Subversion Repositories HelenOS-historic

Rev

Rev 1757 | Go to most recent revision | Details | Compare with Previous | Last modification | View Log | RSS feed

Rev Author Line No. Line
759 palkovsky 1
/*
2
 * Copyright (C) 2006 Ondrej 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>
759 palkovsky 114
 
115
SPINLOCK_INITIALIZE(slab_cache_lock);
769 palkovsky 116
static LIST_INITIALIZE(slab_cache_list);
759 palkovsky 117
 
769 palkovsky 118
/** Magazine cache */
119
static slab_cache_t mag_cache;
120
/** Cache for cache descriptors */
121
static slab_cache_t slab_cache_cache;
122
/** Cache for external slab descriptors
123
 * This time we want per-cpu cache, so do not make it static
1248 jermar 124
 * - using slab for internal slab structures will not deadlock,
769 palkovsky 125
 *   as all slab structures are 'small' - control structures of
126
 *   their caches do not require further allocation
127
 */
128
static slab_cache_t *slab_extern_cache;
771 palkovsky 129
/** Caches for malloc */
130
static slab_cache_t *malloc_caches[SLAB_MAX_MALLOC_W-SLAB_MIN_MALLOC_W+1];
131
char *malloc_names[] =  {
791 palkovsky 132
    "malloc-16","malloc-32","malloc-64","malloc-128",
771 palkovsky 133
    "malloc-256","malloc-512","malloc-1K","malloc-2K",
134
    "malloc-4K","malloc-8K","malloc-16K","malloc-32K",
1428 palkovsky 135
    "malloc-64K","malloc-128K","malloc-256K"
771 palkovsky 136
};
762 palkovsky 137
 
769 palkovsky 138
/** Slab descriptor */
762 palkovsky 139
typedef struct {
140
    slab_cache_t *cache; /**< Pointer to parent cache */
141
    link_t link;       /* List of full/partial slabs */
142
    void *start;       /**< Start address of first available item */
143
    count_t available; /**< Count of available items in this slab */
144
    index_t nextavail; /**< The index of next available item */
145
}slab_t;
146
 
791 palkovsky 147
#ifdef CONFIG_DEBUG
148
static int _slab_initialized = 0;
149
#endif
150
 
759 palkovsky 151
/**************************************/
1248 jermar 152
/* Slab allocation functions          */
759 palkovsky 153
 
762 palkovsky 154
/**
155
 * Allocate frames for slab space and initialize
156
 *
157
 */
158
static slab_t * slab_space_alloc(slab_cache_t *cache, int flags)
159
{
160
    void *data;
161
    slab_t *slab;
162
    size_t fsize;
163
    int i;
164
    int status;
814 palkovsky 165
    int zone=0;
166
 
1760 palkovsky 167
    data = frame_alloc_rc_zone(cache->order, FRAME_KA | flags, &status, &zone);
764 palkovsky 168
    if (status != FRAME_OK) {
762 palkovsky 169
        return NULL;
764 palkovsky 170
    }
768 palkovsky 171
    if (! (cache->flags & SLAB_CACHE_SLINSIDE)) {
769 palkovsky 172
        slab = slab_alloc(slab_extern_cache, flags);
762 palkovsky 173
        if (!slab) {
1760 palkovsky 174
            frame_free(KA2PA(data));
762 palkovsky 175
            return NULL;
176
        }
177
    } else {
178
        fsize = (PAGE_SIZE << cache->order);
179
        slab = data + fsize - sizeof(*slab);
180
    }
1288 jermar 181
 
762 palkovsky 182
    /* Fill in slab structures */
814 palkovsky 183
    for (i=0; i < (1 << cache->order); i++)
1760 palkovsky 184
        frame_set_parent(ADDR2PFN(KA2PA(data))+i, slab, zone);
762 palkovsky 185
 
186
    slab->start = data;
187
    slab->available = cache->objects;
188
    slab->nextavail = 0;
767 palkovsky 189
    slab->cache = cache;
762 palkovsky 190
 
191
    for (i=0; i<cache->objects;i++)
192
        *((int *) (slab->start + i*cache->size)) = i+1;
764 palkovsky 193
 
