slab.c

Go to the documentation of this file.

/*
 * Copyright (C) 2006 Ondrej Palkovsky
 * All rights reserved.
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions
 * are met:
 *
 * - Redistributions of source code must retain the above copyright
 *   notice, this list of conditions and the following disclaimer.
 * - Redistributions in binary form must reproduce the above copyright
 *   notice, this list of conditions and the following disclaimer in the
 *   documentation and/or other materials provided with the distribution.
 * - The name of the author may not be used to endorse or promote products
 *   derived from this software without specific prior written permission.
 *
 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 */

#include <synch/spinlock.h>
#include <mm/slab.h>
#include <adt/list.h>
#include <memstr.h>
#include <align.h>
#include <mm/frame.h>
#include <config.h>
#include <print.h>
#include <arch.h>
#include <panic.h>
#include <debug.h>
#include <bitops.h>

SPINLOCK_INITIALIZE(slab_cache_lock);
static LIST_INITIALIZE(slab_cache_list);

static slab_cache_t mag_cache;
static slab_cache_t slab_cache_cache;
static slab_cache_t *slab_extern_cache;
static slab_cache_t *malloc_caches[SLAB_MAX_MALLOC_W-SLAB_MIN_MALLOC_W+1];
char *malloc_names[] =  {
        "malloc-16","malloc-32","malloc-64","malloc-128",
        "malloc-256","malloc-512","malloc-1K","malloc-2K",
        "malloc-4K","malloc-8K","malloc-16K","malloc-32K",
        "malloc-64K","malloc-128K","malloc-256K"
};

typedef struct {
        slab_cache_t *cache; 
        link_t link;       /* List of full/partial slabs */
        void *start;       
        count_t available; 
        index_t nextavail; 
}slab_t;

#ifdef CONFIG_DEBUG
static int _slab_initialized = 0;
#endif

/**************************************/
/* Slab allocation functions          */

static slab_t * slab_space_alloc(slab_cache_t *cache, int flags)
{
        void *data;
        slab_t *slab;
        size_t fsize;
        int i;
        int status;
        pfn_t pfn;
        int zone=0;
        
        pfn = frame_alloc_rc_zone(cache->order, FRAME_KA | flags, &status, &zone);
        data = (void *) PA2KA(PFN2ADDR(pfn));
        if (status != FRAME_OK) {
                return NULL;
        }
        if (! (cache->flags & SLAB_CACHE_SLINSIDE)) {
                slab = slab_alloc(slab_extern_cache, flags);
                if (!slab) {
                        frame_free(ADDR2PFN(KA2PA(data)));
                        return NULL;
                }
        } else {
                fsize = (PAGE_SIZE << cache->order);
                slab = data + fsize - sizeof(*slab);
        }
        
        /* Fill in slab structures */
        for (i=0; i < (1 << cache->order); i++)
                frame_set_parent(pfn+i, slab, zone);

        slab->start = data;
        slab->available = cache->objects;
        slab->nextavail = 0;
        slab->cache = cache;

        for (i=0; i<cache->objects;i++)
                *((int *) (slab->start + i*cache->size)) = i+1;

        atomic_inc(&cache->allocated_slabs);
        return slab;
}

static count_t slab_space_free(slab_cache_t *cache, slab_t *slab)
{
        frame_free(ADDR2PFN(KA2PA(slab->start)));
        if (! (cache->flags & SLAB_CACHE_SLINSIDE))
                slab_free(slab_extern_cache, slab);

        atomic_dec(&cache->allocated_slabs);
        
        return 1 << cache->order;
}

static slab_t * obj2slab(void *obj)
{
        return (slab_t *)frame_get_parent(ADDR2PFN(KA2PA(obj)), 0);
}

/**************************************/
/* Slab functions */


static count_t slab_obj_destroy(slab_cache_t *cache, void *obj,
                                slab_t *slab)
{
        int freed = 0;

        if (!slab)
                slab = obj2slab(obj);

        ASSERT(slab->cache == cache);

        if (cache->destructor)
                freed = cache->destructor(obj);
        
        spinlock_lock(&cache->slablock);
        ASSERT(slab->available < cache->objects);

        *((int *)obj) = slab->nextavail;
        slab->nextavail = (obj - slab->start)/cache->size;
        slab->available++;

        /* Move it to correct list */
        if (slab->available == cache->objects) {
                /* Free associated memory */
                list_remove(&slab->link);
                spinlock_unlock(&cache->slablock);

                return freed + slab_space_free(cache, slab);

