lib/axmap: Optimize __axmap_set()
[fio.git] / lib / axmap.c
CommitLineData
7ebd796f
JA
1/*
2 * Bitmap of bitmaps, where each layer is number-of-bits-per-word smaller than
3 * the previous. Hence an 'axmap', since we axe each previous layer into a
4 * much smaller piece. I swear, that is why it's named like that. It has
5 * nothing to do with anything remotely narcissistic.
6 *
7 * A set bit at layer N indicates a full word at layer N-1, and so forth. As
de8f6de9 8 * the bitmap becomes progressively more full, checking for existence
7ebd796f
JA
9 * becomes cheaper (since fewer layers are walked, making it a lot more
10 * cache friendly) and locating the next free space likewise.
11 *
12 * Axmaps get pretty close to optimal (1 bit per block) space usage, since
13 * layers quickly diminish in size. Doing the size math is straight forward,
14 * since we have log64(blocks) layers of maps. For 20000 blocks, overhead
15 * is roughly 1.9%, or 1.019 bits per block. The number quickly converges
16 * towards 1.0158, or 1.58% of overhead.
17 */
18#include <stdio.h>
19#include <stdlib.h>
20#include <string.h>
21#include <assert.h>
22
23#include "../arch/arch.h"
24#include "axmap.h"
7ebd796f
JA
25#include "../minmax.h"
26
27#if BITS_PER_LONG == 64
28#define UNIT_SHIFT 6
29#elif BITS_PER_LONG == 32
30#define UNIT_SHIFT 5
31#else
32#error "Number of arch bits unknown"
33#endif
34
e6f735f0 35#define BLOCKS_PER_UNIT (1U << UNIT_SHIFT)
7ebd796f
JA
36#define BLOCKS_PER_UNIT_MASK (BLOCKS_PER_UNIT - 1)
37
76dd186e
JA
38static const unsigned long bit_masks[] = {
39 0x0000000000000000, 0x0000000000000001, 0x0000000000000003, 0x0000000000000007,
40 0x000000000000000f, 0x000000000000001f, 0x000000000000003f, 0x000000000000007f,
41 0x00000000000000ff, 0x00000000000001ff, 0x00000000000003ff, 0x00000000000007ff,
42 0x0000000000000fff, 0x0000000000001fff, 0x0000000000003fff, 0x0000000000007fff,
43 0x000000000000ffff, 0x000000000001ffff, 0x000000000003ffff, 0x000000000007ffff,
44 0x00000000000fffff, 0x00000000001fffff, 0x00000000003fffff, 0x00000000007fffff,
45 0x0000000000ffffff, 0x0000000001ffffff, 0x0000000003ffffff, 0x0000000007ffffff,
46 0x000000000fffffff, 0x000000001fffffff, 0x000000003fffffff, 0x000000007fffffff,
47 0x00000000ffffffff,
48#if BITS_PER_LONG == 64
49 0x00000001ffffffff, 0x00000003ffffffff, 0x00000007ffffffff, 0x0000000fffffffff,
50 0x0000001fffffffff, 0x0000003fffffffff, 0x0000007fffffffff, 0x000000ffffffffff,
51 0x000001ffffffffff, 0x000003ffffffffff, 0x000007ffffffffff, 0x00000fffffffffff,
52 0x00001fffffffffff, 0x00003fffffffffff, 0x00007fffffffffff, 0x0000ffffffffffff,
53 0x0001ffffffffffff, 0x0003ffffffffffff, 0x0007ffffffffffff, 0x000fffffffffffff,
54 0x001fffffffffffff, 0x003fffffffffffff, 0x007fffffffffffff, 0x00ffffffffffffff,
55 0x01ffffffffffffff, 0x03ffffffffffffff, 0x07ffffffffffffff, 0x0fffffffffffffff,
56 0x1fffffffffffffff, 0x3fffffffffffffff, 0x7fffffffffffffff, 0xffffffffffffffff
57#endif
58};
59
c7143018
BVA
60/**
61 * struct axmap_level - a bitmap used to implement struct axmap
62 * @level: Level index. Each map has at least one level with index zero. The
63 * higher the level index, the fewer bits a struct axmap_level contains.