194
    atomic_inc(&cache->allocated_slabs);
762 palkovsky 195
    return slab;
196
}
197
 
759 palkovsky 198
/**
1248 jermar 199
 * Deallocate space associated with slab
762 palkovsky 200
 *
201
 * @return number of freed frames
202
 */
203
static count_t slab_space_free(slab_cache_t *cache, slab_t *slab)
204
{
1760 palkovsky 205
    frame_free(KA2PA(slab->start));
768 palkovsky 206
    if (! (cache->flags & SLAB_CACHE_SLINSIDE))
769 palkovsky 207
        slab_free(slab_extern_cache, slab);
764 palkovsky 208
 
209
    atomic_dec(&cache->allocated_slabs);
210
 
762 palkovsky 211
    return 1 << cache->order;
212
}
213
 
214
/** Map object to slab structure */
215
static slab_t * obj2slab(void *obj)
216
{
814 palkovsky 217
    return (slab_t *)frame_get_parent(ADDR2PFN(KA2PA(obj)), 0);
762 palkovsky 218
}
219
 
220
/**************************************/
1248 jermar 221
/* Slab functions */
762 palkovsky 222
 
223
 
224
/**
759 palkovsky 225
 * Return object to slab and call a destructor
226
 *
762 palkovsky 227
 * @param slab If the caller knows directly slab of the object, otherwise NULL
228
 *
759 palkovsky 229
 * @return Number of freed pages
230
 */
762 palkovsky 231
static count_t slab_obj_destroy(slab_cache_t *cache, void *obj,
232
                slab_t *slab)
759 palkovsky 233
{
787 palkovsky 234
    int freed = 0;
235
 
762 palkovsky 236
    if (!slab)
237
        slab = obj2slab(obj);
238
 
767 palkovsky 239
    ASSERT(slab->cache == cache);
240
 
787 palkovsky 241
    if (cache->destructor)
242
        freed = cache->destructor(obj);
243
 
776 palkovsky 244
    spinlock_lock(&cache->slablock);
789 palkovsky 245
    ASSERT(slab->available < cache->objects);
776 palkovsky 246
 
762 palkovsky 247
    *((int *)obj) = slab->nextavail;
248
    slab->nextavail = (obj - slab->start)/cache->size;
249
    slab->available++;
250
 
251
    /* Move it to correct list */
252
    if (slab->available == cache->objects) {
253
        /* Free associated memory */
254
        list_remove(&slab->link);
782 palkovsky 255
        spinlock_unlock(&cache->slablock);
256
 
787 palkovsky 257
        return freed + slab_space_free(cache, slab);
782 palkovsky 258
 
780 palkovsky 259
    } else if (slab->available == 1) {
260
        /* It was in full, move to partial */
261
        list_remove(&slab->link);
262
        list_prepend(&slab->link, &cache->partial_slabs);
762 palkovsky 263
    }
783 palkovsky 264
    spinlock_unlock(&cache->slablock);
787 palkovsky 265
    return freed;
759 palkovsky 266
}
267
 
268
/**
269
 * Take new object from slab or create new if needed
270
 *
271
 * @return Object address or null
272
 */
273
static void * slab_obj_create(slab_cache_t *cache, int flags)
274
{
762 palkovsky 275
    slab_t *slab;
276
    void *obj;
277
 
776 palkovsky 278
    spinlock_lock(&cache->slablock);
279
 
762 palkovsky 280
    if (list_empty(&cache->partial_slabs)) {
281
        /* Allow recursion and reclaiming
1248 jermar 282
         * - this should work, as the slab control structures
1288 jermar 283
         *   are small and do not need to allocate with anything
284
         *   other than frame_alloc when they are allocating,
762 palkovsky 285
         *   that's why we should get recursion at most 1-level deep
286
         */
776 palkovsky 287
        spinlock_unlock(&cache->slablock);
762 palkovsky 288
        slab = slab_space_alloc(cache, flags);
780 palkovsky 289
        if (!slab)
290
            return NULL;
776 palkovsky 291
        spinlock_lock(&cache->slablock);
762 palkovsky 292
    } else {
293
        slab = list_get_instance(cache->partial_slabs.next,
294
                     slab_t,
295
                     link);
296
        list_remove(&slab->link);
297
    }
298
    obj = slab->start + slab->nextavail * cache->size;
299
    slab->nextavail = *((int *)obj);
300
    slab->available--;
787 palkovsky 301
 
