Subversion Repositories HelenOS

Rev

Rev 1760 | Rev 1780 | 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;
814 palkovsky 164
    int zone=0;
165
 
1766 palkovsky 166
    data = frame_alloc_generic(cache->order, FRAME_KA | flags, &zone);
167
    if (!data) {
762 palkovsky 168
        return NULL;
764 palkovsky 169
    }
768 palkovsky 170
    if (! (cache->flags & SLAB_CACHE_SLINSIDE)) {
769 palkovsky 171
        slab = slab_alloc(slab_extern_cache, flags);
762 palkovsky 172
        if (!slab) {
1760 palkovsky 173
            frame_free(KA2PA(data));
762 palkovsky 174
            return NULL;
175
        }
176
    } else {
177
        fsize = (PAGE_SIZE << cache->order);
178
        slab = data + fsize - sizeof(*slab);
179
    }
1288 jermar 180
 
762 palkovsky 181
    /* Fill in slab structures */
814 palkovsky 182
    for (i=0; i < (1 << cache->order); i++)
1760 palkovsky 183
        frame_set_parent(ADDR2PFN(KA2PA(data))+i, slab, zone);
762 palkovsky 184
 
185
    slab->start = data;
186
    slab->available = cache->objects;
187
    slab->nextavail = 0;
767 palkovsky 188
    slab->cache = cache;
762 palkovsky 189
 
190
    for (i=0; i<cache->objects;i++)
191
        *((int *) (slab->start + i*cache->size)) = i+1;
764 palkovsky 192
 
193
    atomic_inc(&cache->allocated_slabs);
762 palkovsky 194
    return slab;
195
}
196
 
759 palkovsky 197
/**
1248 jermar 198
 * Deallocate space associated with slab
762 palkovsky 199
 *
200
 * @return number of freed frames
201
 */
202
static count_t slab_space_free(slab_cache_t *cache, slab_t *slab)
203
{
1760 palkovsky 204
    frame_free(KA2PA(slab->start));
768 palkovsky 205
    if (! (cache->flags & SLAB_CACHE_SLINSIDE))
769 palkovsky 206
        slab_free(slab_extern_cache, slab);
764 palkovsky 207
 
208
    atomic_dec(&cache->allocated_slabs);
209
 
762 palkovsky 210
    return 1 << cache->order;
211
}
212
 
213
/** Map object to slab structure */
214
static slab_t * obj2slab(void *obj)
215
{
814 palkovsky 216
    return (slab_t *)frame_get_parent(ADDR2PFN(KA2PA(obj)), 0);
762 palkovsky 217
}
218
 
219
/**************************************/
1248 jermar 220
/* Slab functions */
762 palkovsky 221
 
222
 
223
/**
759 palkovsky 224
 * Return object to slab and call a destructor
225
 *
762 palkovsky 226
 * @param slab If the caller knows directly slab of the object, otherwise NULL
227
 *
759 palkovsky 228
 * @return Number of freed pages
229
 */
762 palkovsky 230
static count_t slab_obj_destroy(slab_cache_t *cache, void *obj,
231
                slab_t *slab)
759 palkovsky 232
{
787 palkovsky 233
    int freed = 0;
234
 
762 palkovsky 235
    if (!slab)
236
        slab = obj2slab(obj);
237
 
767 palkovsky 238
    ASSERT(slab->cache == cache);
239
 
787 palkovsky 240
    if (cache->destructor)
241
        freed = cache->destructor(obj);
242
 
776 palkovsky 243
    spinlock_lock(&cache->slablock);
789 palkovsky 244
    ASSERT(slab->available < cache->objects);
776 palkovsky 245
 
762 palkovsky 246
    *((int *)obj) = slab->nextavail;
247
    slab->nextavail = (obj - slab->start)/cache->size;
248
    slab->available++;
249
 