64 * @map_size: Number of elements of the @map array.
65 * @map: A bitmap with @map_size elements.
66 */
7ebd796f
JA
67struct axmap_level {
68 int level;
69 unsigned long map_size;
70 unsigned long *map;
71};
72
c7143018
BVA
73/**
74 * struct axmap - a set that can store numbers 0 .. @nr_bits - 1
75 * @nr_level: Number of elements of the @levels array.
76 * @levels: struct axmap_level array in which lower levels contain more bits
77 * than higher levels.
78 * @nr_bits: One more than the highest value stored in the set.
79 */
7ebd796f
JA
80struct axmap {
81 unsigned int nr_levels;
82 struct axmap_level *levels;
47d94b0b 83 uint64_t nr_bits;
7ebd796f
JA
84};
85
c7143018 86/* Remove all elements from the @axmap set */
7ebd796f
JA
87void axmap_reset(struct axmap *axmap)
88{
89 int i;
90
91 for (i = 0; i < axmap->nr_levels; i++) {
92 struct axmap_level *al = &axmap->levels[i];
93
94 memset(al->map, 0, al->map_size * sizeof(unsigned long));
95 }
96}
97
98void axmap_free(struct axmap *axmap)
99{
100 unsigned int i;
101
102 if (!axmap)
103 return;
104
105 for (i = 0; i < axmap->nr_levels; i++)
aded30f7 106 free(axmap->levels[i].map);
7ebd796f 107
aded30f7
JA
108 free(axmap->levels);
109 free(axmap);
7ebd796f
JA
110}
111
c7143018 112/* Allocate memory for a set that can store the numbers 0 .. @nr_bits - 1. */
7ebd796f
JA
113struct axmap *axmap_new(unsigned long nr_bits)
114{
115 struct axmap *axmap;
116 unsigned int i, levels;
117
aded30f7 118 axmap = malloc(sizeof(*axmap));
7ebd796f
JA
119 if (!axmap)
120 return NULL;
121
122 levels = 1;
123 i = (nr_bits + BLOCKS_PER_UNIT - 1) >> UNIT_SHIFT;
124 while (i > 1) {
125 i = (i + BLOCKS_PER_UNIT - 1) >> UNIT_SHIFT;
126 levels++;
127 }
128
129 axmap->nr_levels = levels;
2404b2c3 130 axmap->levels = calloc(axmap->nr_levels, sizeof(struct axmap_level));
70c0f356
BVA
131 if (!axmap->levels)
132 goto free_axmap;
47d94b0b 133 axmap->nr_bits = nr_bits;
7ebd796f
JA
134
135 for (i = 0; i < axmap->nr_levels; i++) {
136 struct axmap_level *al = &axmap->levels[i];
137
138 al->level = i;
139 al->map_size = (nr_bits + BLOCKS_PER_UNIT - 1) >> UNIT_SHIFT;
aded30f7 140 al->map = malloc(al->map_size * sizeof(unsigned long));
7ebd796f 141 if (!al->map)
70c0f356 142 goto free_levels;
7ebd796f
JA
143
144 nr_bits = (nr_bits + BLOCKS_PER_UNIT - 1) >> UNIT_SHIFT;
145 }
146
147 axmap_reset(axmap);
148 return axmap;
70c0f356
BVA
149
150free_levels:
7ebd796f 151 for (i = 0; i < axmap->nr_levels; i++)
70c0f356 152 free(axmap->levels[i].map);
7ebd796f 153
aded30f7 154 free(axmap->levels);
70c0f356
BVA
155
156free_axmap:
bb1116fe 157 free(axmap);
7ebd796f
JA
158 return NULL;
159}
160
c7143018
BVA
161/*
162 * Call @func for each level, starting at level zero, until a level is found
163 * for which @func returns true. Return false if none of the @func calls
164 * returns true.
