| 1 |
/*------------------------------------------------------------------------- |
| 2 |
* |
| 3 |
* dsa.c |
| 4 |
* Dynamic shared memory areas. |
| 5 |
* |
| 6 |
* This module provides dynamic shared memory areas which are built on top of |
| 7 |
* DSM segments. While dsm.c allows segments of memory of shared memory to be |
| 8 |
* created and shared between backends, it isn't designed to deal with small |
| 9 |
* objects. A DSA area is a shared memory heap usually backed by one or more |
| 10 |
* DSM segments which can allocate memory using dsa_allocate() and dsa_free(). |
| 11 |
* Alternatively, it can be created in pre-existing shared memory, including a |
| 12 |
* DSM segment, and then create extra DSM segments as required. Unlike the |
| 13 |
* regular system heap, it deals in pseudo-pointers which must be converted to |
| 14 |
* backend-local pointers before they are dereferenced. These pseudo-pointers |
| 15 |
* can however be shared with other backends, and can be used to construct |
| 16 |
* shared data structures. |
| 17 |
* |
| 18 |
* Each DSA area manages a set of DSM segments, adding new segments as |
| 19 |
* required and detaching them when they are no longer needed. Each segment |
| 20 |
* contains a number of 4KB pages, a free page manager for tracking |
| 21 |
* consecutive runs of free pages, and a page map for tracking the source of |
| 22 |
* objects allocated on each page. Allocation requests above 8KB are handled |
| 23 |
* by choosing a segment and finding consecutive free pages in its free page |
| 24 |
* manager. Allocation requests for smaller sizes are handled using pools of |
| 25 |
* objects of a selection of sizes. Each pool consists of a number of 16 page |
| 26 |
* (64KB) superblocks allocated in the same way as large objects. Allocation |
| 27 |
* of large objects and new superblocks is serialized by a single LWLock, but |
| 28 |
* allocation of small objects from pre-existing superblocks uses one LWLock |
| 29 |
* per pool. Currently there is one pool, and therefore one lock, per size |
| 30 |
* class. Per-core pools to increase concurrency and strategies for reducing |
| 31 |
* the resulting fragmentation are areas for future research. Each superblock |
| 32 |
* is managed with a 'span', which tracks the superblock's freelist. Free |
| 33 |
* requests are handled by looking in the page map to find which span an |
| 34 |
* address was allocated from, so that small objects can be returned to the |
| 35 |
* appropriate free list, and large object pages can be returned directly to |
| 36 |
* the free page map. When allocating, simple heuristics for selecting |
| 37 |
* segments and superblocks try to encourage occupied memory to be |
| 38 |
* concentrated, increasing the likelihood that whole superblocks can become |
| 39 |
* empty and be returned to the free page manager, and whole segments can |
| 40 |
* become empty and be returned to the operating system. |
| 41 |
* |
| 42 |
* Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group |
| 43 |
* Portions Copyright (c) 1994, Regents of the University of California |
| 44 |
* |
| 45 |
* IDENTIFICATION |
| 46 |
* src/backend/utils/mmgr/dsa.c |
| 47 |
* |
| 48 |
*------------------------------------------------------------------------- |
| 49 |
*/ |
| 50 |
|
| 51 |
#include "postgres.h" |
| 52 |
|
| 53 |
#include "port/atomics.h" |
| 54 |
#include "port/pg_bitutils.h" |
| 55 |
#include "storage/dsm.h" |
| 56 |
#include "storage/lwlock.h" |
| 57 |
#include "utils/dsa.h" |
| 58 |
#include "utils/freepage.h" |
| 59 |
#include "utils/memutils.h" |
| 60 |
#include "utils/resowner.h" |
| 61 |
|
| 62 |
/* |
| 63 |
* How many segments to create before we double the segment size. If this is |
| 64 |
* low, then there is likely to be a lot of wasted space in the largest |
| 65 |
* segment. If it is high, then we risk running out of segment slots (see |
| 66 |
* dsm.c's limits on total number of segments), or limiting the total size |
| 67 |
* an area can manage when using small pointers. |
| 68 |
*/ |
| 69 |
#define DSA_NUM_SEGMENTS_AT_EACH_SIZE 2 |
| 70 |
|
| 71 |
/* |
| 72 |
* The maximum number of DSM segments that an area can own, determined by |
| 73 |
* the number of bits remaining (but capped at 1024). |
| 74 |
*/ |
| 75 |
#define DSA_MAX_SEGMENTS \ |
| 76 |
Min(1024, (1 << ((SIZEOF_DSA_POINTER * 8) - DSA_OFFSET_WIDTH))) |
| 77 |
|
| 78 |
/* The bitmask for extracting the offset from a dsa_pointer. */ |
| 79 |
#define DSA_OFFSET_BITMASK (((dsa_pointer) 1 << DSA_OFFSET_WIDTH) - 1) |
| 80 |
|
| 81 |
/* Number of pages (see FPM_PAGE_SIZE) per regular superblock. */ |
| 82 |
#define DSA_PAGES_PER_SUPERBLOCK 16 |
| 83 |
|
| 84 |
/* |
| 85 |
* A magic number used as a sanity check for following DSM segments belonging |
| 86 |
* to a DSA area (this number will be XORed with the area handle and |
| 87 |
* the segment index). |
| 88 |
*/ |
| 89 |
#define DSA_SEGMENT_HEADER_MAGIC 0x0ce26608 |
| 90 |
|
| 91 |
/* Build a dsa_pointer given a segment number and offset. */ |
| 92 |
#define DSA_MAKE_POINTER(segment_number, offset) \ |
| 93 |
(((dsa_pointer) (segment_number) << DSA_OFFSET_WIDTH) | (offset)) |
| 94 |
|
| 95 |
/* Extract the segment number from a dsa_pointer. */ |
| 96 |
#define DSA_EXTRACT_SEGMENT_NUMBER(dp) ((dp) >> DSA_OFFSET_WIDTH) |
| 97 |
|
| 98 |
/* Extract the offset from a dsa_pointer. */ |
| 99 |
#define DSA_EXTRACT_OFFSET(dp) ((dp) & DSA_OFFSET_BITMASK) |
| 100 |
|
| 101 |
/* The type used for index segment indexes (zero based). */ |
| 102 |
typedef size_t dsa_segment_index; |
| 103 |
|
| 104 |
/* Sentinel value for dsa_segment_index indicating 'none' or 'end'. */ |
| 105 |
#define DSA_SEGMENT_INDEX_NONE (~(dsa_segment_index)0) |
| 106 |
|
| 107 |
/* |
| 108 |
* How many bins of segments do we have? The bins are used to categorize |
| 109 |
* segments by their largest contiguous run of free pages. |
| 110 |
*/ |
| 111 |
#define DSA_NUM_SEGMENT_BINS 16 |
| 112 |
|
| 113 |
/* |
| 114 |
* What is the lowest bin that holds segments that *might* have n contiguous |
| 115 |
* free pages? There is no point in looking in segments in lower bins; they |
| 116 |
* definitely can't service a request for n free pages. |
| 117 |
*/ |
| 118 |
static inline size_t |
| 119 |
contiguous_pages_to_segment_bin(size_t n) |
| 120 |
{ |
| 121 |
size_t bin; |
| 122 |
|
| 123 |
if (n == 0) |
| 124 |
bin = 0; |
| 125 |
else |
| 126 |
bin = pg_leftmost_one_pos_size_t(n) + 1; |
| 127 |
|
| 128 |
return Min(bin, DSA_NUM_SEGMENT_BINS - 1); |
| 129 |
} |
| 130 |
|
| 131 |
/* Macros for access to locks. */ |
| 132 |
#define DSA_AREA_LOCK(area) (&area->control->lock) |
| 133 |
#define DSA_SCLASS_LOCK(area, sclass) (&area->control->pools[sclass].lock) |
| 134 |
|
| 135 |
/* |
| 136 |
* The header for an individual segment. This lives at the start of each DSM |
| 137 |
* segment owned by a DSA area including the first segment (where it appears |
| 138 |
* as part of the dsa_area_control struct). |
| 139 |
*/ |
| 140 |
typedef struct |
| 141 |
{ |
| 142 |
/* Sanity check magic value. */ |
| 143 |
uint32 magic; |
| 144 |
/* Total number of pages in this segment (excluding metadata area). */ |
| 145 |
size_t usable_pages; |
| 146 |
/* Total size of this segment in bytes. */ |
| 147 |
size_t size; |
| 148 |
|
| 149 |
/* |
| 150 |
* Index of the segment that precedes this one in the same segment bin, or |
| 151 |
* DSA_SEGMENT_INDEX_NONE if this is the first one. |
| 152 |
*/ |
| 153 |
dsa_segment_index prev; |
| 154 |
|
| 155 |
/* |
| 156 |
* Index of the segment that follows this one in the same segment bin, or |
| 157 |
* DSA_SEGMENT_INDEX_NONE if this is the last one. |
| 158 |
*/ |
| 159 |
dsa_segment_index next; |
| 160 |
/* The index of the bin that contains this segment. */ |
| 161 |
size_t bin; |
| 162 |
|
| 163 |
/* |
| 164 |
* A flag raised to indicate that this segment is being returned to the |
| 165 |
* operating system and has been unpinned. |
| 166 |
*/ |
| 167 |
bool freed; |
| 168 |
} dsa_segment_header; |
| 169 |
|
| 170 |
/* |
| 171 |
* Metadata for one superblock. |
| 172 |
* |
| 173 |
* For most blocks, span objects are stored out-of-line; that is, the span |
| 174 |
* object is not stored within the block itself. But, as an exception, for a |
| 175 |
* "span of spans", the span object is stored "inline". The allocation is |
| 176 |
* always exactly one page, and the dsa_area_span object is located at |
| 177 |
* the beginning of that page. The size class is DSA_SCLASS_BLOCK_OF_SPANS, |
| 178 |
* and the remaining fields are used just as they would be in an ordinary |
| 179 |
* block. We can't allocate spans out of ordinary superblocks because |
| 180 |
* creating an ordinary superblock requires us to be able to allocate a span |
| 181 |
* *first*. Doing it this way avoids that circularity. |
| 182 |
*/ |
| 183 |
typedef struct |
| 184 |
{ |
| 185 |
dsa_pointer pool; /* Containing pool. */ |
| 186 |
dsa_pointer prevspan; /* Previous span. */ |
| 187 |
dsa_pointer nextspan; /* Next span. */ |
| 188 |
dsa_pointer start; /* Starting address. */ |
| 189 |
size_t npages; /* Length of span in pages. */ |
| 190 |
uint16 size_class; /* Size class. */ |
| 191 |
uint16 ninitialized; /* Maximum number of objects ever allocated. */ |
| 192 |
uint16 nallocatable; /* Number of objects currently allocatable. */ |
| 193 |
uint16 firstfree; /* First object on free list. */ |
| 194 |
uint16 nmax; /* Maximum number of objects ever possible. */ |
| 195 |
uint16 fclass; /* Current fullness class. */ |
| 196 |
} dsa_area_span; |
| 197 |
|
| 198 |
/* |
| 199 |
* Given a pointer to an object in a span, access the index of the next free |
| 200 |
* object in the same span (ie in the span's freelist) as an L-value. |
| 201 |
*/ |
| 202 |
#define NextFreeObjectIndex(object) (* (uint16 *) (object)) |
| 203 |
|
| 204 |
/* |
| 205 |
* Small allocations are handled by dividing a single block of memory into |
| 206 |
* many small objects of equal size. The possible allocation sizes are |
| 207 |
* defined by the following array. Larger size classes are spaced more widely |
| 208 |
* than smaller size classes. We fudge the spacing for size classes >1kB to |
| 209 |
* avoid space wastage: based on the knowledge that we plan to allocate 64kB |
| 210 |
* blocks, we bump the maximum object size up to the largest multiple of |
| 211 |
* 8 bytes that still lets us fit the same number of objects into one block. |
| 212 |
* |
| 213 |
* NB: Because of this fudging, if we were ever to use differently-sized blocks |
| 214 |
* for small allocations, these size classes would need to be reworked to be |
| 215 |
* optimal for the new size. |
| 216 |
* |
| 217 |
* NB: The optimal spacing for size classes, as well as the size of the blocks |
| 218 |
* out of which small objects are allocated, is not a question that has one |
| 219 |
* right answer. Some allocators (such as tcmalloc) use more closely-spaced |
| 220 |
* size classes than we do here, while others (like aset.c) use more |
| 221 |
* widely-spaced classes. Spacing the classes more closely avoids wasting |
| 222 |
* memory within individual chunks, but also means a larger number of |
| 223 |
* potentially-unfilled blocks. |
| 224 |
*/ |
| 225 |
static const uint16 dsa_size_classes[] = { |
| 226 |
sizeof(dsa_area_span), 0, /* special size classes */ |
| 227 |
8, 16, 24, 32, 40, 48, 56, 64, /* 8 classes separated by 8 bytes */ |
| 228 |
80, 96, 112, 128, /* 4 classes separated by 16 bytes */ |
| 229 |
160, 192, 224, 256, /* 4 classes separated by 32 bytes */ |
| 230 |
320, 384, 448, 512, /* 4 classes separated by 64 bytes */ |
| 231 |
640, 768, 896, 1024, /* 4 classes separated by 128 bytes */ |
| 232 |
1280, 1560, 1816, 2048, /* 4 classes separated by ~256 bytes */ |
| 233 |
2616, 3120, 3640, 4096, /* 4 classes separated by ~512 bytes */ |
| 234 |
5456, 6552, 7280, 8192 /* 4 classes separated by ~1024 bytes */ |
| 235 |
}; |
| 236 |
#define DSA_NUM_SIZE_CLASSES lengthof(dsa_size_classes) |
| 237 |
|
| 238 |
/* Special size classes. */ |
| 239 |
#define DSA_SCLASS_BLOCK_OF_SPANS 0 |
| 240 |
#define DSA_SCLASS_SPAN_LARGE 1 |
| 241 |
|
| 242 |
/* |
| 243 |
* The following lookup table is used to map the size of small objects |
| 244 |
* (less than 1kB) onto the corresponding size class. To use this table, |
| 245 |
* round the size of the object up to the next multiple of 8 bytes, and then |
| 246 |
* index into this array. |
| 247 |
*/ |
| 248 |
static const uint8 dsa_size_class_map[] = { |
| 249 |
2, 3, 4, 5, 6, 7, 8, 9, 10, 10, 11, 11, 12, 12, 13, 13, |
| 250 |