762 palkovsky 302
    if (! slab->available)
764 palkovsky 303
        list_prepend(&slab->link, &cache->full_slabs);
762 palkovsky 304
    else
764 palkovsky 305
        list_prepend(&slab->link, &cache->partial_slabs);
776 palkovsky 306
 
307
    spinlock_unlock(&cache->slablock);
787 palkovsky 308
 
309
    if (cache->constructor && cache->constructor(obj, flags)) {
310
        /* Bad, bad, construction failed */
311
        slab_obj_destroy(cache, obj, slab);
312
        return NULL;
313
    }
762 palkovsky 314
    return obj;
759 palkovsky 315
}
316
 
317
/**************************************/
318
/* CPU-Cache slab functions */
319
 
320
/**
781 palkovsky 321
 * Finds a full magazine in cache, takes it from list
322
 * and returns it
323
 *
324
 * @param first If true, return first, else last mag
325
 */
326
static slab_magazine_t * get_mag_from_cache(slab_cache_t *cache,
327
                        int first)
328
{
329
    slab_magazine_t *mag = NULL;
330
    link_t *cur;
331
 
332
    spinlock_lock(&cache->maglock);
333
    if (!list_empty(&cache->magazines)) {
334
        if (first)
335
            cur = cache->magazines.next;
336
        else
337
            cur = cache->magazines.prev;
338
        mag = list_get_instance(cur, slab_magazine_t, link);
339
        list_remove(&mag->link);
340
        atomic_dec(&cache->magazine_counter);
341
    }
342
    spinlock_unlock(&cache->maglock);
343
    return mag;
344
}
345
 
346
/** Prepend magazine to magazine list in cache */
347
static void put_mag_to_cache(slab_cache_t *cache, slab_magazine_t *mag)
348
{
349
    spinlock_lock(&cache->maglock);
350
 
351
    list_prepend(&mag->link, &cache->magazines);
352
    atomic_inc(&cache->magazine_counter);
353
 
354
    spinlock_unlock(&cache->maglock);
355
}
356
 
357
/**
759 palkovsky 358
 * Free all objects in magazine and free memory associated with magazine
359
 *
360
 * @return Number of freed pages
361
 */
362
static count_t magazine_destroy(slab_cache_t *cache,
363
                slab_magazine_t *mag)
364
{
365
    int i;
366
    count_t frames = 0;
367
 
767 palkovsky 368
    for (i=0;i < mag->busy; i++) {
762 palkovsky 369
        frames += slab_obj_destroy(cache, mag->objs[i], NULL);
767 palkovsky 370
        atomic_dec(&cache->cached_objs);
371
    }
759 palkovsky 372
 
373
    slab_free(&mag_cache, mag);
374
 
375
    return frames;
376
}
377
 
378
/**
769 palkovsky 379
 * Find full magazine, set it as current and return it
380
 *
381
 * Assume cpu_magazine lock is held
382
 */
383
static slab_magazine_t * get_full_current_mag(slab_cache_t *cache)
384
{
385
    slab_magazine_t *cmag, *lastmag, *newmag;
386
 
387
    cmag = cache->mag_cache[CPU->id].current;
388
    lastmag = cache->mag_cache[CPU->id].last;
389
    if (cmag) { /* First try local CPU magazines */
390
        if (cmag->busy)
391
            return cmag;
392
 
393
        if (lastmag && lastmag->busy) {
394
            cache->mag_cache[CPU->id].current = lastmag;
395
            cache->mag_cache[CPU->id].last = cmag;
396
            return lastmag;
397
        }
398
    }
399
    /* Local magazines are empty, import one from magazine list */
781 palkovsky 400
    newmag = get_mag_from_cache(cache, 1);
401
    if (!newmag)
769 palkovsky 402
        return NULL;
403
 
