Re: pgbench client-side performance issue on large scripts - Mailing list pgsql-hackers

"Daniel Verite" <daniel@manitou-mail.org> writes:
> For the moment I'll stay with my quick fix, then l'll try
> to come up with something to replace expr_scanner_get_lineno() .

I got nerd-sniped by this question and spent some time looking into
it.  ParseScript has got worse problems than just being slow: it's
actively buggy.  Notice that start_offset is set only once before
entering the loop, and doesn't change thereafter.  How is it that
we're getting sane line numbers at all?  The reason is that (1) if
we've not called yylex() at all yet, expr_scanner_offset() gives the
distance to the end of the string, since the yytext-ending NUL it's
looking for isn't there yet; and (2) expr_scanner_get_lineno() treats
the given start_offset as an upper bound, and won't complain if it
finds the NUL earlier than that.  So it gave the desired
line-number-of-the-current-token on all iterations after the first,
but on the first time through we get the line number of the script
end.  You can only see that in the case of \gset as the first command,
and I guess nobody noticed it yet.

Furthermore, it's not only ParseScript that's got O(N^2) problems;
so does process_backslash_command.  Your test case didn't show that
up, but a test with 50K backslash commands would.  We were actually
doing a strlen() of the whole string for each word of a backslash
command.  strlen() is likely faster than expr_scanner_get_lineno(),
but it's not so fast that O(N^2) effects don't matter.

The attached patch gets rid of both expr_scanner_offset() and
expr_scanner_get_lineno() altogether, in favor of using a new
function I added to psqlscan.l.  That uses the idea from plpgsql
of tracking the last-detected line end so that we don't have to
rescan prior lines over and over.  On my machine, parsing 50K-line
scripts goes from more than 10 seconds to perhaps 50 ms.

            regards, tom lane

From 542c081fd2fe2b60713b6fabbc59598c449f4c9e Mon Sep 17 00:00:00 2001
From: Tom Lane <tgl@sss.pgh.pa.us>
Date: Tue, 25 Feb 2025 16:30:48 -0500
Subject: [PATCH v1] Get rid of O(N^2) script-parsing overhead in pgbench.

pgbench wants to record the starting line number of each command
in its scripts.  It was computing that by scanning from the script
start and counting newlines, so that O(N^2) work had to be done
for an N-command script.  In a script with 50K lines, this adds
up to about 10 seconds on my machine.

To add insult to injury, the results were subtly wrong, because
expr_scanner_offset() scanned to find the NUL that flex inserts
at the end of the current token --- and before the first yylex
call, no such NUL has been inserted.  So we ended by computing the
script's last line number not its first one.  This was visible only
in case of \gset at the start of a script, which perhaps accounts
for the lack of complaints.

To fix, steal an idea from plpgsql and track the most recently
identified line start as we advance through the script.  This
is potentially still O(N^2) with sufficiently long lines, but
that seems unlikely to be a problem in practice.  Also adjust
a couple of other places that were invoking scans from script
start when they didn't really need to.  I made a new psqlscan
function psql_scan_get_location() that replaces both
expr_scanner_offset() and expr_scanner_get_lineno(), since
in practice expr_scanner_get_lineno() was only being invoked
to find the line number of the current parsing offset.

Reported-by: Daniel Verite <daniel@manitou-mail.org>
Author: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://postgr.es/m/84a8a89e-adb8-47a9-9d34-c13f7150ee45@manitou-mail.org
---
 src/bin/pgbench/exprscan.l          | 74 ++++++++++-------------------
 src/bin/pgbench/pgbench.c           | 14 +++---
 src/bin/pgbench/pgbench.h           |  4 +-
 src/fe_utils/psqlscan.l             | 54 +++++++++++++++++++++
 src/include/fe_utils/psqlscan.h     |  3 ++
 src/include/fe_utils/psqlscan_int.h |  4 ++
 6 files changed, 93 insertions(+), 60 deletions(-)

diff --git a/src/bin/pgbench/exprscan.l b/src/bin/pgbench/exprscan.l
index 8943a52e9f0..b4800937691 100644
--- a/src/bin/pgbench/exprscan.l
+++ b/src/bin/pgbench/exprscan.l
@@ -271,10 +271,14 @@ void
 expr_yyerror_more(yyscan_t yyscanner, const char *message, const char *more)
 {
     PsqlScanState state = yyget_extra(yyscanner);
-    int            error_detection_offset = expr_scanner_offset(state) - 1;
+    int            lineno;
+    int            error_detection_offset;
     YYSTYPE        lval;
     char       *full_line;

+    psql_scan_get_location(state, &lineno, &error_detection_offset);
+    error_detection_offset--;
+
     /*
      * While parsing an expression, we may not have collected the whole line
      * yet from the input source.  Lex till EOL so we can report whole line.
@@ -289,7 +293,6 @@ expr_yyerror_more(yyscan_t yyscanner, const char *message, const char *more)
     /* Extract the line, trimming trailing newline if any */
     full_line = expr_scanner_get_substring(state,
                                            expr_start_offset,
-                                           expr_scanner_offset(state),
                                            true);

     syntax_error(expr_source, expr_lineno, full_line, expr_command,
@@ -336,12 +339,15 @@ expr_lex_one_word(PsqlScanState state, PQExpBuffer word_buf, int *offset)
     /* And lex. */
     lexresult = yylex(&lval, state->scanner);

-    /*
-     * Save start offset of word, if any.  We could do this more efficiently,
-     * but for now this seems fine.
-     */
+    /* Save start offset of word, if any. */
     if (lexresult)
-        *offset = expr_scanner_offset(state) - word_buf->len;
+    {
+        int            lineno;
+        int            end_offset;
+
+        psql_scan_get_location(state, &lineno, &end_offset);
+        *offset = end_offset - word_buf->len;
+    }
     else
         *offset = -1;

