/*
* heap_recycle_block -- (internal) recycles unused part of the memory block
*/
static void
heap_recycle_block(PMEMobjpool *pop, struct bucket *b, struct memory_block *m,
uint32_t units)
{
if (bucket_is_small(b)) {
struct memory_block r = {m->chunk_id, m->zone_id,
m->size_idx - units, m->block_off + units
};
bucket_insert_block(b, r);
} else {
heap_resize_chunk(pop, m->chunk_id, m->zone_id, units);
}
m->size_idx = units;
}
/*
* heap_recycle_block -- (internal) recycles unused part of the memory block
*/
static void
heap_recycle_block(struct palloc_heap *heap, struct bucket *b,
struct memory_block *m, uint32_t units)
{
if (b->aclass->type == CLASS_RUN) {
ASSERT(units <= UINT16_MAX);
ASSERT(m->block_off + units <= UINT16_MAX);
struct memory_block r = {m->chunk_id, m->zone_id,
m->size_idx - units, (uint16_t)(m->block_off + units),
0, 0, NULL, NULL};
memblock_rebuild_state(heap, &r);
bucket_insert_block(b, &r);
} else {
heap_resize_chunk(heap, m->chunk_id, m->zone_id, units);
}
m->size_idx = units;
}
/*
* heap_resize_chunk -- (internal) splits the chunk into two smaller ones
*/
static void
heap_resize_chunk(struct palloc_heap *heap,
uint32_t chunk_id, uint32_t zone_id, uint32_t new_size_idx)
{
uint32_t new_chunk_id = chunk_id + new_size_idx;
struct zone *z = ZID_TO_ZONE(heap->layout, zone_id);
struct chunk_header *old_hdr = &z->chunk_headers[chunk_id];
struct chunk_header *new_hdr = &z->chunk_headers[new_chunk_id];
uint32_t rem_size_idx = old_hdr->size_idx - new_size_idx;
heap_chunk_init(heap, new_hdr, CHUNK_TYPE_FREE, rem_size_idx);
heap_chunk_init(heap, old_hdr, CHUNK_TYPE_FREE, new_size_idx);
struct bucket *def_bucket = heap->rt->default_bucket;
struct memory_block m = {new_chunk_id, zone_id, rem_size_idx, 0,
0, 0, NULL, NULL};
memblock_rebuild_state(heap, &m);
bucket_insert_block(def_bucket, &m);
}
/*
* heap_resize_chunk -- (internal) splits the chunk into two smaller ones
*/
static void
heap_resize_chunk(PMEMobjpool *pop,
uint32_t chunk_id, uint32_t zone_id, uint32_t new_size_idx)
{
uint32_t new_chunk_id = chunk_id + new_size_idx;
struct zone *z = &pop->heap->layout->zones[zone_id];
struct chunk_header *old_hdr = &z->chunk_headers[chunk_id];
struct chunk_header *new_hdr = &z->chunk_headers[new_chunk_id];
uint32_t rem_size_idx = old_hdr->size_idx - new_size_idx;
heap_chunk_init(pop, new_hdr, CHUNK_TYPE_FREE, rem_size_idx);
heap_chunk_init(pop, old_hdr, CHUNK_TYPE_FREE, new_size_idx);
struct bucket *def_bucket = pop->heap->buckets[DEFAULT_BUCKET];
struct memory_block m = {new_chunk_id, zone_id, rem_size_idx, 0};
if (bucket_insert_block(def_bucket, m) != 0) {
ERR("bucket_insert_block failed during resize");
}
}
/*
* heap_init_free_chunk -- initializes free chunk transient state
*/
static void
heap_init_free_chunk(struct palloc_heap *heap,
struct chunk_header *hdr,
struct memory_block *m)
{
struct operation_context ctx;
operation_init(&ctx, heap->base, NULL, NULL);
ctx.p_ops = &heap->p_ops;
heap_chunk_write_footer(hdr, hdr->size_idx);
/*
* Perform coalescing just in case there
* are any neighbouring free chunks.