165 */
e39c0676
JA
166static bool axmap_handler(struct axmap *axmap, uint64_t bit_nr,
167 bool (*func)(struct axmap_level *, unsigned long, unsigned int,
7ebd796f
JA
168 void *), void *data)
169{
170 struct axmap_level *al;
edfca254 171 uint64_t index = bit_nr;
7ebd796f
JA
172 int i;
173
174 for (i = 0; i < axmap->nr_levels; i++) {
7ebd796f
JA
175 unsigned long offset = index >> UNIT_SHIFT;
176 unsigned int bit = index & BLOCKS_PER_UNIT_MASK;
177
178 al = &axmap->levels[i];
179
180 if (func(al, offset, bit, data))
e39c0676 181 return true;
edfca254
JA
182
183 if (index)
184 index >>= UNIT_SHIFT;
7ebd796f
JA
185 }
186
e39c0676 187 return false;
7ebd796f
JA
188}
189
c7143018
BVA
190/*
191 * Call @func for each level, starting at the highest level, until a level is
192 * found for which @func returns true. Return false if none of the @func calls
193 * returns true.
194 */
e39c0676 195static bool axmap_handler_topdown(struct axmap *axmap, uint64_t bit_nr,
c406445a 196 bool (*func)(struct axmap_level *, unsigned long, unsigned int, void *))
7ebd796f 197{
76dd186e 198 int i;
7ebd796f
JA
199
200 for (i = axmap->nr_levels - 1; i >= 0; i--) {
a1a4ab81 201 unsigned long index = bit_nr >> (UNIT_SHIFT * i);
7ebd796f
JA
202 unsigned long offset = index >> UNIT_SHIFT;
203 unsigned int bit = index & BLOCKS_PER_UNIT_MASK;
204
41264298 205 if (func(&axmap->levels[i], offset, bit, NULL))
e39c0676 206 return true;
7ebd796f
JA
207 }
208
e39c0676 209 return false;
7ebd796f
JA
210}
211
7ebd796f
JA
212struct axmap_set_data {
213 unsigned int nr_bits;
214 unsigned int set_bits;
7ebd796f
JA
215};
216
c7143018
BVA
217/*
218 * Set at most @__data->nr_bits bits in @al at offset @offset. Do not exceed
219 * the boundary of the element at offset @offset. Return the number of bits
220 * that have been set in @__data->set_bits if @al->level == 0.
221 */
e39c0676 222static bool axmap_set_fn(struct axmap_level *al, unsigned long offset,
7ebd796f
JA
223 unsigned int bit, void *__data)
224{
225 struct axmap_set_data *data = __data;
226 unsigned long mask, overlap;
227 unsigned int nr_bits;
228
229 nr_bits = min(data->nr_bits, BLOCKS_PER_UNIT - bit);
230
231 mask = bit_masks[nr_bits] << bit;
232
233 /*
234 * Mask off any potential overlap, only sets contig regions
235 */
236 overlap = al->map[offset] & mask;
bd71edce 237 if (overlap == mask) {
bd71edce 238 data->set_bits = 0;
e39c0676 239 return true;
bd71edce 240 }
7ebd796f 241
8cd5a2fd 242 if (overlap) {
15a4f496 243 nr_bits = ffz(~overlap) - bit;
8cd5a2fd 244 mask = bit_masks[nr_bits] << bit;
7ebd796f
JA
245 }
246
247 assert(mask);
248 assert(!(al->map[offset] & mask));
7ebd796f
JA
249 al->map[offset] |= mask;
250
251 if (!al->level)
252 data->set_bits = nr_bits;
253
c7143018 254 /* For the next level */
7ebd796f 255 data->nr_bits = 1;
c7143018 256
7ebd796f
JA
257 return al->map[offset] != -1UL;
258}
259
c7143018
BVA
260/*
261 * Set up to @data->nr_bits starting from @bit_nr in @axmap. Start at
262 * @bit_nr. If that bit has not yet been set then set it and continue until
263 * either @data->nr_bits have been set or a 1 bit is found. Store the number
264 * of bits that have been set in @data->set_bits. It is guaranteed that all
265 * bits that have been requested to set fit in the same unsigned long word of
266 * level 0 of @axmap.
