Subversion Repositories HelenOS

/*

 * 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.

 */

 /** @addtogroup genericmm

 * @{

 */

/**

 * @file

 * @brief   Slab allocator.

 *

 * The slab allocator is closely modelled after OpenSolaris slab allocator.

 * @see http://www.usenix.org/events/usenix01/full_papers/bonwick/bonwick_html/

 *

 * with the following exceptions:

 * @li empty slabs are deallocated immediately

 *     (in Linux they are kept in linked list, in Solaris ???)

 * @li empty magazines are deallocated when not needed

 *     (in Solaris they are held in linked list in slab cache)

 *

 * Following features are not currently supported but would be easy to do:

 * @li cache coloring

 * @li dynamic magazine growing (different magazine sizes are already

 *     supported, but we would need to adjust allocation strategy)

 *

 * The slab allocator supports per-CPU caches ('magazines') to facilitate

 * good SMP scaling.

 *

 * When a new object is being allocated, it is first checked, if it is

 * available in a CPU-bound magazine. If it is not found there, it is

 * allocated from a CPU-shared slab - if a partially full one is found,

 * it is used, otherwise a new one is allocated.

 *

 * When an object is being deallocated, it is put to a CPU-bound magazine.

 * If there is no such magazine, a new one is allocated (if this fails,

 * the object is deallocated into slab). If the magazine is full, it is

 * put into cpu-shared list of magazines and a new one is allocated.

 *

 * The CPU-bound magazine is actually a pair of magazines in order to avoid

 * thrashing when somebody is allocating/deallocating 1 item at the magazine

 * size boundary. LIFO order is enforced, which should avoid fragmentation

 * as much as possible.

 *  

 * Every cache contains list of full slabs and list of partially full slabs.

 * Empty slabs are immediately freed (thrashing will be avoided because

 * of magazines).

 *

 * The slab information structure is kept inside the data area, if possible.

 * The cache can be marked that it should not use magazines. This is used

 * only for slab related caches to avoid deadlocks and infinite recursion

 * (the slab allocator uses itself for allocating all it's control structures).

 *

 * The slab allocator allocates a lot of space and does not free it. When

 * the frame allocator fails to allocate a frame, it calls slab_reclaim().

 * It tries 'light reclaim' first, then brutal reclaim. The light reclaim

 * releases slabs from cpu-shared magazine-list, until at least 1 slab

 * is deallocated in each cache (this algorithm should probably change).

 * The brutal reclaim removes all cached objects, even from CPU-bound

 * magazines.

 *

 * TODO:@n

 * For better CPU-scaling the magazine allocation strategy should

 * be extended. Currently, if the cache does not have magazine, it asks

 * for non-cpu cached magazine cache to provide one. It might be feasible

 * to add cpu-cached magazine cache (which would allocate it's magazines

 * from non-cpu-cached mag. cache). This would provide a nice per-cpu

 * buffer. The other possibility is to use the per-cache

 * 'empty-magazine-list', which decreases competing for 1 per-system

 * magazine cache.

 *

 * @li it might be good to add granularity of locks even to slab level,

 *     we could then try_spinlock over all partial slabs and thus improve

 *     scalability even on slab level

 */

#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);

/** Magazine cache */

static slab_cache_t mag_cache;

/** Cache for cache descriptors */

static slab_cache_t slab_cache_cache;

/** Cache for external slab descriptors

 * This time we want per-cpu cache, so do not make it static

 * - using slab for internal slab structures will not deadlock,

 *   as all slab structures are 'small' - control structures of

 *   their caches do not require further allocation

 */

static slab_cache_t *slab_extern_cache;

/** Caches for malloc */

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"

};

/** Slab descriptor */

typedef struct {

    slab_cache_t *cache; /**< Pointer to parent cache */

    link_t link;       /* List of full/partial slabs */

    void *start;       /**< Start address of first available item */

    count_t available; /**< Count of available items in this slab */

    index_t nextavail; /**< The index of next available item */

}slab_t;

#ifdef CONFIG_DEBUG

static int _slab_initialized = 0;

#endif

/**************************************/

/* Slab allocation functions          */

/**

 * Allocate frames for slab space and initialize

 *

 */

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;

}

/**

 * Deallocate space associated with slab

 *

 * @return number of freed frames

 */

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;

}

/** Map object to slab structure */

static slab_t * obj2slab(void *obj)

{

    return (slab_t *)frame_get_parent(ADDR2PFN(KA2PA(obj)), 0);

}

/**************************************/

/* Slab functions */

/**

 * Return object to slab and call a destructor

 *

 * @param slab If the caller knows directly slab of the object, otherwise NULL

 *

 * @return Number of freed pages

 */

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;

}

/**

 * Take new object from slab or create new if needed

 *

 * @return Object address or null

 */

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 */

/**

 * Finds a full magazine in cache, takes it from list

 * and returns it

 *

 * @param first If true, return first, else last mag

 */

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;

}

/** Prepend magazine to magazine list in cache */

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);

}

/**

 * Free all objects in magazine and free memory associated with magazine

 *

 * @return Number of freed pages

 */

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;

}

/**

 * Find full magazine, set it as current and return it

 *

 * Assume cpu_magazine lock is held

 */

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;

}

/**

 * Try to find object in CPU-cache magazines

 *

 * @return Pointer to object or NULL if not available

 */

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;

}

/**

 * Assure that the current magazine is empty, return pointer to it, or NULL if

 * no empty magazine is available and cannot be allocated

 *

 * Assume mag_cache[CPU->id].lock is held

 *

 * We have 2 magazines bound to processor.

 * First try the current.

 *  If full, try the last.

 *   If full, put to magazines list.

 *   allocate new, exchange last & current

 *

 */

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;

}

/**

 * Put object into CPU-cache magazine

 *

 * @return 0 - success, -1 - could not get memory

 */

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 */

/** Return number of objects that fit in certain cache size */

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;

}

/** Return wasted space in slab */

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;

}

/**

 * Initialize mag_cache structure in slab cache

 */

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");

    }

}

/** Initialize allocated memory as a slab cache */

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);

}

/** Create slab cache  */

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;

}

/**

 * Reclaim space occupied by objects that are already free

 *

 * @param flags If contains SLAB_RECLAIM_ALL, do aggressive freeing

 * @return Number of freed pages

 */

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;

}

/** Check that there are no slabs and remove cache from system  */

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);

}

/** Allocate new object from cache - if no flags given, always returns

    memory */

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;

}

/** Return object to cache, use slab if known  */

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);

}

/** Return slab object to cache */

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;