        } else if (slab->available == 1) {
                /* It was in full, move to partial */
                list_remove(&slab->link);
                list_prepend(&slab->link, &cache->partial_slabs);
        }
        spinlock_unlock(&cache->slablock);
        return freed;
}

static void * slab_obj_create(slab_cache_t *cache, int flags)
{
        slab_t *slab;
        void *obj;

        spinlock_lock(&cache->slablock);

        if (list_empty(&cache->partial_slabs)) {
                /* Allow recursion and reclaiming
                 * - this should work, as the slab control structures
                 *   are small and do not need to allocate with anything
                 *   other than frame_alloc when they are allocating,
                 *   that's why we should get recursion at most 1-level deep
                 */
                spinlock_unlock(&cache->slablock);
                slab = slab_space_alloc(cache, flags);
                if (!slab)
                        return NULL;
                spinlock_lock(&cache->slablock);
        } else {
                slab = list_get_instance(cache->partial_slabs.next,
                                         slab_t,
                                         link);
                list_remove(&slab->link);
        }
        obj = slab->start + slab->nextavail * cache->size;
        slab->nextavail = *((int *)obj);
        slab->available--;

        if (! slab->available)
                list_prepend(&slab->link, &cache->full_slabs);
        else
                list_prepend(&slab->link, &cache->partial_slabs);

        spinlock_unlock(&cache->slablock);

        if (cache->constructor && cache->constructor(obj, flags)) {
                /* Bad, bad, construction failed */
                slab_obj_destroy(cache, obj, slab);
                return NULL;
        }
        return obj;
}

/**************************************/
/* CPU-Cache slab functions */

static slab_magazine_t * get_mag_from_cache(slab_cache_t *cache,
                                            int first)
{
        slab_magazine_t *mag = NULL;
        link_t *cur;

        spinlock_lock(&cache->maglock);
        if (!list_empty(&cache->magazines)) {
                if (first)
                        cur = cache->magazines.next;
                else
                        cur = cache->magazines.prev;
                mag = list_get_instance(cur, slab_magazine_t, link);
                list_remove(&mag->link);
                atomic_dec(&cache->magazine_counter);
        }
        spinlock_unlock(&cache->maglock);
        return mag;
}

static void put_mag_to_cache(slab_cache_t *cache, slab_magazine_t *mag)
{
        spinlock_lock(&cache->maglock);

        list_prepend(&mag->link, &cache->magazines);
        atomic_inc(&cache->magazine_counter);
        
        spinlock_unlock(&cache->maglock);
}

static count_t magazine_destroy(slab_cache_t *cache, 
                                slab_magazine_t *mag)
{
        int i;
        count_t frames = 0;

        for (i=0;i < mag->busy; i++) {
                frames += slab_obj_destroy(cache, mag->objs[i], NULL);
                atomic_dec(&cache->cached_objs);
        }
        
        slab_free(&mag_cache, mag);

        return frames;
}

static slab_magazine_t * get_full_current_mag(slab_cache_t *cache)
{
        slab_magazine_t *cmag, *lastmag, *newmag;

        cmag = cache->mag_cache[CPU->id].current;
        lastmag = cache->mag_cache[CPU->id].last;
        if (cmag) { /* First try local CPU magazines */
                if (cmag->busy)
                        return cmag;

                if (lastmag && lastmag->busy) {
                        cache->mag_cache[CPU->id].current = lastmag;
                        cache->mag_cache[CPU->id].last = cmag;
                        return lastmag;
                }
        }
        /* Local magazines are empty, import one from magazine list */
        newmag = get_mag_from_cache(cache, 1);
        if (!newmag)
                return NULL;

        if (lastmag)
                magazine_destroy(cache, lastmag);

        cache->mag_cache[CPU->id].last = cmag;
        cache->mag_cache[CPU->id].current = newmag;
        return newmag;
}

static void * magazine_obj_get(slab_cache_t *cache)
{
        slab_magazine_t *mag;
        void *obj;

        if (!CPU)
                return NULL;

        spinlock_lock(&cache->mag_cache[CPU->id].lock);

        mag = get_full_current_mag(cache);
        if (!mag) {
                spinlock_unlock(&cache->mag_cache[CPU->id].lock);
                return NULL;
        }
        obj = mag->objs[--mag->busy];
        spinlock_unlock(&cache->mag_cache[CPU->id].lock);
        atomic_dec(&cache->cached_objs);
        
