/* Decode a date type */
static int
decode_date(const char *buffer, unsigned int buff_size, unsigned int *out_size)
{
const char *new_buffer = (const char *) TYPEALIGN(sizeof(int32), (uintptr_t) buffer);
unsigned int delta = (unsigned int) ((uintptr_t) new_buffer - (uintptr_t) buffer);
int32 jd,
year,
month,
day;
if (buff_size < delta)
return -1;
buff_size -= delta;
buffer = new_buffer;
if (buff_size < sizeof(int32))
return -2;
*out_size = sizeof(int32) + delta;
jd = *(int32 *) buffer + POSTGRES_EPOCH_JDATE;
j2date(jd, &year, &month, &day);
CopyAppendFmt("%04d-%02d-%02d%s", (year <= 0) ? -year + 1 : year, month, day, (year <= 0) ? " BC" : "");
return 0;
}
/* Decode a timetz type */
static int
decode_timetz(const char *buffer, unsigned int buff_size, unsigned int *out_size)
{
const char *new_buffer = (const char *) TYPEALIGN(sizeof(int64), (uintptr_t) buffer);
unsigned int delta = (unsigned int) ((uintptr_t) new_buffer - (uintptr_t) buffer);
int64 timestamp,
timestamp_sec;
int32 tz_sec,
tz_min;
if (buff_size < delta)
return -1;
buff_size -= delta;
buffer = new_buffer;
if (buff_size < (sizeof(int64) + sizeof(int32)))
return -2;
timestamp = *(int64 *) buffer;
tz_sec = *(int32 *) (buffer + sizeof(int64));
timestamp_sec = timestamp / 1000000;
tz_min = -(tz_sec / 60);
*out_size = sizeof(int64) + sizeof(int32) + delta;
CopyAppendFmt("%02ld:%02ld:%02ld.%06ld%c%02d:%02d",
timestamp_sec / 60 / 60, (timestamp_sec / 60) % 60, timestamp_sec % 60,
timestamp % 1000000, (tz_min > 0 ? '+' : '-'), abs(tz_min / 60), abs(tz_min % 60));
return 0;
}
/* Decode a time type */
static int
decode_time(const char *buffer, unsigned int buff_size, unsigned int *out_size)
{
const char *new_buffer = (const char *) TYPEALIGN(sizeof(int64), (uintptr_t) buffer);
unsigned int delta = (unsigned int) ((uintptr_t) new_buffer - (uintptr_t) buffer);
int64 timestamp,
timestamp_sec;
if (buff_size < delta)
return -1;
buff_size -= delta;
buffer = new_buffer;
if (buff_size < sizeof(int64))
return -2;
timestamp = *(int64 *) buffer;
timestamp_sec = timestamp / 1000000;
*out_size = sizeof(int64) + delta;
CopyAppendFmt("%02ld:%02ld:%02ld.%06ld",
timestamp_sec / 60 / 60, (timestamp_sec / 60) % 60, timestamp_sec % 60,
timestamp % 1000000);
return 0;
}
/* Decode an Oid as int type and pass value out. */
static int
DecodeOidBinary(const char *buffer,
unsigned int buff_size,
unsigned int *processed_size,
Oid *result)
{
const char *new_buffer =
(const char*)TYPEALIGN(sizeof(Oid), (uintptr_t)buffer);
unsigned int delta =
(unsigned int)((uintptr_t)new_buffer - (uintptr_t)buffer);
if (buff_size < delta)
return -1;
buff_size -= delta;
buffer = new_buffer;
if (buff_size < sizeof(int32))
return -2;
*result = *(Oid *)buffer;
*processed_size = sizeof(Oid) + delta;
return 0;
}
static void
prepare_buf(void)
{
int ops;
/* write random data into buffer */
for (ops = 0; ops < XLOG_SEG_SIZE; ops++)
full_buf[ops] = random();
buf = (char *) TYPEALIGN(XLOG_BLCKSZ, full_buf);
}
static inline int compute_null_bitmap_extra_size(TupleDesc tupdesc, int col_align)
{
int nbytes = (tupdesc->natts + 7) >> 3;
int avail_bytes = (tupdesc->tdhasoid || col_align == 4) ? 0 : 4;
Assert(col_align == 4 || col_align == 8);
if (nbytes <= avail_bytes)
return 0;
return TYPEALIGN(col_align, (nbytes - avail_bytes));
}
/* Decode a timestamp type */
static int
decode_timestamp(const char *buffer, unsigned int buff_size, unsigned int *out_size)
{
const char *new_buffer = (const char *) TYPEALIGN(sizeof(int64), (uintptr_t) buffer);
unsigned int delta = (unsigned int) ((uintptr_t) new_buffer - (uintptr_t) buffer);
int64 timestamp,
timestamp_sec;
int32 jd,
year,
month,
day;
if (buff_size < delta)
return -1;
buff_size -= delta;
buffer = new_buffer;
if (buff_size < sizeof(int64))
return -2;
*out_size = sizeof(int64) + delta;
timestamp = *(int64 *) buffer;
jd = timestamp / USECS_PER_DAY;
if (jd != 0)
timestamp -= jd * USECS_PER_DAY;
if (timestamp < INT64CONST(0))
{
timestamp += USECS_PER_DAY;
jd -= 1;
}
/* add offset to go from J2000 back to standard Julian date */
jd += POSTGRES_EPOCH_JDATE;
j2date(jd, &year, &month, &day);
timestamp_sec = timestamp / 1000000;
CopyAppendFmt("%04d-%02d-%02d %02ld:%02ld:%02ld.%06ld%s",
(year <= 0) ? -year + 1 : year, month, day,
timestamp_sec / 60 / 60, (timestamp_sec / 60) % 60, timestamp_sec % 60,
timestamp % 1000000,
(year <= 0) ? " BC" : "");
return 0;
}
/* Decode a float8 type */
static int
decode_float8(const char *buffer, unsigned int buff_size, unsigned int *out_size)
{
const char *new_buffer = (const char *) TYPEALIGN(sizeof(double), (uintptr_t) buffer);
unsigned int delta = (unsigned int) ((uintptr_t) new_buffer - (uintptr_t) buffer);
if (buff_size < delta)
return -1;
buff_size -= delta;
buffer = new_buffer;
if (buff_size < sizeof(double))
return -2;
CopyAppendFmt("%.12lf", *(double *) buffer);
*out_size = sizeof(double) + delta;
return 0;
}
/* Decode a bigint type */
static int
decode_bigint(const char *buffer, unsigned int buff_size, unsigned int *out_size)
{
const char *new_buffer = (const char *) TYPEALIGN(sizeof(int64), (uintptr_t) buffer);
unsigned int delta = (unsigned int) ((uintptr_t) new_buffer - (uintptr_t) buffer);
if (buff_size < delta)
return -1;
buff_size -= delta;
buffer = new_buffer;
if (buff_size < sizeof(int64))
return -2;
CopyAppendFmt("%ld", *(int64 *) buffer);
*out_size = sizeof(int64) + delta;
return 0;
}
/* Decode a name type (used mostly in catalog tables) */
static int
decode_name(const char *buffer, unsigned int buff_size, unsigned int *out_size)
{
const char *new_buffer = (const char *) TYPEALIGN(sizeof(uint32), (uintptr_t) buffer);
unsigned int delta = (unsigned int) ((uintptr_t) new_buffer - (uintptr_t) buffer);
if (buff_size < delta)
return -1;
buff_size -= delta;
buffer = new_buffer;
if (buff_size < NAMEDATALEN)
return -2;
CopyAppendEncode(buffer, strnlen(buffer, NAMEDATALEN));
*out_size = NAMEDATALEN + delta;
return 0;
}
/* Decode a macaddr type */
static int
decode_macaddr(const char *buffer, unsigned int buff_size, unsigned int *out_size)
{
unsigned char macaddr[6];
const char *new_buffer = (const char *) TYPEALIGN(sizeof(int32), (uintptr_t) buffer);
unsigned int delta = (unsigned int) ((uintptr_t) new_buffer - (uintptr_t) buffer);
if (buff_size < delta)
return -1;
buff_size -= delta;
buffer = new_buffer;
if (buff_size < sizeof(macaddr))
return -2;
memcpy(macaddr, buffer, sizeof(macaddr));
CopyAppendFmt("%02x:%02x:%02x:%02x:%02x:%02x",
macaddr[0], macaddr[1], macaddr[2], macaddr[3], macaddr[4], macaddr[5]
);
*out_size = sizeof(macaddr) + delta;
return 0;
}
/*
* Serialize a tuple directly into a buffer.
