From fd0366e1b57df0bb0ad30140acf2c4225cae017b Mon Sep 17 00:00:00 2001
From: Tom Lane <tgl@sss.pgh.pa.us>
Date: Wed, 18 Nov 1998 00:47:28 +0000
Subject: [PATCH] Performance improvement for libpq: avoid calling malloc
separately for each field of each tuple. Makes more difference than you'd
think...
---
src/interfaces/libpq/fe-exec.c | 335 ++++++++++++++++++++++---------
src/interfaces/libpq/libpq-int.h | 64 ++++--
2 files changed, 283 insertions(+), 116 deletions(-)
diff --git a/src/interfaces/libpq/fe-exec.c b/src/interfaces/libpq/fe-exec.c
index 82c697ef05b..e6742a4769f 100644
--- a/src/interfaces/libpq/fe-exec.c
+++ b/src/interfaces/libpq/fe-exec.c
@@ -7,7 +7,7 @@
*
*
* IDENTIFICATION
- * $Header: /cvsroot/pgsql/src/interfaces/libpq/fe-exec.c,v 1.69 1998/10/01 01:40:21 tgl Exp $
+ * $Header: /cvsroot/pgsql/src/interfaces/libpq/fe-exec.c,v 1.70 1998/11/18 00:47:28 tgl Exp $
*
*-------------------------------------------------------------------------
*/
@@ -28,10 +28,6 @@
#include <ctype.h>
-/* the rows array in a PGresGroup has to grow to accommodate the rows */
-/* returned. Each time, we grow by this much: */
-#define TUPARR_GROW_BY 100
-
/* keep this in same order as ExecStatusType in libpq-fe.h */
const char *const pgresStatus[] = {
"PGRES_EMPTY_QUERY",
@@ -49,7 +45,6 @@ const char *const pgresStatus[] = {
((*(conn)->noticeHook) ((conn)->noticeArg, (message)))
-static void freeTuple(PGresAttValue *tuple, int numAttributes);
static int addTuple(PGresult *res, PGresAttValue *tup);
static void parseInput(PGconn *conn);
static int getRowDescriptions(PGconn *conn);
@@ -58,6 +53,63 @@ static int getNotify(PGconn *conn);
static int getNotice(PGconn *conn);
+/* ----------------
+ * Space management for PGresult.
+ *
+ * Formerly, libpq did a separate malloc() for each field of each tuple
+ * returned by a query. This was remarkably expensive --- malloc/free
+ * consumed a sizable part of the application's runtime. And there is
+ * no real need to keep track of the fields separately, since they will
+ * all be freed together when the PGresult is released. So now, we grab
+ * large blocks of storage from malloc and allocate space for query data
+ * within these blocks, using a trivially simple allocator. This reduces
+ * the number of malloc/free calls dramatically, and it also avoids
+ * fragmentation of the malloc storage arena.
+ * The PGresult structure itself is still malloc'd separately. We could
+ * combine it with the first allocation block, but that would waste space
+ * for the common case that no extra storage is actually needed (that is,
+ * the SQL command did not return tuples).
+ * We also malloc the top-level array of tuple pointers separately, because
+ * we need to be able to enlarge it via realloc, and our trivial space
+ * allocator doesn't handle that effectively. (Too bad the FE/BE protocol
+ * doesn't tell us up front how many tuples will be returned.)
+ * All other subsidiary storage for a PGresult is kept in PGresult_data blocks
+ * of size PGRESULT_DATA_BLOCKSIZE. The overhead at the start of each block
+ * is just a link to the next one, if any. Free-space management info is
+ * kept in the owning PGresult.
+ * A query returning a small amount of data will thus require three malloc
+ * calls: one for the PGresult, one for the tuples pointer array, and one
+ * PGresult_data block.
+ * Only the most recently allocated PGresult_data block is a candidate to
+ * have more stuff added to it --- any extra space left over in older blocks
+ * is wasted. We could be smarter and search the whole chain, but the point
+ * here is to be simple and fast. Typical applications do not keep a PGresult
+ * around very long anyway, so some wasted space within one is not a problem.