14, 14, 14, 14, 15, 15, 15, 15, 16, 16, 16, 16, 17, 17, 17, 17, |
| 251 |
18, 18, 18, 18, 18, 18, 18, 18, 19, 19, 19, 19, 19, 19, 19, 19, |
| 252 |
20, 20, 20, 20, 20, 20, 20, 20, 21, 21, 21, 21, 21, 21, 21, 21, |
| 253 |
22, 22, 22, 22, 22, 22, 22, 22, 22, 22, 22, 22, 22, 22, 22, 22, |
| 254 |
23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, |
| 255 |
24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, |
| 256 |
25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25 |
| 257 |
}; |
| 258 |
#define DSA_SIZE_CLASS_MAP_QUANTUM 8 |
| 259 |
|
| 260 |
/* |
| 261 |
* Superblocks are binned by how full they are. Generally, each fullness |
| 262 |
* class corresponds to one quartile, but the block being used for |
| 263 |
* allocations is always at the head of the list for fullness class 1, |
| 264 |
* regardless of how full it really is. |
| 265 |
*/ |
| 266 |
#define DSA_FULLNESS_CLASSES 4 |
| 267 |
|
| 268 |
/* |
| 269 |
* A dsa_area_pool represents a set of objects of a given size class. |
| 270 |
* |
| 271 |
* Perhaps there should be multiple pools for the same size class for |
| 272 |
* contention avoidance, but for now there is just one! |
| 273 |
*/ |
| 274 |
typedef struct |
| 275 |
{ |
| 276 |
/* A lock protecting access to this pool. */ |
| 277 |
LWLock lock; |
| 278 |
/* A set of linked lists of spans, arranged by fullness. */ |
| 279 |
dsa_pointer spans[DSA_FULLNESS_CLASSES]; |
| 280 |
/* Should we pad this out to a cacheline boundary? */ |
| 281 |
} dsa_area_pool; |
| 282 |
|
| 283 |
/* |
| 284 |
* The control block for an area. This lives in shared memory, at the start of |
| 285 |
* the first DSM segment controlled by this area. |
| 286 |
*/ |
| 287 |
typedef struct |
| 288 |
{ |
| 289 |
/* The segment header for the first segment. */ |
| 290 |
dsa_segment_header segment_header; |
| 291 |
/* The handle for this area. */ |
| 292 |
dsa_handle handle; |
| 293 |
/* The handles of the segments owned by this area. */ |
| 294 |
dsm_handle segment_handles[DSA_MAX_SEGMENTS]; |
| 295 |
/* Lists of segments, binned by maximum contiguous run of free pages. */ |
| 296 |
dsa_segment_index segment_bins[DSA_NUM_SEGMENT_BINS]; |
| 297 |
/* The object pools for each size class. */ |
| 298 |
dsa_area_pool pools[DSA_NUM_SIZE_CLASSES]; |
| 299 |
/* initial allocation segment size */ |
| 300 |
size_t init_segment_size; |
| 301 |
/* maximum allocation segment size */ |
| 302 |
size_t max_segment_size; |
| 303 |
/* The total size of all active segments. */ |
| 304 |
size_t total_segment_size; |
| 305 |
/* The maximum total size of backing storage we are allowed. */ |
| 306 |
size_t max_total_segment_size; |
| 307 |
/* Highest used segment index in the history of this area. */ |
| 308 |
dsa_segment_index high_segment_index; |
| 309 |
/* The reference count for this area. */ |
| 310 |
int refcnt; |
| 311 |
/* A flag indicating that this area has been pinned. */ |
| 312 |
bool pinned; |
| 313 |
/* The number of times that segments have been freed. */ |
| 314 |
size_t freed_segment_counter; |
| 315 |
/* The LWLock tranche ID. */ |
| 316 |
int lwlock_tranche_id; |
| 317 |
/* The general lock (protects everything except object pools). */ |
| 318 |
LWLock lock; |
| 319 |
} dsa_area_control; |
| 320 |
|
| 321 |
/* Given a pointer to a pool, find a dsa_pointer. */ |
| 322 |
#define DsaAreaPoolToDsaPointer(area, p) \ |
| 323 |
DSA_MAKE_POINTER(0, (char *) p - (char *) area->control) |
| 324 |
|
| 325 |
/* |
| 326 |
* A dsa_segment_map is stored within the backend-private memory of each |
| 327 |
* individual backend. It holds the base address of the segment within that |
| 328 |
* backend, plus the addresses of key objects within the segment. Those |
| 329 |
* could instead be derived from the base address but it's handy to have them |
| 330 |
* around. |
| 331 |
*/ |
| 332 |
typedef struct |
| 333 |
{ |
| 334 |
dsm_segment *segment; /* DSM segment */ |
| 335 |
char *mapped_address; /* Address at which segment is mapped */ |
| 336 |
dsa_segment_header *header; /* Header (same as mapped_address) */ |
| 337 |
FreePageManager *fpm; /* Free page manager within segment. */ |
| 338 |
dsa_pointer *pagemap; /* Page map within segment. */ |
| 339 |
} dsa_segment_map; |
| 340 |
|
| 341 |
/* |
| 342 |
* Per-backend state for a storage area. Backends obtain one of these by |
| 343 |
* creating an area or attaching to an existing one using a handle. Each |
| 344 |
* process that needs to use an area uses its own object to track where the |
| 345 |
* segments are mapped. |
| 346 |
*/ |
| 347 |
struct dsa_area |
| 348 |
{ |
| 349 |
/* Pointer to the control object in shared memory. */ |
| 350 |
dsa_area_control *control; |
| 351 |
|
| 352 |
/* |
| 353 |
* All the mappings are owned by this. The dsa_area itself is not |
| 354 |
* directly tracked by the ResourceOwner, but the effect is the same. NULL |
| 355 |
* if the attachment has session lifespan, i.e if dsa_pin_mapping() has |
| 356 |
* been called. |
| 357 |
*/ |
| 358 |
ResourceOwner resowner; |
| 359 |
|
| 360 |
/* |
| 361 |
* This backend's array of segment maps, ordered by segment index |
| 362 |
* corresponding to control->segment_handles. Some of the area's segments |
| 363 |
* may not be mapped in this backend yet, and some slots may have been |
| 364 |
* freed and need to be detached; these operations happen on demand. |
| 365 |
*/ |
| 366 |
dsa_segment_map segment_maps[DSA_MAX_SEGMENTS]; |
| 367 |
|
| 368 |
/* The highest segment index this backend has ever mapped. */ |
| 369 |
dsa_segment_index high_segment_index; |
| 370 |
|
| 371 |
/* The last observed freed_segment_counter. */ |
| 372 |
size_t freed_segment_counter; |
| 373 |
}; |
| 374 |
|
| 375 |
#define DSA_SPAN_NOTHING_FREE ((uint16) -1) |
| 376 |
#define DSA_SUPERBLOCK_SIZE (DSA_PAGES_PER_SUPERBLOCK * FPM_PAGE_SIZE) |
| 377 |
|
| 378 |
/* Given a pointer to a segment_map, obtain a segment index number. */ |
| 379 |
#define get_segment_index(area, segment_map_ptr) \ |
| 380 |
(segment_map_ptr - &area->segment_maps[0]) |
| 381 |
|
| 382 |
static void init_span(dsa_area *area, dsa_pointer span_pointer, |
| 383 |
dsa_area_pool *pool, dsa_pointer start, size_t npages, |
| 384 |
uint16 size_class); |
| 385 |
static bool transfer_first_span(dsa_area *area, dsa_area_pool *pool, |
| 386 |
int fromclass, int toclass); |
| 387 |
static inline dsa_pointer alloc_object(dsa_area *area, int size_class); |
| 388 |
static bool ensure_active_superblock(dsa_area *area, dsa_area_pool *pool, |
| 389 |
int size_class); |
| 390 |
static dsa_segment_map *get_segment_by_index(dsa_area *area, |
| 391 |
dsa_segment_index index); |
| 392 |
static void destroy_superblock(dsa_area *area, dsa_pointer span_pointer); |
| 393 |
static void unlink_span(dsa_area *area, dsa_area_span *span); |
| 394 |
static void add_span_to_fullness_class(dsa_area *area, dsa_area_span *span, |
| 395 |
dsa_pointer span_pointer, int fclass); |
| 396 |
static void unlink_segment(dsa_area *area, dsa_segment_map *segment_map); |
| 397 |
static dsa_segment_map *get_best_segment(dsa_area *area, size_t npages); |
| 398 |
static dsa_segment_map *make_new_segment(dsa_area *area, size_t requested_pages); |
| 399 |
static dsa_area *create_internal(void *place, size_t size, |
| 400 |
int tranche_id, |
| 401 |
dsm_handle control_handle, |
| 402 |
dsm_segment *control_segment, |
| 403 |
size_t init_segment_size, |
| 404 |
size_t max_segment_size); |
| 405 |
static dsa_area *attach_internal(void *place, dsm_segment *segment, |
| 406 |
dsa_handle handle); |
| 407 |
static void check_for_freed_segments(dsa_area *area); |
| 408 |
static void check_for_freed_segments_locked(dsa_area *area); |
| 409 |
static void rebin_segment(dsa_area *area, dsa_segment_map *segment_map); |
| 410 |
|
| 411 |
/* |
| 412 |
* Create a new shared area in a new DSM segment. Further DSM segments will |
| 413 |
* be allocated as required to extend the available space. |
| 414 |
* |
| 415 |
* We can't allocate a LWLock tranche_id within this function, because tranche |
| 416 |
* IDs are a scarce resource; there are only 64k available, using low numbers |
| 417 |
* when possible matters, and we have no provision for recycling them. So, |
| 418 |
* we require the caller to provide one. |
| 419 |
*/ |
| 420 |
dsa_area * |
| 421 |
dsa_create_ext(int tranche_id, size_t init_segment_size, size_t max_segment_size) |
| 422 |
{ |
| 423 |
dsm_segment *segment; |
| 424 |
dsa_area *area; |
| 425 |
|
| 426 |
/* |
| 427 |
* Create the DSM segment that will hold the shared control object and the |
| 428 |
* first segment of usable space. |
| 429 |
*/ |
| 430 |
segment = dsm_create(init_segment_size, 0); |
| 431 |
|
| 432 |
/* |
| 433 |
* All segments backing this area are pinned, so that DSA can explicitly |
| 434 |
* control their lifetime (otherwise a newly created segment belonging to |
| 435 |
* this area might be freed when the only backend that happens to have it |
| 436 |
* mapped in ends, corrupting the area). |
| 437 |
*/ |
| 438 |
dsm_pin_segment(segment); |
| 439 |
|
| 440 |
/* Create a new DSA area with the control object in this segment. */ |
| 441 |
area = create_internal(dsm_segment_address(segment), |
| 442 |
init_segment_size, |
| 443 |
tranche_id, |
| 444 |
dsm_segment_handle(segment), segment, |
| 445 |
init_segment_size, max_segment_size); |
| 446 |
|
| 447 |
/* Clean up when the control segment detaches. */ |
| 448 |
on_dsm_detach(segment, &dsa_on_dsm_detach_release_in_place, |
| 449 |
PointerGetDatum(dsm_segment_address(segment))); |
| 450 |
|
| 451 |
return area; |
| 452 |
} |
| 453 |
|
| 454 |
/* |
| 455 |
* Create a new shared area in an existing shared memory space, which may be |
| 456 |
* either DSM or Postmaster-initialized memory. DSM segments will be |
| 457 |
* allocated as required to extend the available space, though that can be |
| 458 |
* prevented with dsa_set_size_limit(area, size) using the same size provided |
| 459 |
* to dsa_create_in_place. |
| 460 |
* |
| 461 |
* Areas created in-place must eventually be released by the backend that |
| 462 |
* created them and all backends that attach to them. This can be done |
| 463 |
* explicitly with dsa_release_in_place, or, in the special case that 'place' |
| 464 |
* happens to be in a pre-existing DSM segment, by passing in a pointer to the |
| 465 |
* segment so that a detach hook can be registered with the containing DSM |
| 466 |
* segment. |
| 467 |
* |
| 468 |
* See dsa_create() for a note about the tranche arguments. |
| 469 |
*/ |
| 470 |
dsa_area * |
| 471 |
dsa_create_in_place_ext(void *place, size_t size, |
| 472 |
int tranche_id, dsm_segment *segment, |
| 473 |
size_t init_segment_size, size_t max_segment_size) |
| 474 |
{ |
| 475 |
dsa_area *area; |
| 476 |
|
| 477 |
area = create_internal(place, size, tranche_id, |
| 478 |
DSM_HANDLE_INVALID, NULL, |
| 479 |
init_segment_size, max_segment_size); |
| 480 |
|
| 481 |
/* |
| 482 |
* Clean up when the control segment detaches, if a containing DSM segment |
| 483 |
* was provided. |
| 484 |
*/ |
| 485 |
if (segment != NULL) |
| 486 |
on_dsm_detach(segment, &dsa_on_dsm_detach_release_in_place, |
| 487 |
PointerGetDatum(place)); |
| 488 |
|
| 489 |
return area; |
| 490 |
} |
| 491 |
|
| 492 |
/* |
| 493 |
* Obtain a handle that can be passed to other processes so that they can |
| 494 |
* attach to the given area. Cannot be called for areas created with |
| 495 |
* dsa_create_in_place. |
| 496 |
*/ |
| 497 |
dsa_handle |
| 498 |
dsa_get_handle(dsa_area *area) |
| 499 |
{ |
| 500 |
Assert(area->control->handle != DSA_HANDLE_INVALID); |
| 501 |
return area->control->handle; |
| 502 |
} |
| 503 |
|
| 504 |
/* |
| 505 |
* Attach to an area given a handle generated (possibly in another process) by |
| 506 |
* dsa_get_handle. The area must have been created with dsa_create (not |
| 507 |
* dsa_create_in_place). |
| 508 |
*/ |
| 509 |
dsa_area * |
| 510 |
dsa_attach(dsa_handle handle) |
| 511 |
{ |
| 512 |
dsm_segment *segment; |
| 513 |
dsa_area *area; |
| 514 |
|
| 515 |
/* |
| 516 |
* An area handle is really a DSM segment handle for the first segment, so |
| 517 |
* we go ahead and attach to that. |
| 518 |
*/ |
| 519 |
segment = dsm_attach(handle); |
| 520 |
if (segment == NULL) |
| 521 |
ereport(ERROR, |
| 522 |
(errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE), |
| 523 |
errmsg("could not attach to dynamic shared area"))); |
| 524 |
|
| 525 |
area = attach_internal(dsm_segment_address(segment), segment, handle); |
| 526 |
|
| 527 |
/* Clean up when the control segment detaches. */ |
| 528 |
on_dsm_detach(segment, &dsa_on_dsm_detach_release_in_place, |
| 529 |
