (SST bytes), id (1 byte), pad (SST-1 bytes) */ 1499 write_size_t(p, nbytes); 1500 p[SST] = (uchar)id; 1501 memset(p + SST + 1 , FORBIDDENBYTE, SST-1); 1502 1503 if (nbytes > 0) 1504 memset(p + 2*SST, CLEANBYTE, nbytes); 1505 1506 /* at tail, write pad (SST bytes) and serialno (SST bytes) */ 1507 tail = p + 2*SST + nbytes; 1508 memset(tail, FORBIDDENBYTE, SST); 1509 write_size_t(tail + SST, serialno); 1510 1511 return p + 2*SST; 1512 } 1513 1514 /* The debug free first checks the 2*SST bytes on each end for sanity (in 1515 particular, that the FORBIDDENBYTEs with the api ID are still intact). 1516 Then fills the original bytes with DEADBYTE. 1517 Then calls the underlying free. 1518 */ 1519 void 1520 _PyObject_DebugFreeApi(char api, void *p) 1521 { 1522 uchar *q = (uchar *)p - 2*SST; /* address returned from malloc */ 1523 size_t nbytes; 1524 1525 if (p == NULL) 1526 return; 1527 _PyObject_DebugCheckAddressApi(api, p); 1528 nbytes = read_size_t(q); 1529 nbytes += 4*SST; 1530 if (nbytes > 0) 1531 memset(q, DEADBYTE, nbytes); 1532 PyObject_Free(q); 1533 } 1534 1535 void * 1536 _PyObject_DebugReallocApi(char api, void *p, size_t nbytes) 1537 { 1538 uchar *q = (uchar *)p; 1539 uchar *tail; 1540 size_t total; /* nbytes + 4*SST */ 1541 size_t original_nbytes; 1542 int i; 1543 1544 if (p == NULL) 1545 return _PyObject_DebugMallocApi(api, nbytes); 1546 1547 _PyObject_DebugCheckAddressApi(api, p); 1548 bumpserialno(); 1549 original_nbytes = read_size_t(q - 2*SST); 1550 total = nbytes + 4*SST; 1551 if (total < nbytes) 1552 /* overflow: can't represent total as a size_t */ 1553 return NULL; 1554 1555 if (nbytes < original_nbytes) { 1556 /* shrinking: mark old extra memory dead */ 1557 memset(q + nbytes, DEADBYTE, original_nbytes - nbytes + 2*SST); 1558 } 1559 1560 /* Resize and add decorations. We may get a new pointer here, in which 1561 * case we didn't get the chance to mark the old memory with DEADBYTE, 1562 * but we live with that. 1563 */ 1564 q = (uchar *)PyObject_Realloc(q - 2*SST, total); 1565 if (q == NULL) 1566 return NULL; 1567 1568 write_size_t(q, nbytes); 1569 assert(q[SST] == (uchar)api); 1570 for (i = 1; i < SST; ++i) 1571 assert(q[SST + i] == FORBIDDENBYTE); 1572 q += 2*SST; 1573 tail = q + nbytes; 1574 memset(tail, FORBIDDENBYTE, SST); 1575 write_size_t(tail + SST, serialno); 1576 1577 if (nbytes > original_nbytes) { 1578 /* growing: mark new extra memory clean */ 1579 memset(q + original_nbytes, CLEANBYTE, 1580 nbytes - original_nbytes); 1581 } 1582 1583 return q; 1584 } 1585 1586 /* Check the forbidden bytes on both ends of the memory allocated for p. 1587 * If anything is wrong, print info to stderr via _PyObject_DebugDumpAddress, 1588 * and call Py_FatalError to kill the program. 1589 * The API id, is also checked. 