*
* We're called with at least enough space for a tuple-chunk-header.
*/
int
SerializeTupleDirect(HeapTuple tuple, SerTupInfo * pSerInfo, struct directTransportBuffer *b)
{
int natts;
int dataSize = TUPLE_CHUNK_HEADER_SIZE;
TupleDesc tupdesc;
AssertArg(tuple != NULL);
AssertArg(pSerInfo != NULL);
AssertArg(b != NULL);
tupdesc = pSerInfo->tupdesc;
natts = tupdesc->natts;
do
{
if (natts == 0)
{
/* TC_EMTPY is just one chunk */
SetChunkType(b->pri, TC_EMPTY);
SetChunkDataSize(b->pri, 0);
break;
}
/* easy case */
if (is_heaptuple_memtuple(tuple))
{
int tupleSize;
int paddedSize;
tupleSize = memtuple_get_size((MemTuple)tuple, NULL);
paddedSize = TYPEALIGN(TUPLE_CHUNK_ALIGN, tupleSize);
if (paddedSize + TUPLE_CHUNK_HEADER_SIZE > b->prilen)
return 0;
/* will fit. */
memcpy(b->pri + TUPLE_CHUNK_HEADER_SIZE, tuple, tupleSize);
memset(b->pri + TUPLE_CHUNK_HEADER_SIZE + tupleSize, 0, paddedSize - tupleSize);
dataSize += paddedSize;
SetChunkType(b->pri, TC_WHOLE);
SetChunkDataSize(b->pri, dataSize - TUPLE_CHUNK_HEADER_SIZE);
break;
}
else
{
TupSerHeader tsh;
unsigned int datalen;
unsigned int nullslen;
HeapTupleHeader t_data = tuple->t_data;
unsigned char *pos;
datalen = tuple->t_len - t_data->t_hoff;
if (HeapTupleHasNulls(tuple))
nullslen = BITMAPLEN(HeapTupleHeaderGetNatts(t_data));
else
nullslen = 0;
tsh.tuplen = sizeof(TupSerHeader) + TYPEALIGN(TUPLE_CHUNK_ALIGN, nullslen) + TYPEALIGN(TUPLE_CHUNK_ALIGN, datalen);
tsh.natts = HeapTupleHeaderGetNatts(t_data);
tsh.infomask = t_data->t_infomask;
if (dataSize + tsh.tuplen > b->prilen ||
(tsh.infomask & HEAP_HASEXTERNAL) != 0)
return 0;
pos = b->pri + TUPLE_CHUNK_HEADER_SIZE;
memcpy(pos, (char *)&tsh, sizeof(TupSerHeader));
pos += sizeof(TupSerHeader);
if (nullslen)
{
memcpy(pos, (char *)t_data->t_bits, nullslen);
pos += nullslen;
memset(pos, 0, TYPEALIGN(TUPLE_CHUNK_ALIGN, nullslen) - nullslen);
pos += TYPEALIGN(TUPLE_CHUNK_ALIGN, nullslen) - nullslen;
}
memcpy(pos, (char *)t_data + t_data->t_hoff, datalen);
pos += datalen;
memset(pos, 0, TYPEALIGN(TUPLE_CHUNK_ALIGN, datalen) - datalen);
pos += TYPEALIGN(TUPLE_CHUNK_ALIGN, datalen) - datalen;
dataSize += tsh.tuplen;
SetChunkType(b->pri, TC_WHOLE);
SetChunkDataSize(b->pri, dataSize - TUPLE_CHUNK_HEADER_SIZE);
break;
}
/* tuple that we can't handle here (big ?) -- do the older "out-of-line" serialization */
return 0;
}
while (0);
return dataSize;
}
static inline void
skipPadding(StringInfo serialTup)
{
serialTup->cursor = TYPEALIGN(TUPLE_CHUNK_ALIGN,serialTup->cursor);
}
/*
* Convert a HeapTuple into a byte-sequence, and store it directly
* into a chunklist for transmission.
*
* This code is based on the printtup_internal_20() function in printtup.c.
*/
void
SerializeTupleIntoChunks(HeapTuple tuple, SerTupInfo * pSerInfo, TupleChunkList tcList)
{
TupleChunkListItem tcItem = NULL;
MemoryContext oldCtxt;
TupleDesc tupdesc;
int i,
natts;
bool fHandled;
AssertArg(tcList != NULL);
AssertArg(tuple != NULL);
AssertArg(pSerInfo != NULL);
tupdesc = pSerInfo->tupdesc;
natts = tupdesc->natts;
/* get ready to go */
tcList->p_first = NULL;
tcList->p_last = NULL;
tcList->num_chunks = 0;
tcList->serialized_data_length = 0;
tcList->max_chunk_length = Gp_max_tuple_chunk_size;
if (natts == 0)
{
tcItem = getChunkFromCache(&pSerInfo->chunkCache);
if (tcItem == NULL)
{
ereport(FATAL, (errcode(ERRCODE_OUT_OF_MEMORY),
errmsg("Could not allocate space for first chunk item in new chunk list.")));
}
/* TC_EMTPY is just one chunk */
SetChunkType(tcItem->chunk_data, TC_EMPTY);
tcItem->chunk_length = TUPLE_CHUNK_HEADER_SIZE;
appendChunkToTCList(tcList, tcItem);
return;
}
tcItem = getChunkFromCache(&pSerInfo->chunkCache);
if (tcItem == NULL)
{
ereport(FATAL, (errcode(ERRCODE_OUT_OF_MEMORY),
errmsg("Could not allocate space for first chunk item in new chunk list.")));
}
/* assume that we'll take a single chunk */
SetChunkType(tcItem->chunk_data, TC_WHOLE);
tcItem->chunk_length = TUPLE_CHUNK_HEADER_SIZE;
appendChunkToTCList(tcList, tcItem);
AssertState(s_tupSerMemCtxt != NULL);
if (is_heaptuple_memtuple(tuple))
{
addByteStringToChunkList(tcList, (char *)tuple, memtuple_get_size((MemTuple)tuple, NULL), &pSerInfo->chunkCache);
addPadding(tcList, &pSerInfo->chunkCache, memtuple_get_size((MemTuple)tuple, NULL));
}
else
{
TupSerHeader tsh;
unsigned int datalen;
unsigned int nullslen;
HeapTupleHeader t_data = tuple->t_data;
datalen = tuple->t_len - t_data->t_hoff;
if (HeapTupleHasNulls(tuple))
nullslen = BITMAPLEN(HeapTupleHeaderGetNatts(t_data));
else
nullslen = 0;
tsh.tuplen = sizeof(TupSerHeader) + TYPEALIGN(TUPLE_CHUNK_ALIGN,nullslen) + datalen;
tsh.natts = HeapTupleHeaderGetNatts(t_data);
tsh.infomask = t_data->t_infomask;
addByteStringToChunkList(tcList, (char *)&tsh, sizeof(TupSerHeader), &pSerInfo->chunkCache);
/* If we don't have any attributes which have been toasted, we
* can be very very simple: just send the raw data. */
if ((tsh.infomask & HEAP_HASEXTERNAL) == 0)
{
if (nullslen)
{
addByteStringToChunkList(tcList, (char *)t_data->t_bits, nullslen, &pSerInfo->chunkCache);
addPadding(tcList,&pSerInfo->chunkCache,nullslen);
}
addByteStringToChunkList(tcList, (char *)t_data + t_data->t_hoff, datalen, &pSerInfo->chunkCache);
addPadding(tcList,&pSerInfo->chunkCache,datalen);
}
else
{
/* We have to be more careful when we have tuples that
* have been toasted. Ideally we'd like to send the
* untoasted attributes in as "raw" a format as possible
* but that makes rebuilding the tuple harder .