404
    if (lastmag)
781 palkovsky 405
        magazine_destroy(cache, lastmag);
406
 
769 palkovsky 407
    cache->mag_cache[CPU->id].last = cmag;
408
    cache->mag_cache[CPU->id].current = newmag;
409
    return newmag;
410
}
411
 
412
/**
759 palkovsky 413
 * Try to find object in CPU-cache magazines
414
 *
415
 * @return Pointer to object or NULL if not available
416
 */
417
static void * magazine_obj_get(slab_cache_t *cache)
418
{
419
    slab_magazine_t *mag;
767 palkovsky 420
    void *obj;
759 palkovsky 421
 
772 palkovsky 422
    if (!CPU)
423
        return NULL;
424
 
759 palkovsky 425
    spinlock_lock(&cache->mag_cache[CPU->id].lock);
426
 
769 palkovsky 427
    mag = get_full_current_mag(cache);
428
    if (!mag) {
429
        spinlock_unlock(&cache->mag_cache[CPU->id].lock);
430
        return NULL;
759 palkovsky 431
    }
767 palkovsky 432
    obj = mag->objs[--mag->busy];
759 palkovsky 433
    spinlock_unlock(&cache->mag_cache[CPU->id].lock);
767 palkovsky 434
    atomic_dec(&cache->cached_objs);
435
 
436
    return obj;
759 palkovsky 437
}
438
 
439
/**
768 palkovsky 440
 * Assure that the current magazine is empty, return pointer to it, or NULL if
769 palkovsky 441
 * no empty magazine is available and cannot be allocated
759 palkovsky 442
 *
773 palkovsky 443
 * Assume mag_cache[CPU->id].lock is held
444
 *
759 palkovsky 445
 * We have 2 magazines bound to processor.
446
 * First try the current.
447
 *  If full, try the last.
448
 *   If full, put to magazines list.
449
 *   allocate new, exchange last & current
450
 *
768 palkovsky 451
 */
452
static slab_magazine_t * make_empty_current_mag(slab_cache_t *cache)
453
{
454
    slab_magazine_t *cmag,*lastmag,*newmag;
455
 
456
    cmag = cache->mag_cache[CPU->id].current;
457
    lastmag = cache->mag_cache[CPU->id].last;
458
 
459
    if (cmag) {
460
        if (cmag->busy < cmag->size)
461
            return cmag;
462
        if (lastmag && lastmag->busy < lastmag->size) {
463
            cache->mag_cache[CPU->id].last = cmag;
464
            cache->mag_cache[CPU->id].current = lastmag;
465
            return lastmag;
466
        }
467
    }
468
    /* current | last are full | nonexistent, allocate new */
469
    /* We do not want to sleep just because of caching */
470
    /* Especially we do not want reclaiming to start, as
471
     * this would deadlock */
472
    newmag = slab_alloc(&mag_cache, FRAME_ATOMIC | FRAME_NO_RECLAIM);
473
    if (!newmag)
474
        return NULL;
475
    newmag->size = SLAB_MAG_SIZE;
476
    newmag->busy = 0;
477
 
478
    /* Flush last to magazine list */
781 palkovsky 479
    if (lastmag)
480
        put_mag_to_cache(cache, lastmag);
481
 
768 palkovsky 482
    /* Move current as last, save new as current */
483
    cache->mag_cache[CPU->id].last = cmag; 
484
    cache->mag_cache[CPU->id].current = newmag;
485
 
486
    return newmag;
487
}
488
 
489
/**
490
 * Put object into CPU-cache magazine
491
 *
759 palkovsky 492
 * @return 0 - success, -1 - could not get memory
493
 */
494
static int magazine_obj_put(slab_cache_t *cache, void *obj)
495
{
496
    slab_magazine_t *mag;
497
 
772 palkovsky 498
    if (!CPU)
499
        return -1;
500
 
759 palkovsky 501
    spinlock_lock(&cache->mag_cache[CPU->id].lock);
768 palkovsky 502
 
503
    mag = make_empty_current_mag(cache);
769 palkovsky 504
    if (!mag) {
505
        spinlock_unlock(&cache->mag_cache[CPU->id].lock);
506
        return -1;
507
    }
759 palkovsky 508
 
