dymelor.c

Go to the documentation of this file.

#include <core/init.h>
#include <mm/mm.h>
#include <scheduler/scheduler.h>

static void malloc_area_init(malloc_area * m_area, size_t size, int num_chunks)
{
    m_area->chunk_size = size;
    m_area->alloc_chunks = 0;
    m_area->dirty_chunks = 0;
    m_area->next_chunk = 0;
    m_area->num_chunks = num_chunks;
    m_area->state_changed = 0;
    m_area->last_access = -1;
    m_area->use_bitmap = NULL;
    m_area->dirty_bitmap = NULL;
    m_area->self_pointer = NULL;
    m_area->area = NULL;

#ifndef NDEBUG
    atomic_set(&m_area->presence, 0);
#endif

    m_area->prev = -1;
    m_area->next = -1;

    SET_AREA_LOCK_BIT(m_area);

}

malloc_state *malloc_state_init(void)
{
    int i, num_chunks;
    size_t chunk_size;
    malloc_state *state;

    state = rsalloc(sizeof(malloc_state));

    state->total_log_size = 0;
    state->total_inc_size = 0;
    state->busy_areas = 0;
    state->dirty_areas = 0;
    state->num_areas = NUM_AREAS;
    state->max_num_areas = MAX_NUM_AREAS;
    state->bitmap_size = 0;
    state->dirty_bitmap_size = 0;
    state->timestamp = -1;
    state->is_incremental = false;

    state->areas = (malloc_area *) rsalloc(state->max_num_areas * sizeof(malloc_area));
    if (unlikely(state->areas == NULL)) {
        rootsim_error(true, "Unable to allocate memory.\n");
    }

    chunk_size = MIN_CHUNK_SIZE;
    num_chunks = MIN_NUM_CHUNKS;

    for (i = 0; i < NUM_AREAS; i++) {

        malloc_area_init(&state->areas[i], chunk_size, num_chunks);
        state->areas[i].idx = i;
        chunk_size = chunk_size << 1;
    }

    return state;
}

void malloc_state_wipe(malloc_state **state_ptr)
{
    int i;
    malloc_state *state = *state_ptr;

    for (i = 0; i < NUM_AREAS; i++) {
        rsfree(state->areas[i].self_pointer); // TODO: when reintroducing the buddy, this must be changed
    }

    rsfree(*state_ptr);
    *state_ptr = NULL;
}

static size_t compute_size(size_t size)
{
    // Account for the space needed to keep the pointer to the malloc area
    size += sizeof(long long);
    size_t size_new;

    size_new = POWEROF2(size);
    if (unlikely(size_new < MIN_CHUNK_SIZE))
        size_new = MIN_CHUNK_SIZE;

    return size_new;
}

static void find_next_free(malloc_area * m_area)
{
    m_area->next_chunk++;

    while (m_area->next_chunk < m_area->num_chunks) {
        if (!bitmap_check(m_area->use_bitmap, m_area->next_chunk))
            break;

        m_area->next_chunk++;
    }

}

void *do_malloc(struct lp_struct *lp, size_t size)
{
    malloc_area *m_area, *prev_area = NULL;
    void *ptr;
    size_t area_size, bitmap_size;

    size = compute_size(size);

    if (unlikely(size > MAX_CHUNK_SIZE)) {
        rootsim_error(false, "Requested a memory allocation of %d but the limit is %d. Reconfigure MAX_CHUNK_SIZE. Returning NULL.\n",
                  size, MAX_CHUNK_SIZE);
        return NULL;
    }

    m_area = &lp->mm->m_state->areas[(int)log2(size) -
                     (int)log2(MIN_CHUNK_SIZE)];

#ifndef NDEBUG
    atomic_inc(&m_area->presence);
    assert(atomic_read(&m_area->presence) == 1);
#endif

    while (m_area != NULL && m_area->alloc_chunks == m_area->num_chunks) {
        prev_area = m_area;
        if (m_area->next == -1) {
            m_area = NULL;
        } else {
            m_area = &(lp->mm->m_state->areas[m_area->next]);
#ifndef NDEBUG
            atomic_inc(&m_area->presence);
            assert(atomic_read(&m_area->presence) == 1);
#endif
        }
    }