+ *
+ * Tuning constants for the space allocator are:
+ * PGRESULT_DATA_BLOCKSIZE: size of a standard allocation block, in bytes
+ * PGRESULT_ALIGN_BOUNDARY: assumed alignment requirement for binary data
+ * PGRESULT_SEP_ALLOC_THRESHOLD: objects bigger than this are given separate
+ * blocks, instead of being crammed into a regular allocation block.
+ * Requirements for correct function are:
+ * PGRESULT_ALIGN_BOUNDARY >= sizeof(pointer)
+ * to ensure the initial pointer in a block is not overwritten.
+ * PGRESULT_ALIGN_BOUNDARY must be a multiple of the alignment requirements
+ * of all machine data types.
+ * PGRESULT_SEP_ALLOC_THRESHOLD + PGRESULT_ALIGN_BOUNDARY <=
+ * PGRESULT_DATA_BLOCKSIZE
+ * pqResultAlloc assumes an object smaller than the threshold will fit
+ * in a new block.
+ * The amount of space wasted at the end of a block could be as much as
+ * PGRESULT_SEP_ALLOC_THRESHOLD, so it doesn't pay to make that too large.
+ * ----------------
+ */
+
+#define PGRESULT_DATA_BLOCKSIZE 2048
+#define PGRESULT_ALIGN_BOUNDARY 16 /* 8 is probably enough, really */
+#define PGRESULT_SEP_ALLOC_THRESHOLD (PGRESULT_DATA_BLOCKSIZE / 2)
+
+
/*
* PQmakeEmptyPGresult
* returns a newly allocated, initialized PGresult with given status.
@@ -76,7 +128,7 @@ PQmakeEmptyPGresult(PGconn *conn, ExecStatusType status)
result = (PGresult *) malloc(sizeof(PGresult));
- result->conn = conn; /* should go away eventually */
+ result->conn = conn; /* might be NULL */
result->ntups = 0;
result->numAttributes = 0;
result->attDescs = NULL;
@@ -86,6 +138,11 @@ PQmakeEmptyPGresult(PGconn *conn, ExecStatusType status)
result->cmdStatus[0] = '\0';
result->binary = 0;
result->errMsg = NULL;
+ result->null_field[0] = '\0';
+ result->curBlock = NULL;
+ result->curOffset = 0;
+ result->spaceLeft = 0;
+
if (conn) /* consider copying conn's errorMessage */
{
switch (status)
@@ -105,6 +162,117 @@ PQmakeEmptyPGresult(PGconn *conn, ExecStatusType status)
return result;
}
+/*
+ * pqResultAlloc -
+ * Allocate subsidiary storage for a PGresult.
+ *
+ * nBytes is the amount of space needed for the object.
+ * If isBinary is true, we assume that we need to align the object on
+ * a machine allocation boundary.
+ * If isBinary is false, we assume the object is a char string and can
+ * be allocated on any byte boundary.
+ */
+void *
+pqResultAlloc(PGresult *res, int nBytes, int isBinary)
+{
+ char *space;
+ PGresult_data *block;
+
+ if (! res)
+ return NULL;
+
+ if (nBytes <= 0)
+ return res->null_field;
+
+ /* If alignment is needed, round up the current position to an
+ * alignment boundary.
+ */
+ if (isBinary)
+ {
+ int offset = res->curOffset % PGRESULT_ALIGN_BOUNDARY;
+ if (offset)
+ {
+ res->curOffset += PGRESULT_ALIGN_BOUNDARY - offset;
+ res->spaceLeft -= PGRESULT_ALIGN_BOUNDARY - offset;
+ }
+ }
+
+ /* If there's enough space in the current block, no problem. */
+ if (nBytes <= res->spaceLeft)
+ {
+ space = res->curBlock->space + res->curOffset;
+ res->curOffset += nBytes;
+ res->spaceLeft -= nBytes;
+ return space;
+ }
+
+ /* If the requested object is very large, give it its own block; this
+ * avoids wasting what might be most of the current block to start a new
+ * block. (We'd have to special-case requests bigger than the block size
+ * anyway.) The object is always given binary alignment in this case.
+ */
+ if (nBytes >= PGRESULT_SEP_ALLOC_THRESHOLD)
+ {
+ block = (PGresult_data *) malloc(nBytes + PGRESULT_ALIGN_BOUNDARY);
+ if (! block)
+ return NULL;
+ space = block->space + PGRESULT_ALIGN_BOUNDARY;
+ if (res->curBlock)
+ {
+ /* Tuck special block below the active block, so that we don't
+ * have to waste the free space in the active block.