509
    mag->objs[mag->busy++] = obj;
510
 
511
    spinlock_unlock(&cache->mag_cache[CPU->id].lock);
767 palkovsky 512
    atomic_inc(&cache->cached_objs);
759 palkovsky 513
    return 0;
514
}
515
 
516
 
517
/**************************************/
1248 jermar 518
/* Slab cache functions */
759 palkovsky 519
 
762 palkovsky 520
/** Return number of objects that fit in certain cache size */
521
static int comp_objects(slab_cache_t *cache)
522
{
523
    if (cache->flags & SLAB_CACHE_SLINSIDE)
524
        return ((PAGE_SIZE << cache->order) - sizeof(slab_t)) / cache->size;
525
    else
526
        return (PAGE_SIZE << cache->order) / cache->size;
527
}
528
 
529
/** Return wasted space in slab */
530
static int badness(slab_cache_t *cache)
531
{
532
    int objects;
533
    int ssize;
534
 
535
    objects = comp_objects(cache);
536
    ssize = PAGE_SIZE << cache->order;
537
    if (cache->flags & SLAB_CACHE_SLINSIDE)
538
        ssize -= sizeof(slab_t);
539
    return ssize - objects*cache->size;
540
}
541
 
789 palkovsky 542
/**
543
 * Initialize mag_cache structure in slab cache
544
 */
545
static void make_magcache(slab_cache_t *cache)
546
{
547
    int i;
791 palkovsky 548
 
549
    ASSERT(_slab_initialized >= 2);
789 palkovsky 550
 
822 palkovsky 551
    cache->mag_cache = malloc(sizeof(slab_mag_cache_t)*config.cpu_count,0);
789 palkovsky 552
    for (i=0; i < config.cpu_count; i++) {
553
        memsetb((__address)&cache->mag_cache[i],
554
            sizeof(cache->mag_cache[i]), 0);
555
        spinlock_initialize(&cache->mag_cache[i].lock,
556
                    "slab_maglock_cpu");
557
    }
558
}
559
 
759 palkovsky 560
/** Initialize allocated memory as a slab cache */
561
static void
562
_slab_cache_create(slab_cache_t *cache,
563
           char *name,
564
           size_t size,
565
           size_t align,
566
           int (*constructor)(void *obj, int kmflag),
787 palkovsky 567
           int (*destructor)(void *obj),
759 palkovsky 568
           int flags)
569
{
771 palkovsky 570
    int pages;
783 palkovsky 571
    ipl_t ipl;
759 palkovsky 572
 
573
    memsetb((__address)cache, sizeof(*cache), 0);
574
    cache->name = name;
575
 
766 palkovsky 576
    if (align < sizeof(__native))
577
        align = sizeof(__native);
578
    size = ALIGN_UP(size, align);
579
 
762 palkovsky 580
    cache->size = size;
759 palkovsky 581
 
582
    cache->constructor = constructor;
583
    cache->destructor = destructor;
584
    cache->flags = flags;
585
 
586
    list_initialize(&cache->full_slabs);
587
    list_initialize(&cache->partial_slabs);
588
    list_initialize(&cache->magazines);
776 palkovsky 589
    spinlock_initialize(&cache->slablock, "slab_lock");
590
    spinlock_initialize(&cache->maglock, "slab_maglock");
789 palkovsky 591
    if (! (cache->flags & SLAB_CACHE_NOMAGAZINE))
592
        make_magcache(cache);
759 palkovsky 593
 
594
    /* Compute slab sizes, object counts in slabs etc. */
595
    if (cache->size < SLAB_INSIDE_SIZE)
596
        cache->flags |= SLAB_CACHE_SLINSIDE;
597
 
762 palkovsky 598
    /* Minimum slab order */
1682 palkovsky 599
    pages = SIZE2FRAMES(cache->size);
1677 palkovsky 600
    /* We need the 2^order >= pages */
601
    if (pages == 1)
602
        cache->order = 0;
603
    else
604
        cache->order = fnzb(pages-1)+1;
766 palkovsky 605
 
