/*
* operation_process -- processes registered operations
*
* The order of processing is important: persistent, transient.
* This is because the transient entries that reside on persistent memory might
* require write to a location that is currently occupied by a valid persistent
* state but becomes a transient state after operation is processed.
*/
void
operation_process(struct operation_context *ctx)
{
struct operation_entry *e;
/*
* If there's exactly one persistent entry there's no need to involve
* the redo log. We can simply assign the value, the operation will be
* atomic.
*/
if (ctx->nentries[ENTRY_PERSISTENT] == 1) {
e = &ctx->entries[ENTRY_PERSISTENT][0];
VALGRIND_ADD_TO_TX(e->ptr, sizeof(uint64_t));
PM_EQU((*e->ptr), (e->value));
pmemobj_persist(ctx->pop, e->ptr, sizeof(uint64_t));
VALGRIND_REMOVE_FROM_TX(e->ptr, sizeof(uint64_t));
} else if (ctx->nentries[ENTRY_PERSISTENT] != 0) {
operation_process_persistent_redo(ctx);
}
for (size_t i = 0; i < ctx->nentries[ENTRY_TRANSIENT]; ++i) {
e = &ctx->entries[ENTRY_TRANSIENT][i];
PM_EQU((*e->ptr), (e->value));
/*
* Just in case that the entry was transient but in reality
* the variable is on persistent memory. This is true for
* chunk footers.
*/
VALGRIND_SET_CLEAN(e->ptr, sizeof(e->value));
}
}
/*
* alloc_prep_block -- (internal) prepares a memory block for allocation
*
* Once the block is fully reserved and it's guaranteed that no one else will
* be able to write to this memory region it is safe to write the allocation
* header and call the object construction function.
*
* Because the memory block at this stage is only reserved in transient state
* there's no need to worry about fail-safety of this method because in case
* of a crash the memory will be back in the free blocks collection.
*/
static int
alloc_prep_block(struct palloc_heap *heap, struct memory_block m,
palloc_constr constructor, void *arg, uint64_t *offset_value)
{
void *block_data = heap_get_block_data(heap, m);
void *userdatap = (char *)block_data + ALLOC_OFF;
uint64_t unit_size = MEMBLOCK_OPS(AUTO, &m)->
block_size(&m, heap->layout);
uint64_t real_size = unit_size * m.size_idx;
ASSERT((uint64_t)block_data % ALLOC_BLOCK_SIZE == 0);
ASSERT((uint64_t)userdatap % ALLOC_BLOCK_SIZE == 0);
/* mark everything (including headers) as accessible */
VALGRIND_DO_MAKE_MEM_UNDEFINED(block_data, real_size);
/* mark space as allocated */
VALGRIND_DO_MEMPOOL_ALLOC(heap->layout, userdatap,
real_size - ALLOC_OFF);
alloc_write_header(heap, block_data, m, real_size);
int ret;
if (constructor != NULL &&
(ret = constructor(heap->base, userdatap,
real_size - ALLOC_OFF, arg)) != 0) {
/*
* If canceled, revert the block back to the free state in vg
* machinery. Because the free operation is only performed on
* the user data, the allocation header is made inaccessible
* in a separate call.
*/
VALGRIND_DO_MEMPOOL_FREE(heap->layout, userdatap);
VALGRIND_DO_MAKE_MEM_NOACCESS(block_data, ALLOC_OFF);
/*
* During this method there are several stores to pmem that are
* not immediately flushed and in case of a cancellation those
* stores are no longer relevant anyway.
*/
VALGRIND_SET_CLEAN(block_data, ALLOC_OFF);
return ret;
}
/* flushes both the alloc and oob headers */
pmemops_persist(&heap->p_ops, block_data, ALLOC_OFF);
/*
* To avoid determining the user data pointer twice this method is also
* responsible for calculating the offset of the object in the pool that
* will be used to set the offset destination pointer provided by the
* caller.