267 */
7ebd796f
JA
268static void __axmap_set(struct axmap *axmap, uint64_t bit_nr,
269 struct axmap_set_data *data)
270{
1cf97788 271 unsigned int nr_bits = data->nr_bits;
7ebd796f 272
47d94b0b
JA
273 if (bit_nr > axmap->nr_bits)
274 return;
275 else if (bit_nr + nr_bits > axmap->nr_bits)
276 nr_bits = axmap->nr_bits - bit_nr;
277
1cf97788 278 assert(nr_bits <= BLOCKS_PER_UNIT);
7ebd796f 279
1cf97788 280 axmap_handler(axmap, bit_nr, axmap_set_fn, data);
7ebd796f
JA
281}
282
283void axmap_set(struct axmap *axmap, uint64_t bit_nr)
284{
285 struct axmap_set_data data = { .nr_bits = 1, };
286
287 __axmap_set(axmap, bit_nr, &data);
288}
289
c7143018
BVA
290/*
291 * Set up to @nr_bits starting from @bit in @axmap. Start at @bit. If that
292 * bit has not yet been set then set it and continue until either @nr_bits
293 * have been set or a 1 bit is found. Return the number of bits that have been
294 * set.
295 */
e39c0676
JA
296unsigned int axmap_set_nr(struct axmap *axmap, uint64_t bit_nr,
297 unsigned int nr_bits)
7ebd796f 298{
0bfdf9f1
JA
299 unsigned int set_bits = 0;
300
301 do {
09d6bf09 302 struct axmap_set_data data = { .nr_bits = nr_bits, };
0bfdf9f1
JA
303 unsigned int max_bits, this_set;
304
305 max_bits = BLOCKS_PER_UNIT - (bit_nr & BLOCKS_PER_UNIT_MASK);
bd71edce 306 if (nr_bits > max_bits)
0bfdf9f1
JA
307 data.nr_bits = max_bits;
308
309 this_set = data.nr_bits;
310 __axmap_set(axmap, bit_nr, &data);
311 set_bits += data.set_bits;
312 if (data.set_bits != this_set)
313 break;
7ebd796f 314
0bfdf9f1
JA
315 nr_bits -= data.set_bits;
316 bit_nr += data.set_bits;
317 } while (nr_bits);
318
319 return set_bits;
7ebd796f
JA
320}
321
e39c0676
JA
322static bool axmap_isset_fn(struct axmap_level *al, unsigned long offset,
323 unsigned int bit, void *unused)
7ebd796f
JA
324{
325 return (al->map[offset] & (1UL << bit)) != 0;
326}
327
e39c0676 328bool axmap_isset(struct axmap *axmap, uint64_t bit_nr)
7ebd796f 329{
ac569176 330 if (bit_nr <= axmap->nr_bits)
c406445a 331 return axmap_handler_topdown(axmap, bit_nr, axmap_isset_fn);
47d94b0b 332
e39c0676 333 return false;
7ebd796f
JA
334}
335
b9304c5b
MK
336/*
337 * Find the first free bit that is at least as large as bit_nr. Return
338 * -1 if no free bit is found before the end of the map.