1590 */ 1591 void 1592 _PyObject_DebugCheckAddressApi(char api, const void *p) 1593 { 1594 const uchar *q = (const uchar *)p; 1595 char msgbuf[64]; 1596 char *msg; 1597 size_t nbytes; 1598 const uchar *tail; 1599 int i; 1600 char id; 1601 1602 if (p == NULL) { 1603 msg = "didn't expect a NULL pointer"; 1604 goto error; 1605 } 1606 1607 /* Check the API id */ 1608 id = (char)q[-SST]; 1609 if (id != api) { 1610 msg = msgbuf; 1611 snprintf(msg, sizeof(msgbuf), "bad ID: Allocated using API '%c', verified using API '%c'", id, api); 1612 msgbuf[sizeof(msgbuf)-1] = 0; 1613 goto error; 1614 } 1615 1616 /* Check the stuff at the start of p first: if there's underwrite 1617 * corruption, the number-of-bytes field may be nuts, and checking 1618 * the tail could lead to a segfault then. 1619 */ 1620 for (i = SST-1; i >= 1; --i) { 1621 if (*(q-i) != FORBIDDENBYTE) { 1622 msg = "bad leading pad byte"; 1623 goto error; 1624 } 1625 } 1626 1627 nbytes = read_size_t(q - 2*SST); 1628 tail = q + nbytes; 1629 for (i = 0; i < SST; ++i) { 1630 if (tail[i] != FORBIDDENBYTE) { 1631 msg = "bad trailing pad byte"; 1632 goto error; 1633 } 1634 } 1635 1636 return; 1637 1638 error: 1639 _PyObject_DebugDumpAddress(p); 1640 Py_FatalError(msg); 1641 } 1642 1643 /* Display info to stderr about the memory block at p. */ 1644 void 1645 _PyObject_DebugDumpAddress(const void *p) 1646 { 1647 const uchar *q = (const uchar *)p; 1648 const uchar *tail; 1649 size_t nbytes, serial; 1650 int i; 1651 int ok; 1652 char id; 1653 1654 fprintf(stderr, "Debug memory block at address p=%p:", p); 1655 if (p == NULL) { 1656 fprintf(stderr, "\n"); 1657 return; 1658 } 1659 id = (char)q[-SST]; 1660 fprintf(stderr, " API '%c'\n", id); 1661 1662 nbytes = read_size_t(q - 2*SST); 1663 fprintf(stderr, " %" PY_FORMAT_SIZE_T "u bytes originally " 1664 "requested\n", nbytes); 1665 1666 /* In case this is nuts, check the leading pad bytes first. */ 1667 fprintf(stderr, " The %d pad bytes at p-%d are ", SST-1, SST-1); 1668 ok = 1; 1669 for (i = 1; i <= SST-1; ++i) { 1670 if (*(q-i) != FORBIDDENBYTE) { 1671 ok = 0; 1672 break; 1673 } 1674 } 1675 if (ok) 1676 fputs("FORBIDDENBYTE, as expected.\n", stderr); 1677 else { 1678 fprintf(stderr, "not all FORBIDDENBYTE (0x%02x):\n", 1679 FORBIDDENBYTE); 1680 for (i = SST-1; i >= 1; --i) { 1681 const uchar byte = *(q-i); 1682 fprintf(stderr, " at p-%d: 0x%02x", i, byte); 1683 if (byte != FORBIDDENBYTE) 1684 fputs(" *** OUCH", stderr); 1685 fputc('\n', stderr); 1686 } 1687 1688 fputs(" Because memory is corrupted at the start, the " 1689 "count of bytes requested\n" 1690 " may be bogus, and checking the trailing pad " 1691 "bytes may segfault.\n", stderr); 1692 } 1693 1694 tail = q + nbytes; 1695 fprintf(stderr, " The %d pad bytes at tail=%p are ", SST, tail); 1696 ok = 1; 1697 for (i = 0; i < SST; ++i) { 1698 if (tail[i] != FORBIDDENBYTE) { 1699 ok = 0; 1700 break; 1701 } 1702 } 1703 if (ok) 1704 fputs("FORBIDDENBYTE, as expected.\n", stderr); 1705 else { 1706 fprintf(stderr, "not all FORBIDDENBYTE (0x%02x):\n", 1707 FORBIDDENBYTE); 1708 for (i = 0; i < SST; ++i) { 1709 const uchar byte = tail[i]; 1710 fprintf(stderr, " at tail+%d: 0x%02x", 1711 i, byte); 1712 if (byte != FORBIDDENBYTE) 1713 fputs(" *** OUCH", stderr); 1714 fputc('\n', stderr); 1715 } 1716 } 1717 1718 serial = read_size_t(tail + SST); 1719 fprintf(stderr, " The block was made by call #%" PY_FORMAT_SIZE_T 1720 "u to debug malloc/realloc.