*/
*offset_value = PMALLOC_PTR_TO_OFF(heap, userdatap);
return 0;
}
/*
* block_invalidate -- invalidates allocation data and header
*/
static void
block_invalidate(const struct memory_block *m)
{
void *data = m->m_ops->get_user_data(m);
size_t size = m->m_ops->get_user_size(m);
VALGRIND_SET_CLEAN(data, size);
memblock_header_ops[m->header_type].invalidate(m);
}
/*
* huge_prep_operation_hdr -- prepares the new value of a chunk header that will
* be set after the operation concludes.
*/
static void
huge_prep_operation_hdr(const struct memory_block *m, enum memblock_state op,
struct operation_context *ctx)
{
struct chunk_header *hdr = heap_get_chunk_hdr(m->heap, m);
/*
* Depending on the operation that needs to be performed a new chunk
* header needs to be prepared with the new chunk state.
*/
uint64_t val = chunk_get_chunk_hdr_value(
op == MEMBLOCK_ALLOCATED ? CHUNK_TYPE_USED : CHUNK_TYPE_FREE,
hdr->flags,
m->size_idx);
if (ctx == NULL) {
util_atomic_store_explicit64((uint64_t *)hdr, val,
memory_order_relaxed);
pmemops_persist(&m->heap->p_ops, hdr, sizeof(*hdr));
} else {
operation_add_entry(ctx, hdr, val, ULOG_OPERATION_SET);
}
VALGRIND_DO_MAKE_MEM_NOACCESS(hdr + 1,
(hdr->size_idx - 1) * sizeof(struct chunk_header));
/*
* In the case of chunks larger than one unit the footer must be
* created immediately AFTER the persistent state is safely updated.
*/
if (m->size_idx == 1)
return;
struct chunk_header *footer = hdr + m->size_idx - 1;
VALGRIND_DO_MAKE_MEM_UNDEFINED(footer, sizeof(*footer));
val = chunk_get_chunk_hdr_value(CHUNK_TYPE_FOOTER, 0, m->size_idx);
/*
* It's only safe to write the footer AFTER the persistent part of
* the operation have been successfully processed because the footer
* pointer might point to a currently valid persistent state
* of a different chunk.
* The footer entry change is updated as transient because it will
* be recreated at heap boot regardless - it's just needed for runtime
* operations.
*/
if (ctx == NULL) {
util_atomic_store_explicit64((uint64_t *)footer, val,
memory_order_relaxed);
VALGRIND_SET_CLEAN(footer, sizeof(*footer));
} else {
operation_add_typed_entry(ctx,
footer, val, ULOG_OPERATION_SET, LOG_TRANSIENT);
}
}
/*
* heap_chunk_write_footer -- (internal) writes a chunk footer
*/
static void
heap_chunk_write_footer(struct chunk_header *hdr, uint32_t size_idx)
{
if (size_idx == 1) /* that would overwrite the header */
return;
struct chunk_header f = *hdr;
f.type = CHUNK_TYPE_FOOTER;
f.size_idx = size_idx;
*(hdr + size_idx - 1) = f;
/* no need to persist, footers are recreated in heap_populate_buckets */
VALGRIND_SET_CLEAN(hdr + size_idx - 1, sizeof (f));
}
/*
* container_ravl_insert_block -- (internal) inserts a new memory block
* into the container
*/
static int
container_ravl_insert_block(struct block_container *bc,
const struct memory_block *m)
{
struct block_container_ravl *c =
(struct block_container_ravl *)bc;
struct memory_block *e = m->m_ops->get_user_data(m);
VALGRIND_DO_MAKE_MEM_DEFINED(e, sizeof(*e));
VALGRIND_ADD_TO_TX(e, sizeof(*e));
*e = *m;
VALGRIND_SET_CLEAN(e, sizeof(*e));
VALGRIND_REMOVE_FROM_TX(e, sizeof(*e));
return ravl_insert(c->tree, e);
}
/*
* 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);
}
/*
* memblock_header_compact_invalidate --
* (internal) invalidates a compact header
*/
static void
memblock_header_compact_invalidate(const struct memory_block *m)
{
struct allocation_header_compact *hdr = m->m_ops->get_real_data(m);
VALGRIND_SET_CLEAN(hdr, sizeof(*hdr));
}