*/
struct memory_block nm = heap_coalesce_huge(heap, m);
if (nm.chunk_id != m->chunk_id) {
m->m_ops->prep_hdr(&nm, MEMBLOCK_FREE, &ctx);
operation_process(&ctx);
}
*m = nm;
bucket_insert_block(heap->rt->default_bucket, m);
}
/*
* heap_run_insert -- (internal) inserts and splits a block of memory into a run
*/
static void
heap_run_insert(struct bucket *b, uint32_t chunk_id, uint32_t zone_id,
uint32_t size_idx, uint16_t block_off)
{
ASSERT(size_idx <= BITS_PER_VALUE);
ASSERT(block_off + size_idx <= bucket_bitmap_nallocs(b));
size_t unit_max = bucket_unit_max(b);
struct memory_block m = {chunk_id, zone_id,
unit_max - (block_off % 4), block_off};
if (m.size_idx > size_idx)
m.size_idx = size_idx;
do {
bucket_insert_block(b, m);
m.block_off += m.size_idx;
size_idx -= m.size_idx;
m.size_idx = size_idx > unit_max ? unit_max : size_idx;
} while (size_idx != 0);
}
/*
* heap_recycle_block -- (internal) recycles unused part of the memory block
*/
static void
heap_recycle_block(PMEMobjpool *pop, struct bucket *b, struct memory_block *m,
uint32_t units)
{
#ifdef _EAP_ALLOC_OPTIMIZE
//m->size_idx = units;
//return;
#endif
//printf("bucket_insert_block heap_recycle_block");
if (bucket_is_small(b)) {
struct memory_block r = {m->chunk_id, m->zone_id,
m->size_idx - units, m->block_off + units};
bucket_insert_block(b, r);
} else {
heap_resize_chunk(pop, m->chunk_id, m->zone_id, units);
}
m->size_idx = units;
}
/*
* heap_populate_buckets -- (internal) creates volatile state of memory blocks
*/
static void
heap_populate_buckets(PMEMobjpool *pop)
{
struct pmalloc_heap *h = pop->heap;
if (h->zones_exhausted == h->max_zone)
return;
uint32_t zone_id = h->zones_exhausted++;
struct zone *z = &h->layout->zones[zone_id];
/* ignore zone and chunk headers */
VALGRIND_ADD_TO_GLOBAL_TX_IGNORE(z, sizeof (z->header) +
sizeof (z->chunk_headers));
if (z->header.magic != ZONE_HEADER_MAGIC)
heap_zone_init(pop, zone_id);
struct bucket *def_bucket = h->buckets[DEFAULT_BUCKET];
for (uint32_t i = 0; i < z->header.size_idx; ) {
struct chunk_header *hdr = &z->chunk_headers[i];
heap_chunk_write_footer(hdr, hdr->size_idx);
if (hdr->type == CHUNK_TYPE_RUN) {
struct chunk_run *run =
(struct chunk_run *)&z->chunks[i];
heap_populate_run_bucket(pop,
h->bucket_map[run->block_size], i, zone_id);
} else if (hdr->type == CHUNK_TYPE_FREE) {
struct memory_block m = {i, zone_id, hdr->size_idx, 0};
bucket_insert_block(def_bucket, m);
}
i += hdr->size_idx;
}
}
/*
* pmalloc_construct -- allocates a new block of memory with a constructor
*
* The block offset is written persistently into the off variable, but only
* after the constructor function has been called.
*
* If successful function returns zero. Otherwise an error number is returned.
