fio: fix aio trim completion latencies
[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. */
604d3f5b 113struct axmap *axmap_new(uint64_t nr_bits)
7ebd796f
JA
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
604d3f5b
JA
138 nr_bits = (nr_bits + BLOCKS_PER_UNIT - 1) >> UNIT_SHIFT;
139
7ebd796f 140 al->level = i;
604d3f5b 141 al->map_size = nr_bits;
aded30f7 142 al->map = malloc(al->map_size * sizeof(unsigned long));
7ebd796f 143 if (!al->map)
70c0f356 144 goto free_levels;
7ebd796f 145
7ebd796f
JA
146 }
147
148 axmap_reset(axmap);
149 return axmap;
70c0f356
BVA
150
151free_levels:
7ebd796f 152 for (i = 0; i < axmap->nr_levels; i++)
70c0f356 153 free(axmap->levels[i].map);
7ebd796f 154
aded30f7 155 free(axmap->levels);
70c0f356
BVA
156
157free_axmap:
bb1116fe 158 free(axmap);
7ebd796f
JA
159 return NULL;
160}
161
c7143018
BVA
162/*
163 * Call @func for each level, starting at level zero, until a level is found
164 * for which @func returns true. Return false if none of the @func calls
165 * returns true.
166 */
e39c0676 167static bool axmap_handler(struct axmap *axmap, uint64_t bit_nr,
39c56bc0 168 bool (*func)(struct axmap_level *, uint64_t, unsigned int,
7ebd796f
JA
169 void *), void *data)
170{
171 struct axmap_level *al;
edfca254 172 uint64_t index = bit_nr;
7ebd796f
JA
173 int i;
174
175 for (i = 0; i < axmap->nr_levels; i++) {
7ebd796f
JA
176 unsigned long offset = index >> UNIT_SHIFT;
177 unsigned int bit = index & BLOCKS_PER_UNIT_MASK;
178
179 al = &axmap->levels[i];
180
181 if (func(al, offset, bit, data))
e39c0676 182 return true;
edfca254
JA
183
184 if (index)
185 index >>= UNIT_SHIFT;
7ebd796f
JA
186 }
187
e39c0676 188 return false;
7ebd796f
JA
189}
190
c7143018
BVA
191/*
192 * Call @func for each level, starting at the highest level, until a level is
193 * found for which @func returns true. Return false if none of the @func calls
194 * returns true.
195 */
e39c0676 196static bool axmap_handler_topdown(struct axmap *axmap, uint64_t bit_nr,
39c56bc0 197 bool (*func)(struct axmap_level *, uint64_t, unsigned int, void *))
7ebd796f 198{
76dd186e 199 int i;
7ebd796f
JA
200
201 for (i = axmap->nr_levels - 1; i >= 0; i--) {
39c56bc0 202 uint64_t index = bit_nr >> (UNIT_SHIFT * i);
7ebd796f
JA
203 unsigned long offset = index >> UNIT_SHIFT;
204 unsigned int bit = index & BLOCKS_PER_UNIT_MASK;
205
41264298 206 if (func(&axmap->levels[i], offset, bit, NULL))
e39c0676 207 return true;
7ebd796f
JA
208 }
209
e39c0676 210 return false;
7ebd796f
JA
211}
212
7ebd796f
JA
213struct axmap_set_data {
214 unsigned int nr_bits;
215 unsigned int set_bits;
7ebd796f
JA
216};
217
c7143018
BVA
218/*
219 * Set at most @__data->nr_bits bits in @al at offset @offset. Do not exceed
220 * the boundary of the element at offset @offset. Return the number of bits
221 * that have been set in @__data->set_bits if @al->level == 0.
