ndb - bug#21754 ndb api events: fix sizes for blobs + remove expensive loop

parent 454d5454
......@@ -65,6 +65,39 @@ print_std(const SubTableData * sdata, LinearSectionPtr ptr[3])
}
#endif
// EventBufData
Uint32
EventBufData::get_blob_part_no() const
{
assert(ptr[0].sz > 2);
Uint32 pos = AttributeHeader(ptr[0].p[0]).getDataSize() +
AttributeHeader(ptr[0].p[1]).getDataSize();
Uint32 no = ptr[1].p[pos];
return no;
}
void
EventBufData::add_part_size(Uint32 & full_count, Uint32 & full_sz) const
{
Uint32 tmp_count = 0;
Uint32 tmp_sz = 0;
const EventBufData* data2 = m_next_blob;
while (data2 != 0) {
tmp_count++;
tmp_sz += data2->sz;
const EventBufData* data3 = data2->m_next;
while (data3 != 0) {
tmp_count++;
tmp_sz += data3->sz;
data3 = data3->m_next;
}
data2 = data2->m_next_blob;
}
full_count += tmp_count;
full_sz += tmp_sz;
}
/*
* Class NdbEventOperationImpl
*
......@@ -1162,11 +1195,12 @@ NdbEventBuffer::nextEvent()
// set NdbEventOperation data
op->m_data_item= data;
// remove item from m_available_data
m_available_data.remove_first();
// remove item from m_available_data and return size
Uint32 full_count, full_sz;
m_available_data.remove_first(full_count, full_sz);
// add it to used list
m_used_data.append_used_data(data);
m_used_data.append_used_data(data, full_count, full_sz);
#ifdef VM_TRACE
op->m_data_done_count++;
......@@ -1840,7 +1874,7 @@ NdbEventBuffer::insertDataL(NdbEventOperationImpl *op,
op->m_has_error = 2;
DBUG_RETURN_EVENT(-1);
}
if (unlikely(copy_data(sdata, ptr, data)))
if (unlikely(copy_data(sdata, ptr, data, NULL)))
{
op->m_has_error = 3;
DBUG_RETURN_EVENT(-1);
......@@ -1872,7 +1906,7 @@ NdbEventBuffer::insertDataL(NdbEventOperationImpl *op,
}
}
// link blob event under main event
add_blob_data(main_data, data);
add_blob_data(bucket, main_data, data);
}
if (use_hash)
{
......@@ -1886,7 +1920,7 @@ NdbEventBuffer::insertDataL(NdbEventOperationImpl *op,
else
{
// event with same op, PK found, merge into old buffer
if (unlikely(merge_data(sdata, ptr, data)))
if (unlikely(merge_data(sdata, ptr, data, & bucket->m_data.m_sz)))
{
op->m_has_error = 3;
DBUG_RETURN_EVENT(-1);
......@@ -1909,6 +1943,9 @@ NdbEventBuffer::insertDataL(NdbEventOperationImpl *op,
}
}
}
#ifdef NDB_EVENT_VERIFY_SIZE
verify_size(bucket->m_data);
#endif
DBUG_RETURN_EVENT(0);
}
......@@ -1962,8 +1999,21 @@ NdbEventBuffer::alloc_data()
}
// remove data from free list
m_free_data = data->m_next;
if (data->m_next_blob == 0)
m_free_data = data->m_next;
else {
EventBufData* data2 = data->m_next_blob;
if (data2->m_next == 0) {
data->m_next_blob = data2->m_next_blob;
data = data2;
} else {
EventBufData* data3 = data2->m_next;
data2->m_next = data3->m_next;
data = data3;
}
}
data->m_next = 0;
data->m_next_blob = 0;
#ifdef VM_TRACE
m_free_data_count--;
assert(m_free_data_sz >= data->sz);
......@@ -1975,7 +2025,9 @@ NdbEventBuffer::alloc_data()
// allocate initial or bigger memory area in EventBufData
// takes sizes from given ptr and sets up data->ptr
int
NdbEventBuffer::alloc_mem(EventBufData* data, LinearSectionPtr ptr[3])
NdbEventBuffer::alloc_mem(EventBufData* data,
LinearSectionPtr ptr[3],
Uint32 * change_sz)
{
DBUG_ENTER("NdbEventBuffer::alloc_mem");
DBUG_PRINT("info", ("ptr sz %u + %u + %u", ptr[0].sz, ptr[1].sz, ptr[2].sz));
......@@ -1988,6 +2040,8 @@ NdbEventBuffer::alloc_mem(EventBufData* data, LinearSectionPtr ptr[3])