PointerGetDatum(dsm_segment_address(segment))); |
| 530 |
|
| 531 |
return area; |
| 532 |
} |
| 533 |
|
| 534 |
/* |
| 535 |
* Attach to an area that was created with dsa_create_in_place. The caller |
| 536 |
* must somehow know the location in memory that was used when the area was |
| 537 |
* created, though it may be mapped at a different virtual address in this |
| 538 |
* process. |
| 539 |
* |
| 540 |
* See dsa_create_in_place for note about releasing in-place areas, and the |
| 541 |
* optional 'segment' argument which can be provided to allow automatic |
| 542 |
* release if the containing memory happens to be a DSM segment. |
| 543 |
*/ |
| 544 |
dsa_area * |
| 545 |
dsa_attach_in_place(void *place, dsm_segment *segment) |
| 546 |
{ |
| 547 |
dsa_area *area; |
| 548 |
|
| 549 |
area = attach_internal(place, NULL, DSA_HANDLE_INVALID); |
| 550 |
|
| 551 |
/* |
| 552 |
* Clean up when the control segment detaches, if a containing DSM segment |
| 553 |
* was provided. |
| 554 |
*/ |
| 555 |
if (segment != NULL) |
| 556 |
on_dsm_detach(segment, &dsa_on_dsm_detach_release_in_place, |
| 557 |
PointerGetDatum(place)); |
| 558 |
|
| 559 |
return area; |
| 560 |
} |
| 561 |
|
| 562 |
/* |
| 563 |
* Release a DSA area that was produced by dsa_create_in_place or |
| 564 |
* dsa_attach_in_place. The 'segment' argument is ignored but provides an |
| 565 |
* interface suitable for on_dsm_detach, for the convenience of users who want |
| 566 |
* to create a DSA segment inside an existing DSM segment and have it |
| 567 |
* automatically released when the containing DSM segment is detached. |
| 568 |
* 'place' should be the address of the place where the area was created. |
| 569 |
* |
| 570 |
* This callback is automatically registered for the DSM segment containing |
| 571 |
* the control object of in-place areas when a segment is provided to |
| 572 |
* dsa_create_in_place or dsa_attach_in_place, and also for all areas created |
| 573 |
* with dsa_create. |
| 574 |
*/ |
| 575 |
void |
| 576 |
dsa_on_dsm_detach_release_in_place(dsm_segment *segment, Datum place) |
| 577 |
{ |
| 578 |
dsa_release_in_place(DatumGetPointer(place)); |
| 579 |
} |
| 580 |
|
| 581 |
/* |
| 582 |
* Release a DSA area that was produced by dsa_create_in_place or |
| 583 |
* dsa_attach_in_place. The 'code' argument is ignored but provides an |
| 584 |
* interface suitable for on_shmem_exit or before_shmem_exit, for the |
| 585 |
* convenience of users who want to create a DSA segment inside shared memory |
| 586 |
* other than a DSM segment and have it automatically release at backend exit. |
| 587 |
* 'place' should be the address of the place where the area was created. |
| 588 |
*/ |
| 589 |
void |
| 590 |
dsa_on_shmem_exit_release_in_place(int code, Datum place) |
| 591 |
{ |
| 592 |
dsa_release_in_place(DatumGetPointer(place)); |
| 593 |
} |
| 594 |
|
| 595 |
/* |
| 596 |
* Release a DSA area that was produced by dsa_create_in_place or |
| 597 |
* dsa_attach_in_place. It is preferable to use one of the 'dsa_on_XXX' |
| 598 |
* callbacks so that this is managed automatically, because failure to release |
| 599 |
* an area created in-place leaks its segments permanently. |
| 600 |
* |
| 601 |
* This is also called automatically for areas produced by dsa_create or |
| 602 |
* dsa_attach as an implementation detail. |
| 603 |
*/ |
| 604 |
void |
| 605 |
dsa_release_in_place(void *place) |
| 606 |
{ |
| 607 |
dsa_area_control *control = (dsa_area_control *) place; |
| 608 |
int i; |
| 609 |
|
| 610 |
LWLockAcquire(&control->lock, LW_EXCLUSIVE); |
| 611 |
Assert(control->segment_header.magic == |
| 612 |
(DSA_SEGMENT_HEADER_MAGIC ^ control->handle ^ 0)); |
| 613 |
Assert(control->refcnt > 0); |
| 614 |
if (--control->refcnt == 0) |
| 615 |
{ |
| 616 |
for (i = 0; i <= control->high_segment_index; ++i) |
| 617 |
{ |
| 618 |
dsm_handle handle; |
| 619 |
|
| 620 |
handle = control->segment_handles[i]; |
| 621 |
if (handle != DSM_HANDLE_INVALID) |
| 622 |
dsm_unpin_segment(handle); |
| 623 |
} |
| 624 |
} |
| 625 |
LWLockRelease(&control->lock); |
| 626 |
} |
| 627 |
|
| 628 |
/* |
| 629 |
* Keep a DSA area attached until end of session or explicit detach. |
| 630 |
* |
| 631 |
* By default, areas are owned by the current resource owner, which means they |
| 632 |
* are detached automatically when that scope ends. |
| 633 |
*/ |
| 634 |
void |
| 635 |
dsa_pin_mapping(dsa_area *area) |
| 636 |
{ |
| 637 |
int i; |
| 638 |
|
| 639 |
if (area->resowner != NULL) |
| 640 |
{ |
| 641 |
area->resowner = NULL; |
| 642 |
|
| 643 |
for (i = 0; i <= area->high_segment_index; ++i) |
| 644 |
if (area->segment_maps[i].segment != NULL) |
| 645 |
dsm_pin_mapping(area->segment_maps[i].segment); |
| 646 |
} |
| 647 |
} |
| 648 |
|
| 649 |
/* |
| 650 |
* Allocate memory in this storage area. The return value is a dsa_pointer |
| 651 |
* that can be passed to other processes, and converted to a local pointer |
| 652 |
* with dsa_get_address. 'flags' is a bitmap which should be constructed |
| 653 |
* from the following values: |
| 654 |
* |
| 655 |
* DSA_ALLOC_HUGE allows allocations >= 1GB. Otherwise, such allocations |
| 656 |
* will result in an ERROR. |
| 657 |
* |
| 658 |
* DSA_ALLOC_NO_OOM causes this function to return InvalidDsaPointer when |
| 659 |
* no memory is available or a size limit established by dsa_set_size_limit |
| 660 |
* would be exceeded. Otherwise, such allocations will result in an ERROR. |
| 661 |
* |
| 662 |
* DSA_ALLOC_ZERO causes the allocated memory to be zeroed. Otherwise, the |
| 663 |
* contents of newly-allocated memory are indeterminate. |
| 664 |
* |
| 665 |
* These flags correspond to similarly named flags used by |
| 666 |
* MemoryContextAllocExtended(). See also the macros dsa_allocate and |
| 667 |
* dsa_allocate0 which expand to a call to this function with commonly used |
| 668 |
* flags. |
| 669 |
*/ |
| 670 |
dsa_pointer |
| 671 |
dsa_allocate_extended(dsa_area *area, size_t size, int flags) |
| 672 |
{ |
| 673 |
uint16 size_class; |
| 674 |
dsa_pointer start_pointer; |
| 675 |
dsa_segment_map *segment_map; |
| 676 |
dsa_pointer result; |
| 677 |
|
| 678 |
Assert(size > 0); |
| 679 |
|
| 680 |
/* Sanity check on huge individual allocation size. */ |
| 681 |
if (((flags & DSA_ALLOC_HUGE) != 0 && !AllocHugeSizeIsValid(size)) || |
| 682 |
((flags & DSA_ALLOC_HUGE) == 0 && !AllocSizeIsValid(size))) |
| 683 |
elog(ERROR, "invalid DSA memory alloc request size %zu", size); |
| 684 |
|
| 685 |
/* |
| 686 |
* If bigger than the largest size class, just grab a run of pages from |
| 687 |
* the free page manager, instead of allocating an object from a pool. |
| 688 |
* There will still be a span, but it's a special class of span that |
| 689 |
* manages this whole allocation and simply gives all pages back to the |
| 690 |
* free page manager when dsa_free is called. |
| 691 |
*/ |
| 692 |
if (size > dsa_size_classes[lengthof(dsa_size_classes) - 1]) |
| 693 |
{ |
| 694 |
size_t npages = fpm_size_to_pages(size); |
| 695 |
size_t first_page; |
| 696 |
dsa_pointer span_pointer; |
| 697 |
dsa_area_pool *pool = &area->control->pools[DSA_SCLASS_SPAN_LARGE]; |
| 698 |
|
| 699 |
/* Obtain a span object. */ |
| 700 |
span_pointer = alloc_object(area, DSA_SCLASS_BLOCK_OF_SPANS); |
| 701 |
if (!DsaPointerIsValid(span_pointer)) |
| 702 |
{ |
| 703 |
/* Raise error unless asked not to. */ |
| 704 |
if ((flags & DSA_ALLOC_NO_OOM) == 0) |
| 705 |
ereport(ERROR, |
| 706 |
(errcode(ERRCODE_OUT_OF_MEMORY), |
| 707 |
errmsg("out of memory"), |
| 708 |
errdetail("Failed on DSA request of size %zu.", |
| 709 |
size))); |
| 710 |
return InvalidDsaPointer; |
| 711 |
} |
| 712 |
|
| 713 |
LWLockAcquire(DSA_AREA_LOCK(area), LW_EXCLUSIVE); |
| 714 |
|
| 715 |
/* Find a segment from which to allocate. */ |
| 716 |
segment_map = get_best_segment(area, npages); |
| 717 |
if (segment_map == NULL) |
| 718 |
segment_map = make_new_segment(area, npages); |
| 719 |
if (segment_map == NULL) |
| 720 |
{ |
| 721 |
/* Can't make any more segments: game over. */ |
| 722 |
LWLockRelease(DSA_AREA_LOCK(area)); |
| 723 |
dsa_free(area, span_pointer); |
| 724 |
|
| 725 |
/* Raise error unless asked not to. */ |
| 726 |
if ((flags & DSA_ALLOC_NO_OOM) == 0) |
| 727 |
ereport(ERROR, |
| 728 |
(errcode(ERRCODE_OUT_OF_MEMORY), |
| 729 |
errmsg("out of memory"), |
| 730 |
errdetail("Failed on DSA request of size %zu.", |
| 731 |
size))); |
| 732 |
return InvalidDsaPointer; |
| 733 |
} |
| 734 |
|
| 735 |
/* |
| 736 |
* Ask the free page manager for a run of pages. This should always |
| 737 |
* succeed, since both get_best_segment and make_new_segment should |
| 738 |
* only return a non-NULL pointer if it actually contains enough |
| 739 |
* contiguous freespace. If it does fail, something in our backend |
| 740 |
* private state is out of whack, so use FATAL to kill the process. |
| 741 |
*/ |
| 742 |
if (!FreePageManagerGet(segment_map->fpm, npages, &first_page)) |
| 743 |
elog(FATAL, |
| 744 |
"dsa_allocate could not find %zu free pages", npages); |
| 745 |
LWLockRelease(DSA_AREA_LOCK(area)); |
| 746 |
|
| 747 |
start_pointer = DSA_MAKE_POINTER(get_segment_index(area, segment_map), |
| 748 |
first_page * FPM_PAGE_SIZE); |
| 749 |
|
| 750 |
/* Initialize span and pagemap. */ |
| 751 |
LWLockAcquire(DSA_SCLASS_LOCK(area, DSA_SCLASS_SPAN_LARGE), |
| 752 |
LW_EXCLUSIVE); |
| 753 |
init_span(area, span_pointer, pool, start_pointer, npages, |
| 754 |
DSA_SCLASS_SPAN_LARGE); |
| 755 |
segment_map->pagemap[first_page] = span_pointer; |
| 756 |
LWLockRelease(DSA_SCLASS_LOCK(area, DSA_SCLASS_SPAN_LARGE)); |
| 757 |
|
| 758 |
/* Zero-initialize the memory if requested. */ |
| 759 |
if ((flags & DSA_ALLOC_ZERO) != 0) |
| 760 |
memset(dsa_get_address(area, start_pointer), 0, size); |
| 761 |
|
| 762 |
return start_pointer; |
| 763 |
} |
| 764 |
|
| 765 |
/* Map allocation to a size class. */ |
| 766 |
if (size < lengthof(dsa_size_class_map) * DSA_SIZE_CLASS_MAP_QUANTUM) |
| 767 |
{ |
| 768 |
int mapidx; |
| 769 |
|
| 770 |
/* For smaller sizes we have a lookup table... */ |
| 771 |
mapidx = ((size + DSA_SIZE_CLASS_MAP_QUANTUM - 1) / |
| 772 |
DSA_SIZE_CLASS_MAP_QUANTUM) - 1; |
| 773 |
size_class = dsa_size_class_map[mapidx]; |
| 774 |
} |
| 775 |
else |
| 776 |
{ |
| 777 |
uint16 min; |
| 778 |
uint16 max; |
| 779 |
|
| 780 |
/* ... and for the rest we search by binary chop. */ |
| 781 |
min = dsa_size_class_map[lengthof(dsa_size_class_map) - 1]; |
| 782 |
max = lengthof(dsa_size_classes) - 1; |
| 783 |
|
| 784 |
while (min < max) |
| 785 |
{ |
| 786 |
uint16 mid = (min + max) / 2; |
| 787 |
uint16 class_size = dsa_size_classes[mid]; |
| 788 |
|
| 789 |
if (class_size < size) |
| 790 |
min = mid + 1; |
| 791 |
else |
| 792 |
max = mid; |
| 793 |
} |
| 794 |
|
| 795 |
size_class = min; |
| 796 |
} |
| 797 |
Assert(size <= dsa_size_classes[size_class]); |
| 798 |
Assert(size_class == 0 || size > dsa_size_classes[size_class - 1]); |
| 799 |
|
| 800 |
/* Attempt to allocate an object from the appropriate pool. */ |
| 801 |
result = alloc_object(area, size_class); |
| 802 |
|
| 803 |
/* Check for failure to allocate. */ |
| 804 |
if (!DsaPointerIsValid(result)) |
| 805 |
{ |
| 806 |
/* Raise error unless asked not to. */ |
| 807 |
if ((flags & DSA_ALLOC_NO_OOM) == 0) |
| 808 |
ereport(ERROR, |
| 809 |
(errcode(ERRCODE_OUT_OF_MEMORY), |
| 810 |
errmsg("out of memory"), |
| 811 |
errdetail("Failed on DSA request of size %zu.", size))); |
| 812 |
return InvalidDsaPointer; |
| 813 |
} |
| 814 |
|
| 815 |
/* Zero-initialize the memory if requested. */ |
| 816 |
if ((flags & DSA_ALLOC_ZERO) != 0) |
| 817 |
memset(dsa_get_address(area, result), 0, size); |
| 818 |
|
| 819 |
return result; |
| 820 |
} |
| 821 |
|
| 822 |
/* |
| 823 |
* Free memory obtained with dsa_allocate. |
| 824 |
*/ |
| 825 |
void |
| 826 |
dsa_free(dsa_area *area, dsa_pointer dp) |
| 827 |
{ |
| 828 |
dsa_segment_map *segment_map; |
| 829 |
int pageno; |
| 830 |
dsa_pointer span_pointer; |
| 831 |
dsa_area_span *span; |
| 832 |
char *superblock; |
| 833 |
char *object; |
| 834 |
size_t size; |
| 835 |
int size_class; |
| 836 |
|
| 837 |
/* Make sure we don't have a stale segment in the slot 'dp' refers to. */ |
| 838 |
check_for_freed_segments(area); |
| 839 |
|
| 840 |
/* Locate the object, span and pool. */ |
| 841 |
segment_map = get_segment_by_index(area, DSA_EXTRACT_SEGMENT_NUMBER(dp)); |
| 842 |
pageno = DSA_EXTRACT_OFFSET(dp) / FPM_PAGE_SIZE; |
| 843 |
span_pointer = segment_map->pagemap[pageno]; |
| 844 |
span = dsa_get_address(area, span_pointer); |
| 845 |