*/
int
pmalloc_construct(PMEMobjpool *pop, uint64_t *off, size_t size,
void (*constructor)(PMEMobjpool *pop, void *ptr, void *arg),
void *arg, uint64_t data_off)
{
size_t sizeh = size + sizeof (struct allocation_header);
struct bucket *b = heap_get_best_bucket(pop, sizeh);
int err = 0;
uint32_t units = bucket_calc_units(b, sizeh);
struct memory_block m = {0, 0, units, 0};
if ((err = heap_get_bestfit_block(pop, b, &m)) != 0)
return err;
uint64_t op_result = 0;
void *block_data = heap_get_block_data(pop, m);
void *datap = block_data + sizeof (struct allocation_header);
ASSERT((uint64_t)block_data % _POBJ_CL_ALIGNMENT == 0);
uint64_t real_size = bucket_unit_size(b) * units;
alloc_write_header(pop, block_data, m.chunk_id, m.zone_id, real_size);
if (constructor != NULL)
constructor(pop, datap + data_off, arg);
if ((err = heap_lock_if_run(pop, m)) != 0)
return err;
void *hdr = heap_get_block_header(pop, m, HEAP_OP_ALLOC, &op_result);
struct lane_section *lane;
if ((err = lane_hold(pop, &lane, LANE_SECTION_ALLOCATOR)) != 0)
goto err_lane_hold;
struct allocator_lane_section *sec =
(struct allocator_lane_section *)lane->layout;
redo_log_store(pop, sec->redo, ALLOC_OP_REDO_PTR_OFFSET,
pop_offset(pop, off), pop_offset(pop, datap));
redo_log_store_last(pop, sec->redo, ALLOC_OP_REDO_HEADER,
pop_offset(pop, hdr), op_result);
redo_log_process(pop, sec->redo, MAX_ALLOC_OP_REDO);
if (lane_release(pop) != 0) {
ERR("Failed to release the lane");
ASSERT(0);
}
if (heap_unlock_if_run(pop, m) != 0) {
ERR("Failed to release run lock");
ASSERT(0);
}
return 0;
err_lane_hold:
if (heap_unlock_if_run(pop, m) != 0) {
ERR("Failed to release run lock");
ASSERT(0);
}
if (bucket_insert_block(b, m) != 0) {
ERR("Failed to recover heap volatile state");
ASSERT(0);
}
return err;
}
/*
* pfree -- deallocates a memory block previously allocated by pmalloc
*
* A zero value is written persistently into the off variable.
*
* If successful function returns zero. Otherwise an error number is returned.
*/
int
pfree(PMEMobjpool *pop, uint64_t *off, uint64_t data_off)
{
struct allocation_header *alloc = alloc_get_header(pop, *off);
int err = 0;
struct lane_section *lane;
if ((err = lane_hold(pop, &lane, LANE_SECTION_ALLOCATOR)) != 0)
return err;
struct bucket *b = heap_get_chunk_bucket(pop,
alloc->chunk_id, alloc->zone_id);
struct memory_block m = get_mblock_from_alloc(pop, b, alloc);
#ifdef DEBUG
if (!heap_block_is_allocated(pop, m)) {
ERR("Double free or heap corruption");
ASSERT(0);
}
#endif /* DEBUG */
if ((err = heap_lock_if_run(pop, m)) != 0)
goto out;
uint64_t op_result;
void *hdr;
struct memory_block res = heap_free_block(pop, b, m, &hdr, &op_result);
struct allocator_lane_section *sec =
(struct allocator_lane_section *)lane->layout;
redo_log_store(pop, sec->redo, ALLOC_OP_REDO_PTR_OFFSET,
pop_offset(pop, off), 0);
redo_log_store_last(pop, sec->redo, ALLOC_OP_REDO_HEADER,
pop_offset(pop, hdr), op_result);
redo_log_process(pop, sec->redo, MAX_ALLOC_OP_REDO);
if (heap_unlock_if_run(pop, m) != 0) {
ERR("Failed to release run lock");
ASSERT(0);
}
VALGRIND_DO_MEMPOOL_FREE(pop,
(char *)alloc + sizeof (*alloc) + data_off);
bucket_insert_block(pop, b, res);
if (bucket_is_small(b) && heap_degrade_run_if_empty(pop, b, res) != 0) {
ERR("Failed to degrade run");
ASSERT(0);
}
out:
if (lane_release(pop) != 0) {
ERR("Failed to release the lane");
ASSERT(0);
}
return err;
}
/*
* heap_ensure_run_bucket_filled -- (internal) refills the bucket if needed
*/
static int
heap_ensure_run_bucket_filled(struct palloc_heap *heap, struct bucket *b,
uint32_t units)
{
ASSERTeq(b->aclass->type, CLASS_RUN);
if (b->is_active) {
b->c_ops->rm_all(b->container);
b->active_memory_block.m_ops
->claim_revoke(&b->active_memory_block);
b->is_active = 0;
}
struct heap_rt *h = heap->rt;
struct memory_block m = MEMORY_BLOCK_NONE;
if (recycler_get(h->recyclers[b->aclass->id], &m) == 0) {
pthread_mutex_t *lock = m.m_ops->get_lock(&m);
util_mutex_lock(lock);
heap_reuse_run(heap, b, &m);
util_mutex_unlock(lock);
b->active_memory_block = m;
b->is_active = 1;
return 0;
}
m.size_idx = b->aclass->run.size_idx;
/* cannot reuse an existing run, create a new one */
struct bucket *defb = heap_get_default_bucket(heap);
util_mutex_lock(&defb->lock);
if (heap_get_bestfit_block(heap, defb, &m) == 0) {
ASSERTeq(m.block_off, 0);
heap_create_run(heap, b, &m);
b->active_memory_block = m;
b->is_active = 1;
util_mutex_unlock(&defb->lock);
return 0;
}
util_mutex_unlock(&defb->lock);
/*
* Try the recycler again, the previous call to the bestfit_block for
* huge chunks might have reclaimed some unused runs.