if (data->sz < alloc_size)
{
Uint32 add_sz = alloc_size - data->sz;
NdbMem_Free((char*)data->memory);
assert(m_total_alloc >= data->sz);
m_total_alloc -= data->sz;
......@@ -1999,6 +2053,9 @@ NdbEventBuffer::alloc_mem(EventBufData* data, LinearSectionPtr ptr[3])
DBUG_RETURN(-1);
data->sz = alloc_size;
m_total_alloc += data->sz;
if (change_sz != NULL)
*change_sz += add_sz;
}
Uint32* memptr = data->memory;
......@@ -2014,14 +2071,30 @@ NdbEventBuffer::alloc_mem(EventBufData* data, LinearSectionPtr ptr[3])
DBUG_RETURN(0);
}
void
NdbEventBuffer::dealloc_mem(EventBufData* data,
Uint32 * change_sz)
{
NdbMem_Free((char*)data->memory);
assert(m_total_alloc >= data->sz);
m_total_alloc -= data->sz;
if (change_sz != NULL) {
assert(*change_sz >= data->sz);
*change_sz -= data->sz;
}
data->memory = 0;
data->sz = 0;
}
int
NdbEventBuffer::copy_data(const SubTableData * const sdata,
LinearSectionPtr ptr[3],
EventBufData* data)
EventBufData* data,
Uint32 * change_sz)
{
DBUG_ENTER_EVENT("NdbEventBuffer::copy_data");
if (alloc_mem(data, ptr) != 0)
if (alloc_mem(data, ptr, change_sz) != 0)
DBUG_RETURN_EVENT(-1);
memcpy(data->sdata, sdata, sizeof(SubTableData));
int i;
......@@ -2093,7 +2166,8 @@ copy_attr(AttributeHeader ah,
int
NdbEventBuffer::merge_data(const SubTableData * const sdata,
LinearSectionPtr ptr2[3],
EventBufData* data)
EventBufData* data,
Uint32 * change_sz)
{
DBUG_ENTER_EVENT("NdbEventBuffer::merge_data");
......@@ -2102,7 +2176,7 @@ NdbEventBuffer::merge_data(const SubTableData * const sdata,
int t1 = SubTableData::getOperation(data->sdata->requestInfo);
int t2 = SubTableData::getOperation(sdata->requestInfo);
if (t1 == Ev_t::enum_NUL)
DBUG_RETURN_EVENT(copy_data(sdata, ptr2, data));
DBUG_RETURN_EVENT(copy_data(sdata, ptr2, data, change_sz));
Ev_t* tp = 0;
int i;
......@@ -2142,6 +2216,8 @@ NdbEventBuffer::merge_data(const SubTableData * const sdata,
DBUG_RETURN_EVENT(0);
}
// TODO: use old data items, avoid malloc/free on each merge
// save old data
EventBufData olddata = *data;
data->memory = 0;
......@@ -2158,7 +2234,7 @@ NdbEventBuffer::merge_data(const SubTableData * const sdata,
{
if (loop == 1)
{
if (alloc_mem(data, ptr) != 0)
if (alloc_mem(data, ptr, change_sz) != 0)
{
result = -1;
goto end;
......@@ -2277,11 +2353,7 @@ NdbEventBuffer::merge_data(const SubTableData * const sdata,
}
end:
// free old data
NdbMem_Free((char*)olddata.memory);
assert(m_total_alloc >= olddata.sz);
m_total_alloc -= olddata.sz;
dealloc_mem(&olddata, change_sz);
DBUG_RETURN_EVENT(result);
}
......@@ -2357,7 +2429,7 @@ NdbEventBuffer::get_main_data(Gci_container* bucket,
SubTableData sdata = *blob_data->sdata;
sdata.tableId = main_op->m_eventImpl->m_tableImpl->m_id;
SubTableData::setOperation(sdata.requestInfo, NdbDictionary::Event::_TE_NUL);
if (copy_data(&sdata, ptr, main_data) != 0)
if (copy_data(&sdata, ptr, main_data, NULL) != 0)
DBUG_RETURN_EVENT(-1);
hpos.data = main_data;
......@@ -2365,7 +2437,8 @@ NdbEventBuffer::get_main_data(Gci_container* bucket,
}
void
NdbEventBuffer::add_blob_data(EventBufData* main_data,
NdbEventBuffer::add_blob_data(Gci_container* bucket,
EventBufData* main_data,
EventBufData* blob_data)
{
DBUG_ENTER_EVENT("NdbEventBuffer::add_blob_data");
......@@ -2389,6 +2462,9 @@ NdbEventBuffer::add_blob_data(EventBufData* main_data,