        return obj;
}

static slab_magazine_t * make_empty_current_mag(slab_cache_t *cache)
{
        slab_magazine_t *cmag,*lastmag,*newmag;

        cmag = cache->mag_cache[CPU->id].current;
        lastmag = cache->mag_cache[CPU->id].last;

        if (cmag) {
                if (cmag->busy < cmag->size)
                        return cmag;
                if (lastmag && lastmag->busy < lastmag->size) {
                        cache->mag_cache[CPU->id].last = cmag;
                        cache->mag_cache[CPU->id].current = lastmag;
                        return lastmag;
                }
        }
        /* current | last are full | nonexistent, allocate new */
        /* We do not want to sleep just because of caching */
        /* Especially we do not want reclaiming to start, as 
         * this would deadlock */
        newmag = slab_alloc(&mag_cache, FRAME_ATOMIC | FRAME_NO_RECLAIM);
        if (!newmag)
                return NULL;
        newmag->size = SLAB_MAG_SIZE;
        newmag->busy = 0;

        /* Flush last to magazine list */
        if (lastmag)
                put_mag_to_cache(cache, lastmag);

        /* Move current as last, save new as current */
        cache->mag_cache[CPU->id].last = cmag;  
        cache->mag_cache[CPU->id].current = newmag;     

        return newmag;
}

static int magazine_obj_put(slab_cache_t *cache, void *obj)
{
        slab_magazine_t *mag;

        if (!CPU)
                return -1;

        spinlock_lock(&cache->mag_cache[CPU->id].lock);

        mag = make_empty_current_mag(cache);
        if (!mag) {
                spinlock_unlock(&cache->mag_cache[CPU->id].lock);
                return -1;
        }
        
        mag->objs[mag->busy++] = obj;

        spinlock_unlock(&cache->mag_cache[CPU->id].lock);
        atomic_inc(&cache->cached_objs);
        return 0;
}


/**************************************/
/* Slab cache functions */

static int comp_objects(slab_cache_t *cache)
{
        if (cache->flags & SLAB_CACHE_SLINSIDE)
                return ((PAGE_SIZE << cache->order) - sizeof(slab_t)) / cache->size;
        else 
                return (PAGE_SIZE << cache->order) / cache->size;
}

static int badness(slab_cache_t *cache)
{
        int objects;
        int ssize;

        objects = comp_objects(cache);
        ssize = PAGE_SIZE << cache->order;
        if (cache->flags & SLAB_CACHE_SLINSIDE)
                ssize -= sizeof(slab_t);
        return ssize - objects*cache->size;
}

static void make_magcache(slab_cache_t *cache)
{
        int i;
        
        ASSERT(_slab_initialized >= 2);

        cache->mag_cache = malloc(sizeof(slab_mag_cache_t)*config.cpu_count,0);
        for (i=0; i < config.cpu_count; i++) {
                memsetb((__address)&cache->mag_cache[i],
                        sizeof(cache->mag_cache[i]), 0);
                spinlock_initialize(&cache->mag_cache[i].lock, 
                                    "slab_maglock_cpu");
        }
}

static void
_slab_cache_create(slab_cache_t *cache,
                   char *name,
                   size_t size,
                   size_t align,
                   int (*constructor)(void *obj, int kmflag),
                   int (*destructor)(void *obj),
                   int flags)
{
        int pages;
        ipl_t ipl;

        memsetb((__address)cache, sizeof(*cache), 0);
        cache->name = name;

        if (align < sizeof(__native))
                align = sizeof(__native);
        size = ALIGN_UP(size, align);
                
        cache->size = size;

        cache->constructor = constructor;
        cache->destructor = destructor;
        cache->flags = flags;

        list_initialize(&cache->full_slabs);
        list_initialize(&cache->partial_slabs);
        list_initialize(&cache->magazines);
        spinlock_initialize(&cache->slablock, "slab_lock");
        spinlock_initialize(&cache->maglock, "slab_maglock");
        if (! (cache->flags & SLAB_CACHE_NOMAGAZINE))
                make_magcache(cache);

        /* Compute slab sizes, object counts in slabs etc. */
        if (cache->size < SLAB_INSIDE_SIZE)
                cache->flags |= SLAB_CACHE_SLINSIDE;

        /* Minimum slab order */
        pages = SIZE2FRAMES(cache->size);
        /* We need the 2^order >= pages */
        if (pages == 1)
                cache->order = 0;
        else
                cache->order = fnzb(pages-1)+1;

        while (badness(cache) > SLAB_MAX_BADNESS(cache)) {
                cache->order += 1;
        }
        cache->objects = comp_objects(cache);
        /* If info fits in, put it inside */
        if (badness(cache) > sizeof(slab_t))
                cache->flags |= SLAB_CACHE_SLINSIDE;