superblock = dsa_get_address(area, span->start); |
| 846 |
object = dsa_get_address(area, dp); |
| 847 |
size_class = span->size_class; |
| 848 |
size = dsa_size_classes[size_class]; |
| 849 |
|
| 850 |
/* |
| 851 |
* Special case for large objects that live in a special span: we return |
| 852 |
* those pages directly to the free page manager and free the span. |
| 853 |
*/ |
| 854 |
if (span->size_class == DSA_SCLASS_SPAN_LARGE) |
| 855 |
{ |
| 856 |
|
| 857 |
#ifdef CLOBBER_FREED_MEMORY |
| 858 |
memset(object, 0x7f, span->npages * FPM_PAGE_SIZE); |
| 859 |
#endif |
| 860 |
|
| 861 |
/* Give pages back to free page manager. */ |
| 862 |
LWLockAcquire(DSA_AREA_LOCK(area), LW_EXCLUSIVE); |
| 863 |
FreePageManagerPut(segment_map->fpm, |
| 864 |
DSA_EXTRACT_OFFSET(span->start) / FPM_PAGE_SIZE, |
| 865 |
span->npages); |
| 866 |
|
| 867 |
/* Move segment to appropriate bin if necessary. */ |
| 868 |
rebin_segment(area, segment_map); |
| 869 |
LWLockRelease(DSA_AREA_LOCK(area)); |
| 870 |
|
| 871 |
/* Unlink span. */ |
| 872 |
LWLockAcquire(DSA_SCLASS_LOCK(area, DSA_SCLASS_SPAN_LARGE), |
| 873 |
LW_EXCLUSIVE); |
| 874 |
unlink_span(area, span); |
| 875 |
LWLockRelease(DSA_SCLASS_LOCK(area, DSA_SCLASS_SPAN_LARGE)); |
| 876 |
/* Free the span object so it can be reused. */ |
| 877 |
dsa_free(area, span_pointer); |
| 878 |
return; |
| 879 |
} |
| 880 |
|
| 881 |
#ifdef CLOBBER_FREED_MEMORY |
| 882 |
memset(object, 0x7f, size); |
| 883 |
#endif |
| 884 |
|
| 885 |
LWLockAcquire(DSA_SCLASS_LOCK(area, size_class), LW_EXCLUSIVE); |
| 886 |
|
| 887 |
/* Put the object on the span's freelist. */ |
| 888 |
Assert(object >= superblock); |
| 889 |
Assert(object < superblock + DSA_SUPERBLOCK_SIZE); |
| 890 |
Assert((object - superblock) % size == 0); |
| 891 |
NextFreeObjectIndex(object) = span->firstfree; |
| 892 |
span->firstfree = (object - superblock) / size; |
| 893 |
++span->nallocatable; |
| 894 |
|
| 895 |
/* |
| 896 |
* See if the span needs to moved to a different fullness class, or be |
| 897 |
* freed so its pages can be given back to the segment. |
| 898 |
*/ |
| 899 |
if (span->nallocatable == 1 && span->fclass == DSA_FULLNESS_CLASSES - 1) |
| 900 |
{ |
| 901 |
/* |
| 902 |
* The block was completely full and is located in the |
| 903 |
* highest-numbered fullness class, which is never scanned for free |
| 904 |
* chunks. We must move it to the next-lower fullness class. |
| 905 |
*/ |
| 906 |
unlink_span(area, span); |
| 907 |
add_span_to_fullness_class(area, span, span_pointer, |
| 908 |
DSA_FULLNESS_CLASSES - 2); |
| 909 |
|
| 910 |
/* |
| 911 |
* If this is the only span, and there is no active span, then we |
| 912 |
* should probably move this span to fullness class 1. (Otherwise if |
| 913 |
* you allocate exactly all the objects in the only span, it moves to |
| 914 |
* class 3, then you free them all, it moves to 2, and then is given |
| 915 |
* back, leaving no active span). |
| 916 |
*/ |
| 917 |
} |
| 918 |
else if (span->nallocatable == span->nmax && |
| 919 |
(span->fclass != 1 || span->prevspan != InvalidDsaPointer)) |
| 920 |
{ |
| 921 |
/* |
| 922 |
* This entire block is free, and it's not the active block for this |
| 923 |
* size class. Return the memory to the free page manager. We don't |
| 924 |
* do this for the active block to prevent hysteresis: if we |
| 925 |
* repeatedly allocate and free the only chunk in the active block, it |
| 926 |
* will be very inefficient if we deallocate and reallocate the block |
| 927 |
* every time. |
| 928 |
*/ |
| 929 |
destroy_superblock(area, span_pointer); |
| 930 |
} |
| 931 |
|
| 932 |
LWLockRelease(DSA_SCLASS_LOCK(area, size_class)); |
| 933 |
} |
| 934 |
|
| 935 |
/* |
| 936 |
* Obtain a backend-local address for a dsa_pointer. 'dp' must point to |
| 937 |
* memory allocated by the given area (possibly in another process) that |
| 938 |
* hasn't yet been freed. This may cause a segment to be mapped into the |
| 939 |
* current process if required, and may cause freed segments to be unmapped. |
| 940 |
*/ |
| 941 |
void * |
| 942 |
dsa_get_address(dsa_area *area, dsa_pointer dp) |
| 943 |
{ |
| 944 |
dsa_segment_index index; |
| 945 |
size_t offset; |
| 946 |
|
| 947 |
/* Convert InvalidDsaPointer to NULL. */ |
| 948 |
if (!DsaPointerIsValid(dp)) |
| 949 |
return NULL; |
| 950 |
|
| 951 |
/* Process any requests to detach from freed segments. */ |
| 952 |
check_for_freed_segments(area); |
| 953 |
|
| 954 |
/* Break the dsa_pointer into its components. */ |
| 955 |
index = DSA_EXTRACT_SEGMENT_NUMBER(dp); |
| 956 |
offset = DSA_EXTRACT_OFFSET(dp); |
| 957 |
Assert(index < DSA_MAX_SEGMENTS); |
| 958 |
|
| 959 |
/* Check if we need to cause this segment to be mapped in. */ |
| 960 |
if (unlikely(area->segment_maps[index].mapped_address == NULL)) |
| 961 |
{ |
| 962 |
/* Call for effect (we don't need the result). */ |
| 963 |
get_segment_by_index(area, index); |
| 964 |
} |
| 965 |
|
| 966 |
return area->segment_maps[index].mapped_address + offset; |
| 967 |
} |
| 968 |
|
| 969 |
/* |
| 970 |
* Pin this area, so that it will continue to exist even if all backends |
| 971 |
* detach from it. In that case, the area can still be reattached to if a |
| 972 |
* handle has been recorded somewhere. |
| 973 |
*/ |
| 974 |
void |
| 975 |
dsa_pin(dsa_area *area) |
| 976 |
{ |
| 977 |
LWLockAcquire(DSA_AREA_LOCK(area), LW_EXCLUSIVE); |
| 978 |
if (area->control->pinned) |
| 979 |
{ |
| 980 |
LWLockRelease(DSA_AREA_LOCK(area)); |
| 981 |
elog(ERROR, "dsa_area already pinned"); |
| 982 |
} |
| 983 |
area->control->pinned = true; |
| 984 |
++area->control->refcnt; |
| 985 |
LWLockRelease(DSA_AREA_LOCK(area)); |
| 986 |
} |
| 987 |
|
| 988 |
/* |
| 989 |
* Undo the effects of dsa_pin, so that the given area can be freed when no |
| 990 |
* backends are attached to it. May be called only if dsa_pin has been |
| 991 |
* called. |
| 992 |
*/ |
| 993 |
void |
| 994 |
dsa_unpin(dsa_area *area) |
| 995 |
{ |
| 996 |
LWLockAcquire(DSA_AREA_LOCK(area), LW_EXCLUSIVE); |
| 997 |
Assert(area->control->refcnt > 1); |
| 998 |
if (!area->control->pinned) |
| 999 |
{ |
| 1000 |
LWLockRelease(DSA_AREA_LOCK(area)); |
| 1001 |
elog(ERROR, "dsa_area not pinned"); |
| 1002 |
} |
| 1003 |
area->control->pinned = false; |
| 1004 |
--area->control->refcnt; |
| 1005 |
LWLockRelease(DSA_AREA_LOCK(area)); |
| 1006 |
} |
| 1007 |
|
| 1008 |
/* |
| 1009 |
* Set the total size limit for this area. This limit is checked whenever new |
| 1010 |
* segments need to be allocated from the operating system. If the new size |
| 1011 |
* limit is already exceeded, this has no immediate effect. |
| 1012 |
* |
| 1013 |
* Note that the total virtual memory usage may be temporarily larger than |
| 1014 |
* this limit when segments have been freed, but not yet detached by all |
| 1015 |
* backends that have attached to them. |
| 1016 |
*/ |
| 1017 |
void |
| 1018 |
dsa_set_size_limit(dsa_area *area, size_t limit) |
| 1019 |
{ |
| 1020 |
LWLockAcquire(DSA_AREA_LOCK(area), LW_EXCLUSIVE); |
| 1021 |
area->control->max_total_segment_size = limit; |
| 1022 |
LWLockRelease(DSA_AREA_LOCK(area)); |
| 1023 |
} |
| 1024 |
|
| 1025 |
/* Return the total size of all active segments */ |
| 1026 |
size_t |
| 1027 |
dsa_get_total_size(dsa_area *area) |
| 1028 |
{ |
| 1029 |
size_t size; |
| 1030 |
|
| 1031 |
LWLockAcquire(DSA_AREA_LOCK(area), LW_EXCLUSIVE); |
| 1032 |
size = area->control->total_segment_size; |
| 1033 |
LWLockRelease(DSA_AREA_LOCK(area)); |
| 1034 |
|
| 1035 |
return size; |
| 1036 |
} |
| 1037 |
|
| 1038 |
/* |
| 1039 |
* Aggressively free all spare memory in the hope of returning DSM segments to |
| 1040 |
* the operating system. |
| 1041 |
*/ |
| 1042 |
void |
| 1043 |
dsa_trim(dsa_area *area) |
| 1044 |
{ |
| 1045 |
int size_class; |
| 1046 |
|
| 1047 |
/* |
| 1048 |
* Trim in reverse pool order so we get to the spans-of-spans last, just |
| 1049 |
* in case any become entirely free while processing all the other pools. |
| 1050 |
*/ |
| 1051 |
for (size_class = DSA_NUM_SIZE_CLASSES - 1; size_class >= 0; --size_class) |
| 1052 |
{ |
| 1053 |
dsa_area_pool *pool = &area->control->pools[size_class]; |
| 1054 |
dsa_pointer span_pointer; |
| 1055 |
|
| 1056 |
if (size_class == DSA_SCLASS_SPAN_LARGE) |
| 1057 |
{ |
| 1058 |
/* Large object frees give back segments aggressively already. */ |
| 1059 |
continue; |
| 1060 |
} |
| 1061 |
|
| 1062 |
/* |
| 1063 |
* Search fullness class 1 only. That is where we expect to find an |
| 1064 |
* entirely empty superblock (entirely empty superblocks in other |
| 1065 |
* fullness classes are returned to the free page map by dsa_free). |
| 1066 |
*/ |
| 1067 |
LWLockAcquire(DSA_SCLASS_LOCK(area, size_class), LW_EXCLUSIVE); |
| 1068 |
span_pointer = pool->spans[1]; |
| 1069 |
while (DsaPointerIsValid(span_pointer)) |
| 1070 |
{ |
| 1071 |
dsa_area_span *span = dsa_get_address(area, span_pointer); |
| 1072 |
dsa_pointer next = span->nextspan; |
| 1073 |
|
| 1074 |
if (span->nallocatable == span->nmax) |
| 1075 |
destroy_superblock(area, span_pointer); |
| 1076 |
|
| 1077 |
span_pointer = next; |
| 1078 |
} |
| 1079 |
LWLockRelease(DSA_SCLASS_LOCK(area, size_class)); |
| 1080 |
} |
| 1081 |
} |
| 1082 |
|
| 1083 |
/* |
| 1084 |
* Print out debugging information about the internal state of the shared |
| 1085 |
* memory area. |
| 1086 |
*/ |
| 1087 |
void |
| 1088 |
dsa_dump(dsa_area *area) |
| 1089 |
{ |
| 1090 |
size_t i, |
| 1091 |
j; |
| 1092 |
|
| 1093 |
/* |
| 1094 |
* Note: This gives an inconsistent snapshot as it acquires and releases |
| 1095 |
* individual locks as it goes... |
| 1096 |
*/ |
| 1097 |
|
| 1098 |
LWLockAcquire(DSA_AREA_LOCK(area), LW_EXCLUSIVE); |
| 1099 |
check_for_freed_segments_locked(area); |
| 1100 |
fprintf(stderr, "dsa_area handle %x:\n", area->control->handle); |
| 1101 |
fprintf(stderr, " max_total_segment_size: %zu\n", |
| 1102 |
area->control->max_total_segment_size); |
| 1103 |
fprintf(stderr, " total_segment_size: %zu\n", |
| 1104 |
area->control->total_segment_size); |
| 1105 |
fprintf(stderr, " refcnt: %d\n", area->control->refcnt); |
| 1106 |
fprintf(stderr, " pinned: %c\n", area->control->pinned ? 't' : 'f'); |
| 1107 |
fprintf(stderr, " segment bins:\n"); |
| 1108 |
for (i = 0; i < DSA_NUM_SEGMENT_BINS; ++i) |
| 1109 |
{ |
| 1110 |
if (area->control->segment_bins[i] != DSA_SEGMENT_INDEX_NONE) |
| 1111 |
{ |
| 1112 |
dsa_segment_index segment_index; |
| 1113 |
|
| 1114 |
if (i == 0) |
| 1115 |
fprintf(stderr, |
| 1116 |
" segment bin %zu (no contiguous free pages):\n", i); |
| 1117 |
else |
| 1118 |
fprintf(stderr, |
| 1119 |
" segment bin %zu (at least %d contiguous pages free):\n", |
| 1120 |
i, 1 << (i - 1)); |
| 1121 |
segment_index = area->control->segment_bins[i]; |
| 1122 |
while (segment_index != DSA_SEGMENT_INDEX_NONE) |
| 1123 |
{ |
| 1124 |
dsa_segment_map *segment_map; |
| 1125 |
|
| 1126 |
segment_map = |
| 1127 |
get_segment_by_index(area, segment_index); |
| 1128 |
|
| 1129 |
fprintf(stderr, |
| 1130 |
" segment index %zu, usable_pages = %zu, " |
| 1131 |
"contiguous_pages = %zu, mapped at %p\n", |
| 1132 |
segment_index, |
| 1133 |
segment_map->header->usable_pages, |
| 1134 |
fpm_largest(segment_map->fpm), |
| 1135 |
segment_map->mapped_address); |
| 1136 |
segment_index = segment_map->header->next; |
| 1137 |
} |
| 1138 |
} |
| 1139 |
} |
| 1140 |
LWLockRelease(DSA_AREA_LOCK(area)); |
| 1141 |
|
| 1142 |
fprintf(stderr, " pools:\n"); |
| 1143 |
for (i = 0; i < DSA_NUM_SIZE_CLASSES; ++i) |
| 1144 |
{ |
| 1145 |
bool found = false; |
| 1146 |
|
| 1147 |
LWLockAcquire(DSA_SCLASS_LOCK(area, i), LW_EXCLUSIVE); |
| 1148 |
for (j = 0; j < DSA_FULLNESS_CLASSES; ++j) |
| 1149 |
if (DsaPointerIsValid(area->control->pools[i].spans[j])) |
| 1150 |
found = true; |
| 1151 |
if (found) |
| 1152 |
{ |
| 1153 |
if (i == DSA_SCLASS_BLOCK_OF_SPANS) |
| 1154 |
fprintf(stderr, " pool for blocks of span objects:\n"); |
| 1155 |
else if (i == DSA_SCLASS_SPAN_LARGE) |
| 1156 |
fprintf(stderr, " pool for large object spans:\n"); |
| 1157 |
else |
| 1158 |
fprintf(stderr, |
| 1159 |
" pool for size class %zu (object size %hu bytes):\n", |
| 1160 |
i, dsa_size_classes[i]); |
| 1161 |
for (j = 0; j < DSA_FULLNESS_CLASSES; ++j) |
| 1162 |
{ |
| 1163 |
if (!DsaPointerIsValid(area->control->pools[i].spans[j])) |
| 1164 |
fprintf(stderr, " fullness class %zu is empty\n", j); |
| 1165 |
else |
| 1166 |
{ |
| 1167 |
dsa_pointer span_pointer = area->control->pools[i].spans[j]; |
| 1168 |
|
| 1169 |
fprintf(stderr, " fullness class %zu:\n", j); |
| 1170 |
while (DsaPointerIsValid(span_pointer)) |
| 1171 |
{ |
| 1172 |