#ifndef NDEBUG
    if (prev_area != NULL) {
        atomic_dec(&prev_area->presence);
    }
#endif

    if (m_area == NULL) {

        printf("Initializing an additional area\n");
        fflush(stdout);

        if (lp->mm->m_state->num_areas == lp->mm->m_state->max_num_areas) {

            malloc_area *tmp = NULL;

            if ((lp->mm->m_state->max_num_areas << 1) > MAX_LIMIT_NUM_AREAS) {
#ifndef NDEBUG
                atomic_dec(&m_area->presence);
#endif
                return NULL;
            }

            lp->mm->m_state->max_num_areas <<= 1;

            rootsim_error(true, "To reimplement\n");
//                      tmp = (malloc_area *)pool_realloc_memory(lp->mm->m_state->areas, lp->mm->m_state->max_num_areas * sizeof(malloc_area));
            if (tmp == NULL) {

                rootsim_error(false,
                          "DyMeLoR: cannot reallocate the block of malloc_area.\n");
                lp->mm->m_state->max_num_areas >>= 1;

#ifndef NDEBUG
                atomic_dec(&m_area->presence);
#endif
                return NULL;
            }

            lp->mm->m_state->areas = tmp;
        }
        // Allocate a new malloc area
        m_area = &lp->mm->m_state->areas[lp->mm->m_state->num_areas];

        // The malloc area to be instantiated has twice the number of chunks wrt the last full malloc area for the same chunks size
        malloc_area_init(m_area, size, prev_area->num_chunks << 1);

#ifndef NDEBUG
        atomic_inc(&m_area->presence);
        assert(atomic_read(&m_area->presence) == 1);
#endif

        m_area->idx = lp->mm->m_state->num_areas;
        lp->mm->m_state->num_areas++;
        prev_area->next = m_area->idx;
        m_area->prev = prev_area->idx;

    }

    if (m_area->area == NULL) {

        bitmap_size = bitmap_required_size(m_area->num_chunks);

        area_size = sizeof(malloc_area *) + bitmap_size * 2 + m_area->num_chunks * size;

//              m_area->self_pointer = (malloc_area *)allocate_lp_memory(lp, area_size);
        m_area->self_pointer = rsalloc(area_size);
        bzero(m_area->self_pointer, area_size);

        if (unlikely(m_area->self_pointer == NULL)) {
            rootsim_error(true, "Error while allocating memory.\n");
        }

        m_area->dirty_chunks = 0;
        *(unsigned long long *)(m_area->self_pointer) =
            (unsigned long long)m_area;

        m_area->use_bitmap =
            ((unsigned char *)m_area->self_pointer +
             sizeof(malloc_area *));

        m_area->dirty_bitmap =
            ((unsigned char *)m_area->use_bitmap + bitmap_size);

        m_area->area =
            (void *)((char *)m_area->dirty_bitmap + bitmap_size);
    }

    if (unlikely(m_area->area == NULL)) {
        rootsim_error(true, "Error while allocating memory.\n");
    }
#ifndef NDEBUG
    if (bitmap_check(m_area->use_bitmap, m_area->next_chunk)) {
        rootsim_error(true, "Error: reallocating an already allocated chunk at %s:%d\n");
    }
#endif

    ptr = (void *)((char *)m_area->area + (m_area->next_chunk * size));

    bitmap_set(m_area->use_bitmap, m_area->next_chunk);

    bitmap_size = bitmap_required_size(m_area->num_chunks);

    if (m_area->alloc_chunks == 0) {
        lp->mm->m_state->bitmap_size += bitmap_size;
        lp->mm->m_state->busy_areas++;
    }

    if (m_area->state_changed == 0) {
        lp->mm->m_state->dirty_bitmap_size += bitmap_size;
        lp->mm->m_state->dirty_areas++;
    }

    m_area->state_changed = 1;
    m_area->last_access = lvt(current);

    if (!CHECK_LOG_MODE_BIT(m_area)) {
        if ((double)m_area->alloc_chunks / (double)m_area->num_chunks > MAX_LOG_THRESHOLD) {
            SET_LOG_MODE_BIT(m_area);
            lp->mm->m_state->total_log_size += (m_area->num_chunks - (m_area->alloc_chunks - 1)) * size;
        } else
            lp->mm->m_state->total_log_size += size;
    }
    //~ int chk_size = m_area->chunk_size;
    //~ RESET_BIT_AT(chk_size, 0);
    //~ RESET_BIT_AT(chk_size, 1);

    m_area->alloc_chunks++;
    find_next_free(m_area);

    // TODO: togliere
    memset(ptr, 0xe8, size);

    // Keep track of the malloc_area which this chunk belongs to
    *(unsigned long long *)ptr = (unsigned long long)m_area->self_pointer;