        /* Add cache to cache list */
        ipl = interrupts_disable();
        spinlock_lock(&slab_cache_lock);

        list_append(&cache->link, &slab_cache_list);

        spinlock_unlock(&slab_cache_lock);
        interrupts_restore(ipl);
}

slab_cache_t * slab_cache_create(char *name,
                                 size_t size,
                                 size_t align,
                                 int (*constructor)(void *obj, int kmflag),
                                 int (*destructor)(void *obj),
                                 int flags)
{
        slab_cache_t *cache;

        cache = slab_alloc(&slab_cache_cache, 0);
        _slab_cache_create(cache, name, size, align, constructor, destructor,
                           flags);
        return cache;
}

static count_t _slab_reclaim(slab_cache_t *cache, int flags)
{
        int i;
        slab_magazine_t *mag;
        count_t frames = 0;
        int magcount;
        
        if (cache->flags & SLAB_CACHE_NOMAGAZINE)
                return 0; /* Nothing to do */

        /* We count up to original magazine count to avoid
         * endless loop 
         */
        magcount = atomic_get(&cache->magazine_counter);
        while (magcount-- && (mag=get_mag_from_cache(cache,0))) {
                frames += magazine_destroy(cache,mag);
                if (!(flags & SLAB_RECLAIM_ALL) && frames)
                        break;
        }
        
        if (flags & SLAB_RECLAIM_ALL) {
                /* Free cpu-bound magazines */
                /* Destroy CPU magazines */
                for (i=0; i<config.cpu_count; i++) {
                        spinlock_lock(&cache->mag_cache[i].lock);

                        mag = cache->mag_cache[i].current;
                        if (mag)
                                frames += magazine_destroy(cache, mag);
                        cache->mag_cache[i].current = NULL;
                        
                        mag = cache->mag_cache[i].last;
                        if (mag)
                                frames += magazine_destroy(cache, mag);
                        cache->mag_cache[i].last = NULL;

                        spinlock_unlock(&cache->mag_cache[i].lock);
                }
        }

        return frames;
}

void slab_cache_destroy(slab_cache_t *cache)
{
        ipl_t ipl;

        /* First remove cache from link, so that we don't need
         * to disable interrupts later
         */

        ipl = interrupts_disable();
        spinlock_lock(&slab_cache_lock);

        list_remove(&cache->link);

        spinlock_unlock(&slab_cache_lock);
        interrupts_restore(ipl);

        /* Do not lock anything, we assume the software is correct and
         * does not touch the cache when it decides to destroy it */
        
        /* Destroy all magazines */
        _slab_reclaim(cache, SLAB_RECLAIM_ALL);

        /* All slabs must be empty */
        if (!list_empty(&cache->full_slabs) \
            || !list_empty(&cache->partial_slabs))
                panic("Destroying cache that is not empty.");

        if (!(cache->flags & SLAB_CACHE_NOMAGAZINE))
                free(cache->mag_cache);
        slab_free(&slab_cache_cache, cache);
}

void * slab_alloc(slab_cache_t *cache, int flags)
{
        ipl_t ipl;
        void *result = NULL;
        
        /* Disable interrupts to avoid deadlocks with interrupt handlers */
        ipl = interrupts_disable();

        if (!(cache->flags & SLAB_CACHE_NOMAGAZINE)) {
                result = magazine_obj_get(cache);
        }
        if (!result)
                result = slab_obj_create(cache, flags);

        interrupts_restore(ipl);

        if (result)
                atomic_inc(&cache->allocated_objs);

        return result;
}

static void _slab_free(slab_cache_t *cache, void *obj, slab_t *slab)
{
        ipl_t ipl;

        ipl = interrupts_disable();

        if ((cache->flags & SLAB_CACHE_NOMAGAZINE) \
            || magazine_obj_put(cache, obj)) {

                slab_obj_destroy(cache, obj, slab);

        }
        interrupts_restore(ipl);
        atomic_dec(&cache->allocated_objs);
}

void slab_free(slab_cache_t *cache, void *obj)
{
        _slab_free(cache,obj,NULL);
}

/* Go through all caches and reclaim what is possible */
count_t slab_reclaim(int flags)
{
        slab_cache_t *cache;
        link_t *cur;
        count_t frames = 0;

        spinlock_lock(&slab_cache_lock);