*/
if (recycler_get(h->recyclers[b->aclass->id], &m) == 0) {
pthread_mutex_t *lock = m.m_ops->get_lock(&m);
util_mutex_lock(lock);
heap_reuse_run(heap, b, &m);
util_mutex_unlock(lock);
/*
* To verify that the recycler run is not able to satisfy our
* request we attempt to retrieve a block. This is not ideal,
* and should be replaced by a different heuristic once proper
* memory block scoring is implemented.
*/
struct memory_block tmp = MEMORY_BLOCK_NONE;
tmp.size_idx = units;
if (b->c_ops->get_rm_bestfit(b->container, &tmp) != 0) {
b->c_ops->rm_all(b->container);
m.m_ops->claim_revoke(&m);
return ENOMEM;
} else {
bucket_insert_block(b, &tmp);
}
b->active_memory_block = m;
b->is_active = 1;
return 0;
}
return ENOMEM;
}
/*
* pfree -- deallocates a memory block previously allocated by pmalloc
*
* A zero value is written persistently into the off variable.
*
* If successful function returns zero. Otherwise an error number is returned.
*/
int
pfree(PMEMobjpool *pop, uint64_t *off)
{
struct allocation_header *alloc = alloc_get_header(pop, *off);
int err = 0;
struct lane_section *lane;
if ((err = lane_hold(pop, &lane, LANE_SECTION_ALLOCATOR)) != 0)
return err;
struct bucket *b = heap_get_best_bucket(pop, alloc->size);
struct memory_block m = get_mblock_from_alloc(pop, b, alloc);
#ifdef DEBUG
if (!heap_block_is_allocated(pop, m)) {
ERR("Double free or heap corruption");
ASSERT(0);
}
#endif /* DEBUG */
if ((err = heap_lock_if_run(pop, m)) != 0)
goto out;
uint64_t op_result;
void *hdr;
struct memory_block res = heap_free_block(pop, b, m, &hdr, &op_result);
struct allocator_lane_section *sec =
(struct allocator_lane_section *)lane->layout;
redo_log_store(pop, sec->redo, ALLOC_OP_REDO_PTR_OFFSET,
pop_offset(pop, off), 0);
redo_log_store_last(pop, sec->redo, ALLOC_OP_REDO_HEADER,
pop_offset(pop, hdr), op_result);
redo_log_process(pop, sec->redo, MAX_ALLOC_OP_REDO);
/*
* There's no point in rolling back redo log changes because the
* volatile errors don't break the persistent state.
*/
if (bucket_insert_block(b, res) != 0) {
ERR("Failed to update the heap volatile state");
ASSERT(0);
}
if (heap_unlock_if_run(pop, m) != 0) {
ERR("Failed to release run lock");
ASSERT(0);
}
if (bucket_is_small(b) && heap_degrade_run_if_empty(pop, b, res) != 0) {
ERR("Failed to degrade run");
ASSERT(0);
}
out:
if (lane_release(pop) != 0) {
ERR("Failed to release the lane");
ASSERT(0);
}
return err;
}
/*
* heap_chunk_init -- (internal) writes chunk header
*/
static void
heap_chunk_init(struct palloc_heap *heap, struct chunk_header *hdr,
uint16_t type, uint32_t size_idx)
{
struct chunk_header nhdr = {
.type = type,
.flags = 0,
.size_idx = size_idx
};
VALGRIND_DO_MAKE_MEM_UNDEFINED(hdr, sizeof(*hdr));
*hdr = nhdr; /* write the entire header (8 bytes) at once */
pmemops_persist(&heap->p_ops, hdr, sizeof(*hdr));
heap_chunk_write_footer(hdr, size_idx);
}
/*
* heap_zone_init -- (internal) writes zone's first chunk and header
*/
static void
heap_zone_init(struct palloc_heap *heap, uint32_t zone_id)
{
struct zone *z = ZID_TO_ZONE(heap->layout, zone_id);
uint32_t size_idx = get_zone_size_idx(zone_id, heap->rt->max_zone,