blob_data->m_next = head->m_next;
head->m_next = blob_data;
}
// adjust data list size
bucket->m_data.m_count += 1;
bucket->m_data.m_sz += blob_data->sz;
DBUG_VOID_RETURN_EVENT;
}
......@@ -2424,6 +2500,9 @@ NdbEventBuffer::move_data()
void
NdbEventBuffer::free_list(EventBufData_list &list)
{
#ifdef NDB_EVENT_VERIFY_SIZE
verify_size(list);
#endif
// return list to m_free_data
list.m_tail->m_next= m_free_data;
m_free_data= list.m_head;
......@@ -2432,38 +2511,15 @@ NdbEventBuffer::free_list(EventBufData_list &list)
#endif
m_free_data_sz+= list.m_sz;
// free blobs XXX unacceptable performance, fix later
{
EventBufData* data = list.m_head;
while (1) {
while (data->m_next_blob != NULL) {
EventBufData* blob_head = data->m_next_blob;
data->m_next_blob = blob_head->m_next_blob;
blob_head->m_next_blob = NULL;
while (blob_head != NULL) {
EventBufData* blob_part = blob_head;
blob_head = blob_head->m_next;
blob_part->m_next = m_free_data;
m_free_data = blob_part;
#ifdef VM_TRACE
m_free_data_count++;
#endif
m_free_data_sz += blob_part->sz;
}
}
if (data == list.m_tail)
break;
data = data->m_next;
}
}
// list returned to m_free_data
list.m_head = list.m_tail = NULL;
list.m_count = list.m_sz = 0;
}
void EventBufData_list::append_list(EventBufData_list *list, Uint64 gci)
{
#ifdef NDB_EVENT_VERIFY_SIZE
NdbEventBuffer::verify_size(*list);
#endif
move_gci_ops(list, gci);
if (m_tail)
......@@ -2702,6 +2758,29 @@ send_report:
#endif
}
#ifdef VM_TRACE
void
NdbEventBuffer::verify_size(const EventBufData* data, Uint32 count, Uint32 sz)
{
Uint32 tmp_count = 0;
Uint32 tmp_sz = 0;
while (data != 0) {
Uint32 full_count, full_sz;
data->get_full_size(full_count, full_sz);
tmp_count += full_count;
tmp_sz += full_sz;
data = data->m_next;
}
assert(tmp_count == count);
assert(tmp_sz == sz);
}
void
NdbEventBuffer::verify_size(const EventBufData_list & list)
{
verify_size(list.m_head, list.m_count, list.m_sz);
}
#endif
// hash table routines
// could optimize the all-fixed case
......
......@@ -40,6 +40,12 @@
#define DBUG_DUMP_EVENT(A,B,C)
#endif
#undef NDB_EVENT_VERIFY_SIZE
#ifdef VM_TRACE
// not much effect on performance, leave on
#define NDB_EVENT_VERIFY_SIZE
#endif
class NdbEventOperationImpl;
struct EventBufData
......@@ -54,9 +60,13 @@ struct EventBufData
/*
* Blobs are stored in blob list (m_next_blob) where each entry
* is list of parts (m_next) in part number order.
* is list of parts (m_next). TODO order by part number
*
* Processed data (m_used_data, m_free_data) keeps the old blob
* list intact. It is reconsumed when new data items are needed.
*
* TODO order by part no and link for fast read and free_list
* Data item lists keep track of item count and sum(sz) and
* these include both main items and blob parts.
*/
EventBufData *m_next; // Next wrt to global order or Next blob part
......@@ -66,14 +76,22 @@ struct EventBufData
Uint32 m_pkhash; // PK hash (without op) for fast compare
EventBufData() {}
// Get blob part number from blob data
Uint32 get_blob_part_no() {
assert(ptr[0].sz > 2);
Uint32 pos = AttributeHeader(ptr[0].p[0]).getDataSize() +
AttributeHeader(ptr[0].p[1]).getDataSize();
Uint32 no = ptr[1].p[pos];
return no;
Uint32 get_blob_part_no() const;
/*
* Main item does not include summary of parts (space / performance
* tradeoff). The summary is needed when moving single data item.