250
    /* Move it to correct list */
251
    if (slab->available == cache->objects) {
252
        /* Free associated memory */
253
        list_remove(&slab->link);
782 palkovsky 254
        spinlock_unlock(&cache->slablock);
255
 
787 palkovsky 256
        return freed + slab_space_free(cache, slab);
782 palkovsky 257
 
780 palkovsky 258
    } else if (slab->available == 1) {
259
        /* It was in full, move to partial */
260
        list_remove(&slab->link);
261
        list_prepend(&slab->link, &cache->partial_slabs);
762 palkovsky 262
    }
783 palkovsky 263
    spinlock_unlock(&cache->slablock);
787 palkovsky 264
    return freed;
759 palkovsky 265
}
266
 
267
/**
268
 * Take new object from slab or create new if needed
269
 *
270
 * @return Object address or null
271
 */
272
static void * slab_obj_create(slab_cache_t *cache, int flags)
273
{
762 palkovsky 274
    slab_t *slab;
275
    void *obj;
276
 
776 palkovsky 277
    spinlock_lock(&cache->slablock);
278
 
762 palkovsky 279
    if (list_empty(&cache->partial_slabs)) {
280
        /* Allow recursion and reclaiming
1248 jermar 281
         * - this should work, as the slab control structures
1288 jermar 282
         *   are small and do not need to allocate with anything
283
         *   other than frame_alloc when they are allocating,
762 palkovsky 284
         *   that's why we should get recursion at most 1-level deep
285
         */
776 palkovsky 286
        spinlock_unlock(&cache->slablock);
762 palkovsky 287
        slab = slab_space_alloc(cache, flags);
780 palkovsky 288
        if (!slab)
289
            return NULL;
776 palkovsky 290
        spinlock_lock(&cache->slablock);
762 palkovsky 291
    } else {
292
        slab = list_get_instance(cache->partial_slabs.next,
293
                     slab_t,
294
                     link);
295
        list_remove(&slab->link);
296
    }
297
    obj = slab->start + slab->nextavail * cache->size;
298
    slab->nextavail = *((int *)obj);
299
    slab->available--;
787 palkovsky 300
 
762 palkovsky 301
    if (! slab->available)
764 palkovsky 302
        list_prepend(&slab->link, &cache->full_slabs);
762 palkovsky 303
    else
764 palkovsky 304
        list_prepend(&slab->link, &cache->partial_slabs);
776 palkovsky 305
 
306
    spinlock_unlock(&cache->slablock);
787 palkovsky 307
 
308
    if (cache->constructor && cache->constructor(obj, flags)) {
309
        /* Bad, bad, construction failed */
310
        slab_obj_destroy(cache, obj, slab);
311
        return NULL;
312
    }
762 palkovsky 313
    return obj;
759 palkovsky 314
}
315
 
316
/**************************************/
317
/* CPU-Cache slab functions */
318
 
319
/**
781 palkovsky 320
 * Finds a full magazine in cache, takes it from list
321
 * and returns it
322
 *
323
 * @param first If true, return first, else last mag
324
 */
325
static slab_magazine_t * get_mag_from_cache(slab_cache_t *cache,
326
                        int first)
327
{
328
    slab_magazine_t *mag = NULL;
329
    link_t *cur;
330
 
331
    spinlock_lock(&cache->maglock);
332
    if (!list_empty(&cache->magazines)) {
333
        if (first)
334
            cur = cache->magazines.next;
335
        else
336
            cur = cache->magazines.prev;
337
        mag = list_get_instance(cur, slab_magazine_t, link);
338
        list_remove(&mag->link);
339
        atomic_dec(&cache->magazine_counter);
340
    }
341
    spinlock_unlock(&cache->maglock);
342
    return mag;
343
}
344
 
345
/** Prepend magazine to magazine list in cache */
346
static void put_mag_to_cache(slab_cache_t *cache, slab_magazine_t *mag)
347
{
348
    spinlock_lock(&cache->maglock);
349
 