*/
oldCtxt = MemoryContextSwitchTo(s_tupSerMemCtxt);
/* deconstruct the tuple (faster than a heap_getattr loop) */
heap_deform_tuple(tuple, tupdesc, pSerInfo->values, pSerInfo->nulls);
MemoryContextSwitchTo(oldCtxt);
/* Send the nulls character-array. */
addByteStringToChunkList(tcList, pSerInfo->nulls, natts, &pSerInfo->chunkCache);
addPadding(tcList,&pSerInfo->chunkCache,natts);
/*
* send the attributes of this tuple: NOTE anything which allocates
* temporary space (e.g. could result in a PG_DETOAST_DATUM) should be
* executed with the memory context set to s_tupSerMemCtxt
*/
for (i = 0; i < natts; ++i)
{
SerAttrInfo *attrInfo = pSerInfo->myinfo + i;
Datum origattr = pSerInfo->values[i],
attr;
bytea *outputbytes=0;
/* skip null attributes (already taken care of above) */
if (pSerInfo->nulls[i])
continue;
/*
* If we have a toasted datum, forcibly detoast it here to avoid
* memory leakage: we want to force the detoast allocation(s) to
* happen in our reset-able serialization context.
*/
if (attrInfo->typisvarlena)
{
oldCtxt = MemoryContextSwitchTo(s_tupSerMemCtxt);
/* we want to detoast but leave compressed, if
* possible, but we have to handle varlena
* attributes (and others ?) differently than we
* currently do (first step is to use
* heap_tuple_fetch_attr() instead of
* PG_DETOAST_DATUM()). */
attr = PointerGetDatum(PG_DETOAST_DATUM(origattr));
MemoryContextSwitchTo(oldCtxt);
}
else
attr = origattr;
/*
* Assume that the data's output will be handled by the special IO
* code, and if not then we can handle it the slow way.
*/
fHandled = true;
switch (attrInfo->atttypid)
{
case INT4OID:
addInt32ToChunkList(tcList, DatumGetInt32(attr), &pSerInfo->chunkCache);
break;
case CHAROID:
addCharToChunkList(tcList, DatumGetChar(attr), &pSerInfo->chunkCache);
addPadding(tcList,&pSerInfo->chunkCache,1);
break;
case BPCHAROID:
case VARCHAROID:
case INT2VECTOROID: /* postgres serialization logic broken, use our own */
case OIDVECTOROID: /* postgres serialization logic broken, use our own */
case ANYARRAYOID:
{
text *pText = DatumGetTextP(attr);
int32 textSize = VARSIZE(pText) - VARHDRSZ;
addInt32ToChunkList(tcList, textSize, &pSerInfo->chunkCache);
addByteStringToChunkList(tcList, (char *) VARDATA(pText), textSize, &pSerInfo->chunkCache);
addPadding(tcList,&pSerInfo->chunkCache,textSize);
break;
}
case DATEOID:
{
DateADT date = DatumGetDateADT(attr);
addByteStringToChunkList(tcList, (char *) &date, sizeof(DateADT), &pSerInfo->chunkCache);
break;
}
case NUMERICOID:
{
/*
* Treat the numeric as a varlena variable, and just push
* the whole shebang to the output-buffer. We don't care
* about the guts of the numeric.
*/
Numeric num = DatumGetNumeric(attr);
int32 numSize = VARSIZE(num) - VARHDRSZ;
addInt32ToChunkList(tcList, numSize, &pSerInfo->chunkCache);
addByteStringToChunkList(tcList, (char *) VARDATA(num), numSize, &pSerInfo->chunkCache);
addPadding(tcList,&pSerInfo->chunkCache,numSize);
break;
}
case ACLITEMOID:
{
AclItem *aip = DatumGetAclItemP(attr);
char *outputstring;
int32 aclSize ;
outputstring = DatumGetCString(DirectFunctionCall1(aclitemout,
PointerGetDatum(aip)));
aclSize = strlen(outputstring);
addInt32ToChunkList(tcList, aclSize, &pSerInfo->chunkCache);
addByteStringToChunkList(tcList, outputstring,aclSize, &pSerInfo->chunkCache);
addPadding(tcList,&pSerInfo->chunkCache,aclSize);
break;
}
case 210: /* storage manager */
{
char *smgrstr;
int32 strsize;
smgrstr = DatumGetCString(DirectFunctionCall1(smgrout, 0));
strsize = strlen(smgrstr);
addInt32ToChunkList(tcList, strsize, &pSerInfo->chunkCache);
addByteStringToChunkList(tcList, smgrstr, strsize, &pSerInfo->chunkCache);
addPadding(tcList,&pSerInfo->chunkCache,strsize);
break;
}
default:
fHandled = false;
}
if (fHandled)
continue;
/*
* the FunctionCall2 call into the send function may result in some
* allocations which we'd like to have contained by our reset-able
* context
*/
oldCtxt = MemoryContextSwitchTo(s_tupSerMemCtxt);
/* Call the attribute type's binary input converter. */
if (attrInfo->send_finfo.fn_nargs == 1)
outputbytes =
DatumGetByteaP(FunctionCall1(&attrInfo->send_finfo,
attr));
else if (attrInfo->send_finfo.fn_nargs == 2)
outputbytes =
DatumGetByteaP(FunctionCall2(&attrInfo->send_finfo,
attr,
ObjectIdGetDatum(attrInfo->send_typio_param)));
else if (attrInfo->send_finfo.fn_nargs == 3)
outputbytes =
DatumGetByteaP(FunctionCall3(&attrInfo->send_finfo,
attr,
ObjectIdGetDatum(attrInfo->send_typio_param),
Int32GetDatum(tupdesc->attrs[i]->atttypmod)));
else
{
ereport(ERROR,
(errcode(ERRCODE_INVALID_BINARY_REPRESENTATION),
errmsg("Conversion function takes %d args",attrInfo->recv_finfo.fn_nargs)));
}
MemoryContextSwitchTo(oldCtxt);
/* We assume the result will not have been toasted */
addInt32ToChunkList(tcList, VARSIZE(outputbytes) - VARHDRSZ, &pSerInfo->chunkCache);
addByteStringToChunkList(tcList, VARDATA(outputbytes),
VARSIZE(outputbytes) - VARHDRSZ, &pSerInfo->chunkCache);
addPadding(tcList,&pSerInfo->chunkCache,VARSIZE(outputbytes) - VARHDRSZ);
/*
* this was allocated in our reset-able context, but we *are* done
* with it; and for tuples with several large columns it'd be nice to
* free the memory back to the context
*/
pfree(outputbytes);
}
MemoryContextReset(s_tupSerMemCtxt);
}
}
/*