heap->size);
heap_chunk_init(heap, &z->chunk_headers[0], CHUNK_TYPE_FREE, size_idx);
struct zone_header nhdr = {
.size_idx = size_idx,
.magic = ZONE_HEADER_MAGIC,
};
z->header = nhdr; /* write the entire header (8 bytes) at once */
pmemops_persist(&heap->p_ops, &z->header, sizeof(z->header));
}
/*
* heap_run_init -- (internal) creates a run based on a chunk
*/
static void
heap_run_init(struct palloc_heap *heap, struct bucket *b,
const struct memory_block *m)
{
struct alloc_class *c = b->aclass;
ASSERTeq(c->type, CLASS_RUN);
struct zone *z = ZID_TO_ZONE(heap->layout, m->zone_id);
struct chunk_run *run = (struct chunk_run *)&z->chunks[m->chunk_id];
ASSERTne(m->size_idx, 0);
size_t runsize = SIZEOF_RUN(run, m->size_idx);
VALGRIND_DO_MAKE_MEM_UNDEFINED(run, runsize);
/* add/remove chunk_run and chunk_header to valgrind transaction */
VALGRIND_ADD_TO_TX(run, runsize);
run->block_size = c->unit_size;
pmemops_persist(&heap->p_ops, &run->block_size,
sizeof(run->block_size));
/* set all the bits */
memset(run->bitmap, 0xFF, sizeof(run->bitmap));
unsigned nval = c->run.bitmap_nval;
ASSERT(nval > 0);
/* clear only the bits available for allocations from this bucket */
memset(run->bitmap, 0, sizeof(uint64_t) * (nval - 1));
run->bitmap[nval - 1] = c->run.bitmap_lastval;
run->incarnation_claim = heap->run_id;
VALGRIND_SET_CLEAN(&run->incarnation_claim,
sizeof(run->incarnation_claim));
VALGRIND_REMOVE_FROM_TX(run, runsize);
pmemops_persist(&heap->p_ops, run->bitmap, sizeof(run->bitmap));
struct chunk_header run_data_hdr;
run_data_hdr.type = CHUNK_TYPE_RUN_DATA;
run_data_hdr.flags = 0;
struct chunk_header *data_hdr;
for (unsigned i = 1; i < m->size_idx; ++i) {
data_hdr = &z->chunk_headers[m->chunk_id + i];
VALGRIND_DO_MAKE_MEM_UNDEFINED(data_hdr, sizeof(*data_hdr));
VALGRIND_ADD_TO_TX(data_hdr, sizeof(*data_hdr));
run_data_hdr.size_idx = i;
*data_hdr = run_data_hdr;
VALGRIND_REMOVE_FROM_TX(data_hdr, sizeof(*data_hdr));
}
pmemops_persist(&heap->p_ops,
&z->chunk_headers[m->chunk_id + 1],
sizeof(struct chunk_header) * (m->size_idx - 1));
struct chunk_header *hdr = &z->chunk_headers[m->chunk_id];
ASSERT(hdr->type == CHUNK_TYPE_FREE);
VALGRIND_ADD_TO_TX(hdr, sizeof(*hdr));
struct chunk_header run_hdr;
run_hdr.size_idx = hdr->size_idx;
run_hdr.type = CHUNK_TYPE_RUN;
run_hdr.flags = header_type_to_flag[c->header_type];
*hdr = run_hdr;
VALGRIND_REMOVE_FROM_TX(hdr, sizeof(*hdr));
pmemops_persist(&heap->p_ops, hdr, sizeof(*hdr));
}
/*
* heap_run_insert -- (internal) inserts and splits a block of memory into a run
*/
static void
heap_run_insert(struct palloc_heap *heap, struct bucket *b,
const struct memory_block *m, uint32_t size_idx, uint16_t block_off)
{
struct alloc_class *c = b->aclass;
ASSERTeq(c->type, CLASS_RUN);
ASSERT(size_idx <= BITS_PER_VALUE);
ASSERT(block_off + size_idx <= c->run.bitmap_nallocs);
uint32_t unit_max = c->run.unit_max;
struct memory_block nm = *m;
nm.size_idx = unit_max - (block_off % unit_max);
nm.block_off = block_off;
if (nm.size_idx > size_idx)
nm.size_idx = size_idx;
do {
bucket_insert_block(b, &nm);