dsa_area_span *span; |
| 1173 |
|
| 1174 |
span = dsa_get_address(area, span_pointer); |
| 1175 |
fprintf(stderr, |
| 1176 |
" span descriptor at " |
| 1177 |
DSA_POINTER_FORMAT ", superblock at " |
| 1178 |
DSA_POINTER_FORMAT |
| 1179 |
", pages = %zu, objects free = %hu/%hu\n", |
| 1180 |
span_pointer, span->start, span->npages, |
| 1181 |
span->nallocatable, span->nmax); |
| 1182 |
span_pointer = span->nextspan; |
| 1183 |
} |
| 1184 |
} |
| 1185 |
} |
| 1186 |
} |
| 1187 |
LWLockRelease(DSA_SCLASS_LOCK(area, i)); |
| 1188 |
} |
| 1189 |
} |
| 1190 |
|
| 1191 |
/* |
| 1192 |
* Return the smallest size that you can successfully provide to |
| 1193 |
* dsa_create_in_place. |
| 1194 |
*/ |
| 1195 |
size_t |
| 1196 |
dsa_minimum_size(void) |
| 1197 |
{ |
| 1198 |
size_t size; |
| 1199 |
int pages = 0; |
| 1200 |
|
| 1201 |
size = MAXALIGN(sizeof(dsa_area_control)) + |
| 1202 |
MAXALIGN(sizeof(FreePageManager)); |
| 1203 |
|
| 1204 |
/* Figure out how many pages we need, including the page map... */ |
| 1205 |
while (((size + FPM_PAGE_SIZE - 1) / FPM_PAGE_SIZE) > pages) |
| 1206 |
{ |
| 1207 |
++pages; |
| 1208 |
size += sizeof(dsa_pointer); |
| 1209 |
} |
| 1210 |
|
| 1211 |
return pages * FPM_PAGE_SIZE; |
| 1212 |
} |
| 1213 |
|
| 1214 |
/* |
| 1215 |
* Workhorse function for dsa_create and dsa_create_in_place. |
| 1216 |
*/ |
| 1217 |
static dsa_area * |
| 1218 |
create_internal(void *place, size_t size, |
| 1219 |
int tranche_id, |
| 1220 |
dsm_handle control_handle, |
| 1221 |
dsm_segment *control_segment, |
| 1222 |
size_t init_segment_size, size_t max_segment_size) |
| 1223 |
{ |
| 1224 |
dsa_area_control *control; |
| 1225 |
dsa_area *area; |
| 1226 |
dsa_segment_map *segment_map; |
| 1227 |
size_t usable_pages; |
| 1228 |
size_t total_pages; |
| 1229 |
size_t metadata_bytes; |
| 1230 |
int i; |
| 1231 |
|
| 1232 |
/* Check the initial and maximum block sizes */ |
| 1233 |
Assert(init_segment_size >= DSA_MIN_SEGMENT_SIZE); |
| 1234 |
Assert(max_segment_size >= init_segment_size); |
| 1235 |
Assert(max_segment_size <= DSA_MAX_SEGMENT_SIZE); |
| 1236 |
|
| 1237 |
/* Sanity check on the space we have to work in. */ |
| 1238 |
if (size < dsa_minimum_size()) |
| 1239 |
elog(ERROR, "dsa_area space must be at least %zu, but %zu provided", |
| 1240 |
dsa_minimum_size(), size); |
| 1241 |
|
| 1242 |
/* Now figure out how much space is usable */ |
| 1243 |
total_pages = size / FPM_PAGE_SIZE; |
| 1244 |
metadata_bytes = |
| 1245 |
MAXALIGN(sizeof(dsa_area_control)) + |
| 1246 |
MAXALIGN(sizeof(FreePageManager)) + |
| 1247 |
total_pages * sizeof(dsa_pointer); |
| 1248 |
/* Add padding up to next page boundary. */ |
| 1249 |
if (metadata_bytes % FPM_PAGE_SIZE != 0) |
| 1250 |
metadata_bytes += FPM_PAGE_SIZE - (metadata_bytes % FPM_PAGE_SIZE); |
| 1251 |
Assert(metadata_bytes <= size); |
| 1252 |
usable_pages = (size - metadata_bytes) / FPM_PAGE_SIZE; |
| 1253 |
|
| 1254 |
/* |
| 1255 |
* Initialize the dsa_area_control object located at the start of the |
| 1256 |
* space. |
| 1257 |
*/ |
| 1258 |
control = (dsa_area_control *) place; |
| 1259 |
memset(place, 0, sizeof(*control)); |
| 1260 |
control->segment_header.magic = |
| 1261 |
DSA_SEGMENT_HEADER_MAGIC ^ control_handle ^ 0; |
| 1262 |
control->segment_header.next = DSA_SEGMENT_INDEX_NONE; |
| 1263 |
control->segment_header.prev = DSA_SEGMENT_INDEX_NONE; |
| 1264 |
control->segment_header.usable_pages = usable_pages; |
| 1265 |
control->segment_header.freed = false; |
| 1266 |
control->segment_header.size = size; |
| 1267 |
control->handle = control_handle; |
| 1268 |
control->init_segment_size = init_segment_size; |
| 1269 |
control->max_segment_size = max_segment_size; |
| 1270 |
control->max_total_segment_size = (size_t) -1; |
| 1271 |
control->total_segment_size = size; |
| 1272 |
control->segment_handles[0] = control_handle; |
| 1273 |
for (i = 0; i < DSA_NUM_SEGMENT_BINS; ++i) |
| 1274 |
control->segment_bins[i] = DSA_SEGMENT_INDEX_NONE; |
| 1275 |
control->refcnt = 1; |
| 1276 |
control->lwlock_tranche_id = tranche_id; |
| 1277 |
|
| 1278 |
/* |
| 1279 |
* Create the dsa_area object that this backend will use to access the |
| 1280 |
* area. Other backends will need to obtain their own dsa_area object by |
| 1281 |
* attaching. |
| 1282 |
*/ |
| 1283 |
area = palloc(sizeof(dsa_area)); |
| 1284 |
area->control = control; |
| 1285 |
area->resowner = CurrentResourceOwner; |
| 1286 |
memset(area->segment_maps, 0, sizeof(dsa_segment_map) * DSA_MAX_SEGMENTS); |
| 1287 |
area->high_segment_index = 0; |
| 1288 |
area->freed_segment_counter = 0; |
| 1289 |
LWLockInitialize(&control->lock, control->lwlock_tranche_id); |
| 1290 |
for (i = 0; i < DSA_NUM_SIZE_CLASSES; ++i) |
| 1291 |
LWLockInitialize(DSA_SCLASS_LOCK(area, i), |
| 1292 |
control->lwlock_tranche_id); |
| 1293 |
|
| 1294 |
/* Set up the segment map for this process's mapping. */ |
| 1295 |
segment_map = &area->segment_maps[0]; |
| 1296 |
segment_map->segment = control_segment; |
| 1297 |
segment_map->mapped_address = place; |
| 1298 |
segment_map->header = (dsa_segment_header *) place; |
| 1299 |
segment_map->fpm = (FreePageManager *) |
| 1300 |
(segment_map->mapped_address + |
| 1301 |
MAXALIGN(sizeof(dsa_area_control))); |
| 1302 |
segment_map->pagemap = (dsa_pointer *) |
| 1303 |
(segment_map->mapped_address + |
| 1304 |
MAXALIGN(sizeof(dsa_area_control)) + |
| 1305 |
MAXALIGN(sizeof(FreePageManager))); |
| 1306 |
|
| 1307 |
/* Set up the free page map. */ |
| 1308 |
FreePageManagerInitialize(segment_map->fpm, segment_map->mapped_address); |
| 1309 |
/* There can be 0 usable pages if size is dsa_minimum_size(). */ |
| 1310 |
|
| 1311 |
if (usable_pages > 0) |
| 1312 |
FreePageManagerPut(segment_map->fpm, metadata_bytes / FPM_PAGE_SIZE, |
| 1313 |
usable_pages); |
| 1314 |
|
| 1315 |
/* Put this segment into the appropriate bin. */ |
| 1316 |
control->segment_bins[contiguous_pages_to_segment_bin(usable_pages)] = 0; |
| 1317 |
segment_map->header->bin = contiguous_pages_to_segment_bin(usable_pages); |
| 1318 |
|
| 1319 |
return area; |
| 1320 |
} |
| 1321 |
|
| 1322 |
/* |
| 1323 |
* Workhorse function for dsa_attach and dsa_attach_in_place. |
| 1324 |
*/ |
| 1325 |
static dsa_area * |
| 1326 |
attach_internal(void *place, dsm_segment *segment, dsa_handle handle) |
| 1327 |
{ |
| 1328 |
dsa_area_control *control; |
| 1329 |
dsa_area *area; |
| 1330 |
dsa_segment_map *segment_map; |
| 1331 |
|
| 1332 |
control = (dsa_area_control *) place; |
| 1333 |
Assert(control->handle == handle); |
| 1334 |
Assert(control->segment_handles[0] == handle); |
| 1335 |
Assert(control->segment_header.magic == |
| 1336 |
(DSA_SEGMENT_HEADER_MAGIC ^ handle ^ 0)); |
| 1337 |
|
| 1338 |
/* Build the backend-local area object. */ |
| 1339 |
area = palloc(sizeof(dsa_area)); |
| 1340 |
area->control = control; |
| 1341 |
area->resowner = CurrentResourceOwner; |
| 1342 |
memset(&area->segment_maps[0], 0, |
| 1343 |
sizeof(dsa_segment_map) * DSA_MAX_SEGMENTS); |
| 1344 |
area->high_segment_index = 0; |
| 1345 |
|
| 1346 |
/* Set up the segment map for this process's mapping. */ |
| 1347 |
segment_map = &area->segment_maps[0]; |
| 1348 |
segment_map->segment = segment; /* NULL for in-place */ |
| 1349 |
segment_map->mapped_address = place; |
| 1350 |
segment_map->header = (dsa_segment_header *) segment_map->mapped_address; |
| 1351 |
segment_map->fpm = (FreePageManager *) |
| 1352 |
(segment_map->mapped_address + MAXALIGN(sizeof(dsa_area_control))); |
| 1353 |
segment_map->pagemap = (dsa_pointer *) |
| 1354 |
(segment_map->mapped_address + MAXALIGN(sizeof(dsa_area_control)) + |
| 1355 |
MAXALIGN(sizeof(FreePageManager))); |
| 1356 |
|
| 1357 |
/* Bump the reference count. */ |
| 1358 |
LWLockAcquire(DSA_AREA_LOCK(area), LW_EXCLUSIVE); |
| 1359 |
if (control->refcnt == 0) |
| 1360 |
{ |
| 1361 |
/* We can't attach to a DSA area that has already been destroyed. */ |
| 1362 |
ereport(ERROR, |
| 1363 |
(errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE), |
| 1364 |
errmsg("could not attach to dynamic shared area"))); |
| 1365 |
} |
| 1366 |
++control->refcnt; |
| 1367 |
area->freed_segment_counter = area->control->freed_segment_counter; |
| 1368 |
LWLockRelease(DSA_AREA_LOCK(area)); |
| 1369 |
|
| 1370 |
return area; |
| 1371 |
} |
| 1372 |
|
| 1373 |
/* |
| 1374 |
* Add a new span to fullness class 1 of the indicated pool. |
| 1375 |
*/ |
| 1376 |
static void |
| 1377 |
init_span(dsa_area *area, |
| 1378 |
dsa_pointer span_pointer, |
| 1379 |
dsa_area_pool *pool, dsa_pointer start, size_t npages, |
| 1380 |
uint16 size_class) |
| 1381 |
{ |
| 1382 |
dsa_area_span *span = dsa_get_address(area, span_pointer); |
| 1383 |
size_t obsize = dsa_size_classes[size_class]; |
| 1384 |
|
| 1385 |
/* |
| 1386 |
* The per-pool lock must be held because we manipulate the span list for |
| 1387 |
* this pool. |
| 1388 |
*/ |
| 1389 |
Assert(LWLockHeldByMe(DSA_SCLASS_LOCK(area, size_class))); |
| 1390 |
|
| 1391 |
/* Push this span onto the front of the span list for fullness class 1. */ |
| 1392 |
if (DsaPointerIsValid(pool->spans[1])) |
| 1393 |
{ |
| 1394 |
dsa_area_span *head = (dsa_area_span *) |
| 1395 |
dsa_get_address(area, pool->spans[1]); |
| 1396 |
|
| 1397 |
head->prevspan = span_pointer; |
| 1398 |
} |
| 1399 |
span->pool = DsaAreaPoolToDsaPointer(area, pool); |
| 1400 |
span->nextspan = pool->spans[1]; |
| 1401 |
span->prevspan = InvalidDsaPointer; |
| 1402 |
pool->spans[1] = span_pointer; |
| 1403 |
|
| 1404 |
span->start = start; |
| 1405 |
span->npages = npages; |
| 1406 |
span->size_class = size_class; |
| 1407 |
span->ninitialized = 0; |
| 1408 |
if (size_class == DSA_SCLASS_BLOCK_OF_SPANS) |
| 1409 |
{ |
| 1410 |
/* |
| 1411 |
* A block-of-spans contains its own descriptor, so mark one object as |
| 1412 |
* initialized and reduce the count of allocatable objects by one. |
| 1413 |
* Doing this here has the side effect of also reducing nmax by one, |
| 1414 |
* which is important to make sure we free this object at the correct |
| 1415 |
* time. |
| 1416 |
*/ |
| 1417 |
span->ninitialized = 1; |
| 1418 |
span->nallocatable = FPM_PAGE_SIZE / obsize - 1; |
| 1419 |
} |
| 1420 |
else if (size_class != DSA_SCLASS_SPAN_LARGE) |
| 1421 |
span->nallocatable = DSA_SUPERBLOCK_SIZE / obsize; |
| 1422 |
span->firstfree = DSA_SPAN_NOTHING_FREE; |
| 1423 |
span->nmax = span->nallocatable; |
| 1424 |
span->fclass = 1; |
| 1425 |
} |
| 1426 |
|
| 1427 |
/* |
| 1428 |
* Transfer the first span in one fullness class to the head of another |
| 1429 |
* fullness class. |
| 1430 |
*/ |
| 1431 |
static bool |
| 1432 |
transfer_first_span(dsa_area *area, |
| 1433 |
dsa_area_pool *pool, int fromclass, int toclass) |
| 1434 |
{ |
| 1435 |
dsa_pointer span_pointer; |
| 1436 |
dsa_area_span *span; |
| 1437 |
dsa_area_span *nextspan; |
| 1438 |
|
| 1439 |
/* Can't do it if source list is empty. */ |
| 1440 |
span_pointer = pool->spans[fromclass]; |
| 1441 |
if (!DsaPointerIsValid(span_pointer)) |
| 1442 |
return false; |
| 1443 |
|
| 1444 |
/* Remove span from head of source list. */ |
| 1445 |
span = dsa_get_address(area, span_pointer); |
| 1446 |
pool->spans[fromclass] = span->nextspan; |
| 1447 |
if (DsaPointerIsValid(span->nextspan)) |
| 1448 |
{ |
| 1449 |
nextspan = (dsa_area_span *) |
| 1450 |
dsa_get_address(area, span->nextspan); |
| 1451 |
nextspan->prevspan = InvalidDsaPointer; |
| 1452 |
} |
| 1453 |
|
| 1454 |
/* Add span to head of target list. */ |
| 1455 |
span->nextspan = pool->spans[toclass]; |
| 1456 |
pool->spans[toclass] = span_pointer; |
| 1457 |
if (DsaPointerIsValid(span->nextspan)) |
| 1458 |
{ |
| 1459 |
nextspan = (dsa_area_span *) |
| 1460 |
dsa_get_address(area, span->nextspan); |
| 1461 |
nextspan->prevspan = span_pointer; |
| 1462 |
} |
| 1463 |
span->fclass = toclass; |
| 1464 |
|
| 1465 |
return true; |
| 1466 |
} |
| 1467 |
|
| 1468 |
/* |
| 1469 |
* Allocate one object of the requested size class from the given area. |
| 1470 |
*/ |
| 1471 |
static inline dsa_pointer |
| 1472 |
alloc_object(dsa_area *area, int size_class) |
| 1473 |
{ |
| 1474 |
dsa_area_pool *pool = &area->control->pools[size_class]; |
| 1475 |
dsa_area_span *span; |
| 1476 |
dsa_pointer block; |
| 1477 |
dsa_pointer result; |
| 1478 |
char *object; |
| 1479 |
size_t size; |
| 1480 |
|
| 1481 |
/* |
| 1482 |
* Even though ensure_active_superblock can in turn call alloc_object if |
| 1483 |
* it needs to allocate a new span, that's always from a different pool, |
| 1484 |
* and the order of lock acquisition is always the same, so it's OK that |
| 1485 |
* we hold this lock for the duration of this function. |
| 1486 |