762 palkovsky 606
    while (badness(cache) > SLAB_MAX_BADNESS(cache)) {
607
        cache->order += 1;
608
    }
609
    cache->objects = comp_objects(cache);
766 palkovsky 610
    /* If info fits in, put it inside */
611
    if (badness(cache) > sizeof(slab_t))
612
        cache->flags |= SLAB_CACHE_SLINSIDE;
762 palkovsky 613
 
783 palkovsky 614
    /* Add cache to cache list */
615
    ipl = interrupts_disable();
759 palkovsky 616
    spinlock_lock(&slab_cache_lock);
617
 
618
    list_append(&cache->link, &slab_cache_list);
619
 
620
    spinlock_unlock(&slab_cache_lock);
783 palkovsky 621
    interrupts_restore(ipl);
759 palkovsky 622
}
623
 
624
/** Create slab cache  */
625
slab_cache_t * slab_cache_create(char *name,
626
                 size_t size,
627
                 size_t align,
628
                 int (*constructor)(void *obj, int kmflag),
787 palkovsky 629
                 int (*destructor)(void *obj),
759 palkovsky 630
                 int flags)
631
{
632
    slab_cache_t *cache;
633
 
769 palkovsky 634
    cache = slab_alloc(&slab_cache_cache, 0);
759 palkovsky 635
    _slab_cache_create(cache, name, size, align, constructor, destructor,
636
               flags);
637
    return cache;
638
}
639
 
640
/**
641
 * Reclaim space occupied by objects that are already free
642
 *
643
 * @param flags If contains SLAB_RECLAIM_ALL, do aggressive freeing
644
 * @return Number of freed pages
645
 */
646
static count_t _slab_reclaim(slab_cache_t *cache, int flags)
647
{
648
    int i;
649
    slab_magazine_t *mag;
650
    count_t frames = 0;
781 palkovsky 651
    int magcount;
759 palkovsky 652
 
653
    if (cache->flags & SLAB_CACHE_NOMAGAZINE)
654
        return 0; /* Nothing to do */
781 palkovsky 655
 
656
    /* We count up to original magazine count to avoid
657
     * endless loop
658
     */
659
    magcount = atomic_get(&cache->magazine_counter);
660
    while (magcount-- && (mag=get_mag_from_cache(cache,0))) {
661
        frames += magazine_destroy(cache,mag);
662
        if (!(flags & SLAB_RECLAIM_ALL) && frames)
663
            break;
769 palkovsky 664
    }
759 palkovsky 665
 
666
    if (flags & SLAB_RECLAIM_ALL) {
781 palkovsky 667
        /* Free cpu-bound magazines */
759 palkovsky 668
        /* Destroy CPU magazines */
669
        for (i=0; i<config.cpu_count; i++) {
781 palkovsky 670
            spinlock_lock(&cache->mag_cache[i].lock);
671
 
759 palkovsky 672
            mag = cache->mag_cache[i].current;
673
            if (mag)
674
                frames += magazine_destroy(cache, mag);
675
            cache->mag_cache[i].current = NULL;
676
 
677
            mag = cache->mag_cache[i].last;
678
            if (mag)
679
                frames += magazine_destroy(cache, mag);
680
            cache->mag_cache[i].last = NULL;
781 palkovsky 681
 
682
            spinlock_unlock(&cache->mag_cache[i].lock);
759 palkovsky 683
        }
684
    }
767 palkovsky 685
 
759 palkovsky 686
    return frames;
687
}
688
 
689
/** Check that there are no slabs and remove cache from system  */
690
void slab_cache_destroy(slab_cache_t *cache)
691
{
781 palkovsky 692
    ipl_t ipl;
693
 
694
    /* First remove cache from link, so that we don't need
695
     * to disable interrupts later
696
     */
697
 
698
    ipl = interrupts_disable();
699
    spinlock_lock(&slab_cache_lock);
700
 
701
    list_remove(&cache->link);
702
 
703
    spinlock_unlock(&slab_cache_lock);
704
    interrupts_restore(ipl);
705
 
759 palkovsky 706
    /* Do not lock anything, we assume the software is correct and
707
     * does not touch the cache when it decides to destroy it */
708
 