350
    list_prepend(&mag->link, &cache->magazines);
351
    atomic_inc(&cache->magazine_counter);
352
 
353
    spinlock_unlock(&cache->maglock);
354
}
355
 
356
/**
759 palkovsky 357
 * Free all objects in magazine and free memory associated with magazine
358
 *
359
 * @return Number of freed pages
360
 */
361
static count_t magazine_destroy(slab_cache_t *cache,
362
                slab_magazine_t *mag)
363
{
364
    int i;
365
    count_t frames = 0;
366
 
767 palkovsky 367
    for (i=0;i < mag->busy; i++) {
762 palkovsky 368
        frames += slab_obj_destroy(cache, mag->objs[i], NULL);
767 palkovsky 369
        atomic_dec(&cache->cached_objs);
370
    }
759 palkovsky 371
 
372
    slab_free(&mag_cache, mag);
373
 
374
    return frames;
375
}
376
 
377
/**
769 palkovsky 378
 * Find full magazine, set it as current and return it
379
 *
380
 * Assume cpu_magazine lock is held
381
 */
382
static slab_magazine_t * get_full_current_mag(slab_cache_t *cache)
383
{
384
    slab_magazine_t *cmag, *lastmag, *newmag;
385
 
386
    cmag = cache->mag_cache[CPU->id].current;
387
    lastmag = cache->mag_cache[CPU->id].last;
388
    if (cmag) { /* First try local CPU magazines */
389
        if (cmag->busy)
390
            return cmag;
391
 
392
        if (lastmag && lastmag->busy) {
393
            cache->mag_cache[CPU->id].current = lastmag;
394
            cache->mag_cache[CPU->id].last = cmag;
395
            return lastmag;
396
        }
397
    }
398
    /* Local magazines are empty, import one from magazine list */
781 palkovsky 399
    newmag = get_mag_from_cache(cache, 1);
400
    if (!newmag)
769 palkovsky 401
        return NULL;
402
 
403
    if (lastmag)
781 palkovsky 404
        magazine_destroy(cache, lastmag);
405
 
769 palkovsky 406
    cache->mag_cache[CPU->id].last = cmag;
407
    cache->mag_cache[CPU->id].current = newmag;
408
    return newmag;
409
}
410
 
411
/**
759 palkovsky 412
 * Try to find object in CPU-cache magazines
413
 *
414
 * @return Pointer to object or NULL if not available
415
 */
416
static void * magazine_obj_get(slab_cache_t *cache)
417
{
418
    slab_magazine_t *mag;
767 palkovsky 419
    void *obj;
759 palkovsky 420
 
772 palkovsky 421
    if (!CPU)
422
        return NULL;
423
 
759 palkovsky 424
    spinlock_lock(&cache->mag_cache[CPU->id].lock);
425
 
769 palkovsky 426
    mag = get_full_current_mag(cache);
427
    if (!mag) {
428
        spinlock_unlock(&cache->mag_cache[CPU->id].lock);
429
        return NULL;
759 palkovsky 430
    }
767 palkovsky 431
    obj = mag->objs[--mag->busy];
759 palkovsky 432
    spinlock_unlock(&cache->mag_cache[CPU->id].lock);
767 palkovsky 433
    atomic_dec(&cache->cached_objs);
434
 
435
    return obj;
759 palkovsky 436
}
437
 
438
/**
768 palkovsky 439
 * Assure that the current magazine is empty, return pointer to it, or NULL if
769 palkovsky 440
 * no empty magazine is available and cannot be allocated
759 palkovsky 441
 *
773 palkovsky 442
 * Assume mag_cache[CPU->id].lock is held
443
 *
759 palkovsky 444
 * We have 2 magazines bound to processor.
445
 * First try the current.
446
 *  If full, try the last.
447
 *   If full, put to magazines list.
448
 *   allocate new, exchange last & current
449
 *
768 palkovsky 450
 */
451
static slab_magazine_t * make_empty_current_mag(slab_cache_t *cache)
452
{
453
    slab_magazine_t *cmag,*lastmag,*newmag;
454
 