ASSERT(nm.size_idx <= UINT16_MAX);
ASSERT(nm.block_off + nm.size_idx <= UINT16_MAX);
nm.block_off = (uint16_t)(nm.block_off + (uint16_t)nm.size_idx);
size_idx -= nm.size_idx;
nm.size_idx = size_idx > unit_max ? unit_max : size_idx;
} while (size_idx != 0);
}
/*
* heap_process_run_metadata -- (internal) parses the run bitmap
*/
static uint32_t
heap_process_run_metadata(struct palloc_heap *heap, struct bucket *b,
const struct memory_block *m)
{
struct alloc_class *c = b->aclass;
ASSERTeq(c->type, CLASS_RUN);
uint16_t block_off = 0;
uint16_t block_size_idx = 0;
uint32_t inserted_blocks = 0;
struct zone *z = ZID_TO_ZONE(heap->layout, m->zone_id);
struct chunk_run *run = (struct chunk_run *)&z->chunks[m->chunk_id];
for (unsigned i = 0; i < c->run.bitmap_nval; ++i) {
ASSERT(i < MAX_BITMAP_VALUES);
uint64_t v = run->bitmap[i];
ASSERT(BITS_PER_VALUE * i <= UINT16_MAX);
block_off = (uint16_t)(BITS_PER_VALUE * i);
if (v == 0) {
heap_run_insert(heap, b, m, BITS_PER_VALUE, block_off);
inserted_blocks += BITS_PER_VALUE;
continue;
} else if (v == UINT64_MAX) {
continue;
}
for (unsigned j = 0; j < BITS_PER_VALUE; ++j) {
if (BIT_IS_CLR(v, j)) {
block_size_idx++;
} else if (block_size_idx != 0) {
ASSERT(block_off >= block_size_idx);
heap_run_insert(heap, b, m,
block_size_idx,
(uint16_t)(block_off - block_size_idx));
inserted_blocks += block_size_idx;
block_size_idx = 0;
}
if ((block_off++) == c->run.bitmap_nallocs) {
i = MAX_BITMAP_VALUES;
break;
}
}
if (block_size_idx != 0) {
ASSERT(block_off >= block_size_idx);
heap_run_insert(heap, b, m,
block_size_idx,
(uint16_t)(block_off - block_size_idx));
inserted_blocks += block_size_idx;
block_size_idx = 0;
}
}
return inserted_blocks;
}
/*
* heap_create_run -- (internal) initializes a new run on an existing free chunk
*/
static void
heap_create_run(struct palloc_heap *heap, struct bucket *b,
struct memory_block *m)
{
heap_run_init(heap, b, m);
memblock_rebuild_state(heap, m);
heap_process_run_metadata(heap, b, m);
}
/*
* heap_degrade_run_if_empty -- makes a chunk out of an empty run
*/
int
heap_degrade_run_if_empty(PMEMobjpool *pop, struct bucket *b,
struct memory_block m)
{
struct zone *z = &pop->heap->layout->zones[m.zone_id];
struct chunk_header *hdr = &z->chunk_headers[m.chunk_id];
ASSERT(hdr->type == CHUNK_TYPE_RUN);
struct chunk_run *run = (struct chunk_run *)&z->chunks[m.chunk_id];
int err = 0;
if ((err = pthread_mutex_lock(heap_get_run_lock(pop, m))) != 0)
return err;
int i;
for (i = 0; i < bucket_bitmap_nval(b) - 1; ++i)
if (run->bitmap[i] != 0)
goto out;
if (run->bitmap[i] != bucket_bitmap_lastval(b))
goto out;
m.block_off = 0;
m.size_idx = RUN_UNIT_MAX;
uint32_t size_idx_sum = 0;
while (size_idx_sum != bucket_bitmap_nallocs(b)) {
if (bucket_get_rm_block_exact(b, m) != 0) {
ERR("persistent and volatile state mismatched");
ASSERT(0);
}
size_idx_sum += m.size_idx;
m.block_off += RUN_UNIT_MAX;
if (m.block_off + RUN_UNIT_MAX > bucket_bitmap_nallocs(b))
m.size_idx = bucket_bitmap_nallocs(b) - m.block_off;
else
m.size_idx = RUN_UNIT_MAX;
}
struct bucket *defb = pop->heap->buckets[DEFAULT_BUCKET];
if ((err = bucket_lock(defb)) != 0) {