* It is not needed when moving entire list.
*/
void get_full_size(Uint32 & full_count, Uint32 & full_sz) const {
full_count = 1;
full_sz = sz;
if (m_next_blob != 0)
add_part_size(full_count, full_sz);
}
void add_part_size(Uint32 & full_count, Uint32 & full_sz) const;
};
class EventBufData_list
......@@ -82,19 +100,22 @@ public:
EventBufData_list();
~EventBufData_list();
void remove_first();
// append data and insert data into Gci_op list with add_gci_op
void append_data(EventBufData *data);
// remove first and return its size
void remove_first(Uint32 & full_count, Uint32 & full_sz);
// for remove+append avoid double call to get_full_size()
void append_used_data(EventBufData *data, Uint32 full_count, Uint32 full_sz);
// append data and insert data but ignore Gci_op list
void append_used_data(EventBufData *data);
// append data and insert data into Gci_op list with add_gci_op
void append_data(EventBufData *data);
// append list to another, will call move_gci_ops
void append_list(EventBufData_list *list, Uint64 gci);
int is_empty();
EventBufData *m_head, *m_tail;
unsigned m_count;
unsigned m_sz;
Uint32 m_count;
Uint32 m_sz;
/*
distinct ops per gci (assume no hash needed)
......@@ -193,33 +214,47 @@ int EventBufData_list::is_empty()
}
inline
void EventBufData_list::remove_first()
void EventBufData_list::remove_first(Uint32 & full_count, Uint32 & full_sz)
{
m_count--;
m_sz-= m_head->sz;
m_head= m_head->m_next;
m_head->get_full_size(full_count, full_sz);
#ifdef VM_TRACE
assert(m_count >= full_count);
assert(m_sz >= full_sz);
#endif
m_count -= full_count;
m_sz -= full_sz;
m_head = m_head->m_next;
if (m_head == 0)
m_tail= 0;
m_tail = 0;
}
inline
void EventBufData_list::append_used_data(EventBufData *data)
void EventBufData_list::append_used_data(EventBufData *data, Uint32 full_count, Uint32 full_sz)
{
data->m_next= 0;
data->m_next = 0;
if (m_tail)
m_tail->m_next= data;
m_tail->m_next = data;
else
{
#ifdef VM_TRACE
assert(m_head == 0);
assert(m_count == 0);
assert(m_sz == 0);
#endif
m_head= data;
m_head = data;
}
m_tail= data;
m_tail = data;
m_count++;
m_sz+= data->sz;
m_count += full_count;
m_sz += full_sz;
}
inline
void EventBufData_list::append_used_data(EventBufData *data)
{
Uint32 full_count, full_sz;
data->get_full_size(full_count, full_sz);
append_used_data(data, full_count, full_sz);
}
inline
......@@ -442,17 +477,24 @@ public:
// routines to copy/merge events
EventBufData* alloc_data();
int alloc_mem(EventBufData* data, LinearSectionPtr ptr[3]);
int alloc_mem(EventBufData* data,
LinearSectionPtr ptr[3],
Uint32 * change_sz);
void dealloc_mem(EventBufData* data,
Uint32 * change_sz);
int copy_data(const SubTableData * const sdata,
LinearSectionPtr ptr[3],
EventBufData* data);
EventBufData* data,
Uint32 * change_sz);
int merge_data(const SubTableData * const sdata,
LinearSectionPtr ptr[3],
EventBufData* data);
EventBufData* data,
Uint32 * change_sz);
int get_main_data(Gci_container* bucket,
EventBufData_hash::Pos& hpos,
EventBufData* blob_data);
void add_blob_data(EventBufData* main_data,
void add_blob_data(Gci_container* bucket,
EventBufData* main_data,
EventBufData* blob_data);
void free_list(EventBufData_list &list);
......@@ -478,9 +520,9 @@ public:
Gci_container m_complete_data;
EventBufData *m_free_data;
#ifdef VM_TRACE
unsigned m_free_data_count;
Uint32 m_free_data_count;
#endif
unsigned m_free_data_sz;
Uint32 m_free_data_sz;
// user thread
EventBufData_list m_available_data;
......@@ -493,6 +535,12 @@ public:
unsigned m_gci_slip_thresh;
NdbError m_error;
#ifdef VM_TRACE
static void verify_size(const EventBufData* data, Uint32 count, Uint32 sz);
static void verify_size(const EventBufData_list & list);
#endif
private:
int expand(unsigned sz);
......
