Commit | Line | Data |
---|---|---|
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 |
38 | static 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 | ||
7ebd796f JA |
60 | struct axmap_level { |
61 | int level; | |
62 | unsigned long map_size; | |
63 | unsigned long *map; | |
64 | }; | |
65 | ||
66 | struct axmap { | |
67 | unsigned int nr_levels; | |
68 | struct axmap_level *levels; | |
47d94b0b | 69 | uint64_t nr_bits; |
7ebd796f JA |
70 | }; |
71 | ||
76dd186e | 72 | static inline unsigned long ulog64(unsigned long val, unsigned int log) |
7ebd796f JA |
73 | { |
74 | while (log-- && val) | |
75 | val >>= UNIT_SHIFT; | |
76 | ||
77 | return val; | |
78 | } | |
79 | ||
80 | void axmap_reset(struct axmap *axmap) | |
81 | { | |
82 | int i; | |
83 | ||
84 | for (i = 0; i < axmap->nr_levels; i++) { | |
85 | struct axmap_level *al = &axmap->levels[i]; | |
86 | ||
87 | memset(al->map, 0, al->map_size * sizeof(unsigned long)); | |
88 | } | |
89 | } | |
90 | ||
91 | void axmap_free(struct axmap *axmap) | |
92 | { | |
93 | unsigned int i; | |
94 | ||
95 | if (!axmap) | |
96 | return; | |
97 | ||
98 | for (i = 0; i < axmap->nr_levels; i++) | |
aded30f7 | 99 | free(axmap->levels[i].map); |
7ebd796f | 100 | |
aded30f7 JA |
101 | free(axmap->levels); |
102 | free(axmap); | |
7ebd796f JA |
103 | } |
104 | ||
105 | struct axmap *axmap_new(unsigned long nr_bits) | |
106 | { | |
107 | struct axmap *axmap; | |
108 | unsigned int i, levels; | |
109 | ||
aded30f7 | 110 | axmap = malloc(sizeof(*axmap)); |
7ebd796f JA |
111 | if (!axmap) |
112 | return NULL; | |
113 | ||
114 | levels = 1; | |
115 | i = (nr_bits + BLOCKS_PER_UNIT - 1) >> UNIT_SHIFT; | |
116 | while (i > 1) { | |
117 | i = (i + BLOCKS_PER_UNIT - 1) >> UNIT_SHIFT; | |
118 | levels++; | |
119 | } | |
120 | ||
121 | axmap->nr_levels = levels; | |
2404b2c3 | 122 | axmap->levels = calloc(axmap->nr_levels, sizeof(struct axmap_level)); |
47d94b0b | 123 | axmap->nr_bits = nr_bits; |
7ebd796f JA |
124 | |
125 | for (i = 0; i < axmap->nr_levels; i++) { | |
126 | struct axmap_level *al = &axmap->levels[i]; | |
127 | ||
128 | al->level = i; | |
129 | al->map_size = (nr_bits + BLOCKS_PER_UNIT - 1) >> UNIT_SHIFT; | |
aded30f7 | 130 | al->map = malloc(al->map_size * sizeof(unsigned long)); |
7ebd796f JA |
131 | if (!al->map) |
132 | goto err; | |
133 | ||
134 | nr_bits = (nr_bits + BLOCKS_PER_UNIT - 1) >> UNIT_SHIFT; | |
135 | } | |
136 | ||
137 | axmap_reset(axmap); | |
138 | return axmap; | |
139 | err: | |
140 | for (i = 0; i < axmap->nr_levels; i++) | |
141 | if (axmap->levels[i].map) | |
aded30f7 | 142 | free(axmap->levels[i].map); |
7ebd796f | 143 | |
aded30f7 | 144 | free(axmap->levels); |
bb1116fe | 145 | free(axmap); |
7ebd796f JA |
146 | return NULL; |
147 | } | |
148 | ||
e39c0676 JA |
149 | static bool axmap_handler(struct axmap *axmap, uint64_t bit_nr, |
150 | bool (*func)(struct axmap_level *, unsigned long, unsigned int, | |
7ebd796f JA |
151 | void *), void *data) |
152 | { | |
153 | struct axmap_level *al; | |
edfca254 | 154 | uint64_t index = bit_nr; |
7ebd796f JA |
155 | int i; |
156 | ||