+ */
+ block->next = res->curBlock->next;
+ res->curBlock->next = block;
+ }
+ else
+ {
+ /* Must set up the new block as the first active block. */
+ block->next = NULL;
+ res->curBlock = block;
+ res->spaceLeft = 0; /* be sure it's marked full */
+ }
+ return space;
+ }
+
+ /* Otherwise, start a new block. */
+ block = (PGresult_data *) malloc(PGRESULT_DATA_BLOCKSIZE);
+ if (! block)
+ return NULL;
+ block->next = res->curBlock;
+ res->curBlock = block;
+ if (isBinary)
+ {
+ /* object needs full alignment */
+ res->curOffset = PGRESULT_ALIGN_BOUNDARY;
+ res->spaceLeft = PGRESULT_DATA_BLOCKSIZE - PGRESULT_ALIGN_BOUNDARY;
+ }
+ else
+ {
+ /* we can cram it right after the overhead pointer */
+ res->curOffset = sizeof(PGresult_data);
+ res->spaceLeft = PGRESULT_DATA_BLOCKSIZE - sizeof(PGresult_data);
+ }
+
+ space = block->space + res->curOffset;
+ res->curOffset += nBytes;
+ res->spaceLeft -= nBytes;
+ return space;
+}
+
+/*
+ * pqResultStrdup -
+ * Like strdup, but the space is subsidiary PGresult space.
+ */
+char *
+pqResultStrdup(PGresult *res, const char *str)
+{
+ char *space = (char*) pqResultAlloc(res, strlen(str)+1, FALSE);
+ if (space)
+ strcpy(space, str);
+ return space;
+}
+
/*
* pqSetResultError -
* assign a new error message to a PGresult
@@ -114,11 +282,10 @@ pqSetResultError(PGresult *res, const char *msg)
{
if (!res)
return;
- if (res->errMsg)
- free(res->errMsg);
- res->errMsg = NULL;
if (msg && *msg)
- res->errMsg = strdup(msg);
+ res->errMsg = pqResultStrdup(res, msg);
+ else
+ res->errMsg = NULL;
}
/*
@@ -128,58 +295,25 @@ pqSetResultError(PGresult *res, const char *msg)
void
PQclear(PGresult *res)
{
- int i;
+ PGresult_data *block;
if (!res)
return;
- /* free all the rows */
- if (res->tuples)
- {
- for (i = 0; i < res->ntups; i++)
- freeTuple(res->tuples[i], res->numAttributes);
- free(res->tuples);
- }
-
- /* free all the attributes */
- if (res->attDescs)
- {
- for (i = 0; i < res->numAttributes; i++)
- {
- if (res->attDescs[i].name)
- free(res->attDescs[i].name);
- }
- free(res->attDescs);
+ /* Free all the subsidiary blocks */
+ while ((block = res->curBlock) != NULL) {
+ res->curBlock = block->next;
+ free(block);
}
- /* free the error text */
- if (res->errMsg)
- free(res->errMsg);
+ /* Free the top-level tuple pointer array */
+ if (res->tuples)
+ free(res->tuples);
- /* free the structure itself */
+ /* Free the PGresult structure itself */
free(res);
}
-/*
- * Free a single tuple structure.
- */
-
-static void
-freeTuple(PGresAttValue *tuple, int numAttributes)
-{
- int i;
-
- if (tuple)
- {
- for (i = 0; i < numAttributes; i++)
- {
- if (tuple[i].value)
- free(tuple[i].value);
- }
- free(tuple);
- }
-}
-
/*
* Handy subroutine to deallocate any partially constructed async result.
*/
@@ -187,13 +321,8 @@ freeTuple(PGresAttValue *tuple, int numAttributes)
void
pqClearAsyncResult(PGconn *conn)
{
- /* Get rid of incomplete result and any not-yet-added tuple */
if (conn->result)
- {
- if (conn->curTuple)
- freeTuple(conn->curTuple, conn->result->numAttributes);
PQclear(conn->result);
- }
conn->result = NULL;
conn->curTuple = NULL;
}
@@ -201,7 +330,7 @@ pqClearAsyncResult(PGconn *conn)
/*
* addTuple
- * add a row to the PGresult structure, growing it if necessary
+ * add a row pointer to the PGresult structure, growing it if necessary
* Returns TRUE if OK, FALSE if not enough memory to add the row
*/
static int
@@ -220,7 +349,7 @@ addTuple(PGresult *res, PGresAttValue *tup)
* Note that the positions beyond res->ntups are garbage, not
* necessarily NULL.