        /* TODO: Add assert, that interrupts are disabled, otherwise
         * memory allocation from interrupts can deadlock.
         */

        for (cur = slab_cache_list.next;cur!=&slab_cache_list; cur=cur->next) {
                cache = list_get_instance(cur, slab_cache_t, link);
                frames += _slab_reclaim(cache, flags);
        }

        spinlock_unlock(&slab_cache_lock);

        return frames;
}


/* Print list of slabs */
void slab_print_list(void)
{
        slab_cache_t *cache;
        link_t *cur;
        ipl_t ipl;
        
        ipl = interrupts_disable();
        spinlock_lock(&slab_cache_lock);
        printf("slab name\t  Osize\t  Pages\t Obj/pg\t  Slabs\t Cached\tAllocobjs\tCtl\n");
        for (cur = slab_cache_list.next;cur!=&slab_cache_list; cur=cur->next) {
                cache = list_get_instance(cur, slab_cache_t, link);
                printf("%s\t%7zd\t%7zd\t%7zd\t%7zd\t%7zd\t%7zd\t\t%s\n", cache->name, cache->size, 
                       (1 << cache->order), cache->objects,
                       atomic_get(&cache->allocated_slabs),
                       atomic_get(&cache->cached_objs),
                       atomic_get(&cache->allocated_objs),
                       cache->flags & SLAB_CACHE_SLINSIDE ? "In" : "Out");
        }
        spinlock_unlock(&slab_cache_lock);
        interrupts_restore(ipl);
}

void slab_cache_init(void)
{
        int i, size;

        /* Initialize magazine cache */
        _slab_cache_create(&mag_cache,
                           "slab_magazine",
                           sizeof(slab_magazine_t)+SLAB_MAG_SIZE*sizeof(void*),
                           sizeof(__address),
                           NULL, NULL,
                           SLAB_CACHE_NOMAGAZINE | SLAB_CACHE_SLINSIDE);
        /* Initialize slab_cache cache */
        _slab_cache_create(&slab_cache_cache,
                           "slab_cache",
                           sizeof(slab_cache_cache),
                           sizeof(__address),
                           NULL, NULL,
                           SLAB_CACHE_NOMAGAZINE | SLAB_CACHE_SLINSIDE);
        /* Initialize external slab cache */
        slab_extern_cache = slab_cache_create("slab_extern",
                                              sizeof(slab_t),
                                              0, NULL, NULL,
                                              SLAB_CACHE_SLINSIDE | SLAB_CACHE_MAGDEFERRED);

        /* Initialize structures for malloc */
        for (i=0, size=(1<<SLAB_MIN_MALLOC_W);
             i < (SLAB_MAX_MALLOC_W-SLAB_MIN_MALLOC_W+1);
             i++, size <<= 1) {
                malloc_caches[i] = slab_cache_create(malloc_names[i],
                                                     size, 0,
                                                     NULL,NULL, SLAB_CACHE_MAGDEFERRED);
        }
#ifdef CONFIG_DEBUG       
        _slab_initialized = 1;
#endif
}

void slab_enable_cpucache(void)
{
        link_t *cur;
        slab_cache_t *s;

#ifdef CONFIG_DEBUG
        _slab_initialized = 2;
#endif

        spinlock_lock(&slab_cache_lock);
        
        for (cur=slab_cache_list.next; cur != &slab_cache_list;cur=cur->next){
                s = list_get_instance(cur, slab_cache_t, link);
                if ((s->flags & SLAB_CACHE_MAGDEFERRED) != SLAB_CACHE_MAGDEFERRED)
                        continue;
                make_magcache(s);
                s->flags &= ~SLAB_CACHE_MAGDEFERRED;
        }

        spinlock_unlock(&slab_cache_lock);
}

/**************************************/
/* kalloc/kfree functions             */
void * malloc(unsigned int size, int flags)
{
        int idx;

        ASSERT(_slab_initialized);
        ASSERT(size && size <= (1 << SLAB_MAX_MALLOC_W));
        
        if (size < (1 << SLAB_MIN_MALLOC_W))
                size = (1 << SLAB_MIN_MALLOC_W);

        idx = fnzb(size-1) - SLAB_MIN_MALLOC_W + 1;

        return slab_alloc(malloc_caches[idx], flags);
}

void free(void *obj)
{
        slab_t *slab;

        if (!obj) return;

        slab = obj2slab(obj);
        _slab_free(slab->cache, obj, slab);
}

---