ERR("Failed to lock default bucket");
ASSERT(0);
}
m.block_off = 0;
m.size_idx = 1;
heap_chunk_init(pop, hdr, CHUNK_TYPE_FREE, m.size_idx);
uint64_t *mhdr;
uint64_t op_result;
struct memory_block fm =
heap_free_block(pop, defb, m, &mhdr, &op_result);
VALGRIND_ADD_TO_TX(mhdr, sizeof (*mhdr));
*mhdr = op_result;
VALGRIND_REMOVE_FROM_TX(mhdr, sizeof (*mhdr));
pop->persist(mhdr, sizeof (*mhdr));
if ((err = bucket_insert_block(defb, fm)) != 0) {
ERR("Failed to update heap volatile state");
}
bucket_unlock(defb);
out:
if (pthread_mutex_unlock(heap_get_run_lock(pop, m)) != 0) {
ERR("Failed to release run lock");
ASSERT(0);
}
return err;
}
int
pfree(PMEMobjpool *pop, uint64_t *off)
{
struct allocation_header *alloc = alloc_get_header(pop, *off);
struct bucket *b = heap_get_best_bucket(pop, alloc->size);
int err = 0;
struct memory_block m = get_mblock_from_alloc(pop, b, alloc);
if ((err = heap_lock_if_run(pop, m)) != 0)
return err;
#ifdef _EAP_ALLOC_OPTIMIZE
//fprintf(stderr,"_EAP_ALLOC_OPTIMIZE\n");
if(is_alloc_free_opt_enable(alloc->size))
{
goto error_lane_hold;
//goto temphere;
}else {
//printf("Relaxing allocs %zu\n", alloc->size);
}
#endif
uint64_t op_result;
void *hdr;
struct memory_block res = heap_free_block(pop, b, m, &hdr, &op_result);
struct lane_section *lane;
if ((err = lane_hold(pop, &lane, LANE_SECTION_ALLOCATOR)) != 0)
goto error_lane_hold;
struct allocator_lane_section *sec =
(struct allocator_lane_section *)lane->layout;
redo_log_store(pop, sec->redo, ALLOC_OP_REDO_PTR_OFFSET,
pop_offset(pop, off), 0);
redo_log_store_last(pop, sec->redo, ALLOC_OP_REDO_HEADER,
pop_offset(pop, hdr), op_result);
redo_log_process(pop, sec->redo, MAX_ALLOC_OP_REDO);
if (lane_release(pop) != 0) {
ERR("Failed to release the lane");
ASSERT(0);
}
#ifdef _EAP_ALLOC_OPTIMIZE
goto temphere;
temphere:
// if(is_alloc_free_opt_enable(alloc->size))
// goto error_lane_hold;
#endif
/*
* There's no point in rolling back redo log changes because the
* volatile errors don't break the persistent state.
*/
if (bucket_insert_block(b, res)
!= 0) {
ERR("Failed to update the heap volatile state");
ASSERT(0);
}
if (heap_unlock_if_run(pop, m) != 0) {
ERR("Failed to release run lock");
ASSERT(0);
}
if (bucket_is_small(b) && heap_degrade_run_if_empty(pop, b, res) != 0) {
ERR("Failed to degrade run");
ASSERT(0);
}
return 0;
error_lane_hold:
if (heap_unlock_if_run(pop, m) != 0) {
ERR("Failed to release run lock");
ASSERT(0);
}
return err;
}
/*
* palloc_reservation_create -- creates a volatile reservation of a
* memory block.
*
* The first step in the allocation of a new block is reserving it in
* the transient heap - which is represented by the bucket abstraction.
*
* To provide optimal scaling for multi-threaded applications and reduce
* fragmentation the appropriate bucket is chosen depending on the
* current thread context and to which allocation class the requested
* size falls into.
*
* Once the bucket is selected, just enough memory is reserved for the
* requested size. The underlying block allocation algorithm
* (best-fit, next-fit, ...) varies depending on the bucket container.