*/
- int newSize = res->tupArrSize + TUPARR_GROW_BY;
+ int newSize = (res->tupArrSize > 0) ? res->tupArrSize * 2 : 128;
PGresAttValue ** newTuples = (PGresAttValue **)
realloc(res->tuples, newSize * sizeof(PGresAttValue *));
if (! newTuples)
@@ -564,7 +693,7 @@ getRowDescriptions(PGconn *conn)
if (nfields > 0)
{
result->attDescs = (PGresAttDesc *)
- malloc(nfields * sizeof(PGresAttDesc));
+ pqResultAlloc(result, nfields * sizeof(PGresAttDesc), TRUE);
MemSet((char *) result->attDescs, 0, nfields * sizeof(PGresAttDesc));
}
@@ -574,7 +703,7 @@ getRowDescriptions(PGconn *conn)
char typName[MAX_MESSAGE_LEN];
int typid;
int typlen;
- int atttypmod = -1;
+ int atttypmod;
if (pqGets(typName, MAX_MESSAGE_LEN, conn) ||
pqGetInt(&typid, 4, conn) ||
@@ -594,7 +723,7 @@ getRowDescriptions(PGconn *conn)
*/
if (typlen == 0xFFFF)
typlen = -1;
- result->attDescs[i].name = strdup(typName);
+ result->attDescs[i].name = pqResultStrdup(result, typName);
result->attDescs[i].typid = typid;
result->attDescs[i].typlen = typlen;
result->attDescs[i].atttypmod = atttypmod;
@@ -618,7 +747,8 @@ getRowDescriptions(PGconn *conn)
static int
getAnotherTuple(PGconn *conn, int binary)
{
- int nfields = conn->result->numAttributes;
+ PGresult *result = conn->result;
+ int nfields = result->numAttributes;
PGresAttValue *tup;
char bitmap[MAX_FIELDS]; /* the backend sends us a bitmap
* of which attributes are null */
@@ -629,13 +759,13 @@ getAnotherTuple(PGconn *conn, int binary)
int bitcnt; /* number of bits examined in current byte */
int vlen; /* length of the current field value */
- conn->result->binary = binary;
+ result->binary = binary;
/* Allocate tuple space if first time for this data message */
if (conn->curTuple == NULL)
{
conn->curTuple = (PGresAttValue *)
- malloc(nfields * sizeof(PGresAttValue));
+ pqResultAlloc(result, nfields * sizeof(PGresAttValue), TRUE);
if (conn->curTuple == NULL)
goto outOfMemory;
MemSet((char *) conn->curTuple, 0, nfields * sizeof(PGresAttValue));
@@ -670,12 +800,7 @@ getAnotherTuple(PGconn *conn, int binary)
if (!(bmap & 0200))
{
/* if the field value is absent, make it a null string */
- if (tup[i].value == NULL)
- {
- tup[i].value = strdup("");
- if (tup[i].value == NULL)
- goto outOfMemory;
- }
+ tup[i].value = result->null_field;
tup[i].len = NULL_LEN;
}
else
@@ -689,7 +814,7 @@ getAnotherTuple(PGconn *conn, int binary)
vlen = 0;
if (tup[i].value == NULL)
{
- tup[i].value = (char *) malloc(vlen + 1);
+ tup[i].value = (char *) pqResultAlloc(result, vlen+1, binary);
if (tup[i].value == NULL)
goto outOfMemory;
}
@@ -714,14 +839,8 @@ getAnotherTuple(PGconn *conn, int binary)
}
/* Success! Store the completed tuple in the result */
- if (! addTuple(conn->result, tup))
- {
- /* Oops, not enough memory to add the tuple to conn->result,
- * so must free it ourselves...