*/ |
| 1487 |
Assert(!LWLockHeldByMe(DSA_SCLASS_LOCK(area, size_class))); |
| 1488 |
LWLockAcquire(DSA_SCLASS_LOCK(area, size_class), LW_EXCLUSIVE); |
| 1489 |
|
| 1490 |
/* |
| 1491 |
* If there's no active superblock, we must successfully obtain one or |
| 1492 |
* fail the request. |
| 1493 |
*/ |
| 1494 |
if (!DsaPointerIsValid(pool->spans[1]) && |
| 1495 |
!ensure_active_superblock(area, pool, size_class)) |
| 1496 |
{ |
| 1497 |
result = InvalidDsaPointer; |
| 1498 |
} |
| 1499 |
else |
| 1500 |
{ |
| 1501 |
/* |
| 1502 |
* There should be a block in fullness class 1 at this point, and it |
| 1503 |
* should never be completely full. Thus we can either pop an object |
| 1504 |
* from the free list or, failing that, initialize a new object. |
| 1505 |
*/ |
| 1506 |
Assert(DsaPointerIsValid(pool->spans[1])); |
| 1507 |
span = (dsa_area_span *) |
| 1508 |
dsa_get_address(area, pool->spans[1]); |
| 1509 |
Assert(span->nallocatable > 0); |
| 1510 |
block = span->start; |
| 1511 |
Assert(size_class < DSA_NUM_SIZE_CLASSES); |
| 1512 |
size = dsa_size_classes[size_class]; |
| 1513 |
if (span->firstfree != DSA_SPAN_NOTHING_FREE) |
| 1514 |
{ |
| 1515 |
result = block + span->firstfree * size; |
| 1516 |
object = dsa_get_address(area, result); |
| 1517 |
span->firstfree = NextFreeObjectIndex(object); |
| 1518 |
} |
| 1519 |
else |
| 1520 |
{ |
| 1521 |
result = block + span->ninitialized * size; |
| 1522 |
++span->ninitialized; |
| 1523 |
} |
| 1524 |
--span->nallocatable; |
| 1525 |
|
| 1526 |
/* If it's now full, move it to the highest-numbered fullness class. */ |
| 1527 |
if (span->nallocatable == 0) |
| 1528 |
transfer_first_span(area, pool, 1, DSA_FULLNESS_CLASSES - 1); |
| 1529 |
} |
| 1530 |
|
| 1531 |
Assert(LWLockHeldByMe(DSA_SCLASS_LOCK(area, size_class))); |
| 1532 |
LWLockRelease(DSA_SCLASS_LOCK(area, size_class)); |
| 1533 |
|
| 1534 |
return result; |
| 1535 |
} |
| 1536 |
|
| 1537 |
/* |
| 1538 |
* Ensure an active (i.e. fullness class 1) superblock, unless all existing |
| 1539 |
* superblocks are completely full and no more can be allocated. |
| 1540 |
* |
| 1541 |
* Fullness classes K of 0..N are loosely intended to represent blocks whose |
| 1542 |
* utilization percentage is at least K/N, but we only enforce this rigorously |
| 1543 |
* for the highest-numbered fullness class, which always contains exactly |
| 1544 |
* those blocks that are completely full. It's otherwise acceptable for a |
| 1545 |
* block to be in a higher-numbered fullness class than the one to which it |
| 1546 |
* logically belongs. In addition, the active block, which is always the |
| 1547 |
* first block in fullness class 1, is permitted to have a higher allocation |
| 1548 |
* percentage than would normally be allowable for that fullness class; we |
| 1549 |
* don't move it until it's completely full, and then it goes to the |
| 1550 |
* highest-numbered fullness class. |
| 1551 |
* |
| 1552 |
* It might seem odd that the active block is the head of fullness class 1 |
| 1553 |
* rather than fullness class 0, but experience with other allocators has |
| 1554 |
* shown that it's usually better to allocate from a block that's moderately |
| 1555 |
* full rather than one that's nearly empty. Insofar as is reasonably |
| 1556 |
* possible, we want to avoid performing new allocations in a block that would |
| 1557 |
* otherwise become empty soon. |
| 1558 |
*/ |
| 1559 |
static bool |
| 1560 |
ensure_active_superblock(dsa_area *area, dsa_area_pool *pool, |
| 1561 |
int size_class) |
| 1562 |
{ |
| 1563 |
dsa_pointer span_pointer; |
| 1564 |
dsa_pointer start_pointer; |
| 1565 |
size_t obsize = dsa_size_classes[size_class]; |
| 1566 |
size_t nmax; |
| 1567 |
int fclass; |
| 1568 |
size_t npages = 1; |
| 1569 |
size_t first_page; |
| 1570 |
size_t i; |
| 1571 |
dsa_segment_map *segment_map; |
| 1572 |
|
| 1573 |
Assert(LWLockHeldByMe(DSA_SCLASS_LOCK(area, size_class))); |
| 1574 |
|
| 1575 |
/* |
| 1576 |
* Compute the number of objects that will fit in a block of this size |
| 1577 |
* class. Span-of-spans blocks are just a single page, and the first |
| 1578 |
* object isn't available for use because it describes the block-of-spans |
| 1579 |
* itself. |
| 1580 |
*/ |
| 1581 |
if (size_class == DSA_SCLASS_BLOCK_OF_SPANS) |
| 1582 |
nmax = FPM_PAGE_SIZE / obsize - 1; |
| 1583 |
else |
| 1584 |
nmax = DSA_SUPERBLOCK_SIZE / obsize; |
| 1585 |
|
| 1586 |
/* |
| 1587 |
* If fullness class 1 is empty, try to find a span to put in it by |
| 1588 |
* scanning higher-numbered fullness classes (excluding the last one, |
| 1589 |
* whose blocks are certain to all be completely full). |
| 1590 |
*/ |
| 1591 |
for (fclass = 2; fclass < DSA_FULLNESS_CLASSES - 1; ++fclass) |
| 1592 |
{ |
| 1593 |
span_pointer = pool->spans[fclass]; |
| 1594 |
|
| 1595 |
while (DsaPointerIsValid(span_pointer)) |
| 1596 |
{ |
| 1597 |
int tfclass; |
| 1598 |
dsa_area_span *span; |
| 1599 |
dsa_area_span *nextspan; |
| 1600 |
dsa_area_span *prevspan; |
| 1601 |
dsa_pointer next_span_pointer; |
| 1602 |
|
| 1603 |
span = (dsa_area_span *) |
| 1604 |
dsa_get_address(area, span_pointer); |
| 1605 |
next_span_pointer = span->nextspan; |
| 1606 |
|
| 1607 |
/* Figure out what fullness class should contain this span. */ |
| 1608 |
tfclass = (nmax - span->nallocatable) |
| 1609 |
* (DSA_FULLNESS_CLASSES - 1) / nmax; |
| 1610 |
|
| 1611 |
/* Look up next span. */ |
| 1612 |
if (DsaPointerIsValid(span->nextspan)) |
| 1613 |
nextspan = (dsa_area_span *) |
| 1614 |
dsa_get_address(area, span->nextspan); |
| 1615 |
else |
| 1616 |
nextspan = NULL; |
| 1617 |
|
| 1618 |
/* |
| 1619 |
* If utilization has dropped enough that this now belongs in some |
| 1620 |
* other fullness class, move it there. |
| 1621 |
*/ |
| 1622 |
if (tfclass < fclass) |
| 1623 |
{ |
| 1624 |
/* Remove from the current fullness class list. */ |
| 1625 |
if (pool->spans[fclass] == span_pointer) |
| 1626 |
{ |
| 1627 |
/* It was the head; remove it. */ |
| 1628 |
Assert(!DsaPointerIsValid(span->prevspan)); |
| 1629 |
pool->spans[fclass] = span->nextspan; |
| 1630 |
if (nextspan != NULL) |
| 1631 |
nextspan->prevspan = InvalidDsaPointer; |
| 1632 |
} |
| 1633 |
else |
| 1634 |
{ |
| 1635 |
/* It was not the head. */ |
| 1636 |
Assert(DsaPointerIsValid(span->prevspan)); |
| 1637 |
prevspan = (dsa_area_span *) |
| 1638 |
dsa_get_address(area, span->prevspan); |
| 1639 |
prevspan->nextspan = span->nextspan; |
| 1640 |
} |
| 1641 |
if (nextspan != NULL) |
| 1642 |
nextspan->prevspan = span->prevspan; |
| 1643 |
|
| 1644 |
/* Push onto the head of the new fullness class list. */ |
| 1645 |
span->nextspan = pool->spans[tfclass]; |
| 1646 |
pool->spans[tfclass] = span_pointer; |
| 1647 |
span->prevspan = InvalidDsaPointer; |
| 1648 |
if (DsaPointerIsValid(span->nextspan)) |
| 1649 |
{ |
| 1650 |
nextspan = (dsa_area_span *) |
| 1651 |
dsa_get_address(area, span->nextspan); |
| 1652 |
nextspan->prevspan = span_pointer; |
| 1653 |
} |
| 1654 |
span->fclass = tfclass; |
| 1655 |
} |
| 1656 |
|
| 1657 |
/* Advance to next span on list. */ |
| 1658 |
span_pointer = next_span_pointer; |
| 1659 |
} |
| 1660 |
|
| 1661 |
/* Stop now if we found a suitable block. */ |
| 1662 |
if (DsaPointerIsValid(pool->spans[1])) |
| 1663 |
return true; |
| 1664 |
} |
| 1665 |
|
| 1666 |
/* |
| 1667 |
* If there are no blocks that properly belong in fullness class 1, pick |
| 1668 |
* one from some other fullness class and move it there anyway, so that we |
| 1669 |
* have an allocation target. Our last choice is to transfer a block |
| 1670 |
* that's almost empty (and might become completely empty soon if left |
| 1671 |
* alone), but even that is better than failing, which is what we must do |
| 1672 |
* if there are no blocks at all with freespace. |
| 1673 |
*/ |
| 1674 |
Assert(!DsaPointerIsValid(pool->spans[1])); |
| 1675 |
for (fclass = 2; fclass < DSA_FULLNESS_CLASSES - 1; ++fclass) |
| 1676 |
if (transfer_first_span(area, pool, fclass, 1)) |
| 1677 |
return true; |
| 1678 |
if (!DsaPointerIsValid(pool->spans[1]) && |
| 1679 |
transfer_first_span(area, pool, 0, 1)) |
| 1680 |
return true; |
| 1681 |
|
| 1682 |
/* |
| 1683 |
* We failed to find an existing span with free objects, so we need to |
| 1684 |
* allocate a new superblock and construct a new span to manage it. |
| 1685 |
* |
| 1686 |
* First, get a dsa_area_span object to describe the new superblock block |
| 1687 |
* ... unless this allocation is for a dsa_area_span object, in which case |
| 1688 |
* that's surely not going to work. We handle that case by storing the |
| 1689 |
* span describing a block-of-spans inline. |
| 1690 |
*/ |
| 1691 |
if (size_class != DSA_SCLASS_BLOCK_OF_SPANS) |
| 1692 |
{ |
| 1693 |
span_pointer = alloc_object(area, DSA_SCLASS_BLOCK_OF_SPANS); |
| 1694 |
if (!DsaPointerIsValid(span_pointer)) |
| 1695 |
return false; |
| 1696 |
npages = DSA_PAGES_PER_SUPERBLOCK; |
| 1697 |
} |
| 1698 |
|
| 1699 |
/* Find or create a segment and allocate the superblock. */ |
| 1700 |
LWLockAcquire(DSA_AREA_LOCK(area), LW_EXCLUSIVE); |
| 1701 |
segment_map = get_best_segment(area, npages); |
| 1702 |
if (segment_map == NULL) |
| 1703 |
{ |
| 1704 |
segment_map = make_new_segment(area, npages); |
| 1705 |
if (segment_map == NULL) |
| 1706 |
{ |
| 1707 |
LWLockRelease(DSA_AREA_LOCK(area)); |
| 1708 |
return false; |
| 1709 |
} |
| 1710 |
} |
| 1711 |
|
| 1712 |
/* |
| 1713 |
* This shouldn't happen: get_best_segment() or make_new_segment() |
| 1714 |
* promised that we can successfully allocate npages. |
| 1715 |
*/ |
| 1716 |
if (!FreePageManagerGet(segment_map->fpm, npages, &first_page)) |
| 1717 |
elog(FATAL, |
| 1718 |
"dsa_allocate could not find %zu free pages for superblock", |
| 1719 |
npages); |
| 1720 |
LWLockRelease(DSA_AREA_LOCK(area)); |
| 1721 |
|
| 1722 |
/* Compute the start of the superblock. */ |
| 1723 |
start_pointer = |
| 1724 |
DSA_MAKE_POINTER(get_segment_index(area, segment_map), |
| 1725 |
first_page * FPM_PAGE_SIZE); |
| 1726 |
|
| 1727 |
/* |
| 1728 |
* If this is a block-of-spans, carve the descriptor right out of the |
| 1729 |
* allocated space. |
| 1730 |
*/ |
| 1731 |
if (size_class == DSA_SCLASS_BLOCK_OF_SPANS) |
| 1732 |
{ |
| 1733 |
/* |
| 1734 |
* We have a pointer into the segment. We need to build a dsa_pointer |
| 1735 |
* from the segment index and offset into the segment. |
| 1736 |
*/ |
| 1737 |
span_pointer = start_pointer; |
| 1738 |
} |
| 1739 |
|
| 1740 |
/* Initialize span and pagemap. */ |
| 1741 |
init_span(area, span_pointer, pool, start_pointer, npages, size_class); |
| 1742 |
for (i = 0; i < npages; ++i) |
| 1743 |
segment_map->pagemap[first_page + i] = span_pointer; |
| 1744 |
|
| 1745 |
return true; |
| 1746 |
} |
| 1747 |
|
| 1748 |
/* |
| 1749 |
* Return the segment map corresponding to a given segment index, mapping the |
| 1750 |
* segment in if necessary. For internal segment book-keeping, this is called |
| 1751 |
* with the area lock held. It is also called by dsa_free and dsa_get_address |
| 1752 |
* without any locking, relying on the fact they have a known live segment |
| 1753 |
* index and they always call check_for_freed_segments to ensures that any |
| 1754 |
* freed segment occupying the same slot is detached first. |
| 1755 |
*/ |
| 1756 |
static dsa_segment_map * |
| 1757 |
get_segment_by_index(dsa_area *area, dsa_segment_index index) |
| 1758 |
{ |
| 1759 |
if (unlikely(area->segment_maps[index].mapped_address == NULL)) |
| 1760 |
{ |
| 1761 |
dsm_handle handle; |
| 1762 |
dsm_segment *segment; |
| 1763 |
dsa_segment_map *segment_map; |
| 1764 |
ResourceOwner oldowner; |
| 1765 |
|
| 1766 |
/* |
| 1767 |
* If we are reached by dsa_free or dsa_get_address, there must be at |
| 1768 |
* least one object allocated in the referenced segment. Otherwise, |
| 1769 |
* their caller has a double-free or access-after-free bug, which we |
| 1770 |
* have no hope of detecting. So we know it's safe to access this |
| 1771 |
* array slot without holding a lock; it won't change underneath us. |
| 1772 |
* Furthermore, we know that we can see the latest contents of the |
| 1773 |
* slot, as explained in check_for_freed_segments, which those |
| 1774 |
* functions call before arriving here. |
| 1775 |
*/ |
| 1776 |
handle = area->control->segment_handles[index]; |
| 1777 |
|
| 1778 |
/* It's an error to try to access an unused slot. */ |
| 1779 |
if (handle == DSM_HANDLE_INVALID) |
| 1780 |
elog(ERROR, |
| 1781 |
"dsa_area could not attach to a segment that has been freed"); |
| 1782 |
|
| 1783 |
oldowner = CurrentResourceOwner; |
| 1784 |