455
    cmag = cache->mag_cache[CPU->id].current;
456
    lastmag = cache->mag_cache[CPU->id].last;
457
 
458
    if (cmag) {
459
        if (cmag->busy < cmag->size)
460
            return cmag;
461
        if (lastmag && lastmag->busy < lastmag->size) {
462
            cache->mag_cache[CPU->id].last = cmag;
463
            cache->mag_cache[CPU->id].current = lastmag;
464
            return lastmag;
465
        }
466
    }
467
    /* current | last are full | nonexistent, allocate new */
468
    /* We do not want to sleep just because of caching */
469
    /* Especially we do not want reclaiming to start, as
470
     * this would deadlock */
471
    newmag = slab_alloc(&mag_cache, FRAME_ATOMIC | FRAME_NO_RECLAIM);
472
    if (!newmag)
473
        return NULL;
474
    newmag->size = SLAB_MAG_SIZE;
475
    newmag->busy = 0;
476
 
477
    /* Flush last to magazine list */
781 palkovsky 478
    if (lastmag)
479
        put_mag_to_cache(cache, lastmag);
480
 
768 palkovsky 481
    /* Move current as last, save new as current */
482
    cache->mag_cache[CPU->id].last = cmag; 
483
    cache->mag_cache[CPU->id].current = newmag;
484
 
485
    return newmag;
486
}
487
 
488
/**
489
 * Put object into CPU-cache magazine
490
 *
759 palkovsky 491
 * @return 0 - success, -1 - could not get memory
492
 */
493
static int magazine_obj_put(slab_cache_t *cache, void *obj)
494
{
495
    slab_magazine_t *mag;
496
 
772 palkovsky 497
    if (!CPU)
498
        return -1;
499
 
759 palkovsky 500
    spinlock_lock(&cache->mag_cache[CPU->id].lock);
768 palkovsky 501
 
502
    mag = make_empty_current_mag(cache);
769 palkovsky 503
    if (!mag) {
504
        spinlock_unlock(&cache->mag_cache[CPU->id].lock);
505
        return -1;
506
    }
759 palkovsky 507
 
508
    mag->objs[mag->busy++] = obj;
509
 
510
    spinlock_unlock(&cache->mag_cache[CPU->id].lock);
767 palkovsky 511
    atomic_inc(&cache->cached_objs);
759 palkovsky 512
    return 0;
513
}
514
 
515
 
516
/**************************************/
1248 jermar 517
/* Slab cache functions */
759 palkovsky 518
 
762 palkovsky 519
/** Return number of objects that fit in certain cache size */
520
static int comp_objects(slab_cache_t *cache)
521
{
522
    if (cache->flags & SLAB_CACHE_SLINSIDE)
523
        return ((PAGE_SIZE << cache->order) - sizeof(slab_t)) / cache->size;
524
    else
525
        return (PAGE_SIZE << cache->order) / cache->size;
526
}
527
 
528
/** Return wasted space in slab */
529
static int badness(slab_cache_t *cache)
530
{
531
    int objects;
532
    int ssize;
533
 
534
    objects = comp_objects(cache);
535
    ssize = PAGE_SIZE << cache->order;
536
    if (cache->flags & SLAB_CACHE_SLINSIDE)
537
        ssize -= sizeof(slab_t);
538
    return ssize - objects*cache->size;
539
}
540
 
789 palkovsky 541
/**
542
 * Initialize mag_cache structure in slab cache
543
 */
544
static void make_magcache(slab_cache_t *cache)
545
{
546
    int i;
791 palkovsky 547
 