157 | for (i = 0; i < axmap->nr_levels; i++) { | |
7ebd796f JA |
158 | unsigned long offset = index >> UNIT_SHIFT; |
159 | unsigned int bit = index & BLOCKS_PER_UNIT_MASK; | |
160 | ||
161 | al = &axmap->levels[i]; | |
162 | ||
163 | if (func(al, offset, bit, data)) | |
e39c0676 | 164 | return true; |
edfca254 JA |
165 | |
166 | if (index) | |
167 | index >>= UNIT_SHIFT; | |
7ebd796f JA |
168 | } |
169 | ||
e39c0676 | 170 | return false; |
7ebd796f JA |
171 | } |
172 | ||
e39c0676 | 173 | static bool axmap_handler_topdown(struct axmap *axmap, uint64_t bit_nr, |
c406445a | 174 | bool (*func)(struct axmap_level *, unsigned long, unsigned int, void *)) |
7ebd796f | 175 | { |
76dd186e | 176 | int i; |
7ebd796f JA |
177 | |
178 | for (i = axmap->nr_levels - 1; i >= 0; i--) { | |
76dd186e | 179 | unsigned long index = ulog64(bit_nr, i); |
7ebd796f JA |
180 | unsigned long offset = index >> UNIT_SHIFT; |
181 | unsigned int bit = index & BLOCKS_PER_UNIT_MASK; | |
182 | ||
41264298 | 183 | if (func(&axmap->levels[i], offset, bit, NULL)) |
e39c0676 | 184 | return true; |
7ebd796f JA |
185 | } |
186 | ||
e39c0676 | 187 | return false; |
7ebd796f JA |
188 | } |
189 | ||
7ebd796f JA |
190 | struct axmap_set_data { |
191 | unsigned int nr_bits; | |
192 | unsigned int set_bits; | |
7ebd796f JA |
193 | }; |
194 | ||
e39c0676 | 195 | static bool axmap_set_fn(struct axmap_level *al, unsigned long offset, |
7ebd796f JA |
196 | unsigned int bit, void *__data) |
197 | { | |
198 | struct axmap_set_data *data = __data; | |
199 | unsigned long mask, overlap; | |
200 | unsigned int nr_bits; | |
201 | ||
202 | nr_bits = min(data->nr_bits, BLOCKS_PER_UNIT - bit); | |
203 | ||
204 | mask = bit_masks[nr_bits] << bit; | |
205 | ||
206 | /* | |
207 | * Mask off any potential overlap, only sets contig regions | |
208 | */ | |
209 | overlap = al->map[offset] & mask; | |
bd71edce JA |
210 | if (overlap == mask) { |
211 | done: | |
212 | data->set_bits = 0; | |
e39c0676 | 213 | return true; |
bd71edce | 214 | } |
7ebd796f | 215 | |
8cd5a2fd JA |
216 | if (overlap) { |
217 | const int __bit = ffz(~overlap); | |
7ebd796f | 218 | |
c2e48e0e JA |
219 | nr_bits = __bit - bit; |
220 | if (!nr_bits) | |
bd71edce | 221 | goto done; |
8cd5a2fd | 222 | |
8cd5a2fd | 223 | mask = bit_masks[nr_bits] << bit; |
7ebd796f JA |
224 | } |
225 | ||
226 | assert(mask); | |
227 | assert(!(al->map[offset] & mask)); | |
7ebd796f JA |
228 | al->map[offset] |= mask; |
229 | ||
230 | if (!al->level) | |
231 | data->set_bits = nr_bits; | |
232 | ||
233 | data->nr_bits = 1; | |
234 | return al->map[offset] != -1UL; | |
235 | } | |
236 | ||
237 | static void __axmap_set(struct axmap *axmap, uint64_t bit_nr, | |
238 | struct axmap_set_data *data) | |
239 | { | |
240 | unsigned int set_bits, nr_bits = data->nr_bits; | |
241 | ||
47d94b0b JA |
242 | if (bit_nr > axmap->nr_bits) |
243 | return; | |
244 | else if (bit_nr + nr_bits > axmap->nr_bits) | |
245 | nr_bits = axmap->nr_bits - bit_nr; | |
246 | ||
7ebd796f JA |
247 | set_bits = 0; |
248 | while (nr_bits) { | |
249 | axmap_handler(axmap, bit_nr, axmap_set_fn, data); | |