#ifndef NDEBUG
    atomic_dec(&m_area->presence);
#endif

    return (void *)((char *)ptr + sizeof(long long));
}

void do_free(struct lp_struct *lp, void *ptr)
{
    (void)lp;

    malloc_area *m_area = NULL;
    int idx;
    size_t chunk_size, bitmap_size;

    if (unlikely(rootsim_config.serial)) {
        rsfree(ptr);
        return;
    }

    if (unlikely(ptr == NULL)) {
        rootsim_error(false, "Invalid pointer during free\n");
        return;
    }

    m_area = get_area(ptr);
    if (unlikely(m_area == NULL)) {
        rootsim_error(false,
                  "Invalid pointer during free: malloc_area NULL\n");
        return;
    }
#ifndef NDEBUG
    atomic_inc(&m_area->presence);
    assert(atomic_read(&m_area->presence) == 1);
#endif

    chunk_size = UNTAGGED_CHUNK_SIZE(m_area);

    idx = (int)((char *)ptr - (char *)m_area->area) / chunk_size;

    if (!bitmap_check(m_area->use_bitmap, idx)) {
        rootsim_error(false, "double free() corruption or address not malloc'd\n");
        abort();
    }
    bitmap_reset(m_area->use_bitmap, idx);

    bitmap_size = bitmap_required_size(m_area->num_chunks);

    m_area->alloc_chunks--;

    if (m_area->alloc_chunks == 0) {
        lp->mm->m_state->bitmap_size -= bitmap_size;
        lp->mm->m_state->busy_areas--;
    }

    if (m_area->state_changed == 0) {
        lp->mm->m_state->dirty_bitmap_size += bitmap_size;
        lp->mm->m_state->dirty_areas++;
    }

    if (bitmap_check(m_area->dirty_bitmap, idx)) {
        bitmap_reset(m_area->dirty_bitmap, idx);
        m_area->dirty_chunks--;

        if (m_area->state_changed == 1 && m_area->dirty_chunks == 0)
            lp->mm->m_state->dirty_bitmap_size -= bitmap_size;

        lp->mm->m_state->total_inc_size -= chunk_size;

        if (unlikely(m_area->dirty_chunks < 0)) {
            rootsim_error(true, "negative number of chunks\n");
        }
    }

    m_area->state_changed = 1;

    m_area->last_access = lvt(current);

    if (CHECK_LOG_MODE_BIT(m_area)) {
        if ((double)m_area->alloc_chunks / (double)m_area->num_chunks < MIN_LOG_THRESHOLD) {
            RESET_LOG_MODE_BIT(m_area);
            lp->mm->m_state->total_log_size -= (m_area->num_chunks - m_area->alloc_chunks) * chunk_size;
        }
    } else {
        lp->mm->m_state->total_log_size -= chunk_size;
    }

    if (idx < m_area->next_chunk)
        m_area->next_chunk = idx;

#ifndef NDEBUG
    atomic_dec(&m_area->presence);
#endif
    // TODO: when do we free unrecoverable areas?
}

void dirty_mem(void *base, int size)
{

    // TODO: Quando reintegriamo l'incrementale questo qui deve ricomparire!
    (void)base;
    (void)size;

    return;

#if 0
//      unsigned long long current_cost;

    // Sanity check on passed address
/*  if(base == NULL) {
        rootsim_error(false, "Trying to access NULL. Memory interception aborted\n");
        return;
    }
*/
/*  if (rootsim_config.snapshot == AUTONOMIC_SNAPSHOT ||
        rootsim_config.snapshot == AUTONOMIC_INC_SNAPSHOT ||
        rootsim_config.snapshot == AUTONOMIC_FULL_SNAPSHOT)
        add_counter++;
*/
    int first_chunk, last_chunk, i, chk_size;

    size_t bitmap_size;

    malloc_area *m_area = get_area(base);

    if (m_area != NULL) {

        chk_size = UNTAGGED_CHUNK_SIZE(m_area->chunk_size);

        // Compute the number of chunks affected by the write
        first_chunk =
            (int)(((char *)base - (char *)m_area->area) / chk_size);

        // If size == -1, then we adopt a conservative approach: dirty all the chunks from the base to the end
        // of the actual malloc area base address belongs to.
        // This has been inserted to support the wrapping of third-party libraries where the size of the
        // update (or even the actual update) cannot be statically/dynamically determined.
        if (size == -1)
            last_chunk = m_area->num_chunks - 1;
        else
            last_chunk = (int)(((char *)base + size - 1 - (char *)m_area->area) / chk_size);

        bitmap_size = bitmap_required_size(m_area->num_chunks);

        if (m_area->state_changed == 1) {
            if (m_area->dirty_chunks == 0)
                lp->mm->m_state->dirty_bitmap_size += bitmap_size;
        } else {
            lp->mm->m_state->dirty_areas++;
            lp->mm->m_state->dirty_bitmap_size += bitmap_size * 2;
            m_area->state_changed = 1;
        }