\n", serial); 1721 1722 if (nbytes > 0) { 1723 i = 0; 1724 fputs(" Data at p:", stderr); 1725 /* print up to 8 bytes at the start */ 1726 while (q < tail && i < 8) { 1727 fprintf(stderr, " %02x", *q); 1728 ++i; 1729 ++q; 1730 } 1731 /* and up to 8 at the end */ 1732 if (q < tail) { 1733 if (tail - q > 8) { 1734 fputs(" ...", stderr); 1735 q = tail - 8; 1736 } 1737 while (q < tail) { 1738 fprintf(stderr, " %02x", *q); 1739 ++q; 1740 } 1741 } 1742 fputc('\n', stderr); 1743 } 1744 } 1745 1746 static size_t 1747 printone(const char* msg, size_t value) 1748 { 1749 int i, k; 1750 char buf[100]; 1751 size_t origvalue = value; 1752 1753 fputs(msg, stderr); 1754 for (i = (int)strlen(msg); i < 35; ++i) 1755 fputc(' ', stderr); 1756 fputc('=', stderr); 1757 1758 /* Write the value with commas. */ 1759 i = 22; 1760 buf[i--] = '\0'; 1761 buf[i--] = '\n'; 1762 k = 3; 1763 do { 1764 size_t nextvalue = value / 10; 1765 unsigned int digit = (unsigned int)(value - nextvalue * 10); 1766 value = nextvalue; 1767 buf[i--] = (char)(digit + '0'); 1768 --k; 1769 if (k == 0 && value && i >= 0) { 1770 k = 3; 1771 buf[i--] = ','; 1772 } 1773 } while (value && i >= 0); 1774 1775 while (i >= 0) 1776 buf[i--] = ' '; 1777 fputs(buf, stderr); 1778 1779 return origvalue; 1780 } 1781 1782 /* Print summary info to stderr about the state of pymalloc's structures. 1783 * In Py_DEBUG mode, also perform some expensive internal consistency 1784 * checks. 1785 */ 1786 void 1787 _PyObject_DebugMallocStats(void) 1788 { 1789 uint i; 1790 const uint numclasses = SMALL_REQUEST_THRESHOLD >> ALIGNMENT_SHIFT; 1791 /* # of pools, allocated blocks, and free blocks per class index */ 1792 size_t numpools[SMALL_REQUEST_THRESHOLD >> ALIGNMENT_SHIFT]; 1793 size_t numblocks[SMALL_REQUEST_THRESHOLD >> ALIGNMENT_SHIFT]; 1794 size_t numfreeblocks[SMALL_REQUEST_THRESHOLD >> ALIGNMENT_SHIFT]; 1795 /* total # of allocated bytes in used and full pools */ 1796 size_t allocated_bytes = 0; 1797 /* total # of available bytes in used pools */ 1798 size_t available_bytes = 0; 1799 /* # of free pools + pools not yet carved out of current arena */ 1800 uint numfreepools = 0; 1801 /* # of bytes for arena alignment padding */ 1802 size_t arena_alignment = 0; 1803 /* # of bytes in used and full pools used for pool_headers */ 1804 size_t pool_header_bytes = 0; 1805 /* # of bytes in used and full pools wasted due to quantization, 1806 * i.e. the necessarily leftover space at the ends of used and 1807 * full pools. 1808 */ 1809 size_t quantization = 0; 1810 /* # of arenas actually allocated. */ 1811 size_t narenas = 0; 1812 /* running total -- should equal narenas * ARENA_SIZE */ 1813 size_t total; 1814 char buf[128]; 1815 1816 fprintf(stderr, "Small block threshold = %d, in %u size classes.\n", 1817 SMALL_REQUEST_THRESHOLD, numclasses); 1818 1819 for (i = 0; i < numclasses; ++i) 1820 numpools[i] = numblocks[i] = numfreeblocks[i] = 0; 1821 1822 /* Because full pools aren't linked to from anything, it's easiest 1823 * to march over all the arenas. If we're lucky, most of the memory 1824 * will be living in full pools -- would be a shame to miss them. 1825 */ 1826 for (i = 0; i < maxarenas; ++i) { 1827 uint j; 1828 uptr base = arenas[i].address; 1829 1830 /* Skip arenas which are not allocated. */ 1831 if (arenas[i].address == (uptr)NULL) 1832 continue; 1833 narenas += 1; 1834 1835 numfreepools += arenas[i].nfreepools; 1836 1837 /* round up to pool alignment */ 1838 if (base & (uptr)POOL_SIZE_MASK) { 