/*
* palloc_exec_actions -- perform the provided free/alloc operations
*/
static void
palloc_exec_actions(struct palloc_heap *heap,
struct operation_context *ctx,
struct pobj_action_internal *actv,
int actvcnt)
{
/*
* The operations array is sorted so that proper lock ordering is
* ensured.
*/
qsort(actv, (size_t)actvcnt, sizeof(struct pobj_action_internal),
palloc_action_compare);
struct pobj_action_internal *act;
for (int i = 0; i < actvcnt; ++i) {
act = &actv[i];
/*
* This lock must be held for the duration between the creation
* of the allocation metadata updates in the operation context
* and the operation processing. This is because a different
* thread might operate on the same 8-byte value of the run
* bitmap and override allocation performed by this thread.
*/
if (i == 0 || act->lock != actv[i - 1].lock) {
if (act->lock)
util_mutex_lock(act->lock);
}
action_funcs[act->type].exec(heap, act, ctx);
}
/* wait for all the headers to be persistent */
pmemops_drain(&heap->p_ops);
operation_process(ctx);
for (int i = 0; i < actvcnt; ++i) {
act = &actv[i];
action_funcs[act->type].on_process(heap, act);
if (i == 0 || act->lock != actv[i - 1].lock) {
if (act->lock)
util_mutex_unlock(act->lock);
}
}
for (int i = 0; i < actvcnt; ++i) {
act = &actv[i];
action_funcs[act->type].on_unlock(heap, act);
}
}
/*
* 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);
}
static void
test_heap()
{
struct mock_pop *mpop = Malloc(MOCK_POOL_SIZE);
PMEMobjpool *pop = &mpop->p;
memset(pop, 0, MOCK_POOL_SIZE);
pop->size = MOCK_POOL_SIZE;
pop->heap_size = MOCK_POOL_SIZE - sizeof(PMEMobjpool);
pop->heap_offset = (uint64_t)((uint64_t)&mpop->heap - (uint64_t)mpop);
pop->persist = obj_heap_persist;
UT_ASSERT(heap_check(pop) != 0);
UT_ASSERT(heap_init(pop) == 0);
UT_ASSERT(heap_boot(pop) == 0);
UT_ASSERT(pop->heap != NULL);
struct bucket *b_small = heap_get_best_bucket(pop, 1);
struct bucket *b_big = heap_get_best_bucket(pop, 2048);
UT_ASSERT(b_small->unit_size < b_big->unit_size);
struct bucket *b_def = heap_get_best_bucket(pop, CHUNKSIZE);
UT_ASSERT(b_def->unit_size == CHUNKSIZE);
/* new small buckets should be empty */
UT_ASSERT(b_small->type == BUCKET_RUN);
UT_ASSERT(b_big->type == BUCKET_RUN);
struct memory_block blocks[MAX_BLOCKS] = {
{0, 0, 1, 0},
{0, 0, 1, 0},
{0, 0, 1, 0}
};
for (int i = 0; i < MAX_BLOCKS; ++i) {
heap_get_bestfit_block(pop, b_def, &blocks[i]);
UT_ASSERT(blocks[i].block_off == 0);
}
struct memory_block *blocksp[MAX_BLOCKS] = {NULL};
struct memory_block prev;
heap_get_adjacent_free_block(pop, b_def, &prev, blocks[1], 1);
UT_ASSERT(prev.chunk_id == blocks[0].chunk_id);
blocksp[0] = &prev;
struct memory_block cnt;
heap_get_adjacent_free_block(pop, b_def, &cnt, blocks[0], 0);
UT_ASSERT(cnt.chunk_id == blocks[1].chunk_id);
blocksp[1] = &cnt;
struct memory_block next;
heap_get_adjacent_free_block(pop, b_def, &next, blocks[1], 0);
UT_ASSERT(next.chunk_id == blocks[2].chunk_id);
blocksp[2] = &next;
struct operation_context *ctx = operation_init(pop, NULL);
struct memory_block result =
heap_coalesce(pop, blocksp, MAX_BLOCKS, HEAP_OP_FREE, ctx);
operation_process(ctx);
operation_delete(ctx);
UT_ASSERT(result.size_idx == 3);
UT_ASSERT(result.chunk_id == prev.chunk_id);
UT_ASSERT(heap_check(pop) == 0);
heap_cleanup(pop);
UT_ASSERT(pop->heap == NULL);
Free(mpop);
}
/*
* 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;
}
/*
* palloc_operation -- persistent memory operation. Takes a NULL pointer
* or an existing memory block and modifies it to occupy, at least, 'size'
* number of bytes.
*
* The malloc, free and realloc routines are implemented in the context of this
* common operation which encompasses all of the functionality usually done
* separately in those methods.
*
* The first thing that needs to be done is determining which memory blocks
* will be affected by the operation - this varies depending on the whether the
* operation will need to modify or free an existing block and/or allocate
* a new one.
*
* Simplified allocation process flow is as follows:
* - reserve a new block in the transient heap
* - prepare the new block
* - create redo log of required modifications
* - chunk metadata
* - offset of the new object
* - commit and process the redo log
*
* And similarly, the deallocation process:
* - create redo log of required modifications
* - reverse the chunk metadata back to the 'free' state
* - set the destination of the object offset to zero
* - commit and process the redo log
* - return the memory block back to the free blocks transient heap
*
* Reallocation is a combination of the above, which one additional step
* of copying the old content in the meantime.
*/
int
palloc_operation(struct palloc_heap *heap,
uint64_t off, uint64_t *dest_off, size_t size,
palloc_constr constructor, void *arg,
struct operation_context *ctx)
{
struct bucket *b = NULL;
struct allocation_header *alloc = NULL;
struct memory_block existing_block = {0, 0, 0, 0};
struct memory_block new_block = {0, 0, 0, 0};
struct memory_block reclaimed_block = {0, 0, 0, 0};
int ret = 0;
/*
* These two lock are responsible for protecting the metadata for the
* persistent representation of a chunk. Depending on the operation and
* the type of a chunk, they might be NULL.
*/
pthread_mutex_t *existing_block_lock = NULL;
pthread_mutex_t *new_block_lock = NULL;
size_t sizeh = size + sizeof(struct allocation_header);
/*
* The offset of an existing block can be nonzero which means this
* operation is either free or a realloc - either way the offset of the
* object needs to be translated into structure that all of the heap
* methods operate in.
*/
if (off != 0) {
alloc = ALLOC_GET_HEADER(heap, off);
existing_block = get_mblock_from_alloc(heap, alloc);
/*
* This lock must be held until the operation is processed
* successfully, because other threads might operate on the
* same bitmap value.
*/
existing_block_lock = MEMBLOCK_OPS(AUTO, &existing_block)->
get_lock(&existing_block, heap);
if (existing_block_lock != NULL)
util_mutex_lock(existing_block_lock);
#ifdef DEBUG
if (MEMBLOCK_OPS(AUTO,
&existing_block)->get_state(&existing_block, heap) !=
MEMBLOCK_ALLOCATED) {
ERR("Double free or heap corruption");
ASSERT(0);
}
#endif /* DEBUG */
/*
* The memory block must return back to the originating bucket,
* otherwise coalescing of neighbouring blocks will be rendered
* impossible.
*
* If the block was allocated in a different incarnation of the
* heap (i.e. the application was restarted) and the chunk from
* which the allocation comes from was not yet processed, the
* originating bucket does not exists and all of the otherwise
* necessary volatile heap modifications won't be performed for
* this memory block.
*/
b = heap_get_chunk_bucket(heap, alloc->chunk_id,
alloc->zone_id);
}
/* if allocation or reallocation, reserve new memory */
if (size != 0) {
/* reallocation to exactly the same size, which is a no-op */
if (alloc != NULL && alloc->size == sizeh)
goto out;
errno = alloc_reserve_block(heap, &new_block, sizeh);
if (errno != 0) {
ret = -1;
goto out;
}
}
/*
* The offset value which is to be written to the destination pointer
* provided by the caller.
*/
uint64_t offset_value = 0;
/* lock and persistently free the existing memory block */
if (!MEMORY_BLOCK_IS_EMPTY(existing_block)) {
/*
* This method will insert new entries into the operation
* context which will, after processing, update the chunk
* metadata to 'free' - it also takes care of all the necessary
* coalescing of blocks.
* Even though the transient state of the heap is used during
* this method to locate neighbouring blocks, it isn't modified.
*
* The rb block is the coalesced memory block that the free
* resulted in, to prevent volatile memory leak it needs to be
* inserted into the corresponding bucket.
*/
reclaimed_block = heap_free_block(heap, b, existing_block, ctx);
offset_value = 0;
}
if (!MEMORY_BLOCK_IS_EMPTY(new_block)) {
if (alloc_prep_block(heap, new_block, constructor,
arg, &offset_value) != 0) {
/*
* Constructor returned non-zero value which means
* the memory block reservation has to be rolled back.