548
    ASSERT(_slab_initialized >= 2);
789 palkovsky 549
 
822 palkovsky 550
    cache->mag_cache = malloc(sizeof(slab_mag_cache_t)*config.cpu_count,0);
789 palkovsky 551
    for (i=0; i < config.cpu_count; i++) {
552
        memsetb((__address)&cache->mag_cache[i],
553
            sizeof(cache->mag_cache[i]), 0);
554
        spinlock_initialize(&cache->mag_cache[i].lock,
555
                    "slab_maglock_cpu");
556
    }
557
}
558
 
759 palkovsky 559
/** Initialize allocated memory as a slab cache */
560
static void
561
_slab_cache_create(slab_cache_t *cache,
562
           char *name,
563
           size_t size,
564
           size_t align,
565
           int (*constructor)(void *obj, int kmflag),
787 palkovsky 566
           int (*destructor)(void *obj),
759 palkovsky 567
           int flags)
568
{
771 palkovsky 569
    int pages;
783 palkovsky 570
    ipl_t ipl;
759 palkovsky 571
 
572
    memsetb((__address)cache, sizeof(*cache), 0);
573
    cache->name = name;
574
 
766 palkovsky 575
    if (align < sizeof(__native))
576
        align = sizeof(__native);
577
    size = ALIGN_UP(size, align);
578
 
762 palkovsky 579
    cache->size = size;
759 palkovsky 580
 
581
    cache->constructor = constructor;
582
    cache->destructor = destructor;
583
    cache->flags = flags;
584
 
585
    list_initialize(&cache->full_slabs);
586
    list_initialize(&cache->partial_slabs);
587
    list_initialize(&cache->magazines);
776 palkovsky 588
    spinlock_initialize(&cache->slablock, "slab_lock");
589
    spinlock_initialize(&cache->maglock, "slab_maglock");
789 palkovsky 590
    if (! (cache->flags & SLAB_CACHE_NOMAGAZINE))
591
        make_magcache(cache);
759 palkovsky 592
 
593
    /* Compute slab sizes, object counts in slabs etc. */
594
    if (cache->size < SLAB_INSIDE_SIZE)
595
        cache->flags |= SLAB_CACHE_SLINSIDE;
596
 
762 palkovsky 597
    /* Minimum slab order */
1682 palkovsky 598
    pages = SIZE2FRAMES(cache->size);
1677 palkovsky 599
    /* We need the 2^order >= pages */
600
    if (pages == 1)
601
        cache->order = 0;
602
    else
603
        cache->order = fnzb(pages-1)+1;
766 palkovsky 604
 
762 palkovsky 605
    while (badness(cache) > SLAB_MAX_BADNESS(cache)) {
606
        cache->order += 1;
607
    }
608
    cache->objects = comp_objects(cache);
766 palkovsky 609
    /* If info fits in, put it inside */
610
    if (badness(cache) > sizeof(slab_t))
611
        cache->flags |= SLAB_CACHE_SLINSIDE;
762 palkovsky 612
 
783 palkovsky 613
    /* Add cache to cache list */
614
    ipl = interrupts_disable();
759 palkovsky 615
    spinlock_lock(&slab_cache_lock);
616
 
617
    list_append(&cache->link, &slab_cache_list);
618
 
619
    spinlock_unlock(&slab_cache_lock);
783 palkovsky 620
    interrupts_restore(ipl);
759 palkovsky 621
}
622
 
623
/** Create slab cache  */
624
slab_cache_t * slab_cache_create(char *name,
625
                 size_t size,
626
                 size_t align,
627
                 int (*constructor)(void *obj, int kmflag),
787 palkovsky 628
                 int (*destructor)(void *obj),
759 palkovsky 629
                 int flags)
630
{
631
    slab_cache_t *cache;
632
 
769 palkovsky 633
    cache = slab_alloc(&slab_cache_cache, 0);
759 palkovsky 634
    _slab_cache_create(cache, name, size, align, constructor, destructor,
635
               flags);
636
    return cache;
637
}
638
 
