Subversion Repositories HelenOS

Rev

Rev 3183 | Rev 4303 | Go to most recent revision | Details | Compare with Previous | Last modification | View Log | RSS feed

Rev Author Line No. Line
759 palkovsky 1
/*
2071 jermar 2
 * Copyright (c) 2006 Ondrej Palkovsky
759 palkovsky 3
 * All rights reserved.
4
 *
5
 * Redistribution and use in source and binary forms, with or without
6
 * modification, are permitted provided that the following conditions
7
 * are met:
8
 *
9
 * - Redistributions of source code must retain the above copyright
10
 *   notice, this list of conditions and the following disclaimer.
11
 * - Redistributions in binary form must reproduce the above copyright
12
 *   notice, this list of conditions and the following disclaimer in the
13
 *   documentation and/or other materials provided with the distribution.
14
 * - The name of the author may not be used to endorse or promote products
15
 *   derived from this software without specific prior written permission.
16
 *
17
 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
18
 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
19
 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
20
 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
21
 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
22
 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23
 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24
 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25
 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
26
 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27
 */
28
 
1757 jermar 29
/** @addtogroup genericmm
1702 cejka 30
 * @{
31
 */
32
 
1248 jermar 33
/**
1702 cejka 34
 * @file
1248 jermar 35
 * @brief   Slab allocator.
769 palkovsky 36
 *
1248 jermar 37
 * The slab allocator is closely modelled after OpenSolaris slab allocator.
38
 * @see http://www.usenix.org/events/usenix01/full_papers/bonwick/bonwick_html/
39
 *
769 palkovsky 40
 * with the following exceptions:
1248 jermar 41
 * @li empty slabs are deallocated immediately
769 palkovsky 42
 *     (in Linux they are kept in linked list, in Solaris ???)
1248 jermar 43
 * @li empty magazines are deallocated when not needed
769 palkovsky 44
 *     (in Solaris they are held in linked list in slab cache)
45
 *
1248 jermar 46
 * Following features are not currently supported but would be easy to do:
47
 * @li cache coloring
48
 * @li dynamic magazine growing (different magazine sizes are already
1144 jermar 49
 *     supported, but we would need to adjust allocation strategy)
769 palkovsky 50
 *
1248 jermar 51
 * The slab allocator supports per-CPU caches ('magazines') to facilitate
769 palkovsky 52
 * good SMP scaling.
53
 *
54
 * When a new object is being allocated, it is first checked, if it is
1554 jermar 55
 * available in a CPU-bound magazine. If it is not found there, it is
56
 * allocated from a CPU-shared slab - if a partially full one is found,
57
 * it is used, otherwise a new one is allocated.
769 palkovsky 58
 *
1554 jermar 59
 * When an object is being deallocated, it is put to a CPU-bound magazine.
60
 * If there is no such magazine, a new one is allocated (if this fails,
1248 jermar 61
 * the object is deallocated into slab). If the magazine is full, it is
1554 jermar 62
 * put into cpu-shared list of magazines and a new one is allocated.
769 palkovsky 63
 *
1554 jermar 64
 * The CPU-bound magazine is actually a pair of magazines in order to avoid
769 palkovsky 65
 * thrashing when somebody is allocating/deallocating 1 item at the magazine
66
 * size boundary. LIFO order is enforced, which should avoid fragmentation
67
 * as much as possible.
68
 *  
1554 jermar 69
 * Every cache contains list of full slabs and list of partially full slabs.
1248 jermar 70
 * Empty slabs are immediately freed (thrashing will be avoided because
769 palkovsky 71
 * of magazines).
72
 *
1248 jermar 73
 * The slab information structure is kept inside the data area, if possible.
769 palkovsky 74
 * The cache can be marked that it should not use magazines. This is used
1248 jermar 75
 * only for slab related caches to avoid deadlocks and infinite recursion
76
 * (the slab allocator uses itself for allocating all it's control structures).
769 palkovsky 77
 *
1554 jermar 78
 * The slab allocator allocates a lot of space and does not free it. When
79
 * the frame allocator fails to allocate a frame, it calls slab_reclaim().
769 palkovsky 80
 * It tries 'light reclaim' first, then brutal reclaim. The light reclaim
81
 * releases slabs from cpu-shared magazine-list, until at least 1 slab
82
 * is deallocated in each cache (this algorithm should probably change).
83
 * The brutal reclaim removes all cached objects, even from CPU-bound
84
 * magazines.
85
 *
1757 jermar 86
 * @todo
1248 jermar 87
 * For better CPU-scaling the magazine allocation strategy should
775 palkovsky 88
 * be extended. Currently, if the cache does not have magazine, it asks
89
 * for non-cpu cached magazine cache to provide one. It might be feasible
90
 * to add cpu-cached magazine cache (which would allocate it's magazines
91
 * from non-cpu-cached mag. cache). This would provide a nice per-cpu
92
 * buffer. The other possibility is to use the per-cache
93
 * 'empty-magazine-list', which decreases competing for 1 per-system
94
 * magazine cache.
95
 *
1757 jermar 96
 * @todo
97
 * it might be good to add granularity of locks even to slab level,
98
 * we could then try_spinlock over all partial slabs and thus improve
99
 * scalability even on slab level
769 palkovsky 100
 */
101
 