* if we have more than 1 chunk we have to set the chunk types on our
* first chunk and last chunk
*/
if (tcList->num_chunks > 1)
{
TupleChunkListItem first,
last;
first = tcList->p_first;
last = tcList->p_last;
Assert(first != NULL);
Assert(first != last);
Assert(last != NULL);
SetChunkType(first->chunk_data, TC_PARTIAL_START);
SetChunkType(last->chunk_data, TC_PARTIAL_END);
/*
* any intervening chunks are already set to TC_PARTIAL_MID when
* allocated
*/
}
return;
}
HeapTuple
CvtChunksToHeapTup(TupleChunkList tcList, SerTupInfo * pSerInfo)
{
StringInfoData serData;
TupleChunkListItem tcItem;
int i;
HeapTuple htup;
TupleChunkType tcType;
AssertArg(tcList != NULL);
AssertArg(tcList->p_first != NULL);
AssertArg(pSerInfo != NULL);
tcItem = tcList->p_first;
if (tcList->num_chunks == 1)
{
GetChunkType(tcItem, &tcType);
if (tcType == TC_EMPTY)
{
/*
* the sender is indicating that there was a row with no attributes:
* return a NULL tuple
*/
clearTCList(NULL, tcList);
htup = heap_form_tuple(pSerInfo->tupdesc, pSerInfo->values, pSerInfo->nulls);
return htup;
}
}
/*
* Dump all of the data in the tuple chunk list into a single StringInfo,
* so that we can convert it into a HeapTuple. Check chunk types based on
* whether there is only one chunk, or multiple chunks.
*
* We know roughly how much space we'll need, allocate all in one go.
*
*/
initStringInfoOfSize(&serData, tcList->num_chunks * tcList->max_chunk_length);
i = 0;
do
{
/* Make sure that the type of this tuple chunk is correct! */
GetChunkType(tcItem, &tcType);
if (i == 0)
{
if (tcItem->p_next == NULL)
{
if (tcType != TC_WHOLE)
{
ereport(ERROR, (errcode(ERRCODE_PROTOCOL_VIOLATION),
errmsg("Single chunk's type must be TC_WHOLE.")));
}
}
else
/* tcItem->p_next != NULL */
{
if (tcType != TC_PARTIAL_START)
{
ereport(ERROR, (errcode(ERRCODE_PROTOCOL_VIOLATION),
errmsg("First chunk of collection must have type"
" TC_PARTIAL_START.")));
}
}
}
else
/* i > 0 */
{
if (tcItem->p_next == NULL)
{
if (tcType != TC_PARTIAL_END)
{
ereport(ERROR, (errcode(ERRCODE_PROTOCOL_VIOLATION),
errmsg("Last chunk of collection must have type"
" TC_PARTIAL_END.")));
}
}
else
/* tcItem->p_next != NULL */
{
if (tcType != TC_PARTIAL_MID)
{
ereport(ERROR, (errcode(ERRCODE_PROTOCOL_VIOLATION),
errmsg("Last chunk of collection must have type"
" TC_PARTIAL_MID.")));
}
}
}
/* Copy this chunk into the tuple data. Don't include the header! */
appendBinaryStringInfo(&serData,
(const char *) GetChunkDataPtr(tcItem) + TUPLE_CHUNK_HEADER_SIZE,
tcItem->chunk_length - TUPLE_CHUNK_HEADER_SIZE);
/* Go to the next chunk. */
tcItem = tcItem->p_next;
i++;
}
while (tcItem != NULL);
/* we've finished with the TCList, free it now. */
clearTCList(NULL, tcList);
{
TupSerHeader *tshp;
unsigned int datalen;
unsigned int nullslen;
unsigned int hoff;
HeapTupleHeader t_data;
char *pos = (char *)serData.data;
tshp = (TupSerHeader *)pos;
if ((tshp->tuplen & MEMTUP_LEAD_BIT) != 0)
{
uint32 tuplen = memtuple_size_from_uint32(tshp->tuplen);
htup = (HeapTuple) palloc(tuplen);
memcpy(htup, pos, tuplen);
pos += TYPEALIGN(TUPLE_CHUNK_ALIGN,tuplen);
}
else
{
pos += sizeof(TupSerHeader);
/* if the tuple had toasted elements we have to deserialize
* the old slow way. */
if ((tshp->infomask & HEAP_HASEXTERNAL) != 0)
{
serData.cursor += sizeof(TupSerHeader);
htup = DeserializeTuple(pSerInfo, &serData);
/* Free up memory we used. */
pfree(serData.data);
return htup;
}
/* reconstruct lengths of null bitmap and data part */
if (tshp->infomask & HEAP_HASNULL)
nullslen = BITMAPLEN(tshp->natts);
else
nullslen = 0;
if (tshp->tuplen < sizeof(TupSerHeader) + nullslen)
ereport(ERROR, (errcode(ERRCODE_GP_INTERCONNECTION_ERROR),
errmsg("Interconnect error: cannot convert chunks to a heap tuple."),
errdetail("tuple len %d < nullslen %d + headersize (%d)",
tshp->tuplen, nullslen, (int)sizeof(TupSerHeader))));
datalen = tshp->tuplen - sizeof(TupSerHeader) - TYPEALIGN(TUPLE_CHUNK_ALIGN, nullslen);
/* determine overhead size of tuple (should match heap_form_tuple) */
hoff = offsetof(HeapTupleHeaderData, t_bits) + TYPEALIGN(TUPLE_CHUNK_ALIGN, nullslen);
if (tshp->infomask & HEAP_HASOID)
hoff += sizeof(Oid);
hoff = MAXALIGN(hoff);
/* Allocate the space in one chunk, like heap_form_tuple */
htup = (HeapTuple)palloc(HEAPTUPLESIZE + hoff + datalen);
t_data = (HeapTupleHeader) ((char *)htup + HEAPTUPLESIZE);
/* make sure unused header fields are zeroed */
MemSetAligned(t_data, 0, hoff);
/* reconstruct the HeapTupleData fields */
htup->t_len = hoff + datalen;
ItemPointerSetInvalid(&(htup->t_self));
htup->t_data = t_data;
/* reconstruct the HeapTupleHeaderData fields */
ItemPointerSetInvalid(&(t_data->t_ctid));
HeapTupleHeaderSetNatts(t_data, tshp->natts);
t_data->t_infomask = tshp->infomask & ~HEAP_XACT_MASK;
t_data->t_infomask |= HEAP_XMIN_INVALID | HEAP_XMAX_INVALID;
t_data->t_hoff = hoff;
if (nullslen)
{
memcpy((void *)t_data->t_bits, pos, nullslen);
pos += TYPEALIGN(TUPLE_CHUNK_ALIGN,nullslen);
}
/* does the tuple descriptor expect an OID ? Note: we don't
* have to set the oid itself, just the flag! (see heap_formtuple()) */
if (pSerInfo->tupdesc->tdhasoid) /* else leave infomask = 0 */
{
t_data->t_infomask |= HEAP_HASOID;
}
/* and now the data proper (it would be nice if we could just
* point our caller into our existing buffer in-place, but
* we'll leave that for another day) */
memcpy((char *)t_data + hoff, pos, datalen);
}
}
/* Free up memory we used. */
pfree(serData.data);
return htup;
}
int
main(int argc, char *argv[])
{
struct timeval start_t;
struct timeval elapse_t;
int tmpfile,
i,
loops = 1000;
char *full_buf = (char *) malloc(XLOG_SEG_SIZE),
*buf;
char *filename = FSYNC_FILENAME;
if (argc > 2 && strcmp(argv[1], "-f") == 0)
{
filename = argv[2];
argv += 2;
argc -= 2;
}
if (argc > 1)
loops = atoi(argv[1]);
for (i = 0; i < XLOG_SEG_SIZE; i++)
full_buf[i] = 'a';
if ((tmpfile = open(filename, O_RDWR | O_CREAT, S_IRUSR | S_IWUSR, 0)) == -1)
die("Cannot open output file.");
if (write(tmpfile, full_buf, XLOG_SEG_SIZE) != XLOG_SEG_SIZE)
die("write failed");
/* fsync so later fsync's don't have to do it */
if (fsync(tmpfile) != 0)
die("fsync failed");
close(tmpfile);
buf = (char *) TYPEALIGN(ALIGNOF_XLOG_BUFFER, full_buf);
printf("Simple write timing:\n");
/* write only */
gettimeofday(&start_t, NULL);
for (i = 0; i < loops; i++)
{
if ((tmpfile = open(filename, O_RDWR, 0)) == -1)
die("Cannot open output file.");
if (write(tmpfile, buf, WRITE_SIZE / 2) != WRITE_SIZE / 2)
die("write failed");
close(tmpfile);
}
gettimeofday(&elapse_t, NULL);
printf("\twrite ");
print_elapse(start_t, elapse_t);
printf("\n");
printf("\nCompare fsync times on write() and non-write() descriptor:\n");
printf("(If the times are similar, fsync() can sync data written\n on a different descriptor.)\n");
/* write, fsync, close */
gettimeofday(&start_t, NULL);
for (i = 0; i < loops; i++)
{
if ((tmpfile = open(filename, O_RDWR, 0)) == -1)
die("Cannot open output file.");
if (write(tmpfile, buf, WRITE_SIZE / 2) != WRITE_SIZE / 2)
die("write failed");
if (fsync(tmpfile) != 0)
die("fsync failed");
close(tmpfile);
if ((tmpfile = open(filename, O_RDWR, 0)) == -1)
die("Cannot open output file.");
/* do nothing but the open/close the tests are consistent. */
close(tmpfile);
}
gettimeofday(&elapse_t, NULL);
printf("\twrite, fsync, close ");
print_elapse(start_t, elapse_t);
printf("\n");
/* write, close, fsync */
gettimeofday(&start_t, NULL);
for (i = 0; i < loops; i++)
{
if ((tmpfile = open(filename, O_RDWR, 0)) == -1)
die("Cannot open output file.");
if (write(tmpfile, buf, WRITE_SIZE / 2) != WRITE_SIZE / 2)
die("write failed");
close(tmpfile);
/* reopen file */
if ((tmpfile = open(filename, O_RDWR, 0)) == -1)
die("Cannot open output file.");
if (fsync(tmpfile) != 0)
die("fsync failed");
close(tmpfile);
}
gettimeofday(&elapse_t, NULL);
printf("\twrite, close, fsync ");
print_elapse(start_t, elapse_t);
printf("\n");
printf("\nCompare one o_sync write to two:\n");
#ifdef OPEN_SYNC_FLAG
/* 16k o_sync write */
if ((tmpfile = open(filename, O_RDWR | OPEN_SYNC_FLAG, 0)) == -1)
die("Cannot open output file.");
gettimeofday(&start_t, NULL);
for (i = 0; i < loops; i++)
if (write(tmpfile, buf, WRITE_SIZE) != WRITE_SIZE)
die("write failed");
gettimeofday(&elapse_t, NULL);
close(tmpfile);
printf("\tone 16k o_sync write ");
print_elapse(start_t, elapse_t);
printf("\n");
/* 2*8k o_sync writes */
if ((tmpfile = open(filename, O_RDWR | OPEN_SYNC_FLAG, 0)) == -1)
die("Cannot open output file.");
gettimeofday(&start_t, NULL);
for (i = 0; i < loops; i++)
{
if (write(tmpfile, buf, WRITE_SIZE / 2) != WRITE_SIZE / 2)
die("write failed");
if (write(tmpfile, buf, WRITE_SIZE / 2) != WRITE_SIZE / 2)
die("write failed");
}
gettimeofday(&elapse_t, NULL);
close(tmpfile);
printf("\ttwo 8k o_sync writes ");
print_elapse(start_t, elapse_t);
printf("\n");
printf("\nCompare file sync methods with one 8k write:\n");
#else
printf("\t(o_sync unavailable) ");
#endif
printf("\n");
#ifdef OPEN_DATASYNC_FLAG
/* open_dsync, write */
if ((tmpfile = open(filename, O_RDWR | O_DSYNC, 0)) == -1)
die("Cannot open output file.");
gettimeofday(&start_t, NULL);
for (i = 0; i < loops; i++)
if (write(tmpfile, buf, WRITE_SIZE / 2) != WRITE_SIZE / 2)
die("write failed");
gettimeofday(&elapse_t, NULL);
close(tmpfile);
printf("\topen o_dsync, write ");
print_elapse(start_t, elapse_t);
printf("\n");
#ifdef OPEN_SYNC_FLAG
/* open_fsync, write */
if ((tmpfile = open(filename, O_RDWR | OPEN_SYNC_FLAG, 0)) == -1)
die("Cannot open output file.");
gettimeofday(&start_t, NULL);
for (i = 0; i < loops; i++)
if (write(tmpfile, buf, WRITE_SIZE / 2) != WRITE_SIZE / 2)
die("write failed");
gettimeofday(&elapse_t, NULL);
close(tmpfile);
printf("\topen o_sync, write ");
print_elapse(start_t, elapse_t);
#endif
#else
printf("\t(o_dsync unavailable) ");
#endif
printf("\n");
#ifdef HAVE_FDATASYNC
/* write, fdatasync */
if ((tmpfile = open(filename, O_RDWR, 0)) == -1)
die("Cannot open output file.");
gettimeofday(&start_t, NULL);
for (i = 0; i < loops; i++)
{
if (write(tmpfile, buf, WRITE_SIZE / 2) != WRITE_SIZE / 2)
die("write failed");
fdatasync(tmpfile);
}
gettimeofday(&elapse_t, NULL);
close(tmpfile);
printf("\twrite, fdatasync ");
print_elapse(start_t, elapse_t);
#else
printf("\t(fdatasync unavailable)");
#endif
printf("\n");
/* write, fsync, close */
if ((tmpfile = open(filename, O_RDWR, 0)) == -1)
die("Cannot open output file.");
gettimeofday(&start_t, NULL);
for (i = 0; i < loops; i++)
{
if (write(tmpfile, buf, WRITE_SIZE / 2) != WRITE_SIZE / 2)
die("write failed");
if (fsync(tmpfile) != 0)
die("fsync failed");
}
gettimeofday(&elapse_t, NULL);
close(tmpfile);
printf("\twrite, fsync, ");
print_elapse(start_t, elapse_t);
printf("\n");
printf("\nCompare file sync methods with 2 8k writes:\n");
#ifdef OPEN_DATASYNC_FLAG
/* open_dsync, write */
if ((tmpfile = open(filename, O_RDWR | O_DSYNC, 0)) == -1)
die("Cannot open output file.");
gettimeofday(&start_t, NULL);
for (i = 0; i < loops; i++)
{
if (write(tmpfile, buf, WRITE_SIZE / 2) != WRITE_SIZE / 2)
die("write failed");
if (write(tmpfile, buf, WRITE_SIZE / 2) != WRITE_SIZE / 2)
die("write failed");
}
gettimeofday(&elapse_t, NULL);
close(tmpfile);
printf("\topen o_dsync, write ");
print_elapse(start_t, elapse_t);
#else
printf("\t(o_dsync unavailable) ");
#endif
printf("\n");
#ifdef OPEN_SYNC_FLAG
/* open_fsync, write */
if ((tmpfile = open(filename, O_RDWR | OPEN_SYNC_FLAG, 0)) == -1)
die("Cannot open output file.");
gettimeofday(&start_t, NULL);
for (i = 0; i < loops; i++)
{
if (write(tmpfile, buf, WRITE_SIZE / 2) != WRITE_SIZE / 2)
die("write failed");
if (write(tmpfile, buf, WRITE_SIZE / 2) != WRITE_SIZE / 2)
die("write failed");
}
gettimeofday(&elapse_t, NULL);
close(tmpfile);
printf("\topen o_sync, write ");
print_elapse(start_t, elapse_t);
printf("\n");
#endif
#ifdef HAVE_FDATASYNC
/* write, fdatasync */
if ((tmpfile = open(filename, O_RDWR, 0)) == -1)
die("Cannot open output file.");
gettimeofday(&start_t, NULL);
for (i = 0; i < loops; i++)
{
if (write(tmpfile, buf, WRITE_SIZE / 2) != WRITE_SIZE / 2)
die("write failed");
if (write(tmpfile, buf, WRITE_SIZE / 2) != WRITE_SIZE / 2)
die("write failed");
fdatasync(tmpfile);
}
gettimeofday(&elapse_t, NULL);
close(tmpfile);
printf("\twrite, fdatasync ");
print_elapse(start_t, elapse_t);
#else
printf("\t(fdatasync unavailable)");
#endif
printf("\n");
/* write, fsync, close */
if ((tmpfile = open(filename, O_RDWR, 0)) == -1)
die("Cannot open output file.");
gettimeofday(&start_t, NULL);
for (i = 0; i < loops; i++)
{
if (write(tmpfile, buf, WRITE_SIZE / 2) != WRITE_SIZE / 2)
die("write failed");
if (write(tmpfile, buf, WRITE_SIZE / 2) != WRITE_SIZE / 2)
die("write failed");
if (fsync(tmpfile) != 0)
die("fsync failed");
}
gettimeofday(&elapse_t, NULL);
close(tmpfile);
printf("\twrite, fsync, ");
print_elapse(start_t, elapse_t);
printf("\n");
free(full_buf);
unlink(filename);
return 0;
}
/*
* Convert a HeapTuple into a byte-sequence, and store it directly
* into a chunklist for transmission.
*
* This code is based on the printtup_internal_20() function in printtup.c.
*/
void
SerializeTupleIntoChunks(GenericTuple gtuple, SerTupInfo *pSerInfo, TupleChunkList tcList)
{
TupleChunkListItem tcItem = NULL;
MemoryContext oldCtxt;
TupleDesc tupdesc;
int i,
natts;
AssertArg(tcList != NULL);
AssertArg(gtuple != NULL);
AssertArg(pSerInfo != NULL);
tupdesc = pSerInfo->tupdesc;
natts = tupdesc->natts;
/* get ready to go */
tcList->p_first = NULL;
tcList->p_last = NULL;
tcList->num_chunks = 0;
tcList->serialized_data_length = 0;
tcList->max_chunk_length = Gp_max_tuple_chunk_size;
if (natts == 0)
{
tcItem = getChunkFromCache(&pSerInfo->chunkCache);
if (tcItem == NULL)
{
ereport(FATAL, (errcode(ERRCODE_OUT_OF_MEMORY),
errmsg("Could not allocate space for first chunk item in new chunk list.")));
}
/* TC_EMTPY is just one chunk */
SetChunkType(tcItem->chunk_data, TC_EMPTY);
tcItem->chunk_length = TUPLE_CHUNK_HEADER_SIZE;
appendChunkToTCList(tcList, tcItem);
return;
}
tcItem = getChunkFromCache(&pSerInfo->chunkCache);
if (tcItem == NULL)
{
ereport(FATAL, (errcode(ERRCODE_OUT_OF_MEMORY),
errmsg("Could not allocate space for first chunk item in new chunk list.")));
}
/* assume that we'll take a single chunk */
SetChunkType(tcItem->chunk_data, TC_WHOLE);
tcItem->chunk_length = TUPLE_CHUNK_HEADER_SIZE;
appendChunkToTCList(tcList, tcItem);
AssertState(s_tupSerMemCtxt != NULL);
if (is_memtuple(gtuple))
{
MemTuple mtuple = (MemTuple) gtuple;
addByteStringToChunkList(tcList, (char *) mtuple, memtuple_get_size(mtuple), &pSerInfo->chunkCache);
addPadding(tcList, &pSerInfo->chunkCache, memtuple_get_size(mtuple));
}
else
{
HeapTuple tuple = (HeapTuple) gtuple;
HeapTupleHeader t_data = tuple->t_data;
TupSerHeader tsh;
unsigned int datalen;
unsigned int nullslen;
datalen = tuple->t_len - t_data->t_hoff;
if (HeapTupleHasNulls(tuple))
nullslen = BITMAPLEN(HeapTupleHeaderGetNatts(t_data));
else
nullslen = 0;
tsh.tuplen = sizeof(TupSerHeader) + TYPEALIGN(TUPLE_CHUNK_ALIGN,nullslen) + datalen;
tsh.natts = HeapTupleHeaderGetNatts(t_data);
tsh.infomask = t_data->t_infomask;
addByteStringToChunkList(tcList, (char *)&tsh, sizeof(TupSerHeader), &pSerInfo->chunkCache);
/* If we don't have any attributes which have been toasted, we
* can be very very simple: just send the raw data. */
if ((tsh.infomask & HEAP_HASEXTERNAL) == 0)
{
if (nullslen)
{
addByteStringToChunkList(tcList, (char *)t_data->t_bits, nullslen, &pSerInfo->chunkCache);
addPadding(tcList,&pSerInfo->chunkCache,nullslen);
}
addByteStringToChunkList(tcList, (char *)t_data + t_data->t_hoff, datalen, &pSerInfo->chunkCache);
addPadding(tcList,&pSerInfo->chunkCache,datalen);
}
else
{
/* We have to be more careful when we have tuples that
* have been toasted. Ideally we'd like to send the
* untoasted attributes in as "raw" a format as possible
* but that makes rebuilding the tuple harder .
*/
oldCtxt = MemoryContextSwitchTo(s_tupSerMemCtxt);
/* deconstruct the tuple (faster than a heap_getattr loop) */
heap_deform_tuple(tuple, tupdesc, pSerInfo->values, pSerInfo->nulls);
MemoryContextSwitchTo(oldCtxt);
/* Send the nulls character-array. */
addByteStringToChunkList(tcList, pSerInfo->nulls, natts, &pSerInfo->chunkCache);
addPadding(tcList,&pSerInfo->chunkCache,natts);
/*
* send the attributes of this tuple: NOTE anything which allocates
* temporary space (e.g. could result in a PG_DETOAST_DATUM) should be
* executed with the memory context set to s_tupSerMemCtxt
*/
for (i = 0; i < natts; ++i)
{
SerAttrInfo *attrInfo = pSerInfo->myinfo + i;
Datum origattr = pSerInfo->values[i],
attr;
/* skip null attributes (already taken care of above) */
if (pSerInfo->nulls[i])
continue;
if (attrInfo->typlen == -1)
{
int32 sz;
char *data;
/*
* If we have a toasted datum, forcibly detoast it here to avoid
* memory leakage: we want to force the detoast allocation(s) to
* happen in our reset-able serialization context.
*/
oldCtxt = MemoryContextSwitchTo(s_tupSerMemCtxt);
attr = PointerGetDatum(PG_DETOAST_DATUM_PACKED(origattr));
MemoryContextSwitchTo(oldCtxt);
sz = VARSIZE_ANY_EXHDR(attr);
data = VARDATA_ANY(attr);
/* Send length first, then data */
addInt32ToChunkList(tcList, sz, &pSerInfo->chunkCache);
addByteStringToChunkList(tcList, data, sz, &pSerInfo->chunkCache);
addPadding(tcList, &pSerInfo->chunkCache, sz);
}
else if (attrInfo->typlen == -2)
{
int32 sz;
char *data;
/* CString, we would send the string with the terminating '\0' */
data = DatumGetCString(origattr);
sz = strlen(data) + 1;
/* Send length first, then data */
addInt32ToChunkList(tcList, sz, &pSerInfo->chunkCache);
addByteStringToChunkList(tcList, data, sz, &pSerInfo->chunkCache);
addPadding(tcList, &pSerInfo->chunkCache, sz);
}
else if (attrInfo->typbyval)
{
/*
* We send a full-width Datum for all pass-by-value types, regardless of
* the actual size.
*/
addByteStringToChunkList(tcList, (char *) &origattr, sizeof(Datum), &pSerInfo->chunkCache);
addPadding(tcList, &pSerInfo->chunkCache, sizeof(Datum));
}
else
{
addByteStringToChunkList(tcList, DatumGetPointer(origattr), attrInfo->typlen, &pSerInfo->chunkCache);
addPadding(tcList, &pSerInfo->chunkCache, attrInfo->typlen);
attr = origattr;
}
}
MemoryContextReset(s_tupSerMemCtxt);
}
}
/*
* if we have more than 1 chunk we have to set the chunk types on our
* first chunk and last chunk
*/
if (tcList->num_chunks > 1)
{
TupleChunkListItem first,
last;
first = tcList->p_first;
last = tcList->p_last;
Assert(first != NULL);
Assert(first != last);
Assert(last != NULL);
SetChunkType(first->chunk_data, TC_PARTIAL_START);
SetChunkType(last->chunk_data, TC_PARTIAL_END);
/*
* any intervening chunks are already set to TC_PARTIAL_MID when
* allocated
*/
}
return;
}
int
main(int argc, char *argv[])
{
struct timeval start_t;
struct timeval stop_t;
int tmpfile,
i;
char *full_buf = (char *) malloc(XLOG_SEG_SIZE),
*buf;
char *filename = FSYNC_FILENAME;
if (argc > 2 && strcmp(argv[1], "-f") == 0)
{
filename = argv[2];
argv += 2;
argc -= 2;
}
if (argc > 1)
loops = atoi(argv[1]);
for (i = 0; i < XLOG_SEG_SIZE; i++)
full_buf[i] = random();
if ((tmpfile = open(filename, O_RDWR | O_CREAT, S_IRUSR | S_IWUSR, 0)) == -1)
die("Cannot open output file.");
if (write(tmpfile, full_buf, XLOG_SEG_SIZE) != XLOG_SEG_SIZE)
die("write failed");
/* fsync now so later fsync's don't have to do it */
if (fsync(tmpfile) != 0)
die("fsync failed");
close(tmpfile);
buf = (char *) TYPEALIGN(ALIGNOF_XLOG_BUFFER, full_buf);
printf("Loops = %d\n\n", loops);
/*
* Simple write
*/
printf("Simple write:\n");
printf(LABEL_FORMAT, "8k write");
fflush(stdout);
gettimeofday(&start_t, NULL);
for (i = 0; i < loops; i++)
{
if ((tmpfile = open(filename, O_RDWR, 0)) == -1)
die("Cannot open output file.");
if (write(tmpfile, buf, WRITE_SIZE) != WRITE_SIZE)
die("write failed");
close(tmpfile);
}
gettimeofday(&stop_t, NULL);
print_elapse(start_t, stop_t);
/*
* Compare file sync methods with one 8k write
*/
printf("\nCompare file sync methods using one write:\n");
#ifdef OPEN_DATASYNC_FLAG
printf(LABEL_FORMAT, "open_datasync 8k write");
fflush(stdout);
if ((tmpfile = open(filename, O_RDWR | O_DSYNC, 0)) == -1)
die("Cannot open output file.");
gettimeofday(&start_t, NULL);
for (i = 0; i < loops; i++)
{
if (write(tmpfile, buf, WRITE_SIZE) != WRITE_SIZE)
die("write failed");
if (lseek(tmpfile, 0, SEEK_SET) == -1)
die("seek failed");
}
gettimeofday(&stop_t, NULL);
close(tmpfile);
print_elapse(start_t, stop_t);
#else
printf("\t(unavailable: open_datasync)\n");
#endif
#ifdef OPEN_SYNC_FLAG
printf(LABEL_FORMAT, "open_sync 8k write");
fflush(stdout);
if ((tmpfile = open(filename, O_RDWR | OPEN_SYNC_FLAG, 0)) == -1)
die("Cannot open output file.");
gettimeofday(&start_t, NULL);
for (i = 0; i < loops; i++)
{
if (write(tmpfile, buf, WRITE_SIZE) != WRITE_SIZE)
die("write failed");
if (lseek(tmpfile, 0, SEEK_SET) == -1)
die("seek failed");
}
gettimeofday(&stop_t, NULL);
close(tmpfile);
print_elapse(start_t, stop_t);
#else
printf("\t(unavailable: open_sync)\n");
#endif
#ifdef HAVE_FDATASYNC
printf(LABEL_FORMAT, "8k write, fdatasync");
fflush(stdout);
if ((tmpfile = open(filename, O_RDWR, 0)) == -1)
die("Cannot open output file.");
gettimeofday(&start_t, NULL);
for (i = 0; i < loops; i++)
{
if (write(tmpfile, buf, WRITE_SIZE) != WRITE_SIZE)
die("write failed");
fdatasync(tmpfile);
if (lseek(tmpfile, 0, SEEK_SET) == -1)
die("seek failed");
}
gettimeofday(&stop_t, NULL);
close(tmpfile);
print_elapse(start_t, stop_t);
#else
printf("\t(unavailable: fdatasync)\n");
#endif
printf(LABEL_FORMAT, "8k write, fsync");
fflush(stdout);
if ((tmpfile = open(filename, O_RDWR, 0)) == -1)
die("Cannot open output file.");
gettimeofday(&start_t, NULL);
for (i = 0; i < loops; i++)
{
if (write(tmpfile, buf, WRITE_SIZE) != WRITE_SIZE)
die("write failed");
if (fsync(tmpfile) != 0)
die("fsync failed");
if (lseek(tmpfile, 0, SEEK_SET) == -1)
die("seek failed");
}
gettimeofday(&stop_t, NULL);
close(tmpfile);
print_elapse(start_t, stop_t);
/*
* Compare file sync methods with two 8k write
*/
printf("\nCompare file sync methods using two writes:\n");
#ifdef OPEN_DATASYNC_FLAG
printf(LABEL_FORMAT, "2 open_datasync 8k writes");
fflush(stdout);
if ((tmpfile = open(filename, O_RDWR | O_DSYNC, 0)) == -1)
die("Cannot open output file.");
gettimeofday(&start_t, NULL);
for (i = 0; i < loops; i++)
{
if (write(tmpfile, buf, WRITE_SIZE) != WRITE_SIZE)
die("write failed");
if (write(tmpfile, buf, WRITE_SIZE) != WRITE_SIZE)
die("write failed");
if (lseek(tmpfile, 0, SEEK_SET) == -1)
die("seek failed");
}
gettimeofday(&stop_t, NULL);
close(tmpfile);
print_elapse(start_t, stop_t);
#else
printf("\t(unavailable: open_datasync)\n");
#endif
#ifdef OPEN_SYNC_FLAG
printf(LABEL_FORMAT, "2 open_sync 8k writes");
fflush(stdout);
if ((tmpfile = open(filename, O_RDWR | OPEN_SYNC_FLAG, 0)) == -1)
die("Cannot open output file.");
gettimeofday(&start_t, NULL);
for (i = 0; i < loops; i++)
{
if (write(tmpfile, buf, WRITE_SIZE) != WRITE_SIZE)
die("write failed");
if (write(tmpfile, buf, WRITE_SIZE) != WRITE_SIZE)
die("write failed");
if (lseek(tmpfile, 0, SEEK_SET) == -1)
die("seek failed");
}
gettimeofday(&stop_t, NULL);
close(tmpfile);
print_elapse(start_t, stop_t);
#endif
#ifdef HAVE_FDATASYNC
printf(LABEL_FORMAT, "8k write, 8k write, fdatasync");
fflush(stdout);
if ((tmpfile = open(filename, O_RDWR, 0)) == -1)
die("Cannot open output file.");
gettimeofday(&start_t, NULL);
for (i = 0; i < loops; i++)
{
if (write(tmpfile, buf, WRITE_SIZE) != WRITE_SIZE)
die("write failed");
if (write(tmpfile, buf, WRITE_SIZE) != WRITE_SIZE)
die("write failed");
fdatasync(tmpfile);
if (lseek(tmpfile, 0, SEEK_SET) == -1)
die("seek failed");
}
gettimeofday(&stop_t, NULL);
close(tmpfile);
print_elapse(start_t, stop_t);
#else
printf("\t(unavailable: fdatasync)\n");
#endif
printf(LABEL_FORMAT, "8k write, 8k write, fsync");
fflush(stdout);
if ((tmpfile = open(filename, O_RDWR, 0)) == -1)
die("Cannot open output file.");
gettimeofday(&start_t, NULL);
for (i = 0; i < loops; i++)
{
if (write(tmpfile, buf, WRITE_SIZE) != WRITE_SIZE)
die("write failed");
if (write(tmpfile, buf, WRITE_SIZE) != WRITE_SIZE)
die("write failed");
if (fsync(tmpfile) != 0)
die("fsync failed");
if (lseek(tmpfile, 0, SEEK_SET) == -1)
die("seek failed");
}
gettimeofday(&stop_t, NULL);
close(tmpfile);
print_elapse(start_t, stop_t);
/*
* Compare 1 to 2 writes
*/
printf("\nCompare open_sync with different sizes:\n");
#ifdef OPEN_SYNC_FLAG
printf(LABEL_FORMAT, "open_sync 16k write");
fflush(stdout);
if ((tmpfile = open(filename, O_RDWR | OPEN_SYNC_FLAG, 0)) == -1)
die("Cannot open output file.");
gettimeofday(&start_t, NULL);
for (i = 0; i < loops; i++)
{
if (write(tmpfile, buf, WRITE_SIZE * 2) != WRITE_SIZE * 2)
die("write failed");
if (lseek(tmpfile, 0, SEEK_SET) == -1)
die("seek failed");
}
gettimeofday(&stop_t, NULL);
close(tmpfile);
print_elapse(start_t, stop_t);
printf(LABEL_FORMAT, "2 open_sync 8k writes");
fflush(stdout);
if ((tmpfile = open(filename, O_RDWR | OPEN_SYNC_FLAG, 0)) == -1)
die("Cannot open output file.");
gettimeofday(&start_t, NULL);
for (i = 0; i < loops; i++)
{
if (write(tmpfile, buf, WRITE_SIZE) != WRITE_SIZE)
die("write failed");
if (write(tmpfile, buf, WRITE_SIZE) != WRITE_SIZE)
die("write failed");
if (lseek(tmpfile, 0, SEEK_SET) == -1)
die("seek failed");
}
gettimeofday(&stop_t, NULL);
close(tmpfile);
print_elapse(start_t, stop_t);
#else
printf("\t(unavailable: open_sync)\n");
#endif
/*
* Fsync another file descriptor?
*/
printf("\nTest if fsync on non-write file descriptor is honored:\n");
printf("(If the times are similar, fsync() can sync data written\n");
printf("on a different descriptor.)\n");
printf(LABEL_FORMAT, "8k write, fsync, close");
fflush(stdout);
gettimeofday(&start_t, NULL);
for (i = 0; i < loops; i++)
{
if ((tmpfile = open(filename, O_RDWR, 0)) == -1)
die("Cannot open output file.");
if (write(tmpfile, buf, WRITE_SIZE) != WRITE_SIZE)
die("write failed");
if (fsync(tmpfile) != 0)
die("fsync failed");
close(tmpfile);
if ((tmpfile = open(filename, O_RDWR, 0)) == -1)
die("Cannot open output file.");
/* do nothing but the open/close the tests are consistent. */
close(tmpfile);
}
gettimeofday(&stop_t, NULL);
print_elapse(start_t, stop_t);
printf(LABEL_FORMAT, "8k write, close, fsync");
fflush(stdout);
gettimeofday(&start_t, NULL);
for (i = 0; i < loops; i++)
{
if ((tmpfile = open(filename, O_RDWR, 0)) == -1)
die("Cannot open output file.");
if (write(tmpfile, buf, WRITE_SIZE) != WRITE_SIZE)
die("write failed");
close(tmpfile);
/* reopen file */
if ((tmpfile = open(filename, O_RDWR, 0)) == -1)
die("Cannot open output file.");
if (fsync(tmpfile) != 0)
die("fsync failed");
close(tmpfile);
}
gettimeofday(&stop_t, NULL);
print_elapse(start_t, stop_t);
/* cleanup */
free(full_buf);
unlink(filename);
return 0;
}