709
    /* Destroy all magazines */
710
    _slab_reclaim(cache, SLAB_RECLAIM_ALL);
711
 
712
    /* All slabs must be empty */
713
    if (!list_empty(&cache->full_slabs) \
714
        || !list_empty(&cache->partial_slabs))
715
        panic("Destroying cache that is not empty.");
716
 
789 palkovsky 717
    if (!(cache->flags & SLAB_CACHE_NOMAGAZINE))
822 palkovsky 718
        free(cache->mag_cache);
769 palkovsky 719
    slab_free(&slab_cache_cache, cache);
759 palkovsky 720
}
721
 
722
/** Allocate new object from cache - if no flags given, always returns
723
    memory */
724
void * slab_alloc(slab_cache_t *cache, int flags)
725
{
726
    ipl_t ipl;
727
    void *result = NULL;
773 palkovsky 728
 
759 palkovsky 729
    /* Disable interrupts to avoid deadlocks with interrupt handlers */
730
    ipl = interrupts_disable();
771 palkovsky 731
 
814 palkovsky 732
    if (!(cache->flags & SLAB_CACHE_NOMAGAZINE)) {
759 palkovsky 733
        result = magazine_obj_get(cache);
814 palkovsky 734
    }
776 palkovsky 735
    if (!result)
759 palkovsky 736
        result = slab_obj_create(cache, flags);
737
 
769 palkovsky 738
    interrupts_restore(ipl);
739
 
764 palkovsky 740
    if (result)
741
        atomic_inc(&cache->allocated_objs);
742
 
759 palkovsky 743
    return result;
744
}
745
 
771 palkovsky 746
/** Return object to cache, use slab if known  */
747
static void _slab_free(slab_cache_t *cache, void *obj, slab_t *slab)
759 palkovsky 748
{
749
    ipl_t ipl;
750
 
751
    ipl = interrupts_disable();
752
 
762 palkovsky 753
    if ((cache->flags & SLAB_CACHE_NOMAGAZINE) \
754
        || magazine_obj_put(cache, obj)) {
776 palkovsky 755
 
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
{
766
    _slab_free(cache,obj,NULL);
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
 
759 palkovsky 782
    for (cur = slab_cache_list.next;cur!=&slab_cache_list; cur=cur->next) {
783
        cache = list_get_instance(cur, slab_cache_t, link);
784
        frames += _slab_reclaim(cache, flags);
785
    }
786
 
787
    spinlock_unlock(&slab_cache_lock);
788
 
789
    return frames;
790
}
791
 
792
 
793
/* Print list of slabs */
794
void slab_print_list(void)
795
{
796
    slab_cache_t *cache;
797
    link_t *cur;
783 palkovsky 798
    ipl_t ipl;
799
 
800
    ipl = interrupts_disable();
759 palkovsky 801
    spinlock_lock(&slab_cache_lock);
1248 jermar 802
    printf("slab name\t  Osize\t  Pages\t Obj/pg\t  Slabs\t Cached\tAllocobjs\tCtl\n");
759 palkovsky 803
    for (cur = slab_cache_list.next;cur!=&slab_cache_list; cur=cur->next) {
804
        cache = list_get_instance(cur, slab_cache_t, link);
1224 cejka 805
        printf("%s\t%7zd\t%7zd\t%7zd\t%7zd\t%7zd\t%7zd\t\t%s\n", cache->name, cache->size,
766 palkovsky 806
               (1 << cache->order), cache->objects,
767 palkovsky 807
               atomic_get(&cache->allocated_slabs),
808
               atomic_get(&cache->cached_objs),
766 palkovsky 809
               atomic_get(&cache->allocated_objs),
810
               cache->flags & SLAB_CACHE_SLINSIDE ? "In" : "Out");
759 palkovsky 811
    }
812
    spinlock_unlock(&slab_cache_lock);
783 palkovsky 813
    interrupts_restore(ipl);
759 palkovsky 814
}
815
 