759 palkovsky 102
#include <synch/spinlock.h>
103
#include <mm/slab.h>
788 jermar 104
#include <adt/list.h>
759 palkovsky 105
#include <memstr.h>
106
#include <align.h>
762 palkovsky 107
#include <mm/frame.h>
759 palkovsky 108
#include <config.h>
109
#include <print.h>
110
#include <arch.h>
111
#include <panic.h>
762 palkovsky 112
#include <debug.h>
771 palkovsky 113
#include <bitops.h>
2124 decky 114
#include <macros.h>
759 palkovsky 115
 
116
SPINLOCK_INITIALIZE(slab_cache_lock);
769 palkovsky 117
static LIST_INITIALIZE(slab_cache_list);
759 palkovsky 118
 
769 palkovsky 119
/** Magazine cache */
120
static slab_cache_t mag_cache;
121
/** Cache for cache descriptors */
122
static slab_cache_t slab_cache_cache;
123
/** Cache for external slab descriptors
124
 * This time we want per-cpu cache, so do not make it static
1248 jermar 125
 * - using slab for internal slab structures will not deadlock,
769 palkovsky 126
 *   as all slab structures are 'small' - control structures of
127
 *   their caches do not require further allocation
128
 */
129
static slab_cache_t *slab_extern_cache;
771 palkovsky 130
/** Caches for malloc */
2124 decky 131
static slab_cache_t *malloc_caches[SLAB_MAX_MALLOC_W - SLAB_MIN_MALLOC_W + 1];
771 palkovsky 132
char *malloc_names[] =  {
2124 decky 133
    "malloc-16",
134
    "malloc-32",
135
    "malloc-64",
136
    "malloc-128",
137
    "malloc-256",
138
    "malloc-512",
139
    "malloc-1K",
140
    "malloc-2K",
141
    "malloc-4K",
142
    "malloc-8K",
143
    "malloc-16K",
144
    "malloc-32K",
145
    "malloc-64K",
146
    "malloc-128K",
147
    "malloc-256K"
771 palkovsky 148
};
762 palkovsky 149
 
769 palkovsky 150
/** Slab descriptor */
762 palkovsky 151
typedef struct {
1950 jermar 152
    slab_cache_t *cache;    /**< Pointer to parent cache. */
153
    link_t link;        /**< List of full/partial slabs. */
154
    void *start;        /**< Start address of first available item. */
155
    count_t available;  /**< Count of available items in this slab. */
156
    index_t nextavail;  /**< The index of next available item. */
2124 decky 157
} slab_t;
762 palkovsky 158
 
791 palkovsky 159
#ifdef CONFIG_DEBUG
160
static int _slab_initialized = 0;
161
#endif
162
 
759 palkovsky 163
/**************************************/
1248 jermar 164
/* Slab allocation functions          */
759 palkovsky 165
 
762 palkovsky 166
/**
167
 * Allocate frames for slab space and initialize
168
 *
169
 */
3180 jermar 170
static slab_t *slab_space_alloc(slab_cache_t *cache, int flags)
762 palkovsky 171
{
172
    void *data;
173
    slab_t *slab;
174
    size_t fsize;
2745 decky 175
    unsigned int i;
3972 decky 176
    count_t zone = 0;
814 palkovsky 177
 
1766 palkovsky 178
    data = frame_alloc_generic(cache->order, FRAME_KA | flags, &zone);
179
    if (!data) {
762 palkovsky 180
        return NULL;
764 palkovsky 181
    }
3180 jermar 182
    if (!(cache->flags & SLAB_CACHE_SLINSIDE)) {
769 palkovsky 183
        slab = slab_alloc(slab_extern_cache, flags);
762 palkovsky 184
        if (!slab) {
1760 palkovsky 185
            frame_free(KA2PA(data));
762 palkovsky 186
            return NULL;
187
        }
188
    } else {
189
        fsize = (PAGE_SIZE << cache->order);
190
        slab = data + fsize - sizeof(*slab);
191
    }
1288 jermar 192
 
762 palkovsky 193
    /* Fill in slab structures */
2745 decky 194
    for (i = 0; i < ((unsigned int) 1 << cache->order); i++)
195
        frame_set_parent(ADDR2PFN(KA2PA(data)) + i, slab, zone);
762 palkovsky 196
 
197
    slab->start = data;
198
    slab->available = cache->objects;
199
    slab->nextavail = 0;
767 palkovsky 200
    slab->cache = cache;
762 palkovsky 201
 
2745 decky 202
    for (i = 0; i < cache->objects; i++)
3180 jermar 203
        *((int *) (slab->start + i*cache->size)) = i + 1;
764 palkovsky 204
 
205
    atomic_inc(&cache->allocated_slabs);
762 palkovsky 206
    return slab;
207
}
208
 
759 palkovsky 209
/**
1248 jermar 210
 * Deallocate space associated with slab
762 palkovsky 211
 *
212
 * @return number of freed frames
213
 */
214
static count_t slab_space_free(slab_cache_t *cache, slab_t *slab)
215
{
1760 palkovsky 216
    frame_free(KA2PA(slab->start));
768 palkovsky 217
    if (! (cache->flags & SLAB_CACHE_SLINSIDE))
769 palkovsky 218
        slab_free(slab_extern_cache, slab);
764 palkovsky 219
 
220
    atomic_dec(&cache->allocated_slabs);
221
 
762 palkovsky 222
    return 1 << cache->order;
223
}
224
 
225
/** Map object to slab structure */
226
static slab_t * obj2slab(void *obj)
227
{
2124 decky 228
    return (slab_t *) frame_get_parent(ADDR2PFN(KA2PA(obj)), 0);
762 palkovsky 229
}
230
 