- */
- freeTuple(tup, nfields);
+ if (! addTuple(result, tup))
goto outOfMemory;
- }
/* and reset for a new message */
conn->curTuple = NULL;
return 0;
@@ -1437,10 +1556,13 @@ check_field_number(const char *routineName, PGresult *res, int field_num)
return FALSE; /* no way to display error message... */
if (field_num < 0 || field_num >= res->numAttributes)
{
- sprintf(res->conn->errorMessage,
- "%s: ERROR! field number %d is out of range 0..%d\n",
- routineName, field_num, res->numAttributes - 1);
- DONOTICE(res->conn, res->conn->errorMessage);
+ if (res->conn)
+ {
+ sprintf(res->conn->errorMessage,
+ "%s: ERROR! field number %d is out of range 0..%d\n",
+ routineName, field_num, res->numAttributes - 1);
+ DONOTICE(res->conn, res->conn->errorMessage);
+ }
return FALSE;
}
return TRUE;
@@ -1454,18 +1576,24 @@ check_tuple_field_number(const char *routineName, PGresult *res,
return FALSE; /* no way to display error message... */
if (tup_num < 0 || tup_num >= res->ntups)
{
- sprintf(res->conn->errorMessage,
- "%s: ERROR! tuple number %d is out of range 0..%d\n",
- routineName, tup_num, res->ntups - 1);
- DONOTICE(res->conn, res->conn->errorMessage);
+ if (res->conn)
+ {
+ sprintf(res->conn->errorMessage,
+ "%s: ERROR! tuple number %d is out of range 0..%d\n",
+ routineName, tup_num, res->ntups - 1);
+ DONOTICE(res->conn, res->conn->errorMessage);
+ }
return FALSE;
}
if (field_num < 0 || field_num >= res->numAttributes)
{
- sprintf(res->conn->errorMessage,
- "%s: ERROR! field number %d is out of range 0..%d\n",
- routineName, field_num, res->numAttributes - 1);
- DONOTICE(res->conn, res->conn->errorMessage);
+ if (res->conn)
+ {
+ sprintf(res->conn->errorMessage,
+ "%s: ERROR! field number %d is out of range 0..%d\n",
+ routineName, field_num, res->numAttributes - 1);
+ DONOTICE(res->conn, res->conn->errorMessage);
+ }
return FALSE;
}
return TRUE;
@@ -1635,10 +1763,13 @@ PQcmdTuples(PGresult *res)
if (*p == 0)
{
- sprintf(res->conn->errorMessage,
- "PQcmdTuples (%s) -- bad input from server\n",
- res->cmdStatus);
- DONOTICE(res->conn, res->conn->errorMessage);
+ if (res->conn)
+ {
+ sprintf(res->conn->errorMessage,
+ "PQcmdTuples (%s) -- bad input from server\n",
+ res->cmdStatus);
+ DONOTICE(res->conn, res->conn->errorMessage);
+ }
return "";
}
p++;
@@ -1648,9 +1779,12 @@ PQcmdTuples(PGresult *res)
p++; /* INSERT: skip oid */
if (*p == 0)
{
- sprintf(res->conn->errorMessage,
- "PQcmdTuples (INSERT) -- there's no # of tuples\n");
- DONOTICE(res->conn, res->conn->errorMessage);
+ if (res->conn)
+ {
+ sprintf(res->conn->errorMessage,
+ "PQcmdTuples (INSERT) -- there's no # of tuples\n");
+ DONOTICE(res->conn, res->conn->errorMessage);
+ }
return "";
}
p++;
@@ -1680,7 +1814,8 @@ PQgetvalue(PGresult *res, int tup_num, int field_num)
/* PQgetlength:
returns the length of a field value in bytes. If res is binary,
i.e. a result of a binary portal, then the length returned does
- NOT include the size field of the varlena.
+ NOT include the size field of the varlena. (The data returned
+ by PQgetvalue doesn't either.)