639
/**
640
 * Reclaim space occupied by objects that are already free
641
 *
642
 * @param flags If contains SLAB_RECLAIM_ALL, do aggressive freeing
643
 * @return Number of freed pages
644
 */
645
static count_t _slab_reclaim(slab_cache_t *cache, int flags)
646
{
647
    int i;
648
    slab_magazine_t *mag;
649
    count_t frames = 0;
781 palkovsky 650
    int magcount;
759 palkovsky 651
 
652
    if (cache->flags & SLAB_CACHE_NOMAGAZINE)
653
        return 0; /* Nothing to do */
781 palkovsky 654
 
655
    /* We count up to original magazine count to avoid
656
     * endless loop
657
     */
658
    magcount = atomic_get(&cache->magazine_counter);
659
    while (magcount-- && (mag=get_mag_from_cache(cache,0))) {
660
        frames += magazine_destroy(cache,mag);
661
        if (!(flags & SLAB_RECLAIM_ALL) && frames)
662
            break;
769 palkovsky 663
    }
759 palkovsky 664
 
665
    if (flags & SLAB_RECLAIM_ALL) {
781 palkovsky 666
        /* Free cpu-bound magazines */
759 palkovsky 667
        /* Destroy CPU magazines */
668
        for (i=0; i<config.cpu_count; i++) {
781 palkovsky 669
            spinlock_lock(&cache->mag_cache[i].lock);
670
 
759 palkovsky 671
            mag = cache->mag_cache[i].current;
672
            if (mag)
673
                frames += magazine_destroy(cache, mag);
674
            cache->mag_cache[i].current = NULL;
675
 
676
            mag = cache->mag_cache[i].last;
677
            if (mag)
678
                frames += magazine_destroy(cache, mag);
679
            cache->mag_cache[i].last = NULL;
781 palkovsky 680
 
681
            spinlock_unlock(&cache->mag_cache[i].lock);
759 palkovsky 682
        }
683
    }
767 palkovsky 684
 
759 palkovsky 685
    return frames;
686
}
687
 
688
/** Check that there are no slabs and remove cache from system  */
689
void slab_cache_destroy(slab_cache_t *cache)
690
{
781 palkovsky 691
    ipl_t ipl;
692
 
693
    /* First remove cache from link, so that we don't need
694
     * to disable interrupts later
695
     */
696
 
697
    ipl = interrupts_disable();
698
    spinlock_lock(&slab_cache_lock);
699
 
700
    list_remove(&cache->link);
701
 
702
    spinlock_unlock(&slab_cache_lock);
703
    interrupts_restore(ipl);
704
 
759 palkovsky 705
    /* Do not lock anything, we assume the software is correct and
706
     * does not touch the cache when it decides to destroy it */
707
 
708
    /* Destroy all magazines */
709
    _slab_reclaim(cache, SLAB_RECLAIM_ALL);
710
 
711
    /* All slabs must be empty */
712
    if (!list_empty(&cache->full_slabs) \
713
        || !list_empty(&cache->partial_slabs))
714
        panic("Destroying cache that is not empty.");
715
 
789 palkovsky 716
    if (!(cache->flags & SLAB_CACHE_NOMAGAZINE))
822 palkovsky 717
        free(cache->mag_cache);
769 palkovsky 718
    slab_free(&slab_cache_cache, cache);
759 palkovsky 719
}
720
 
721
/** Allocate new object from cache - if no flags given, always returns
722
    memory */
723
void * slab_alloc(slab_cache_t *cache, int flags)
724
{
725
    ipl_t ipl;
726
    void *result = NULL;
773 palkovsky 727
 
759 palkovsky 728
    /* Disable interrupts to avoid deadlocks with interrupt handlers */
729
    ipl = interrupts_disable();
771 palkovsky 730
 
814 palkovsky 731
    if (!(cache->flags & SLAB_CACHE_NOMAGAZINE)) {
759 palkovsky 732
        result = magazine_obj_get(cache);
814 palkovsky 733
    }
776 palkovsky 734
    if (!result)
759 palkovsky 735
        result = slab_obj_create(cache, flags);
736
 