231
/**************************************/
1248 jermar 232
/* Slab functions */
762 palkovsky 233
 
234
 
235
/**
759 palkovsky 236
 * Return object to slab and call a destructor
237
 *
762 palkovsky 238
 * @param slab If the caller knows directly slab of the object, otherwise NULL
239
 *
759 palkovsky 240
 * @return Number of freed pages
241
 */
3180 jermar 242
static count_t slab_obj_destroy(slab_cache_t *cache, void *obj, slab_t *slab)
759 palkovsky 243
{
787 palkovsky 244
    int freed = 0;
245
 
762 palkovsky 246
    if (!slab)
247
        slab = obj2slab(obj);
248
 
767 palkovsky 249
    ASSERT(slab->cache == cache);
250
 
787 palkovsky 251
    if (cache->destructor)
252
        freed = cache->destructor(obj);
253
 
776 palkovsky 254
    spinlock_lock(&cache->slablock);
789 palkovsky 255
    ASSERT(slab->available < cache->objects);
776 palkovsky 256
 
762 palkovsky 257
    *((int *)obj) = slab->nextavail;
3180 jermar 258
    slab->nextavail = (obj - slab->start) / cache->size;
762 palkovsky 259
    slab->available++;
260
 
261
    /* Move it to correct list */
262
    if (slab->available == cache->objects) {
263
        /* Free associated memory */
264
        list_remove(&slab->link);
782 palkovsky 265
        spinlock_unlock(&cache->slablock);
266
 
787 palkovsky 267
        return freed + slab_space_free(cache, slab);
782 palkovsky 268
 
780 palkovsky 269
    } else if (slab->available == 1) {
270
        /* It was in full, move to partial */
271
        list_remove(&slab->link);
272
        list_prepend(&slab->link, &cache->partial_slabs);
762 palkovsky 273
    }
783 palkovsky 274
    spinlock_unlock(&cache->slablock);
787 palkovsky 275
    return freed;
759 palkovsky 276
}
277
 
278
/**
279
 * Take new object from slab or create new if needed
280
 *
281
 * @return Object address or null
282
 */
3180 jermar 283
static void *slab_obj_create(slab_cache_t *cache, int flags)
759 palkovsky 284
{
762 palkovsky 285
    slab_t *slab;
286
    void *obj;
287
 
776 palkovsky 288
    spinlock_lock(&cache->slablock);
289
 
762 palkovsky 290
    if (list_empty(&cache->partial_slabs)) {
291
        /* Allow recursion and reclaiming
1248 jermar 292
         * - this should work, as the slab control structures
1288 jermar 293
         *   are small and do not need to allocate with anything
294
         *   other than frame_alloc when they are allocating,
762 palkovsky 295
         *   that's why we should get recursion at most 1-level deep
296
         */
776 palkovsky 297
        spinlock_unlock(&cache->slablock);
762 palkovsky 298
        slab = slab_space_alloc(cache, flags);
780 palkovsky 299
        if (!slab)
300
            return NULL;
776 palkovsky 301
        spinlock_lock(&cache->slablock);
762 palkovsky 302
    } else {
3180 jermar 303
        slab = list_get_instance(cache->partial_slabs.next, slab_t,
304
            link);
762 palkovsky 305
        list_remove(&slab->link);
306
    }
307
    obj = slab->start + slab->nextavail * cache->size;
308
    slab->nextavail = *((int *)obj);
309
    slab->available--;
787 palkovsky 310
 
1950 jermar 311
    if (!slab->available)
764 palkovsky 312
        list_prepend(&slab->link, &cache->full_slabs);
762 palkovsky 313
    else
764 palkovsky 314
        list_prepend(&slab->link, &cache->partial_slabs);
776 palkovsky 315
 
316
    spinlock_unlock(&cache->slablock);
787 palkovsky 317
 
318
    if (cache->constructor && cache->constructor(obj, flags)) {
319
        /* Bad, bad, construction failed */
320
        slab_obj_destroy(cache, obj, slab);
321
        return NULL;
322
    }
762 palkovsky 323
    return obj;
759 palkovsky 324
}
325
 
326
/**************************************/
327
/* CPU-Cache slab functions */
328
 
329
/**
781 palkovsky 330
 * Finds a full magazine in cache, takes it from list
331
 * and returns it
332
 *
333
 * @param first If true, return first, else last mag
334
 */
3180 jermar 335
static slab_magazine_t *get_mag_from_cache(slab_cache_t *cache, int first)
781 palkovsky 336
{
337
    slab_magazine_t *mag = NULL;
338
    link_t *cur;
339
 
340
    spinlock_lock(&cache->maglock);
341
    if (!list_empty(&cache->magazines)) {
342
        if (first)
343
            cur = cache->magazines.next;
344
        else
345
            cur = cache->magazines.prev;
346
        mag = list_get_instance(cur, slab_magazine_t, link);
347
        list_remove(&mag->link);
348
        atomic_dec(&cache->magazine_counter);
349
    }
350
    spinlock_unlock(&cache->maglock);
351
    return mag;
352
}
353
 
354
/** Prepend magazine to magazine list in cache */
355
static void put_mag_to_cache(slab_cache_t *cache, slab_magazine_t *mag)
356
{
357
    spinlock_lock(&cache->maglock);
358
 