222 */
39c56bc0 223static bool axmap_set_fn(struct axmap_level *al, uint64_t offset,
7ebd796f
JA
224 unsigned int bit, void *__data)
225{
226 struct axmap_set_data *data = __data;
227 unsigned long mask, overlap;
228 unsigned int nr_bits;
229
230 nr_bits = min(data->nr_bits, BLOCKS_PER_UNIT - bit);
231
232 mask = bit_masks[nr_bits] << bit;
233
234 /*
235 * Mask off any potential overlap, only sets contig regions
236 */
237 overlap = al->map[offset] & mask;
bd71edce 238 if (overlap == mask) {
bd71edce 239 data->set_bits = 0;
e39c0676 240 return true;
bd71edce 241 }
7ebd796f 242
8cd5a2fd 243 if (overlap) {
15a4f496 244 nr_bits = ffz(~overlap) - bit;
49e1ade1
JA
245 if (!nr_bits)
246 return true;
8cd5a2fd 247 mask = bit_masks[nr_bits] << bit;
7ebd796f
JA
248 }
249
250 assert(mask);
251 assert(!(al->map[offset] & mask));
7ebd796f
JA
252 al->map[offset] |= mask;
253
254 if (!al->level)
255 data->set_bits = nr_bits;
256
c7143018 257 /* For the next level */
7ebd796f 258 data->nr_bits = 1;
c7143018 259
7ebd796f
JA
260 return al->map[offset] != -1UL;
261}
262
c7143018
BVA
263/*
264 * Set up to @data->nr_bits starting from @bit_nr in @axmap. Start at
265 * @bit_nr. If that bit has not yet been set then set it and continue until
266 * either @data->nr_bits have been set or a 1 bit is found. Store the number
267 * of bits that have been set in @data->set_bits. It is guaranteed that all
268 * bits that have been requested to set fit in the same unsigned long word of
269 * level 0 of @axmap.
270 */
7ebd796f
JA
271static void __axmap_set(struct axmap *axmap, uint64_t bit_nr,
272 struct axmap_set_data *data)
273{
1cf97788 274 unsigned int nr_bits = data->nr_bits;
7ebd796f 275
47d94b0b
JA
276 if (bit_nr > axmap->nr_bits)
277 return;
278 else if (bit_nr + nr_bits > axmap->nr_bits)
279 nr_bits = axmap->nr_bits - bit_nr;
280
1cf97788 281 assert(nr_bits <= BLOCKS_PER_UNIT);
7ebd796f 282
1cf97788 283 axmap_handler(axmap, bit_nr, axmap_set_fn, data);
7ebd796f
JA
284}
285
286void axmap_set(struct axmap *axmap, uint64_t bit_nr)
287{
288 struct axmap_set_data data = { .nr_bits = 1, };
289
290 __axmap_set(axmap, bit_nr, &data);
291}
292
c7143018
BVA
293/*
294 * Set up to @nr_bits starting from @bit in @axmap. Start at @bit. If that
295 * bit has not yet been set then set it and continue until either @nr_bits
296 * have been set or a 1 bit is found. Return the number of bits that have been
297 * set.
298 */
e39c0676
JA
299unsigned int axmap_set_nr(struct axmap *axmap, uint64_t bit_nr,
300 unsigned int nr_bits)
7ebd796f 301{
0bfdf9f1
JA
302 unsigned int set_bits = 0;
303
304 do {
09d6bf09 305 struct axmap_set_data data = { .nr_bits = nr_bits, };
0bfdf9f1
JA
306 unsigned int max_bits, this_set;
307
308 max_bits = BLOCKS_PER_UNIT - (bit_nr & BLOCKS_PER_UNIT_MASK);
bd71edce 309 if (nr_bits > max_bits)
0bfdf9f1
JA
310 data.nr_bits = max_bits;
311
312 this_set = data.nr_bits;
313 __axmap_set(axmap, bit_nr, &data);
314 set_bits += data.set_bits;
315 if (data.set_bits != this_set)
316 break;
7ebd796f 317
0bfdf9f1
JA
318 nr_bits -= data.set_bits;
319 bit_nr += data.set_bits;
320 } while (nr_bits);
321
322 return set_bits;
7ebd796f
JA
323}
324
39c56bc0 325static bool axmap_isset_fn(struct axmap_level *al, uint64_t offset,
e39c0676 326 unsigned int bit, void *unused)
7ebd796f 327{
0f9f1921 328 return (al->map[offset] & (1ULL << bit)) != 0;
7ebd796f
JA
329}
330
e39c0676 331bool axmap_isset(struct axmap *axmap, uint64_t bit_nr)
7ebd796f 332{
ac569176 333 if (bit_nr <= axmap->nr_bits)
c406445a 334 return axmap_handler_topdown(axmap, bit_nr, axmap_isset_fn);
47d94b0b 335
e39c0676 336 return false;
7ebd796f
JA
337}
338
b9304c5b
MK
339/*
340 * Find the first free bit that is at least as large as bit_nr. Return
341 * -1 if no free bit is found before the end of the map.