769 palkovsky 737
    interrupts_restore(ipl);
738
 
764 palkovsky 739
    if (result)
740
        atomic_inc(&cache->allocated_objs);
741
 
759 palkovsky 742
    return result;
743
}
744
 
771 palkovsky 745
/** Return object to cache, use slab if known  */
746
static void _slab_free(slab_cache_t *cache, void *obj, slab_t *slab)
759 palkovsky 747
{
748
    ipl_t ipl;
749
 
750
    ipl = interrupts_disable();
751
 
762 palkovsky 752
    if ((cache->flags & SLAB_CACHE_NOMAGAZINE) \
753
        || magazine_obj_put(cache, obj)) {
776 palkovsky 754
 
771 palkovsky 755
        slab_obj_destroy(cache, obj, slab);
776 palkovsky 756
 
759 palkovsky 757
    }
769 palkovsky 758
    interrupts_restore(ipl);
764 palkovsky 759
    atomic_dec(&cache->allocated_objs);
759 palkovsky 760
}
761
 
771 palkovsky 762
/** Return slab object to cache */
763
void slab_free(slab_cache_t *cache, void *obj)
764
{
765
    _slab_free(cache,obj,NULL);
766
}
767
 
759 palkovsky 768
/* Go through all caches and reclaim what is possible */
769
count_t slab_reclaim(int flags)
770
{
771
    slab_cache_t *cache;
772
    link_t *cur;
773
    count_t frames = 0;
774
 
775
    spinlock_lock(&slab_cache_lock);
776
 
776 palkovsky 777
    /* TODO: Add assert, that interrupts are disabled, otherwise
778
     * memory allocation from interrupts can deadlock.
779
     */
780
 
759 palkovsky 781
    for (cur = slab_cache_list.next;cur!=&slab_cache_list; cur=cur->next) {
782
        cache = list_get_instance(cur, slab_cache_t, link);
783
        frames += _slab_reclaim(cache, flags);
784
    }
785
 
786
    spinlock_unlock(&slab_cache_lock);
787
 
788
    return frames;
789
}
790
 
791
 
792
/* Print list of slabs */
793
void slab_print_list(void)
794
{
795
    slab_cache_t *cache;
796
    link_t *cur;
783 palkovsky 797
    ipl_t ipl;
798
 
799
    ipl = interrupts_disable();
759 palkovsky 800
    spinlock_lock(&slab_cache_lock);
1248 jermar 801
    printf("slab name\t  Osize\t  Pages\t Obj/pg\t  Slabs\t Cached\tAllocobjs\tCtl\n");
759 palkovsky 802
    for (cur = slab_cache_list.next;cur!=&slab_cache_list; cur=cur->next) {
803
        cache = list_get_instance(cur, slab_cache_t, link);
1224 cejka 804
        printf("%s\t%7zd\t%7zd\t%7zd\t%7zd\t%7zd\t%7zd\t\t%s\n", cache->name, cache->size,
766 palkovsky 805
               (1 << cache->order), cache->objects,
767 palkovsky 806
               atomic_get(&cache->allocated_slabs),
807
               atomic_get(&cache->cached_objs),
766 palkovsky 808
               atomic_get(&cache->allocated_objs),
809
               cache->flags & SLAB_CACHE_SLINSIDE ? "In" : "Out");
759 palkovsky 810
    }
811
    spinlock_unlock(&slab_cache_lock);
783 palkovsky 812
    interrupts_restore(ipl);
759 palkovsky 813
}
814
 