359
    list_prepend(&mag->link, &cache->magazines);
360
    atomic_inc(&cache->magazine_counter);
361
 
362
    spinlock_unlock(&cache->maglock);
363
}
364
 
365
/**
759 palkovsky 366
 * Free all objects in magazine and free memory associated with magazine
367
 *
368
 * @return Number of freed pages
369
 */
3180 jermar 370
static count_t magazine_destroy(slab_cache_t *cache, slab_magazine_t *mag)
759 palkovsky 371
{
2745 decky 372
    unsigned int i;
759 palkovsky 373
    count_t frames = 0;
374
 
2745 decky 375
    for (i = 0; i < mag->busy; i++) {
762 palkovsky 376
        frames += slab_obj_destroy(cache, mag->objs[i], NULL);
767 palkovsky 377
        atomic_dec(&cache->cached_objs);
378
    }
759 palkovsky 379
 
380
    slab_free(&mag_cache, mag);
381
 
382
    return frames;
383
}
384
 
385
/**
769 palkovsky 386
 * Find full magazine, set it as current and return it
387
 *
388
 * Assume cpu_magazine lock is held
389
 */
3180 jermar 390
static slab_magazine_t *get_full_current_mag(slab_cache_t *cache)
769 palkovsky 391
{
392
    slab_magazine_t *cmag, *lastmag, *newmag;
393
 
394
    cmag = cache->mag_cache[CPU->id].current;
395
    lastmag = cache->mag_cache[CPU->id].last;
396
    if (cmag) { /* First try local CPU magazines */
397
        if (cmag->busy)
398
            return cmag;
399
 
400
        if (lastmag && lastmag->busy) {
401
            cache->mag_cache[CPU->id].current = lastmag;
402
            cache->mag_cache[CPU->id].last = cmag;
403
            return lastmag;
404
        }
405
    }
406
    /* Local magazines are empty, import one from magazine list */
781 palkovsky 407
    newmag = get_mag_from_cache(cache, 1);
408
    if (!newmag)
769 palkovsky 409
        return NULL;
410
 
411
    if (lastmag)
781 palkovsky 412
        magazine_destroy(cache, lastmag);
413
 
769 palkovsky 414
    cache->mag_cache[CPU->id].last = cmag;
415
    cache->mag_cache[CPU->id].current = newmag;
416
    return newmag;
417
}
418
 
419
/**
759 palkovsky 420
 * Try to find object in CPU-cache magazines
421
 *
422
 * @return Pointer to object or NULL if not available
423
 */
3180 jermar 424
static void *magazine_obj_get(slab_cache_t *cache)
759 palkovsky 425
{
426
    slab_magazine_t *mag;
767 palkovsky 427
    void *obj;
759 palkovsky 428
 
772 palkovsky 429
    if (!CPU)
430
        return NULL;
431
 
759 palkovsky 432
    spinlock_lock(&cache->mag_cache[CPU->id].lock);
433
 
769 palkovsky 434
    mag = get_full_current_mag(cache);
435
    if (!mag) {
436
        spinlock_unlock(&cache->mag_cache[CPU->id].lock);
437
        return NULL;
759 palkovsky 438
    }
767 palkovsky 439
    obj = mag->objs[--mag->busy];
759 palkovsky 440
    spinlock_unlock(&cache->mag_cache[CPU->id].lock);
767 palkovsky 441
    atomic_dec(&cache->cached_objs);
442
 
443
    return obj;
759 palkovsky 444
}
445
 
446
/**
768 palkovsky 447
 * Assure that the current magazine is empty, return pointer to it, or NULL if
769 palkovsky 448
 * no empty magazine is available and cannot be allocated
759 palkovsky 449
 *
773 palkovsky 450
 * Assume mag_cache[CPU->id].lock is held
451
 *
759 palkovsky 452
 * We have 2 magazines bound to processor.
453
 * First try the current.
454
 *  If full, try the last.
455
 *   If full, put to magazines list.
456
 *   allocate new, exchange last & current
457
 *
768 palkovsky 458
 */
3180 jermar 459
static slab_magazine_t *make_empty_current_mag(slab_cache_t *cache)
768 palkovsky 460
{
461
    slab_magazine_t *cmag,*lastmag,*newmag;
462
 
463
    cmag = cache->mag_cache[CPU->id].current;
464
    lastmag = cache->mag_cache[CPU->id].last;
465
 
466
    if (cmag) {
467
        if (cmag->busy < cmag->size)
468
            return cmag;
469
        if (lastmag && lastmag->busy < lastmag->size) {
470
            cache->mag_cache[CPU->id].last = cmag;
471
            cache->mag_cache[CPU->id].current = lastmag;
472
            return lastmag;
473
        }
474
    }
475
    /* current | last are full | nonexistent, allocate new */
476
    /* We do not want to sleep just because of caching */
477
    /* Especially we do not want reclaiming to start, as
478
     * this would deadlock */
479
    newmag = slab_alloc(&mag_cache, FRAME_ATOMIC | FRAME_NO_RECLAIM);
480
    if (!newmag)
481
        return NULL;
482
    newmag->size = SLAB_MAG_SIZE;
483
    newmag->busy = 0;
484
 
485
    /* Flush last to magazine list */
781 palkovsky 486
    if (lastmag)
487
        put_mag_to_cache(cache, lastmag);
488
 