*/
int
PQgetlength(PGresult *res, int tup_num, int field_num)
diff --git a/src/interfaces/libpq/libpq-int.h b/src/interfaces/libpq/libpq-int.h
index e05ea4bebff..76d21b07c9c 100644
--- a/src/interfaces/libpq/libpq-int.h
+++ b/src/interfaces/libpq/libpq-int.h
@@ -11,7 +11,7 @@
*
* Copyright (c) 1994, Regents of the University of California
*
- * $Id: libpq-int.h,v 1.4 1998/10/01 01:40:25 tgl Exp $
+ * $Id: libpq-int.h,v 1.5 1998/11/18 00:47:26 tgl Exp $
*
*-------------------------------------------------------------------------
*/
@@ -49,12 +49,29 @@
#define ERROR_MSG_LENGTH 4096
#define CMDSTATUS_LEN 40
-/* PGresult and the subsidiary types PGresAttDesc, PGresAttValue
+/*
+ * PGresult and the subsidiary types PGresAttDesc, PGresAttValue
* represent the result of a query (or more precisely, of a single SQL
* command --- a query string given to PQexec can contain multiple commands).
* Note we assume that a single command can return at most one tuple group,
* hence there is no need for multiple descriptor sets.
*/
+
+/* Subsidiary-storage management structure for PGresult.
+ * See space management routines in fe-exec.c for details.
+ * Note that space[k] refers to the k'th byte starting from the physical
+ * head of the block.
+ */
+ typedef union pgresult_data PGresult_data;
+
+ union pgresult_data
+ {
+ PGresult_data *next; /* link to next block, or NULL */
+ char space[1]; /* dummy for accessing block as bytes */
+ };
+
+/* Data about a single attribute (column) of a query result */
+
typedef struct pgresAttDesc
{
char *name; /* type name */
@@ -63,11 +80,20 @@
int atttypmod; /* type-specific modifier info */
} PGresAttDesc;
-/* use char* for Attribute values,
- ASCII tuples are guaranteed to be null-terminated
- For binary tuples, the first four bytes of the value is the size,
- and the bytes afterwards are the value. The binary value is
- not guaranteed to be null-terminated. In fact, it can have embedded nulls
+/* Data for a single attribute of a single tuple */
+
+/* We use char* for Attribute values.
+ The value pointer always points to a null-terminated area; we add a
+ null (zero) byte after whatever the backend sends us. This is only
+ particularly useful for ASCII tuples ... with a binary value, the
+ value might have embedded nulls, so the application can't use C string
+ operators on it. But we add a null anyway for consistency.
+ Note that the value itself does not contain a length word.
+
+ A NULL attribute is a special case in two ways: its len field is NULL_LEN
+ and its value field points to null_field in the owning PGresult. All the
+ NULL attributes in a query result point to the same place (there's no need
+ to store a null string separately for each one).
*/
#define NULL_LEN (-1) /* pg_result len for NULL value */
@@ -75,7 +101,7 @@
typedef struct pgresAttValue
{
int len; /* length in bytes of the value */
- char *value; /* actual value */
+ char *value; /* actual value, plus terminating zero byte */
} PGresAttValue;
struct pg_result
@@ -91,15 +117,19 @@
* last insert query */
int binary; /* binary tuple values if binary == 1,
* otherwise ASCII */
- /* NOTE: conn is kept here only for the temporary convenience of
- * applications that rely on it being here. It will go away in a
- * future release, because relying on it is a bad idea --- what if
- * the PGresult has outlived the PGconn? About the only thing it was
- * really good for was fetching the errorMessage, and we stash that
- * here now anyway.
- */
- PGconn *conn; /* connection we did the query on */
+ PGconn *conn; /* connection we did the query on, if any */
char *errMsg; /* error message, or NULL if no error */
+
+ /* All NULL attributes in the query result point to this null string */
+ char null_field[1];
+
+ /* Space management information. Note that attDescs and errMsg,
+ * if not null, point into allocated blocks. But tuples points
+ * to a separately malloc'd block, so that we can realloc it.
+ */
+ PGresult_data *curBlock; /* most recently allocated block */
+ int curOffset; /* start offset of free space in block */
+ int spaceLeft; /* number of free bytes remaining in block */
};
/* PGAsyncStatusType defines the state of the query-execution state machine */
@@ -202,6 +232,8 @@ extern int pqPacketSend(PGconn *conn, const char *buf, size_t len);
/* === in fe-exec.c === */
extern void pqSetResultError(PGresult *res, const char *msg);
+extern void * pqResultAlloc(PGresult *res, int nBytes, int isBinary);
+extern char * pqResultStrdup(PGresult *res, const char *str);
extern void pqClearAsyncResult(PGconn *conn);
/* === in fe-misc.c === */
--
GitLab