CurrentResourceOwner = area->resowner; |
| 1785 |
segment = dsm_attach(handle); |
| 1786 |
CurrentResourceOwner = oldowner; |
| 1787 |
if (segment == NULL) |
| 1788 |
elog(ERROR, "dsa_area could not attach to segment"); |
| 1789 |
segment_map = &area->segment_maps[index]; |
| 1790 |
segment_map->segment = segment; |
| 1791 |
segment_map->mapped_address = dsm_segment_address(segment); |
| 1792 |
segment_map->header = |
| 1793 |
(dsa_segment_header *) segment_map->mapped_address; |
| 1794 |
segment_map->fpm = (FreePageManager *) |
| 1795 |
(segment_map->mapped_address + |
| 1796 |
MAXALIGN(sizeof(dsa_segment_header))); |
| 1797 |
segment_map->pagemap = (dsa_pointer *) |
| 1798 |
(segment_map->mapped_address + |
| 1799 |
MAXALIGN(sizeof(dsa_segment_header)) + |
| 1800 |
MAXALIGN(sizeof(FreePageManager))); |
| 1801 |
|
| 1802 |
/* Remember the highest index this backend has ever mapped. */ |
| 1803 |
if (area->high_segment_index < index) |
| 1804 |
area->high_segment_index = index; |
| 1805 |
|
| 1806 |
Assert(segment_map->header->magic == |
| 1807 |
(DSA_SEGMENT_HEADER_MAGIC ^ area->control->handle ^ index)); |
| 1808 |
} |
| 1809 |
|
| 1810 |
/* |
| 1811 |
* Callers of dsa_get_address() and dsa_free() don't hold the area lock, |
| 1812 |
* but it's a bug in the calling code and undefined behavior if the |
| 1813 |
* address is not live (ie if the segment might possibly have been freed, |
| 1814 |
* they're trying to use a dangling pointer). |
| 1815 |
* |
| 1816 |
* For dsa.c code that holds the area lock to manipulate segment_bins |
| 1817 |
* lists, it would be a bug if we ever reach a freed segment here. After |
| 1818 |
* it's marked as freed, the only thing any backend should do with it is |
| 1819 |
* unmap it, and it should always have done that in |
| 1820 |
* check_for_freed_segments_locked() before arriving here to resolve an |
| 1821 |
* index to a segment_map. |
| 1822 |
* |
| 1823 |
* Either way we can assert that we aren't returning a freed segment. |
| 1824 |
*/ |
| 1825 |
Assert(!area->segment_maps[index].header->freed); |
| 1826 |
|
| 1827 |
return &area->segment_maps[index]; |
| 1828 |
} |
| 1829 |
|
| 1830 |
/* |
| 1831 |
* Return a superblock to the free page manager. If the underlying segment |
| 1832 |
* has become entirely free, then return it to the operating system. |
| 1833 |
* |
| 1834 |
* The appropriate pool lock must be held. |
| 1835 |
*/ |
| 1836 |
static void |
| 1837 |
destroy_superblock(dsa_area *area, dsa_pointer span_pointer) |
| 1838 |
{ |
| 1839 |
dsa_area_span *span = dsa_get_address(area, span_pointer); |
| 1840 |
int size_class = span->size_class; |
| 1841 |
dsa_segment_map *segment_map; |
| 1842 |
|
| 1843 |
|
| 1844 |
/* Remove it from its fullness class list. */ |
| 1845 |
unlink_span(area, span); |
| 1846 |
|
| 1847 |
/* |
| 1848 |
* Note: Here we acquire the area lock while we already hold a per-pool |
| 1849 |
* lock. We never hold the area lock and then take a pool lock, or we |
| 1850 |
* could deadlock. |
| 1851 |
*/ |
| 1852 |
LWLockAcquire(DSA_AREA_LOCK(area), LW_EXCLUSIVE); |
| 1853 |
check_for_freed_segments_locked(area); |
| 1854 |
segment_map = |
| 1855 |
get_segment_by_index(area, DSA_EXTRACT_SEGMENT_NUMBER(span->start)); |
| 1856 |
FreePageManagerPut(segment_map->fpm, |
| 1857 |
DSA_EXTRACT_OFFSET(span->start) / FPM_PAGE_SIZE, |
| 1858 |
span->npages); |
| 1859 |
/* Check if the segment is now entirely free. */ |
| 1860 |
if (fpm_largest(segment_map->fpm) == segment_map->header->usable_pages) |
| 1861 |
{ |
| 1862 |
dsa_segment_index index = get_segment_index(area, segment_map); |
| 1863 |
|
| 1864 |
/* If it's not the segment with extra control data, free it. */ |
| 1865 |
if (index != 0) |
| 1866 |
{ |
| 1867 |
/* |
| 1868 |
* Give it back to the OS, and allow other backends to detect that |
| 1869 |
* they need to detach. |
| 1870 |
*/ |
| 1871 |
unlink_segment(area, segment_map); |
| 1872 |
segment_map->header->freed = true; |
| 1873 |
Assert(area->control->total_segment_size >= |
| 1874 |
segment_map->header->size); |
| 1875 |
area->control->total_segment_size -= |
| 1876 |
segment_map->header->size; |
| 1877 |
dsm_unpin_segment(dsm_segment_handle(segment_map->segment)); |
| 1878 |
dsm_detach(segment_map->segment); |
| 1879 |
area->control->segment_handles[index] = DSM_HANDLE_INVALID; |
| 1880 |
++area->control->freed_segment_counter; |
| 1881 |
segment_map->segment = NULL; |
| 1882 |
segment_map->header = NULL; |
| 1883 |
segment_map->mapped_address = NULL; |
| 1884 |
} |
| 1885 |
} |
| 1886 |
|
| 1887 |
/* Move segment to appropriate bin if necessary. */ |
| 1888 |
if (segment_map->header != NULL) |
| 1889 |
rebin_segment(area, segment_map); |
| 1890 |
|
| 1891 |
LWLockRelease(DSA_AREA_LOCK(area)); |
| 1892 |
|
| 1893 |
/* |
| 1894 |
* Span-of-spans blocks store the span which describes them within the |
| 1895 |
* block itself, so freeing the storage implicitly frees the descriptor |
| 1896 |
* also. If this is a block of any other type, we need to separately free |
| 1897 |
* the span object also. This recursive call to dsa_free will acquire the |
| 1898 |
* span pool's lock. We can't deadlock because the acquisition order is |
| 1899 |
* always some other pool and then the span pool. |
| 1900 |
*/ |
| 1901 |
if (size_class != DSA_SCLASS_BLOCK_OF_SPANS) |
| 1902 |
dsa_free(area, span_pointer); |
| 1903 |
} |
| 1904 |
|
| 1905 |
static void |
| 1906 |
unlink_span(dsa_area *area, dsa_area_span *span) |
| 1907 |
{ |
| 1908 |
if (DsaPointerIsValid(span->nextspan)) |
| 1909 |
{ |
| 1910 |
dsa_area_span *next = dsa_get_address(area, span->nextspan); |
| 1911 |
|
| 1912 |
next->prevspan = span->prevspan; |
| 1913 |
} |
| 1914 |
if (DsaPointerIsValid(span->prevspan)) |
| 1915 |
{ |
| 1916 |
dsa_area_span *prev = dsa_get_address(area, span->prevspan); |
| 1917 |
|
| 1918 |
prev->nextspan = span->nextspan; |
| 1919 |
} |
| 1920 |
else |
| 1921 |
{ |
| 1922 |
dsa_area_pool *pool = dsa_get_address(area, span->pool); |
| 1923 |
|
| 1924 |
pool->spans[span->fclass] = span->nextspan; |
| 1925 |
} |
| 1926 |
} |
| 1927 |
|
| 1928 |
static void |
| 1929 |
add_span_to_fullness_class(dsa_area *area, dsa_area_span *span, |
| 1930 |
dsa_pointer span_pointer, |
| 1931 |
int fclass) |
| 1932 |
{ |
| 1933 |
dsa_area_pool *pool = dsa_get_address(area, span->pool); |
| 1934 |
|
| 1935 |
if (DsaPointerIsValid(pool->spans[fclass])) |
| 1936 |
{ |
| 1937 |
dsa_area_span *head = dsa_get_address(area, |
| 1938 |
pool->spans[fclass]); |
| 1939 |
|
| 1940 |
head->prevspan = span_pointer; |
| 1941 |
} |
| 1942 |
span->prevspan = InvalidDsaPointer; |
| 1943 |
span->nextspan = pool->spans[fclass]; |
| 1944 |
pool->spans[fclass] = span_pointer; |
| 1945 |
span->fclass = fclass; |
| 1946 |
} |
| 1947 |
|
| 1948 |
/* |
| 1949 |
* Detach from an area that was either created or attached to by this process. |
| 1950 |
*/ |
| 1951 |
void |
| 1952 |
dsa_detach(dsa_area *area) |
| 1953 |
{ |
| 1954 |
int i; |
| 1955 |
|
| 1956 |
/* Detach from all segments. */ |
| 1957 |
for (i = 0; i <= area->high_segment_index; ++i) |
| 1958 |
if (area->segment_maps[i].segment != NULL) |
| 1959 |
dsm_detach(area->segment_maps[i].segment); |
| 1960 |
|
| 1961 |
/* |
| 1962 |
* Note that 'detaching' (= detaching from DSM segments) doesn't include |
| 1963 |
* 'releasing' (= adjusting the reference count). It would be nice to |
| 1964 |
* combine these operations, but client code might never get around to |
| 1965 |
* calling dsa_detach because of an error path, and a detach hook on any |
| 1966 |
* particular segment is too late to detach other segments in the area |
| 1967 |
* without risking a 'leak' warning in the non-error path. |
| 1968 |
*/ |
| 1969 |
|
| 1970 |
/* Free the backend-local area object. */ |
| 1971 |
pfree(area); |
| 1972 |
} |
| 1973 |
|
| 1974 |
/* |
| 1975 |
* Unlink a segment from the bin that contains it. |
| 1976 |
*/ |
| 1977 |
static void |
| 1978 |
unlink_segment(dsa_area *area, dsa_segment_map *segment_map) |
| 1979 |
{ |
| 1980 |
if (segment_map->header->prev != DSA_SEGMENT_INDEX_NONE) |
| 1981 |
{ |
| 1982 |
dsa_segment_map *prev; |
| 1983 |
|
| 1984 |
prev = get_segment_by_index(area, segment_map->header->prev); |
| 1985 |
prev->header->next = segment_map->header->next; |
| 1986 |
} |
| 1987 |
else |
| 1988 |
{ |
| 1989 |
Assert(area->control->segment_bins[segment_map->header->bin] == |
| 1990 |
get_segment_index(area, segment_map)); |
| 1991 |
area->control->segment_bins[segment_map->header->bin] = |
| 1992 |
segment_map->header->next; |
| 1993 |
} |
| 1994 |
if (segment_map->header->next != DSA_SEGMENT_INDEX_NONE) |
| 1995 |
{ |
| 1996 |
dsa_segment_map *next; |
| 1997 |
|
| 1998 |
next = get_segment_by_index(area, segment_map->header->next); |
| 1999 |
next->header->prev = segment_map->header->prev; |
| 2000 |
} |
| 2001 |
} |
| 2002 |
|
| 2003 |
/* |
| 2004 |
* Find a segment that could satisfy a request for 'npages' of contiguous |
| 2005 |
* memory, or return NULL if none can be found. This may involve attaching to |
| 2006 |
* segments that weren't previously attached so that we can query their free |
| 2007 |
* pages map. |
| 2008 |
*/ |
| 2009 |
static dsa_segment_map * |
| 2010 |
get_best_segment(dsa_area *area, size_t npages) |
| 2011 |
{ |
| 2012 |
size_t bin; |
| 2013 |
|
| 2014 |
Assert(LWLockHeldByMe(DSA_AREA_LOCK(area))); |
| 2015 |
check_for_freed_segments_locked(area); |
| 2016 |
|
| 2017 |
/* |
| 2018 |
* Start searching from the first bin that *might* have enough contiguous |
| 2019 |
* pages. |
| 2020 |
*/ |
| 2021 |
for (bin = contiguous_pages_to_segment_bin(npages); |
| 2022 |
bin < DSA_NUM_SEGMENT_BINS; |
| 2023 |
++bin) |
| 2024 |
{ |
| 2025 |
/* |
| 2026 |
* The minimum contiguous size that any segment in this bin should |
| 2027 |
* have. We'll re-bin if we see segments with fewer. |
| 2028 |
*/ |
| 2029 |
size_t threshold = (size_t) 1 << (bin - 1); |
| 2030 |
dsa_segment_index segment_index; |
| 2031 |
|
| 2032 |
/* Search this bin for a segment with enough contiguous space. */ |
| 2033 |
segment_index = area->control->segment_bins[bin]; |
| 2034 |
while (segment_index != DSA_SEGMENT_INDEX_NONE) |
| 2035 |
{ |
| 2036 |
dsa_segment_map *segment_map; |
| 2037 |
dsa_segment_index next_segment_index; |
| 2038 |
size_t contiguous_pages; |
| 2039 |
|
| 2040 |
segment_map = get_segment_by_index(area, segment_index); |
| 2041 |
next_segment_index = segment_map->header->next; |
| 2042 |
contiguous_pages = fpm_largest(segment_map->fpm); |
| 2043 |
|
| 2044 |
/* Not enough for the request, still enough for this bin. */ |
| 2045 |
if (contiguous_pages >= threshold && contiguous_pages < npages) |
| 2046 |
{ |
| 2047 |
segment_index = next_segment_index; |
| 2048 |
continue; |
| 2049 |
} |
| 2050 |
|
| 2051 |
/* Re-bin it if it's no longer in the appropriate bin. */ |
| 2052 |
if (contiguous_pages < threshold) |
| 2053 |
{ |
| 2054 |
rebin_segment(area, segment_map); |
| 2055 |
|
| 2056 |
/* |
| 2057 |
* But fall through to see if it's enough to satisfy this |
| 2058 |
* request anyway.... |
| 2059 |
*/ |
| 2060 |
} |
| 2061 |
|
| 2062 |
/* Check if we are done. */ |
| 2063 |
if (contiguous_pages >= npages) |
| 2064 |
return segment_map; |
| 2065 |
|
| 2066 |
/* Continue searching the same bin. */ |
| 2067 |
segment_index = next_segment_index; |
| 2068 |
} |
| 2069 |
} |
| 2070 |
|
| 2071 |
/* Not found. */ |
| 2072 |
return NULL; |
| 2073 |
} |
| 2074 |
|
| 2075 |
/* |
| 2076 |
* Create a new segment that can handle at least requested_pages. Returns |
| 2077 |
* NULL if the requested total size limit or maximum allowed number of |
| 2078 |
* segments would be exceeded. |
| 2079 |
*/ |
| 2080 |
static dsa_segment_map * |
| 2081 |
make_new_segment(dsa_area *area, size_t requested_pages) |
| 2082 |
{ |
| 2083 |
dsa_segment_index new_index; |
| 2084 |
size_t metadata_bytes; |
| 2085 |
size_t total_size; |
| 2086 |
size_t total_pages; |
| 2087 |
size_t usable_pages; |
| 2088 |
dsa_segment_map *segment_map; |
| 2089 |
dsm_segment *segment; |
| 2090 |
ResourceOwner oldowner; |
| 2091 |
|
| 2092 |
Assert(LWLockHeldByMe(DSA_AREA_LOCK(area))); |
| 2093 |
|
| 2094 |
/* Find a segment slot that is not in use (linearly for now). */ |
| 2095 |
for (new_index = 1; new_index < DSA_MAX_SEGMENTS; ++new_index) |
| 2096 |
{ |
| 2097 |
if (area->control->segment_handles[new_index] == DSM_HANDLE_INVALID) |
| 2098 |
break; |
| 2099 |
} |
| 2100 |
if (new_index == DSA_MAX_SEGMENTS) |
| 2101 |
return NULL; |
| 2102 |
|
| 2103 |
/* |
| 2104 |
* If the total size limit is already exceeded, then we exit early and |
| 2105 |
* avoid arithmetic wraparound in the unsigned expressions below. |