768 palkovsky 489
    /* Move current as last, save new as current */
490
    cache->mag_cache[CPU->id].last = cmag; 
491
    cache->mag_cache[CPU->id].current = newmag;
492
 
493
    return newmag;
494
}
495
 
496
/**
497
 * Put object into CPU-cache magazine
498
 *
759 palkovsky 499
 * @return 0 - success, -1 - could not get memory
500
 */
501
static int magazine_obj_put(slab_cache_t *cache, void *obj)
502
{
503
    slab_magazine_t *mag;
504
 
772 palkovsky 505
    if (!CPU)
506
        return -1;
507
 
759 palkovsky 508
    spinlock_lock(&cache->mag_cache[CPU->id].lock);
768 palkovsky 509
 
510
    mag = make_empty_current_mag(cache);
769 palkovsky 511
    if (!mag) {
512
        spinlock_unlock(&cache->mag_cache[CPU->id].lock);
513
        return -1;
514
    }
759 palkovsky 515
 
516
    mag->objs[mag->busy++] = obj;
517
 
518
    spinlock_unlock(&cache->mag_cache[CPU->id].lock);
767 palkovsky 519
    atomic_inc(&cache->cached_objs);
759 palkovsky 520
    return 0;
521
}
522
 
523
 
524
/**************************************/
1248 jermar 525
/* Slab cache functions */
759 palkovsky 526
 
762 palkovsky 527
/** Return number of objects that fit in certain cache size */
2745 decky 528
static unsigned int comp_objects(slab_cache_t *cache)
762 palkovsky 529
{
530
    if (cache->flags & SLAB_CACHE_SLINSIDE)
3180 jermar 531
        return ((PAGE_SIZE << cache->order) - sizeof(slab_t)) /
532
            cache->size;
762 palkovsky 533
    else
534
        return (PAGE_SIZE << cache->order) / cache->size;
535
}
536
 
537
/** Return wasted space in slab */
2745 decky 538
static unsigned int badness(slab_cache_t *cache)
762 palkovsky 539
{
2745 decky 540
    unsigned int objects;
541
    unsigned int ssize;
762 palkovsky 542
 
543
    objects = comp_objects(cache);
544
    ssize = PAGE_SIZE << cache->order;
545
    if (cache->flags & SLAB_CACHE_SLINSIDE)
546
        ssize -= sizeof(slab_t);
2745 decky 547
    return ssize - objects * cache->size;
762 palkovsky 548
}
549
 
789 palkovsky 550
/**
551
 * Initialize mag_cache structure in slab cache
552
 */
553
static void make_magcache(slab_cache_t *cache)
554
{
2745 decky 555
    unsigned int i;
791 palkovsky 556
 
557
    ASSERT(_slab_initialized >= 2);
789 palkovsky 558
 
3180 jermar 559
    cache->mag_cache = malloc(sizeof(slab_mag_cache_t) * config.cpu_count,
560
        0);
2745 decky 561
    for (i = 0; i < config.cpu_count; i++) {
3104 svoboda 562
        memsetb(&cache->mag_cache[i], sizeof(cache->mag_cache[i]), 0);
3180 jermar 563
        spinlock_initialize(&cache->mag_cache[i].lock,
564
            "slab_maglock_cpu");
789 palkovsky 565
    }
566
}
567
 
759 palkovsky 568
/** Initialize allocated memory as a slab cache */
569
static void
3180 jermar 570
_slab_cache_create(slab_cache_t *cache, char *name, size_t size, size_t align,
571
    int (*constructor)(void *obj, int kmflag), int (*destructor)(void *obj),
572
    int flags)
759 palkovsky 573
{
771 palkovsky 574
    int pages;
783 palkovsky 575
    ipl_t ipl;
759 palkovsky 576
 
3104 svoboda 577
    memsetb(cache, sizeof(*cache), 0);
759 palkovsky 578
    cache->name = name;
579
 
1780 jermar 580
    if (align < sizeof(unative_t))
581
        align = sizeof(unative_t);
766 palkovsky 582
    size = ALIGN_UP(size, align);
583
 
762 palkovsky 584
    cache->size = size;
759 palkovsky 585
 
586
    cache->constructor = constructor;
587
    cache->destructor = destructor;
588
    cache->flags = flags;
589
 
590
    list_initialize(&cache->full_slabs);
591
    list_initialize(&cache->partial_slabs);
592
    list_initialize(&cache->magazines);
776 palkovsky 593
    spinlock_initialize(&cache->slablock, "slab_lock");
594
    spinlock_initialize(&cache->maglock, "slab_maglock");
3180 jermar 595
    if (!(cache->flags & SLAB_CACHE_NOMAGAZINE))
789 palkovsky 596
        make_magcache(cache);
759 palkovsky 597
 
598
    /* Compute slab sizes, object counts in slabs etc. */
599
    if (cache->size < SLAB_INSIDE_SIZE)
600
        cache->flags |= SLAB_CACHE_SLINSIDE;
601
 
762 palkovsky 602
    /* Minimum slab order */
1682 palkovsky 603
    pages = SIZE2FRAMES(cache->size);
1677 palkovsky 604
    /* We need the 2^order >= pages */
605
    if (pages == 1)
606
        cache->order = 0;
607
    else
3180 jermar 608
        cache->order = fnzb(pages - 1) + 1;
766 palkovsky 609
 