342 */
343static uint64_t axmap_find_first_free(struct axmap *axmap, uint64_t bit_nr)
7ebd796f 344{
731ef4c7 345 int i;
b9304c5b
MK
346 unsigned long temp;
347 unsigned int bit;
348 uint64_t offset, base_index, index;
349 struct axmap_level *al;
7ebd796f 350
b9304c5b
MK
351 index = 0;
352 for (i = axmap->nr_levels - 1; i >= 0; i--) {
353 al = &axmap->levels[i];
7ebd796f 354
b9304c5b
MK
355 /* Shift previously calculated index for next level */
356 index <<= UNIT_SHIFT;
357
358 /*
359 * Start from an index that's at least as large as the
360 * originally passed in bit number.
361 */
362 base_index = bit_nr >> (UNIT_SHIFT * i);
363 if (index < base_index)
364 index = base_index;
365
366 /* Get the offset and bit for this level */
367 offset = index >> UNIT_SHIFT;
368 bit = index & BLOCKS_PER_UNIT_MASK;
369
370 /*
371 * If the previous level had unused bits in its last
372 * word, the offset could be bigger than the map at
373 * this level. That means no free bits exist before the
374 * end of the map, so return -1.
375 */
376 if (offset >= al->map_size)
377 return -1ULL;
378
379 /* Check the first word starting with the specific bit */
380 temp = ~bit_masks[bit] & ~al->map[offset];
381 if (temp)
382 goto found;
383
384 /*
385 * No free bit in the first word, so iterate
386 * looking for a word with one or more free bits.
387 */
388 for (offset++; offset < al->map_size; offset++) {
389 temp = ~al->map[offset];
390 if (temp)
391 goto found;
7ebd796f 392 }
47d94b0b 393
b9304c5b
MK
394 /* Did not find a free bit */
395 return -1ULL;
7ebd796f 396
b9304c5b
MK
397found:
398 /* Compute the index of the free bit just found */
399 index = (offset << UNIT_SHIFT) + ffz(~temp);
7ebd796f
JA
400 }
401
b9304c5b
MK
402 /* If found an unused bit in the last word of level 0, return -1 */
403 if (index >= axmap->nr_bits)
404 return -1ULL;
405
406 return index;
7ebd796f
JA
407}
408
409/*
410 * 'bit_nr' is already set. Find the next free bit after this one.
b9304c5b 411 * Return -1 if no free bits found.
7ebd796f
JA
412 */
413uint64_t axmap_next_free(struct axmap *axmap, uint64_t bit_nr)
414{
53737ae0 415 uint64_t ret;
b9304c5b
MK
416 uint64_t next_bit = bit_nr + 1;
417 unsigned long temp;
418 uint64_t offset;
419 unsigned int bit;
7ebd796f 420
b9304c5b
MK
421 if (bit_nr >= axmap->nr_bits)
422 return -1ULL;
12bde369 423
b9304c5b
MK
424 /* If at the end of the map, wrap-around */
425 if (next_bit == axmap->nr_bits)
426 next_bit = 0;
7ebd796f 427
b9304c5b
MK
428 offset = next_bit >> UNIT_SHIFT;
429 bit = next_bit & BLOCKS_PER_UNIT_MASK;
7ebd796f 430
53737ae0 431 /*
b9304c5b
MK
432 * As an optimization, do a quick check for a free bit
433 * in the current word at level 0. If not found, do
434 * a topdown search.
53737ae0 435 */
b9304c5b
MK
436 temp = ~bit_masks[bit] & ~axmap->levels[0].map[offset];
437 if (temp) {
438 ret = (offset << UNIT_SHIFT) + ffz(~temp);
53737ae0 439
b9304c5b
MK
440 /* Might have found an unused bit at level 0 */
441 if (ret >= axmap->nr_bits)
442 ret = -1ULL;
443 } else
444 ret = axmap_find_first_free(axmap, next_bit);
445
446 /*
447 * If there are no free bits starting at next_bit and going
448 * to the end of the map, wrap around by searching again
449 * starting at bit 0.
450 */
451 if (ret == -1ULL && next_bit != 0)
452 ret = axmap_find_first_free(axmap, 0);
453 return ret;
7ebd796f 454}