| 2106 |
*/ |
| 2107 |
if (area->control->total_segment_size >= |
| 2108 |
area->control->max_total_segment_size) |
| 2109 |
return NULL; |
| 2110 |
|
| 2111 |
/* |
| 2112 |
* The size should be at least as big as requested, and at least big |
| 2113 |
* enough to follow a geometric series that approximately doubles the |
| 2114 |
* total storage each time we create a new segment. We use geometric |
| 2115 |
* growth because the underlying DSM system isn't designed for large |
| 2116 |
* numbers of segments (otherwise we might even consider just using one |
| 2117 |
* DSM segment for each large allocation and for each superblock, and then |
| 2118 |
* we wouldn't need to use FreePageManager). |
| 2119 |
* |
| 2120 |
* We decide on a total segment size first, so that we produce tidy |
| 2121 |
* power-of-two sized segments. This is a good property to have if we |
| 2122 |
* move to huge pages in the future. Then we work back to the number of |
| 2123 |
* pages we can fit. |
| 2124 |
*/ |
| 2125 |
total_size = area->control->init_segment_size * |
| 2126 |
((size_t) 1 << (new_index / DSA_NUM_SEGMENTS_AT_EACH_SIZE)); |
| 2127 |
total_size = Min(total_size, area->control->max_segment_size); |
| 2128 |
total_size = Min(total_size, |
| 2129 |
area->control->max_total_segment_size - |
| 2130 |
area->control->total_segment_size); |
| 2131 |
|
| 2132 |
total_pages = total_size / FPM_PAGE_SIZE; |
| 2133 |
metadata_bytes = |
| 2134 |
MAXALIGN(sizeof(dsa_segment_header)) + |
| 2135 |
MAXALIGN(sizeof(FreePageManager)) + |
| 2136 |
sizeof(dsa_pointer) * total_pages; |
| 2137 |
|
| 2138 |
/* Add padding up to next page boundary. */ |
| 2139 |
if (metadata_bytes % FPM_PAGE_SIZE != 0) |
| 2140 |
metadata_bytes += FPM_PAGE_SIZE - (metadata_bytes % FPM_PAGE_SIZE); |
| 2141 |
if (total_size <= metadata_bytes) |
| 2142 |
return NULL; |
| 2143 |
usable_pages = (total_size - metadata_bytes) / FPM_PAGE_SIZE; |
| 2144 |
Assert(metadata_bytes + usable_pages * FPM_PAGE_SIZE <= total_size); |
| 2145 |
|
| 2146 |
/* See if that is enough... */ |
| 2147 |
if (requested_pages > usable_pages) |
| 2148 |
{ |
| 2149 |
size_t total_requested_pages PG_USED_FOR_ASSERTS_ONLY; |
| 2150 |
|
| 2151 |
/* |
| 2152 |
* We'll make an odd-sized segment, working forward from the requested |
| 2153 |
* number of pages. |
| 2154 |
*/ |
| 2155 |
usable_pages = requested_pages; |
| 2156 |
metadata_bytes = |
| 2157 |
MAXALIGN(sizeof(dsa_segment_header)) + |
| 2158 |
MAXALIGN(sizeof(FreePageManager)) + |
| 2159 |
usable_pages * sizeof(dsa_pointer); |
| 2160 |
|
| 2161 |
/* |
| 2162 |
* We must also account for pagemap entries needed to cover the |
| 2163 |
* metadata pages themselves. The pagemap must track all pages in the |
| 2164 |
* segment, including the pages occupied by metadata. |
| 2165 |
* |
| 2166 |
* This formula uses integer ceiling division to compute the exact |
| 2167 |
* number of additional entries needed. The divisor (FPM_PAGE_SIZE - |
| 2168 |
* sizeof(dsa_pointer)) accounts for the fact that each metadata page |
| 2169 |
* consumes one pagemap entry of sizeof(dsa_pointer) bytes, leaving |
| 2170 |
* only (FPM_PAGE_SIZE - sizeof(dsa_pointer)) net bytes per metadata |
| 2171 |
* page. |
| 2172 |
*/ |
| 2173 |
metadata_bytes += |
| 2174 |
((metadata_bytes + (FPM_PAGE_SIZE - sizeof(dsa_pointer)) - 1) / |
| 2175 |
(FPM_PAGE_SIZE - sizeof(dsa_pointer))) * |
| 2176 |
sizeof(dsa_pointer); |
| 2177 |
|
| 2178 |
/* Add padding up to next page boundary. */ |
| 2179 |
if (metadata_bytes % FPM_PAGE_SIZE != 0) |
| 2180 |
metadata_bytes += FPM_PAGE_SIZE - (metadata_bytes % FPM_PAGE_SIZE); |
| 2181 |
total_size = metadata_bytes + usable_pages * FPM_PAGE_SIZE; |
| 2182 |
total_requested_pages = total_size / FPM_PAGE_SIZE; |
| 2183 |
|
| 2184 |
/* |
| 2185 |
* Verify that we allocated enough pagemap entries for metadata and |
| 2186 |
* usable pages. This reverse-engineers the new calculation of |
| 2187 |
* "metadata_bytes" done based on the new "requested_pages" for an |
| 2188 |
* odd-sized segment. |
| 2189 |
*/ |
| 2190 |
Assert((metadata_bytes - MAXALIGN(sizeof(dsa_segment_header)) - |
| 2191 |
MAXALIGN(sizeof(FreePageManager))) / sizeof(dsa_pointer) >= total_requested_pages); |
| 2192 |
|
| 2193 |
/* Is that too large for dsa_pointer's addressing scheme? */ |
| 2194 |
if (total_size > DSA_MAX_SEGMENT_SIZE) |
| 2195 |
return NULL; |
| 2196 |
|
| 2197 |
/* Would that exceed the limit? */ |
| 2198 |
if (total_size > area->control->max_total_segment_size - |
| 2199 |
area->control->total_segment_size) |
| 2200 |
return NULL; |
| 2201 |
} |
| 2202 |
|
| 2203 |
/* Create the segment. */ |
| 2204 |
oldowner = CurrentResourceOwner; |
| 2205 |
CurrentResourceOwner = area->resowner; |
| 2206 |
segment = dsm_create(total_size, 0); |
| 2207 |
CurrentResourceOwner = oldowner; |
| 2208 |
if (segment == NULL) |
| 2209 |
return NULL; |
| 2210 |
dsm_pin_segment(segment); |
| 2211 |
|
| 2212 |
/* Store the handle in shared memory to be found by index. */ |
| 2213 |
area->control->segment_handles[new_index] = |
| 2214 |
dsm_segment_handle(segment); |
| 2215 |
/* Track the highest segment index in the history of the area. */ |
| 2216 |
if (area->control->high_segment_index < new_index) |
| 2217 |
area->control->high_segment_index = new_index; |
| 2218 |
/* Track the highest segment index this backend has ever mapped. */ |
| 2219 |
if (area->high_segment_index < new_index) |
| 2220 |
area->high_segment_index = new_index; |
| 2221 |
/* Track total size of all segments. */ |
| 2222 |
area->control->total_segment_size += total_size; |
| 2223 |
Assert(area->control->total_segment_size <= |
| 2224 |
area->control->max_total_segment_size); |
| 2225 |
|
| 2226 |
/* Build a segment map for this segment in this backend. */ |
| 2227 |
segment_map = &area->segment_maps[new_index]; |
| 2228 |
segment_map->segment = segment; |
| 2229 |
segment_map->mapped_address = dsm_segment_address(segment); |
| 2230 |
segment_map->header = (dsa_segment_header *) segment_map->mapped_address; |
| 2231 |
segment_map->fpm = (FreePageManager *) |
| 2232 |
(segment_map->mapped_address + |
| 2233 |
MAXALIGN(sizeof(dsa_segment_header))); |
| 2234 |
segment_map->pagemap = (dsa_pointer *) |
| 2235 |
(segment_map->mapped_address + |
| 2236 |
MAXALIGN(sizeof(dsa_segment_header)) + |
| 2237 |
MAXALIGN(sizeof(FreePageManager))); |
| 2238 |
|
| 2239 |
/* Set up the free page map. */ |
| 2240 |
FreePageManagerInitialize(segment_map->fpm, segment_map->mapped_address); |
| 2241 |
FreePageManagerPut(segment_map->fpm, metadata_bytes / FPM_PAGE_SIZE, |
| 2242 |
usable_pages); |
| 2243 |
|
| 2244 |
/* Set up the segment header and put it in the appropriate bin. */ |
| 2245 |
segment_map->header->magic = |
| 2246 |
DSA_SEGMENT_HEADER_MAGIC ^ area->control->handle ^ new_index; |
| 2247 |
segment_map->header->usable_pages = usable_pages; |
| 2248 |
segment_map->header->size = total_size; |
| 2249 |
segment_map->header->bin = contiguous_pages_to_segment_bin(usable_pages); |
| 2250 |
segment_map->header->prev = DSA_SEGMENT_INDEX_NONE; |
| 2251 |
segment_map->header->next = |
| 2252 |
area->control->segment_bins[segment_map->header->bin]; |
| 2253 |
segment_map->header->freed = false; |
| 2254 |
area->control->segment_bins[segment_map->header->bin] = new_index; |
| 2255 |
if (segment_map->header->next != DSA_SEGMENT_INDEX_NONE) |
| 2256 |
{ |
| 2257 |
dsa_segment_map *next = |
| 2258 |
get_segment_by_index(area, segment_map->header->next); |
| 2259 |
|
| 2260 |
Assert(next->header->bin == segment_map->header->bin); |
| 2261 |
next->header->prev = new_index; |
| 2262 |
} |
| 2263 |
|
| 2264 |
return segment_map; |
| 2265 |
} |
| 2266 |
|
| 2267 |
/* |
| 2268 |
* Check if any segments have been freed by destroy_superblock, so we can |
| 2269 |
* detach from them in this backend. This function is called by |
| 2270 |
* dsa_get_address and dsa_free to make sure that a dsa_pointer they have |
| 2271 |
* received can be resolved to the correct segment. |
| 2272 |
* |
| 2273 |
* The danger we want to defend against is that there could be an old segment |
| 2274 |
* mapped into a given slot in this backend, and the dsa_pointer they have |
| 2275 |
* might refer to some new segment in the same slot. So those functions must |
| 2276 |
* be sure to process all instructions to detach from a freed segment that had |
| 2277 |
* been generated by the time this process received the dsa_pointer, before |
| 2278 |
* they call get_segment_by_index. |
| 2279 |
*/ |
| 2280 |
static void |
| 2281 |
check_for_freed_segments(dsa_area *area) |
| 2282 |
{ |
| 2283 |
size_t freed_segment_counter; |
| 2284 |
|
| 2285 |
/* |
| 2286 |
* Any other process that has freed a segment has incremented |
| 2287 |
* freed_segment_counter while holding an LWLock, and that must precede |
| 2288 |
* any backend creating a new segment in the same slot while holding an |
| 2289 |
* LWLock, and that must precede the creation of any dsa_pointer pointing |
| 2290 |
* into the new segment which might reach us here, and the caller must |
| 2291 |
* have sent the dsa_pointer to this process using appropriate memory |
| 2292 |
* synchronization (some kind of locking or atomic primitive or system |
| 2293 |
* call). So all we need to do on the reading side is ask for the load of |
| 2294 |
* freed_segment_counter to follow the caller's load of the dsa_pointer it |
| 2295 |
* has, and we can be sure to detect any segments that had been freed as |
| 2296 |
* of the time that the dsa_pointer reached this process. |
| 2297 |
*/ |
| 2298 |
pg_read_barrier(); |
| 2299 |
freed_segment_counter = area->control->freed_segment_counter; |
| 2300 |
if (unlikely(area->freed_segment_counter != freed_segment_counter)) |
| 2301 |
{ |
| 2302 |
/* Check all currently mapped segments to find what's been freed. */ |
| 2303 |
LWLockAcquire(DSA_AREA_LOCK(area), LW_EXCLUSIVE); |
| 2304 |
check_for_freed_segments_locked(area); |
| 2305 |
LWLockRelease(DSA_AREA_LOCK(area)); |
| 2306 |
} |
| 2307 |
} |
| 2308 |
|
| 2309 |
/* |
| 2310 |
* Workhorse for check_for_freed_segments(), and also used directly in path |
| 2311 |
* where the area lock is already held. This should be called after acquiring |
| 2312 |
* the lock but before looking up any segment by index number, to make sure we |
| 2313 |
* unmap any stale segments that might have previously had the same index as a |
| 2314 |
* current segment. |
| 2315 |
*/ |
| 2316 |
static void |
| 2317 |
check_for_freed_segments_locked(dsa_area *area) |
| 2318 |
{ |
| 2319 |
size_t freed_segment_counter; |
| 2320 |
int i; |
| 2321 |
|
| 2322 |
Assert(LWLockHeldByMe(DSA_AREA_LOCK(area))); |
| 2323 |
freed_segment_counter = area->control->freed_segment_counter; |
| 2324 |
if (unlikely(area->freed_segment_counter != freed_segment_counter)) |
| 2325 |
{ |
| 2326 |
for (i = 0; i <= area->high_segment_index; ++i) |
| 2327 |
{ |
| 2328 |
if (area->segment_maps[i].header != NULL && |
| 2329 |
area->segment_maps[i].header->freed) |
| 2330 |
{ |
| 2331 |
dsm_detach(area->segment_maps[i].segment); |
| 2332 |
area->segment_maps[i].segment = NULL; |
| 2333 |
area->segment_maps[i].header = NULL; |
| 2334 |
area->segment_maps[i].mapped_address = NULL; |
| 2335 |
} |
| 2336 |
} |
| 2337 |
area->freed_segment_counter = freed_segment_counter; |
| 2338 |
} |
| 2339 |
} |
| 2340 |
|
| 2341 |
/* |
| 2342 |
* Re-bin segment if it's no longer in the appropriate bin. |
| 2343 |
*/ |
| 2344 |
static void |
| 2345 |
rebin_segment(dsa_area *area, dsa_segment_map *segment_map) |
| 2346 |
{ |
| 2347 |
size_t new_bin; |
| 2348 |
dsa_segment_index segment_index; |
| 2349 |
|
| 2350 |
new_bin = contiguous_pages_to_segment_bin(fpm_largest(segment_map->fpm)); |
| 2351 |
if (segment_map->header->bin == new_bin) |
| 2352 |
return; |
| 2353 |
|
| 2354 |
/* Remove it from its current bin. */ |
| 2355 |
unlink_segment(area, segment_map); |
| 2356 |
|
| 2357 |
/* Push it onto the front of its new bin. */ |
| 2358 |
segment_index = get_segment_index(area, segment_map); |
| 2359 |
segment_map->header->prev = DSA_SEGMENT_INDEX_NONE; |
| 2360 |
segment_map->header->next = area->control->segment_bins[new_bin]; |
| 2361 |
segment_map->header->bin = new_bin; |
| 2362 |
area->control->segment_bins[new_bin] = segment_index; |
| 2363 |
if (segment_map->header->next != DSA_SEGMENT_INDEX_NONE) |
| 2364 |
{ |
| 2365 |
dsa_segment_map *next; |
| 2366 |
|
| 2367 |
next = get_segment_by_index(area, segment_map->header->next); |
| 2368 |
Assert(next->header->bin == new_bin); |
| 2369 |
next->header->prev = segment_index; |
| 2370 |
} |
| 2371 |
} |