762 palkovsky 610
    while (badness(cache) > SLAB_MAX_BADNESS(cache)) {
611
        cache->order += 1;
612
    }
613
    cache->objects = comp_objects(cache);
766 palkovsky 614
    /* If info fits in, put it inside */
615
    if (badness(cache) > sizeof(slab_t))
616
        cache->flags |= SLAB_CACHE_SLINSIDE;
762 palkovsky 617
 
783 palkovsky 618
    /* Add cache to cache list */
619
    ipl = interrupts_disable();
759 palkovsky 620
    spinlock_lock(&slab_cache_lock);
621
 
622
    list_append(&cache->link, &slab_cache_list);
623
 
624
    spinlock_unlock(&slab_cache_lock);
783 palkovsky 625
    interrupts_restore(ipl);
759 palkovsky 626
}
627
 
628
/** Create slab cache  */
3180 jermar 629
slab_cache_t *
630
slab_cache_create(char *name, size_t size, size_t align,
631
    int (*constructor)(void *obj, int kmflag), int (*destructor)(void *obj),
632
    int flags)
759 palkovsky 633
{
634
    slab_cache_t *cache;
635
 
769 palkovsky 636
    cache = slab_alloc(&slab_cache_cache, 0);
759 palkovsky 637
    _slab_cache_create(cache, name, size, align, constructor, destructor,
3180 jermar 638
        flags);
759 palkovsky 639
    return cache;
640
}
641
 
642
/**
643
 * Reclaim space occupied by objects that are already free
644
 *
645
 * @param flags If contains SLAB_RECLAIM_ALL, do aggressive freeing
646
 * @return Number of freed pages
647
 */
648
static count_t _slab_reclaim(slab_cache_t *cache, int flags)
649
{
2745 decky 650
    unsigned int i;
759 palkovsky 651
    slab_magazine_t *mag;
652
    count_t frames = 0;
781 palkovsky 653
    int magcount;
759 palkovsky 654
 
655
    if (cache->flags & SLAB_CACHE_NOMAGAZINE)
656
        return 0; /* Nothing to do */
781 palkovsky 657
 
658
    /* We count up to original magazine count to avoid
659
     * endless loop
660
     */
661
    magcount = atomic_get(&cache->magazine_counter);
3180 jermar 662
    while (magcount-- && (mag=get_mag_from_cache(cache, 0))) {
781 palkovsky 663
        frames += magazine_destroy(cache,mag);
664
        if (!(flags & SLAB_RECLAIM_ALL) && frames)
665
            break;
769 palkovsky 666
    }
759 palkovsky 667
 
668
    if (flags & SLAB_RECLAIM_ALL) {
781 palkovsky 669
        /* Free cpu-bound magazines */
759 palkovsky 670
        /* Destroy CPU magazines */
2745 decky 671
        for (i = 0; i < config.cpu_count; i++) {
781 palkovsky 672
            spinlock_lock(&cache->mag_cache[i].lock);
673
 
759 palkovsky 674
            mag = cache->mag_cache[i].current;
675
            if (mag)
676
                frames += magazine_destroy(cache, mag);
677
            cache->mag_cache[i].current = NULL;
678
 
679
            mag = cache->mag_cache[i].last;
680
            if (mag)
681
                frames += magazine_destroy(cache, mag);
682
            cache->mag_cache[i].last = NULL;
781 palkovsky 683
 
684
            spinlock_unlock(&cache->mag_cache[i].lock);
759 palkovsky 685
        }
686
    }
767 palkovsky 687
 
759 palkovsky 688
    return frames;
689
}
690
 
691
/** Check that there are no slabs and remove cache from system  */
692
void slab_cache_destroy(slab_cache_t *cache)
693
{
781 palkovsky 694
    ipl_t ipl;
695
 
696
    /* First remove cache from link, so that we don't need
697
     * to disable interrupts later
698
     */
699
 
700
    ipl = interrupts_disable();
701
    spinlock_lock(&slab_cache_lock);
702
 
703
    list_remove(&cache->link);
704
 
705
    spinlock_unlock(&slab_cache_lock);
706
    interrupts_restore(ipl);
707
 
759 palkovsky 708
    /* Do not lock anything, we assume the software is correct and
709
     * does not touch the cache when it decides to destroy it */
710
 
711
    /* Destroy all magazines */
712
    _slab_reclaim(cache, SLAB_RECLAIM_ALL);
713
 
714
    /* All slabs must be empty */
3180 jermar 715
    if (!list_empty(&cache->full_slabs) ||
716
        !list_empty(&cache->partial_slabs))
759 palkovsky 717
        panic("Destroying cache that is not empty.");
718
 
789 palkovsky 719
    if (!(cache->flags & SLAB_CACHE_NOMAGAZINE))
822 palkovsky 720
        free(cache->mag_cache);
769 palkovsky 721
    slab_free(&slab_cache_cache, cache);
759 palkovsky 722
}
723
 
3180 jermar 724
/** Allocate new object from cache - if no flags given, always returns memory */
725
void *slab_alloc(slab_cache_t *cache, int flags)
759 palkovsky 726
{
727
    ipl_t ipl;
728
    void *result = NULL;
773 palkovsky 729
 
759 palkovsky 730
    /* Disable interrupts to avoid deadlocks with interrupt handlers */
731
    ipl = interrupts_disable();
771 palkovsky 732
 
814 palkovsky 733
    if (!(cache->flags & SLAB_CACHE_NOMAGAZINE)) {
759 palkovsky 734
        result = magazine_obj_get(cache);
814 palkovsky 735
    }
776 palkovsky 736
    if (!result)
759 palkovsky 737
        result = slab_obj_create(cache, flags);
738
 
769 palkovsky 739
    interrupts_restore(ipl);
740
 
764 palkovsky 741
    if (result)
742
        atomic_inc(&cache->allocated_objs);
743
 
759 palkovsky 744
    return result;
745
}
746
 
771 palkovsky 747
/** Return object to cache, use slab if known  */
748
static void _slab_free(slab_cache_t *cache, void *obj, slab_t *slab)
759 palkovsky 749
{
750
    ipl_t ipl;
751
 
752
    ipl = interrupts_disable();
753
 
3180 jermar 754
    if ((cache->flags & SLAB_CACHE_NOMAGAZINE) ||
755
        magazine_obj_put(cache, obj)) {
771 palkovsky 756
        slab_obj_destroy(cache, obj, slab);
776 palkovsky 757
 
759 palkovsky 758
    }
769 palkovsky 759
    interrupts_restore(ipl);
764 palkovsky 760
    atomic_dec(&cache->allocated_objs);
759 palkovsky 761
}
762
 
771 palkovsky 763
/** Return slab object to cache */
764
void slab_free(slab_cache_t *cache, void *obj)
765
{
2124 decky 766
    _slab_free(cache, obj, NULL);
771 palkovsky 767
}
768
 
759 palkovsky 769
/* Go through all caches and reclaim what is possible */
770
count_t slab_reclaim(int flags)
771
{
772
    slab_cache_t *cache;
773
    link_t *cur;
774
    count_t frames = 0;
775
 
776
    spinlock_lock(&slab_cache_lock);
777
 
776 palkovsky 778
    /* TODO: Add assert, that interrupts are disabled, otherwise
779
     * memory allocation from interrupts can deadlock.
780
     */
781
 
3180 jermar 782
    for (cur = slab_cache_list.next; cur != &slab_cache_list;
783
        cur = cur->next) {
759 palkovsky 784
        cache = list_get_instance(cur, slab_cache_t, link);
785
        frames += _slab_reclaim(cache, flags);
786
    }
787
 
788
    spinlock_unlock(&slab_cache_lock);
789
 
790
    return frames;
791
}
792
 
793
 
794
/* Print list of slabs */
795
void slab_print_list(void)
796
{
3183 jermar 797
    int skip = 0;
798
 
3180 jermar 799
    printf("slab name        size     pages  obj/pg slabs  cached allocated"
800
        " ctl\n");
801
    printf("---------------- -------- ------ ------ ------ ------ ---------"
802
        " ---\n");
3183 jermar 803
 
804
    while (true) {
805
        slab_cache_t *cache;
806
        link_t *cur;
807
        ipl_t ipl;
808
        int i;
809
 
810
        /*
811
         * We must not hold the slab_cache_lock spinlock when printing
812
         * the statistics. Otherwise we can easily deadlock if the print
813
         * needs to allocate memory.
814
         *
815
         * Therefore, we walk through the slab cache list, skipping some
816
         * amount of already processed caches during each iteration and
817
         * gathering statistics about the first unprocessed cache. For
818
         * the sake of printing the statistics, we realese the
819
         * slab_cache_lock and reacquire it afterwards. Then the walk
820
         * starts again.
821
         *
822
         * This limits both the efficiency and also accuracy of the
823
         * obtained statistics. The efficiency is decreased because the
824
         * time complexity of the algorithm is quadratic instead of
825
         * linear. The accuracy is impacted because we drop the lock
826
         * after processing one cache. If there is someone else
827
         * manipulating the cache list, we might omit an arbitrary
828
         * number of caches or process one cache multiple times.
829
         * However, we don't bleed for this algorithm for it is only
830
         * statistics.
831
         */
832
 
833
        ipl = interrupts_disable();
834
        spinlock_lock(&slab_cache_lock);
835
 
836
        for (i = 0, cur = slab_cache_list.next;
837
            i < skip && cur != &slab_cache_list;
838
            i++, cur = cur->next)
839
            ;
840
 
841
        if (cur == &slab_cache_list) {
842
            spinlock_unlock(&slab_cache_lock);
843
            interrupts_restore(ipl);
844
            break;
845
        }
846
 
847
        skip++;
848
 
759 palkovsky 849
        cache = list_get_instance(cur, slab_cache_t, link);
3183 jermar 850
 
851
        char *name = cache->name;
852
        uint8_t order = cache->order;
853
        size_t size = cache->size;
854
        unsigned int objects = cache->objects;
855
        long allocated_slabs = atomic_get(&cache->allocated_slabs);
856
        long cached_objs = atomic_get(&cache->cached_objs);
857
        long allocated_objs = atomic_get(&cache->allocated_objs);
858
        int flags = cache->flags;
2052 decky 859
 
3183 jermar 860
        spinlock_unlock(&slab_cache_lock);
861
        interrupts_restore(ipl);
862
 
3057 decky 863
        printf("%-16s %8" PRIs " %6d %6u %6ld %6ld %9ld %-3s\n",
3183 jermar 864
            name, size, (1 << order), objects, allocated_slabs,
865
            cached_objs, allocated_objs,
866
            flags & SLAB_CACHE_SLINSIDE ? "in" : "out");
759 palkovsky 867
    }
868
}
869
 