......@@ -161,6 +161,7 @@ static void
errdb()
{
uint any = 0;
// g_ncc return no error...
if (g_ndb != 0) {
const NdbError& e = g_ndb->getNdbError();
if (e.code != 0)
......@@ -359,9 +360,9 @@ createtable(Tab& t)
}
static int
createtable()
createtables()
{
ll1("createtable");
ll1("createtables");
for (uint i = 0; i < maxtab(); i++)
chkrc(createtable(tab(i)) == 0);
return 0;
......@@ -381,9 +382,9 @@ droptable(Tab& t)
}
static int
droptable()
droptables()
{
ll1("droptable");
ll1("droptables");
for (uint i = 0; i < maxtab(); i++)
chkrc(droptable(tab(i)) == 0);
return 0;
......@@ -419,9 +420,9 @@ createevent(Tab& t)
}
static int
createevent()
createevents()
{
ll1("createevent");
ll1("createevents");
for (uint i = 0; i < maxtab(); i++)
chkrc(createevent(tab(i)) == 0);
return 0;
......@@ -439,11 +440,14 @@ dropevent(Tab& t, bool force = false)
}
static int
dropevent(bool force = false)
dropevents(bool force = false)
{
ll1("dropevent");
for (uint i = 0; i < maxtab(); i++)
ll1("dropevents");
for (uint i = 0; i < maxtab(); i++) {
if (force && g_tablst[i] == 0)
continue;
chkrc(dropevent(tab(i), force) == 0 || force);
}
return 0;
}
......@@ -1173,8 +1177,11 @@ static int
dropeventops(bool force = false)
{
ll1("dropeventops");
for (uint i = 0; i < maxrun(); i++)
for (uint i = 0; i < maxrun(); i++) {
if (force && g_runlst[i] == 0)
continue;
chkrc(dropeventop(run(i), force) == 0 || force);
}
return 0;
}
......@@ -2139,8 +2146,8 @@ runtest()
{
setseed(-1);
initrun();
chkrc(createtable() == 0);
chkrc(createevent() == 0);
chkrc(createtables() == 0);
chkrc(createevents() == 0);
for (g_loop = 0; g_opts.loop == 0 || g_loop < g_opts.loop; g_loop++) {
ll0("=== loop " << g_loop << " ===");
setseed(g_loop);
......@@ -2164,8 +2171,8 @@ runtest()
// time erases everything..
chkrc(waitgci(1) == 0);
}
chkrc(dropevent() == 0);
chkrc(droptable() == 0);
chkrc(dropevents() == 0);
chkrc(droptables() == 0);
resetmem();
deleteops();
return 0;
......@@ -2287,6 +2294,16 @@ checkopts()
return 0;
}
static int
doconnect()
{
g_ncc = new Ndb_cluster_connection();
chkdb(g_ncc->connect(30) == 0);
g_ndb = new Ndb(g_ncc, "TEST_DB");
chkdb(g_ndb->init() == 0 && g_ndb->waitUntilReady(30) == 0);
return 0;
}
int
main(int argc, char** argv)
{
......@@ -2302,19 +2319,13 @@ main(int argc, char** argv)
ret = handle_options(&argc, &argv, my_long_options, ndb_std_get_one_option);
if (ret != 0 || argc != 0 || checkopts() != 0)
return NDBT_ProgramExit(NDBT_WRONGARGS);
g_ncc = new Ndb_cluster_connection();
if (g_ncc->connect(30) == 0) {
g_ndb = new Ndb(g_ncc, "TEST_DB");
if (g_ndb->init() == 0 && g_ndb->waitUntilReady(30) == 0) {
if (runtest() == 0) {
delete g_ndb;
delete g_ncc;
return NDBT_ProgramExit(NDBT_OK);
}
}
if (doconnect() == 0 && runtest() == 0) {
delete g_ndb;
delete g_ncc;
return NDBT_ProgramExit(NDBT_OK);
}
dropeventops(true);
dropevent(true);
dropevents(true);
delete g_ndb;
delete g_ncc;
return NDBT_ProgramExit(NDBT_FAILED);
......
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment