/*
* persist_alloc -- (internal) performs a persistent allocation of the
* memory block previously reserved by volatile bucket
*/
static int
persist_alloc(PMEMobjpool *pop, struct lane_section *lane,
struct memory_block m, uint64_t real_size, uint64_t *off,
void (*constructor)(PMEMobjpool *pop, void *ptr, void *arg),
void *arg, uint64_t data_off)
{
int err;
#ifdef DEBUG
if (heap_block_is_allocated(pop, m)) {
ERR("heap corruption");
ASSERT(0);
}
#endif /* DEBUG */
uint64_t op_result = 0;
void *block_data = heap_get_block_data(pop, m);
void *datap = (char *)block_data + sizeof (struct allocation_header);
void *userdatap = (char *)datap + data_off;
ASSERT((uint64_t)block_data % _POBJ_CL_ALIGNMENT == 0);
/* mark everything (including headers) as accessible */
VALGRIND_DO_MAKE_MEM_UNDEFINED(pop, block_data, real_size);
/* mark space as allocated */
VALGRIND_DO_MEMPOOL_ALLOC(pop, userdatap,
real_size -
sizeof (struct allocation_header) - data_off);
alloc_write_header(pop, block_data, m.chunk_id, m.zone_id, real_size);
if (constructor != NULL)
constructor(pop, userdatap, arg);
if ((err = heap_lock_if_run(pop, m)) != 0) {
VALGRIND_DO_MEMPOOL_FREE(pop, userdatap);
return err;
}
void *hdr = heap_get_block_header(pop, m, HEAP_OP_ALLOC, &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), 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 (heap_unlock_if_run(pop, m) != 0) {
ERR("Failed to release run lock");
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.
*/
void
pfree(PMEMobjpool *pop, uint64_t *off, uint64_t data_off)
{
struct allocation_header *alloc = alloc_get_header(pop, *off);
struct lane_section *lane;
lane_hold(pop, &lane, LANE_SECTION_ALLOCATOR);
struct bucket *b = heap_get_chunk_bucket(pop,
alloc->chunk_id, alloc->zone_id);
struct memory_block m = get_mblock_from_alloc(pop, alloc);
#ifdef DEBUG
if (!heap_block_is_allocated(pop, m)) {
ERR("Double free or heap corruption");
ASSERT(0);
}
#endif /* DEBUG */
heap_lock_if_run(pop, m);
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);
heap_unlock_if_run(pop, m);
VALGRIND_DO_MEMPOOL_FREE(pop,
(char *)alloc + sizeof (*alloc) + data_off);
/* we might have been operating on inactive run */
if (b != NULL) {
CNT_OP(b, insert, pop, res);
if (b->type == BUCKET_RUN)
heap_degrade_run_if_empty(pop, b, res);
}
lane_release(pop);
}
/*
* 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;
}
/*
* 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;
}
/*
* 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};
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);
/*
* 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);
existing_block = get_mblock_from_alloc(heap, alloc);
}
/* 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)
return 0;
errno = alloc_reserve_block(heap, &new_block, sizeh);
if (errno != 0)
return -1;
}
/*
* 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)) {
#ifdef DEBUG
if (!heap_block_is_allocated(heap, existing_block)) {
ERR("Double free or heap corruption");
ASSERT(0);
}
#endif /* DEBUG */
/*
* This lock must be held until the operation is processed
* successfully, because other threads might operate on the
* same bitmap value.
*/
MEMBLOCK_OPS(AUTO, &existing_block)->
lock(&existing_block, heap);
/*
* 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)) {
#ifdef DEBUG
if (heap_block_is_allocated(heap, new_block)) {
ERR("heap corruption");
ASSERT(0);
}
#endif /* DEBUG */
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;
return -1;
}
/*
* 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.
*/
MEMBLOCK_OPS(AUTO, &new_block)->lock(&new_block, heap);
/*
* 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, HDR_OP_ALLOC, 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(new_block)) {
/* new block run lock */
MEMBLOCK_OPS(AUTO, &new_block)->unlock(&new_block, heap);
}
if (!MEMORY_BLOCK_IS_EMPTY(existing_block)) {
/* existing (freed) run lock */
MEMBLOCK_OPS(AUTO, &existing_block)->
unlock(&existing_block, heap);
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);
#ifdef DEBUG
if (heap_block_is_allocated(heap, reclaimed_block)) {
ERR("heap corruption");
ASSERT(0);
}
#endif /* DEBUG */
/*
* 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);
}
}
return 0;
}