*/
struct bucket *new_bucket = heap_get_chunk_bucket(heap,
new_block.chunk_id, new_block.zone_id);
ASSERTne(new_bucket, NULL);
/*
* Omitting the context in this method results in
* coalescing of blocks without affecting the persistent
* heap state.
*/
new_block = heap_free_block(heap, new_bucket,
new_block, NULL);
CNT_OP(new_bucket, insert, heap, new_block);
if (new_bucket->type == BUCKET_RUN)
heap_degrade_run_if_empty(heap,
new_bucket, new_block);
errno = ECANCELED;
ret = -1;
goto out;
}
/*
* This lock must be held for the duration between the creation
* of the allocation metadata updates in the operation context
* and the operation processing. This is because a different
* thread might operate on the same 8-byte value of the run
* bitmap and override allocation performed by this thread.
*/
new_block_lock = MEMBLOCK_OPS(AUTO, &new_block)->
get_lock(&new_block, heap);
/* the locks might be identical in the case of realloc */
if (new_block_lock == existing_block_lock)
new_block_lock = NULL;
if (new_block_lock != NULL)
util_mutex_lock(new_block_lock);
#ifdef DEBUG
if (MEMBLOCK_OPS(AUTO,
&new_block)->get_state(&new_block, heap) !=
MEMBLOCK_FREE) {
ERR("Double free or heap corruption");
ASSERT(0);
}
#endif /* DEBUG */
/*
* The actual required metadata modifications are chunk-type
* dependent, but it always is a modification of a single 8 byte
* value - either modification of few bits in a bitmap or
* changing a chunk type from free to used.
*/
MEMBLOCK_OPS(AUTO, &new_block)->prep_hdr(&new_block,
heap, MEMBLOCK_ALLOCATED, ctx);
}
/* not in-place realloc */
if (!MEMORY_BLOCK_IS_EMPTY(existing_block) &&
!MEMORY_BLOCK_IS_EMPTY(new_block)) {
size_t old_size = alloc->size;
size_t to_cpy = old_size > sizeh ? sizeh : old_size;
VALGRIND_ADD_TO_TX(PMALLOC_OFF_TO_PTR(heap, offset_value),
to_cpy - ALLOC_OFF);
pmemops_memcpy_persist(&heap->p_ops,
PMALLOC_OFF_TO_PTR(heap, offset_value),
PMALLOC_OFF_TO_PTR(heap, off),
to_cpy - ALLOC_OFF);
VALGRIND_REMOVE_FROM_TX(PMALLOC_OFF_TO_PTR(heap, offset_value),
to_cpy - ALLOC_OFF);
}
/*
* If the caller provided a destination value to update, it needs to be
* modified atomically alongside the heap metadata, and so the operation
* context must be used.
* The actual offset value depends on whether the operation type.
*/
if (dest_off != NULL)
operation_add_entry(ctx, dest_off, offset_value, OPERATION_SET);
operation_process(ctx);
/*
* After the operation succeeded, the persistent state is all in order
* but in some cases it might not be in-sync with the its transient
* representation.
*/
if (!MEMORY_BLOCK_IS_EMPTY(existing_block)) {
VALGRIND_DO_MEMPOOL_FREE(heap->layout,
(char *)heap_get_block_data(heap, existing_block)
+ ALLOC_OFF);
/* we might have been operating on inactive run */
if (b != NULL) {
/*
* Even though the initial condition is to check
* whether the existing block exists it's important to
* use the 'reclaimed block' - it is the coalesced one
* and reflects the current persistent heap state,
* whereas the existing block reflects the state from
* before this operation started.
*/
CNT_OP(b, insert, heap, reclaimed_block);
/*
* Degrading of a run means turning it back into a chunk
* in case it's no longer needed.
* It might be tempting to defer this operation until
* such time that the chunk is actually needed, but
* right now the decision is to keep the persistent heap
* state as clean as possible - and that means not
* leaving unused data around.
*/
if (b->type == BUCKET_RUN)
heap_degrade_run_if_empty(heap, b,
reclaimed_block);
}
}
out:
if (new_block_lock != NULL)
util_mutex_unlock(new_block_lock);
if (existing_block_lock != NULL)
util_mutex_unlock(existing_block_lock);
return ret;
}