250 | set_bits += data->set_bits; | |
251 | ||
0bfdf9f1 JA |
252 | if (!data->set_bits || |
253 | data->set_bits != (BLOCKS_PER_UNIT - nr_bits)) | |
7ebd796f JA |
254 | break; |
255 | ||
256 | nr_bits -= data->set_bits; | |
257 | bit_nr += data->set_bits; | |
258 | ||
259 | data->nr_bits = nr_bits; | |
7ebd796f JA |
260 | } |
261 | ||
262 | data->set_bits = set_bits; | |
263 | } | |
264 | ||
265 | void axmap_set(struct axmap *axmap, uint64_t bit_nr) | |
266 | { | |
267 | struct axmap_set_data data = { .nr_bits = 1, }; | |
268 | ||
269 | __axmap_set(axmap, bit_nr, &data); | |
270 | } | |
271 | ||
e39c0676 JA |
272 | unsigned int axmap_set_nr(struct axmap *axmap, uint64_t bit_nr, |
273 | unsigned int nr_bits) | |
7ebd796f | 274 | { |
0bfdf9f1 JA |
275 | unsigned int set_bits = 0; |
276 | ||
277 | do { | |
09d6bf09 | 278 | struct axmap_set_data data = { .nr_bits = nr_bits, }; |
0bfdf9f1 JA |
279 | unsigned int max_bits, this_set; |
280 | ||
281 | max_bits = BLOCKS_PER_UNIT - (bit_nr & BLOCKS_PER_UNIT_MASK); | |
bd71edce | 282 | if (nr_bits > max_bits) |
0bfdf9f1 JA |
283 | data.nr_bits = max_bits; |
284 | ||
285 | this_set = data.nr_bits; | |
286 | __axmap_set(axmap, bit_nr, &data); | |
287 | set_bits += data.set_bits; | |
288 | if (data.set_bits != this_set) | |
289 | break; | |
7ebd796f | 290 | |
0bfdf9f1 JA |
291 | nr_bits -= data.set_bits; |
292 | bit_nr += data.set_bits; | |
293 | } while (nr_bits); | |
294 | ||
295 | return set_bits; | |
7ebd796f JA |
296 | } |
297 | ||
e39c0676 JA |
298 | static bool axmap_isset_fn(struct axmap_level *al, unsigned long offset, |
299 | unsigned int bit, void *unused) | |
7ebd796f JA |
300 | { |
301 | return (al->map[offset] & (1UL << bit)) != 0; | |
302 | } | |
303 | ||
e39c0676 | 304 | bool axmap_isset(struct axmap *axmap, uint64_t bit_nr) |
7ebd796f | 305 | { |
ac569176 | 306 | if (bit_nr <= axmap->nr_bits) |
c406445a | 307 | return axmap_handler_topdown(axmap, bit_nr, axmap_isset_fn); |
47d94b0b | 308 | |
e39c0676 | 309 | return false; |
7ebd796f JA |
310 | } |
311 | ||
b9304c5b MK |
312 | /* |
313 | * Find the first free bit that is at least as large as bit_nr. Return | |
314 | * -1 if no free bit is found before the end of the map. | |
315 | */ | |
316 | static uint64_t axmap_find_first_free(struct axmap *axmap, uint64_t bit_nr) | |
7ebd796f | 317 | { |
731ef4c7 | 318 | int i; |
b9304c5b MK |
319 | unsigned long temp; |
320 | unsigned int bit; | |
321 | uint64_t offset, base_index, index; | |
322 | struct axmap_level *al; | |
7ebd796f | 323 | |
b9304c5b MK |
324 | index = 0; |
325 | for (i = axmap->nr_levels - 1; i >= 0; i--) { | |
326 | al = &axmap->levels[i]; | |
7ebd796f | 327 | |
b9304c5b MK |
328 | /* Shift previously calculated index for next level */ |
329 | index <<= UNIT_SHIFT; | |
330 | ||
331 | /* | |
332 | * Start from an index that's at least as large as the | |
333 | * originally passed in bit number. | |
334 | */ | |
335 | base_index = bit_nr >> (UNIT_SHIFT * i); | |
336 | if (index < base_index) | |
337 | index = base_index; | |
338 | ||
339 | /* Get the offset and bit for this level */ | |
340 | offset = index >> UNIT_SHIFT; | |