816
void slab_cache_init(void)
817
{
771 palkovsky 818
    int i, size;
819
 
759 palkovsky 820
    /* Initialize magazine cache */
821
    _slab_cache_create(&mag_cache,
822
               "slab_magazine",
823
               sizeof(slab_magazine_t)+SLAB_MAG_SIZE*sizeof(void*),
824
               sizeof(__address),
825
               NULL, NULL,
769 palkovsky 826
               SLAB_CACHE_NOMAGAZINE | SLAB_CACHE_SLINSIDE);
827
    /* Initialize slab_cache cache */
828
    _slab_cache_create(&slab_cache_cache,
829
               "slab_cache",
789 palkovsky 830
               sizeof(slab_cache_cache),
769 palkovsky 831
               sizeof(__address),
832
               NULL, NULL,
833
               SLAB_CACHE_NOMAGAZINE | SLAB_CACHE_SLINSIDE);
834
    /* Initialize external slab cache */
835
    slab_extern_cache = slab_cache_create("slab_extern",
836
                          sizeof(slab_t),
837
                          0, NULL, NULL,
789 palkovsky 838
                          SLAB_CACHE_SLINSIDE | SLAB_CACHE_MAGDEFERRED);
759 palkovsky 839
 
840
    /* Initialize structures for malloc */
771 palkovsky 841
    for (i=0, size=(1<<SLAB_MIN_MALLOC_W);
842
         i < (SLAB_MAX_MALLOC_W-SLAB_MIN_MALLOC_W+1);
843
         i++, size <<= 1) {
844
        malloc_caches[i] = slab_cache_create(malloc_names[i],
845
                             size, 0,
789 palkovsky 846
                             NULL,NULL, SLAB_CACHE_MAGDEFERRED);
771 palkovsky 847
    }
778 palkovsky 848
#ifdef CONFIG_DEBUG       
849
    _slab_initialized = 1;
850
#endif
759 palkovsky 851
}
771 palkovsky 852
 
789 palkovsky 853
/** Enable cpu_cache
854
 *
855
 * Kernel calls this function, when it knows the real number of
856
 * processors.
857
 * Allocate slab for cpucache and enable it on all existing
858
 * slabs that are SLAB_CACHE_MAGDEFERRED
859
 */
860
void slab_enable_cpucache(void)
861
{
862
    link_t *cur;
863
    slab_cache_t *s;
864
 
791 palkovsky 865
#ifdef CONFIG_DEBUG
866
    _slab_initialized = 2;
867
#endif
868
 
789 palkovsky 869
    spinlock_lock(&slab_cache_lock);
870
 
871
    for (cur=slab_cache_list.next; cur != &slab_cache_list;cur=cur->next){
872
        s = list_get_instance(cur, slab_cache_t, link);
873
        if ((s->flags & SLAB_CACHE_MAGDEFERRED) != SLAB_CACHE_MAGDEFERRED)
874
            continue;
875
        make_magcache(s);
876
        s->flags &= ~SLAB_CACHE_MAGDEFERRED;
877
    }
878
 
879
    spinlock_unlock(&slab_cache_lock);
880
}
881
 
771 palkovsky 882
/**************************************/
883
/* kalloc/kfree functions             */
822 palkovsky 884
void * malloc(unsigned int size, int flags)
771 palkovsky 885
{
886
    int idx;
778 palkovsky 887
 
888
    ASSERT(_slab_initialized);
1288 jermar 889
    ASSERT(size && size <= (1 << SLAB_MAX_MALLOC_W));
771 palkovsky 890
 
891
    if (size < (1 << SLAB_MIN_MALLOC_W))
892
        size = (1 << SLAB_MIN_MALLOC_W);
893
 
894
    idx = fnzb(size-1) - SLAB_MIN_MALLOC_W + 1;
895
 
896
    return slab_alloc(malloc_caches[idx], flags);
897
}
898
 
822 palkovsky 899
void free(void *obj)
771 palkovsky 900
{
781 palkovsky 901
    slab_t *slab;
902
 
903
    if (!obj) return;
904
 
905
    slab = obj2slab(obj);
771 palkovsky 906
    _slab_free(slab->cache, obj, slab);
907
}
1702 cejka 908
 
1757 jermar 909
/** @}
1702 cejka 910
 */