815
void slab_cache_init(void)
816
{
771 palkovsky 817
    int i, size;
818
 
759 palkovsky 819
    /* Initialize magazine cache */
820
    _slab_cache_create(&mag_cache,
821
               "slab_magazine",
822
               sizeof(slab_magazine_t)+SLAB_MAG_SIZE*sizeof(void*),
823
               sizeof(__address),
824
               NULL, NULL,
769 palkovsky 825
               SLAB_CACHE_NOMAGAZINE | SLAB_CACHE_SLINSIDE);
826
    /* Initialize slab_cache cache */
827
    _slab_cache_create(&slab_cache_cache,
828
               "slab_cache",
789 palkovsky 829
               sizeof(slab_cache_cache),
769 palkovsky 830
               sizeof(__address),
831
               NULL, NULL,
832
               SLAB_CACHE_NOMAGAZINE | SLAB_CACHE_SLINSIDE);
833
    /* Initialize external slab cache */
834
    slab_extern_cache = slab_cache_create("slab_extern",
835
                          sizeof(slab_t),
836
                          0, NULL, NULL,
789 palkovsky 837
                          SLAB_CACHE_SLINSIDE | SLAB_CACHE_MAGDEFERRED);
759 palkovsky 838
 
839
    /* Initialize structures for malloc */
771 palkovsky 840
    for (i=0, size=(1<<SLAB_MIN_MALLOC_W);
841
         i < (SLAB_MAX_MALLOC_W-SLAB_MIN_MALLOC_W+1);
842
         i++, size <<= 1) {
843
        malloc_caches[i] = slab_cache_create(malloc_names[i],
844
                             size, 0,
789 palkovsky 845
                             NULL,NULL, SLAB_CACHE_MAGDEFERRED);
771 palkovsky 846
    }
778 palkovsky 847
#ifdef CONFIG_DEBUG       
848
    _slab_initialized = 1;
849
#endif
759 palkovsky 850
}
771 palkovsky 851
 
789 palkovsky 852
/** Enable cpu_cache
853
 *
854
 * Kernel calls this function, when it knows the real number of
855
 * processors.
856
 * Allocate slab for cpucache and enable it on all existing
857
 * slabs that are SLAB_CACHE_MAGDEFERRED
858
 */
859
void slab_enable_cpucache(void)
860
{
861
    link_t *cur;
862
    slab_cache_t *s;
863
 
791 palkovsky 864
#ifdef CONFIG_DEBUG
865
    _slab_initialized = 2;
866
#endif
867
 
789 palkovsky 868
    spinlock_lock(&slab_cache_lock);
869
 
870
    for (cur=slab_cache_list.next; cur != &slab_cache_list;cur=cur->next){
871
        s = list_get_instance(cur, slab_cache_t, link);
872
        if ((s->flags & SLAB_CACHE_MAGDEFERRED) != SLAB_CACHE_MAGDEFERRED)
873
            continue;
874
        make_magcache(s);
875
        s->flags &= ~SLAB_CACHE_MAGDEFERRED;
876
    }
877
 
878
    spinlock_unlock(&slab_cache_lock);
879
}
880
 
771 palkovsky 881
/**************************************/
882
/* kalloc/kfree functions             */
822 palkovsky 883
void * malloc(unsigned int size, int flags)
771 palkovsky 884
{
885
    int idx;
778 palkovsky 886
 
887
    ASSERT(_slab_initialized);
1288 jermar 888
    ASSERT(size && size <= (1 << SLAB_MAX_MALLOC_W));
771 palkovsky 889
 
890
    if (size < (1 << SLAB_MIN_MALLOC_W))
891
        size = (1 << SLAB_MIN_MALLOC_W);
892
 
893
    idx = fnzb(size-1) - SLAB_MIN_MALLOC_W + 1;
894
 
895
    return slab_alloc(malloc_caches[idx], flags);
896
}
897
 
822 palkovsky 898
void free(void *obj)
771 palkovsky 899
{
781 palkovsky 900
    slab_t *slab;
901
 
902
    if (!obj) return;
903
 
904
    slab = obj2slab(obj);
771 palkovsky 905
    _slab_free(slab->cache, obj, slab);
906
}
1702 cejka 907
 
1757 jermar 908
/** @}
1702 cejka 909
 */