870
void slab_cache_init(void)
871
{
771 palkovsky 872
    int i, size;
873
 
759 palkovsky 874
    /* Initialize magazine cache */
3180 jermar 875
    _slab_cache_create(&mag_cache, "slab_magazine",
876
        sizeof(slab_magazine_t) + SLAB_MAG_SIZE * sizeof(void*),
877
        sizeof(uintptr_t), NULL, NULL, SLAB_CACHE_NOMAGAZINE |
878
        SLAB_CACHE_SLINSIDE);
769 palkovsky 879
    /* Initialize slab_cache cache */
3180 jermar 880
    _slab_cache_create(&slab_cache_cache, "slab_cache",
881
        sizeof(slab_cache_cache), sizeof(uintptr_t), NULL, NULL,
882
        SLAB_CACHE_NOMAGAZINE | SLAB_CACHE_SLINSIDE);
769 palkovsky 883
    /* Initialize external slab cache */
3180 jermar 884
    slab_extern_cache = slab_cache_create("slab_extern", sizeof(slab_t), 0,
885
        NULL, NULL, SLAB_CACHE_SLINSIDE | SLAB_CACHE_MAGDEFERRED);
759 palkovsky 886
 
887
    /* Initialize structures for malloc */
3180 jermar 888
    for (i = 0, size = (1 << SLAB_MIN_MALLOC_W);
889
        i < (SLAB_MAX_MALLOC_W - SLAB_MIN_MALLOC_W + 1);
890
        i++, size <<= 1) {
891
        malloc_caches[i] = slab_cache_create(malloc_names[i], size, 0,
892
            NULL, NULL, SLAB_CACHE_MAGDEFERRED);
771 palkovsky 893
    }
778 palkovsky 894
#ifdef CONFIG_DEBUG       
895
    _slab_initialized = 1;
896
#endif
759 palkovsky 897
}
771 palkovsky 898
 
789 palkovsky 899
/** Enable cpu_cache
900
 *
901
 * Kernel calls this function, when it knows the real number of
902
 * processors.
903
 * Allocate slab for cpucache and enable it on all existing
904
 * slabs that are SLAB_CACHE_MAGDEFERRED
905
 */
906
void slab_enable_cpucache(void)
907
{
908
    link_t *cur;
909
    slab_cache_t *s;
910
 
791 palkovsky 911
#ifdef CONFIG_DEBUG
912
    _slab_initialized = 2;
913
#endif
914
 
789 palkovsky 915
    spinlock_lock(&slab_cache_lock);
916
 
3180 jermar 917
    for (cur = slab_cache_list.next; cur != &slab_cache_list;
918
        cur = cur->next){
789 palkovsky 919
        s = list_get_instance(cur, slab_cache_t, link);
3180 jermar 920
        if ((s->flags & SLAB_CACHE_MAGDEFERRED) !=
921
            SLAB_CACHE_MAGDEFERRED)
789 palkovsky 922
            continue;
923
        make_magcache(s);
924
        s->flags &= ~SLAB_CACHE_MAGDEFERRED;
925
    }
926
 
927
    spinlock_unlock(&slab_cache_lock);
928
}
929
 
771 palkovsky 930
/**************************************/
931
/* kalloc/kfree functions             */
3180 jermar 932
void *malloc(unsigned int size, int flags)
771 palkovsky 933
{
778 palkovsky 934
    ASSERT(_slab_initialized);
1288 jermar 935
    ASSERT(size && size <= (1 << SLAB_MAX_MALLOC_W));
771 palkovsky 936
 
937
    if (size < (1 << SLAB_MIN_MALLOC_W))
938
        size = (1 << SLAB_MIN_MALLOC_W);
939
 
2124 decky 940
    int idx = fnzb(size - 1) - SLAB_MIN_MALLOC_W + 1;
771 palkovsky 941
 
942
    return slab_alloc(malloc_caches[idx], flags);
943
}
944
 
3180 jermar 945
void *realloc(void *ptr, unsigned int size, int flags)
771 palkovsky 946
{
2124 decky 947
    ASSERT(_slab_initialized);
948
    ASSERT(size <= (1 << SLAB_MAX_MALLOC_W));
949
 
950
    void *new_ptr;
951
 
952
    if (size > 0) {
953
        if (size < (1 << SLAB_MIN_MALLOC_W))
954
            size = (1 << SLAB_MIN_MALLOC_W);
955
        int idx = fnzb(size - 1) - SLAB_MIN_MALLOC_W + 1;
956
 
957
        new_ptr = slab_alloc(malloc_caches[idx], flags);
958
    } else
959
        new_ptr = NULL;
960
 
961
    if ((new_ptr != NULL) && (ptr != NULL)) {
962
        slab_t *slab = obj2slab(ptr);
963
        memcpy(new_ptr, ptr, min(size, slab->cache->size));
964
    }
965
 
966
    if (ptr != NULL)
967
        free(ptr);
968
 
969
    return new_ptr;
970
}
781 palkovsky 971
 
2124 decky 972
void free(void *ptr)
973
{
974
    if (!ptr)
1950 jermar 975
        return;
781 palkovsky 976
 
2124 decky 977
    slab_t *slab = obj2slab(ptr);
978
    _slab_free(slab->cache, ptr, slab);
771 palkovsky 979
}
1702 cejka 980
 
1757 jermar 981
/** @}
1702 cejka 982
 */