341 | bit = index & BLOCKS_PER_UNIT_MASK; | |
342 | ||
343 | /* | |
344 | * If the previous level had unused bits in its last | |
345 | * word, the offset could be bigger than the map at | |
346 | * this level. That means no free bits exist before the | |
347 | * end of the map, so return -1. | |
348 | */ | |
349 | if (offset >= al->map_size) | |
350 | return -1ULL; | |
351 | ||
352 | /* Check the first word starting with the specific bit */ | |
353 | temp = ~bit_masks[bit] & ~al->map[offset]; | |
354 | if (temp) | |
355 | goto found; | |
356 | ||
357 | /* | |
358 | * No free bit in the first word, so iterate | |
359 | * looking for a word with one or more free bits. | |
360 | */ | |
361 | for (offset++; offset < al->map_size; offset++) { | |
362 | temp = ~al->map[offset]; | |
363 | if (temp) | |
364 | goto found; | |
7ebd796f | 365 | } |
47d94b0b | 366 | |
b9304c5b MK |
367 | /* Did not find a free bit */ |
368 | return -1ULL; | |
7ebd796f | 369 | |
b9304c5b MK |
370 | found: |
371 | /* Compute the index of the free bit just found */ | |
372 | index = (offset << UNIT_SHIFT) + ffz(~temp); | |
7ebd796f JA |
373 | } |
374 | ||
b9304c5b MK |
375 | /* If found an unused bit in the last word of level 0, return -1 */ |
376 | if (index >= axmap->nr_bits) | |
377 | return -1ULL; | |
378 | ||
379 | return index; | |
7ebd796f JA |
380 | } |
381 | ||
382 | /* | |
383 | * 'bit_nr' is already set. Find the next free bit after this one. | |
b9304c5b | 384 | * Return -1 if no free bits found. |
7ebd796f JA |
385 | */ |
386 | uint64_t axmap_next_free(struct axmap *axmap, uint64_t bit_nr) | |
387 | { | |
53737ae0 | 388 | uint64_t ret; |
b9304c5b MK |
389 | uint64_t next_bit = bit_nr + 1; |
390 | unsigned long temp; | |
391 | uint64_t offset; | |
392 | unsigned int bit; | |
7ebd796f | 393 | |
b9304c5b MK |
394 | if (bit_nr >= axmap->nr_bits) |
395 | return -1ULL; | |
12bde369 | 396 | |
b9304c5b MK |
397 | /* If at the end of the map, wrap-around */ |
398 | if (next_bit == axmap->nr_bits) | |
399 | next_bit = 0; | |
7ebd796f | 400 | |
b9304c5b MK |
401 | offset = next_bit >> UNIT_SHIFT; |
402 | bit = next_bit & BLOCKS_PER_UNIT_MASK; | |
7ebd796f | 403 | |
53737ae0 | 404 | /* |
b9304c5b MK |
405 | * As an optimization, do a quick check for a free bit |
406 | * in the current word at level 0. If not found, do | |
407 | * a topdown search. | |
53737ae0 | 408 | */ |
b9304c5b MK |
409 | temp = ~bit_masks[bit] & ~axmap->levels[0].map[offset]; |
410 | if (temp) { | |
411 | ret = (offset << UNIT_SHIFT) + ffz(~temp); | |
53737ae0 | 412 | |
b9304c5b MK |
413 | /* Might have found an unused bit at level 0 */ |
414 | if (ret >= axmap->nr_bits) | |
415 | ret = -1ULL; | |
416 | } else | |
417 | ret = axmap_find_first_free(axmap, next_bit); | |
418 | ||
419 | /* | |
420 | * If there are no free bits starting at next_bit and going | |
421 | * to the end of the map, wrap around by searching again | |
422 | * starting at bit 0. | |
423 | */ | |
424 | if (ret == -1ULL && next_bit != 0) | |
425 | ret = axmap_find_first_free(axmap, 0); | |
426 | return ret; | |
7ebd796f | 427 | } |