1839 arena_alignment += POOL_SIZE; 1840 base &= ~(uptr)POOL_SIZE_MASK; 1841 base += POOL_SIZE; 1842 } 1843 1844 /* visit every pool in the arena */ 1845 assert(base <= (uptr) arenas[i].pool_address); 1846 for (j = 0; 1847 base < (uptr) arenas[i].pool_address; 1848 ++j, base += POOL_SIZE) { 1849 poolp p = (poolp)base; 1850 const uint sz = p->szidx; 1851 uint freeblocks; 1852 1853 if (p->ref.count == 0) { 1854 /* currently unused */ 1855 assert(pool_is_in_list(p, arenas[i].freepools)); 1856 continue; 1857 } 1858 ++numpools[sz]; 1859 numblocks[sz] += p->ref.count; 1860 freeblocks = NUMBLOCKS(sz) - p->ref.count; 1861 numfreeblocks[sz] += freeblocks; 1862 #ifdef Py_DEBUG 1863 if (freeblocks > 0) 1864 assert(pool_is_in_list(p, usedpools[sz + sz])); 1865 #endif 1866 } 1867 } 1868 assert(narenas == narenas_currently_allocated); 1869 1870 fputc('\n', stderr); 1871 fputs("class size num pools blocks in use avail blocks\n" 1872 "----- ---- --------- ------------- ------------\n", 1873 stderr); 1874 1875 for (i = 0; i < numclasses; ++i) { 1876 size_t p = numpools[i]; 1877 size_t b = numblocks[i]; 1878 size_t f = numfreeblocks[i]; 1879 uint size = INDEX2SIZE(i); 1880 if (p == 0) { 1881 assert(b == 0 && f == 0); 1882 continue; 1883 } 1884 fprintf(stderr, "%5u %6u " 1885 "%11" PY_FORMAT_SIZE_T "u " 1886 "%15" PY_FORMAT_SIZE_T "u " 1887 "%13" PY_FORMAT_SIZE_T "u\n", 1888 i, size, p, b, f); 1889 allocated_bytes += b * size; 1890 available_bytes += f * size; 1891 pool_header_bytes += p * POOL_OVERHEAD; 1892 quantization += p * ((POOL_SIZE - POOL_OVERHEAD) % size); 1893 } 1894 fputc('\n', stderr); 1895 (void)printone("# times object malloc called", serialno); 1896 1897 (void)printone("# arenas allocated total", ntimes_arena_allocated); 1898 (void)printone("# arenas reclaimed", ntimes_arena_allocated - narenas); 1899 (void)printone("# arenas highwater mark", narenas_highwater); 1900 (void)printone("# arenas allocated current", narenas); 1901 1902 PyOS_snprintf(buf, sizeof(buf), 1903 "%" PY_FORMAT_SIZE_T "u arenas * %d bytes/arena", 1904 narenas, ARENA_SIZE); 1905 (void)printone(buf, narenas * ARENA_SIZE); 1906 1907 fputc('\n', stderr); 1908 1909 total = printone("# bytes in allocated blocks", allocated_bytes); 1910 total += printone("# bytes in available blocks", available_bytes); 1911 1912 PyOS_snprintf(buf, sizeof(buf), 1913 "%u unused pools * %d bytes", numfreepools, POOL_SIZE); 1914 total += printone(buf, (size_t)numfreepools * POOL_SIZE); 1915 1916 total += printone("# bytes lost to pool headers", pool_header_bytes); 1917 total += printone("# bytes lost to quantization", quantization); 1918 total += printone("# bytes lost to arena alignment", arena_alignment); 1919 (void)printone("Total", total); 1920 } 1921 1922 #endif /* PYMALLOC_DEBUG */ 1923 1924 #ifdef Py_USING_MEMORY_DEBUGGER 1925 /* Make this function last so gcc won't inline it since the definition is 1926 * after the reference. 1927 */ 1928 int 1929 Py_ADDRESS_IN_RANGE(void *P, poolp pool) 1930 { 1931 uint arenaindex_temp = pool->arenaindex; 1932 1933 return arenaindex_temp < maxarenas && 1934 (uptr)P - arenas[arenaindex_temp].address < (uptr)ARENA_SIZE && 1935 arenas[arenaindex_temp].address != 0; 1936 } 1937 #endif 1938