339 */
340static uint64_t axmap_find_first_free(struct axmap *axmap, uint64_t bit_nr)
7ebd796f 341{
731ef4c7 342 int i;
b9304c5b
MK
343 unsigned long temp;
344 unsigned int bit;
345 uint64_t offset, base_index, index;
346 struct axmap_level *al;
7ebd796f 347
b9304c5b
MK
348 index = 0;
349 for (i = axmap->nr_levels - 1; i >= 0; i--) {
350 al = &axmap->levels[i];
7ebd796f 351
b9304c5b
MK
352 /* Shift previously calculated index for next level */
353 index <<= UNIT_SHIFT;
354
355 /*
356 * Start from an index that's at least as large as the
357 * originally passed in bit number.
358 */
359 base_index = bit_nr >> (UNIT_SHIFT * i);
360 if (index < base_index)
361 index = base_index;
362
363 /* Get the offset and bit for this level */
364 offset = index >> UNIT_SHIFT;
365 bit = index & BLOCKS_PER_UNIT_MASK;
366
367 /*
368 * If the previous level had unused bits in its last
369 * word, the offset could be bigger than the map at
370 * this level. That means no free bits exist before the
371 * end of the map, so return -1.
372 */
373 if (offset >= al->map_size)
374 return -1ULL;
375
376 /* Check the first word starting with the specific bit */
377 temp = ~bit_masks[bit] & ~al->map[offset];
378 if (temp)
379 goto found;
380
381 /*
382 * No free bit in the first word, so iterate
383 * looking for a word with one or more free bits.
384 */
385 for (offset++; offset < al->map_size; offset++) {
386 temp = ~al->map[offset];
387 if (temp)
388 goto found;
7ebd796f 389 }
47d94b0b 390
b9304c5b
MK
391 /* Did not find a free bit */
392 return -1ULL;
7ebd796f 393
b9304c5b
MK
394found:
395 /* Compute the index of the free bit just found */
396 index = (offset << UNIT_SHIFT) + ffz(~temp);
7ebd796f
JA
397 }
398
b9304c5b
MK
399 /* If found an unused bit in the last word of level 0, return -1 */
400 if (index >= axmap->nr_bits)
401 return -1ULL;
402
403 return index;
7ebd796f
JA
404}
405
406/*
407 * 'bit_nr' is already set. Find the next free bit after this one.
b9304c5b 408 * Return -1 if no free bits found.
7ebd796f
JA
409 */
410uint64_t axmap_next_free(struct axmap *axmap, uint64_t bit_nr)
411{
53737ae0 412 uint64_t ret;
b9304c5b
MK
413 uint64_t next_bit = bit_nr + 1;
414 unsigned long temp;
415 uint64_t offset;
416 unsigned int bit;
7ebd796f 417
b9304c5b
MK
418 if (bit_nr >= axmap->nr_bits)
419 return -1ULL;
12bde369 420
b9304c5b
MK
421 /* If at the end of the map, wrap-around */
422 if (next_bit == axmap->nr_bits)
423 next_bit = 0;
7ebd796f 424
b9304c5b
MK
425 offset = next_bit >> UNIT_SHIFT;
426 bit = next_bit & BLOCKS_PER_UNIT_MASK;
7ebd796f 427
53737ae0 428 /*
b9304c5b
MK
429 * As an optimization, do a quick check for a free bit
430 * in the current word at level 0. If not found, do
431 * a topdown search.
53737ae0 432 */
b9304c5b
MK
433 temp = ~bit_masks[bit] & ~axmap->levels[0].map[offset];
434 if (temp) {
435 ret = (offset << UNIT_SHIFT) + ffz(~temp);
53737ae0 436
b9304c5b
MK
437 /* Might have found an unused bit at level 0 */
438 if (ret >= axmap->nr_bits)
439 ret = -1ULL;
440 } else
441 ret = axmap_find_first_free(axmap, next_bit);
442
443 /*
444 * If there are no free bits starting at next_bit and going
445 * to the end of the map, wrap around by searching again
446 * starting at bit 0.
447 */
448 if (ret == -1ULL && next_bit != 0)
449 ret = axmap_find_first_free(axmap, 0);
450 return ret;
7ebd796f 451}