*/
static int
palloc_reservation_create(struct palloc_heap *heap, size_t size,
palloc_constr constructor, void *arg,
uint64_t extra_field, uint16_t object_flags, uint16_t class_id,
struct pobj_action_internal *out)
{
int err = 0;
struct memory_block *new_block = &out->m;
ASSERT(class_id < UINT8_MAX);
struct alloc_class *c = class_id == 0 ?
heap_get_best_class(heap, size) :
alloc_class_by_id(heap_alloc_classes(heap),
(uint8_t)class_id);
if (c == NULL) {
ERR("no allocation class for size %lu bytes", size);
errno = EINVAL;
return -1;
}
/*
* The caller provided size in bytes, but buckets operate in
* 'size indexes' which are multiples of the block size in the
* bucket.
*
* For example, to allocate 500 bytes from a bucket that
* provides 256 byte blocks two memory 'units' are required.
*/
ssize_t size_idx = alloc_class_calc_size_idx(c, size);
if (size_idx < 0) {
ERR("allocation class not suitable for size %lu bytes",
size);
errno = EINVAL;
return -1;
}
ASSERT(size_idx <= UINT32_MAX);
new_block->size_idx = (uint32_t)size_idx;
struct bucket *b = heap_bucket_acquire(heap, c);
err = heap_get_bestfit_block(heap, b, new_block);
if (err != 0)
goto out;
if (alloc_prep_block(heap, new_block, constructor, arg,
extra_field, object_flags, &out->offset) != 0) {
/*
* Constructor returned non-zero value which means
* the memory block reservation has to be rolled back.
*/
if (new_block->type == MEMORY_BLOCK_HUGE) {
bucket_insert_block(b, new_block);
}
err = ECANCELED;
goto out;
}
/*
* Each as of yet unfulfilled reservation needs to be tracked in the
* runtime state.
* The memory block cannot be put back into the global state unless
* there are no active reservations.
*/
if ((out->resvp = bucket_current_resvp(b)) != NULL)
util_fetch_and_add64(out->resvp, 1);
out->lock = new_block->m_ops->get_lock(new_block);
out->new_state = MEMBLOCK_ALLOCATED;
out:
heap_bucket_release(heap, b);
if (err == 0)
return 0;
errno = err;
return -1;
}
/*
* heap_reclaim_run -- checks the run for available memory if unclaimed.
*
* Returns 1 if reclaimed chunk, 0 otherwise.
*/
static int
heap_reclaim_run(struct palloc_heap *heap, struct chunk_run *run,
struct memory_block *m)
{
if (m->m_ops->claim(m) != 0)
return 0; /* this run already has an owner */
struct alloc_class *c = alloc_class_get_create_by_unit_size(
heap->rt->alloc_classes, run->block_size);
if (c == NULL)
return 0;
ASSERTeq(c->type, CLASS_RUN);
pthread_mutex_t *lock = m->m_ops->get_lock(m);
util_mutex_lock(lock);
unsigned i;
unsigned nval = c->run.bitmap_nval;
for (i = 0; nval > 0 && i < nval - 1; ++i)
if (run->bitmap[i] != 0)
break;
int empty = (i == (nval - 1)) &&
(run->bitmap[i] == c->run.bitmap_lastval);
if (empty) {
struct zone *z = ZID_TO_ZONE(heap->layout, m->zone_id);
struct chunk_header *hdr = &z->chunk_headers[m->chunk_id];
struct bucket *defb = heap_get_default_bucket(heap);
/*
* The redo log ptr can be NULL if we are sure that there's only
* one persistent value modification in the entire operation
* context.
*/
struct operation_context ctx;
operation_init(&ctx, heap->base, NULL, NULL);
ctx.p_ops = &heap->p_ops;
struct memory_block nb = MEMORY_BLOCK_NONE;
nb.chunk_id = m->chunk_id;
nb.zone_id = m->zone_id;
nb.block_off = 0;
nb.size_idx = m->size_idx;
heap_chunk_init(heap, hdr, CHUNK_TYPE_FREE, nb.size_idx);
memblock_rebuild_state(heap, &nb);
nb = heap_coalesce_huge(heap, &nb);
nb.m_ops->prep_hdr(&nb, MEMBLOCK_FREE, &ctx);
operation_process(&ctx);
bucket_insert_block(defb, &nb);
*m = nb;
} else {
recycler_put(heap->rt->recyclers[c->id], m);
}
util_mutex_unlock(lock);
return empty;
}