        for (i = first_chunk; i <= last_chunk; i++) {

            // If it is dirtied a clean chunk, set it dirty and increase dirty object count for the malloc_area
            if (!bitmap_check(m_area->dirty_bitmap, i)) {
                bitmap_set(m_area->dirty_bitmap, i);
                lp->mm->m_state->total_inc_size += chk_size;

                m_area->dirty_chunks++;
            }
        }
    }
#endif
}

size_t get_log_size(malloc_state * logged_state)
{
    if (unlikely(logged_state == NULL))
        return 0;

    if (is_incremental(logged_state)) {
        return sizeof(malloc_state) +
            logged_state->dirty_areas * sizeof(malloc_area) +
            logged_state->dirty_bitmap_size +
            logged_state->total_inc_size;
    } else {
        return sizeof(malloc_state) +
            logged_state->busy_areas * sizeof(malloc_area) +
            logged_state->bitmap_size + logged_state->total_log_size;
    }
}

void *__wrap_malloc(size_t size)
{
    void *ptr;

    switch_to_platform_mode();

    if (unlikely(rootsim_config.serial)) {
        ptr = rsalloc(size);
        goto out;
    }

    ptr = do_malloc(current, size);

 out:
    switch_to_application_mode();
    return ptr;
}

void __wrap_free(void *ptr)
{
    switch_to_platform_mode();

    if (unlikely(rootsim_config.serial)) {
        rsfree(ptr);
        goto out;
    }

    do_free(current, ptr);

 out:
    switch_to_application_mode();
}

void *__wrap_realloc(void *ptr, size_t size)
{
    void *new_buffer;
    size_t old_size;
    size_t copy_size;
    malloc_area *m_area;

    if (unlikely(rootsim_config.serial)) {
        return rsrealloc(ptr, size);
    }
    // If ptr is NULL realloc is equivalent to the malloc
    if (unlikely(ptr == NULL)) {
        return __wrap_malloc(size);
    }
    // If ptr is not NULL and the size is 0 realloc is equivalent to the free
    if (unlikely(size == 0)) {
        __wrap_free(ptr);
        return NULL;
    }

    m_area = get_area(ptr);

    // The size could be greater than the real request, but it does not matter since the realloc specific requires that
    // is copied at least the smaller buffer size between the new and the old one
    old_size = UNTAGGED_CHUNK_SIZE(m_area);

    new_buffer = __wrap_malloc(size);

    if (unlikely(new_buffer == NULL))
        return NULL;

    copy_size = min(size, old_size);
    memcpy(new_buffer, ptr, copy_size);
    __wrap_free(ptr);

    return new_buffer;
}

void *__wrap_calloc(size_t nmemb, size_t size)
{
    void *buffer;

    if (unlikely(rootsim_config.serial)) {
        return rscalloc(nmemb, size);
    }

    if (unlikely(nmemb == 0 || size == 0))
        return NULL;

    buffer = __wrap_malloc(nmemb * size);
    if (unlikely(buffer == NULL))
        return NULL;

    bzero(buffer, nmemb * size);

    return buffer;
}

void clean_buffers_on_gvt(struct lp_struct *lp, simtime_t time_barrier)
{
    int i;
    malloc_area *m_area;
    malloc_state *state = lp->mm->m_state;

    // The first NUM_AREAS malloc_areas are placed according to their chunks' sizes. The exceeding malloc_areas can be compacted
    for (i = NUM_AREAS; i < state->num_areas; i++) {
        m_area = &state->areas[i];

        if (m_area->alloc_chunks == 0
            && m_area->last_access < time_barrier
            && !CHECK_AREA_LOCK_BIT(m_area)) {

            if (m_area->self_pointer != NULL) {

                //free_lp_memory(lp, m_area->self_pointer);
                rsfree(m_area->self_pointer);

                m_area->use_bitmap = NULL;
                m_area->dirty_bitmap = NULL;
                m_area->self_pointer = NULL;
                m_area->area = NULL;
                m_area->state_changed = 0;

                // Set the pointers
                if (m_area->prev != -1)
                    state->areas[m_area->prev].next = m_area->next;
                if (m_area->next != -1)
                    state->areas[m_area->next].prev = m_area->prev;

                // Swap, if possible
                if (i < state->num_areas - 1) {
                    memcpy(m_area, &state->areas[state->num_areas - 1], sizeof(malloc_area));
                    m_area->idx = i;
                    if (m_area->prev != -1)
                        state->areas[m_area->prev].next = m_area->idx;
                    if (m_area->next != -1)
                        state->areas[m_area->next].prev = m_area->idx;

                    // Update the self pointer
                    *(long long *)m_area->self_pointer = (long long)m_area;

                    // The swapped area will now be checked
                    i--;
                }
                // Decrement the expected number of areas
                state->num_areas--;
            }
        }
    }
}