@@ -404,65 +410,35 @@ expr_scanner_finish(yyscan_t yyscanner)
 }

 /*
- * Get offset from start of string to end of current lexer token.
+ * Get a malloc'd copy of the lexer input string from start_offset
+ * to end of current lexer token.  If chomp is true, drop any trailing
+ * newline(s).
  *
  * We rely on the knowledge that flex modifies the scan buffer by storing
  * a NUL at the end of the current token (yytext).  Note that this might
  * not work quite right if we were parsing a sub-buffer, but since pgbench
- * never invokes that functionality, it doesn't matter.
- */
-int
-expr_scanner_offset(PsqlScanState state)
-{
-    return strlen(state->scanbuf);
-}
-
-/*
- * Get a malloc'd copy of the lexer input string from start_offset
- * to just before end_offset.  If chomp is true, drop any trailing
- * newline(s).
+ * never invokes that functionality, it doesn't matter.  Also, this will
+ * give the wrong answer (the whole remainder of the input) if called
+ * before any yylex() call has been done.
  */
 char *
 expr_scanner_get_substring(PsqlScanState state,
-                           int start_offset, int end_offset,
+                           int start_offset,
                            bool chomp)
 {
     char       *result;
-    const char *scanptr = state->scanbuf + start_offset;
-    int            slen = end_offset - start_offset;

-    Assert(slen >= 0);
-    Assert(end_offset <= strlen(state->scanbuf));
+    result = pg_strdup(state->scanbuf + start_offset);

     if (chomp)
     {
+        size_t        slen = strlen(result);
+
         while (slen > 0 &&
-               (scanptr[slen - 1] == '\n' || scanptr[slen - 1] == '\r'))
+               (result[slen - 1] == '\n' || result[slen - 1] == '\r'))
             slen--;
+        result[slen] = '\0';
     }

-    result = (char *) pg_malloc(slen + 1);
-    memcpy(result, scanptr, slen);
-    result[slen] = '\0';
-
     return result;
 }
-
-/*
- * Get the line number associated with the given string offset
- * (which must not be past the end of where we've lexed to).
- */
-int
-expr_scanner_get_lineno(PsqlScanState state, int offset)
-{
-    int            lineno = 1;
-    const char *p = state->scanbuf;
-
-    while (*p && offset > 0)
-    {
-        if (*p == '\n')
-            lineno++;
-        p++, offset--;
-    }
-    return lineno;
-}
diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c
index fdc957fa34d..38f8bc11bcd 100644
--- a/src/bin/pgbench/pgbench.c
+++ b/src/bin/pgbench/pgbench.c
@@ -5690,8 +5690,8 @@ process_backslash_command(PsqlScanState sstate, const char *source)
     initPQExpBuffer(&word_buf);

     /* Remember location of the backslash */
-    start_offset = expr_scanner_offset(sstate) - 1;
-    lineno = expr_scanner_get_lineno(sstate, start_offset);
+    psql_scan_get_location(sstate, &lineno, &start_offset);
+    start_offset--;

     /* Collect first word of command */
     if (!expr_lex_one_word(sstate, &word_buf, &word_offset))
@@ -5747,7 +5747,6 @@ process_backslash_command(PsqlScanState sstate, const char *source)
         my_command->first_line =
             expr_scanner_get_substring(sstate,
                                        start_offset,
-                                       expr_scanner_offset(sstate),
                                        true);

         expr_scanner_finish(yyscanner);
@@ -5777,7 +5776,6 @@ process_backslash_command(PsqlScanState sstate, const char *source)
     my_command->first_line =
         expr_scanner_get_substring(sstate,
                                    start_offset,
-                                   expr_scanner_offset(sstate),
                                    true);

     if (my_command->meta == META_SLEEP)
@@ -5952,8 +5950,6 @@ ParseScript(const char *script, const char *desc, int weight)
     PQExpBufferData line_buf;
     int            alloc_num;
     int            index;
-    int            lineno;
-    int            start_offset;

 #define COMMANDS_ALLOC_NUM 128
     alloc_num = COMMANDS_ALLOC_NUM;
@@ -5977,7 +5973,6 @@ ParseScript(const char *script, const char *desc, int weight)
      * stdstrings should be true, which is a bit riskier.
      */
     psql_scan_setup(sstate, script, strlen(script), 0, true);
-    start_offset = expr_scanner_offset(sstate) - 1;

     initPQExpBuffer(&line_buf);

@@ -5985,12 +5980,15 @@ ParseScript(const char *script, const char *desc, int weight)

     for (;;)
     {
+        int            lineno;
+        int            start_offset;
         PsqlScanResult sr;
         promptStatus_t prompt;
         Command    *command = NULL;

         resetPQExpBuffer(&line_buf);
-        lineno = expr_scanner_get_lineno(sstate, start_offset);
+
+        psql_scan_get_location(sstate, &lineno, &start_offset);

         sr = psql_scan(sstate, &line_buf, &prompt);

diff --git a/src/bin/pgbench/pgbench.h b/src/bin/pgbench/pgbench.h
index f6a883611c5..0ba216e5f72 100644
--- a/src/bin/pgbench/pgbench.h
+++ b/src/bin/pgbench/pgbench.h
@@ -149,11 +149,9 @@ extern yyscan_t expr_scanner_init(PsqlScanState state,
                                   const char *source, int lineno, int start_offset,
                                   const char *command);
 extern void expr_scanner_finish(yyscan_t yyscanner);
-extern int    expr_scanner_offset(PsqlScanState state);
 extern char *expr_scanner_get_substring(PsqlScanState state,
-                                        int start_offset, int end_offset,
+                                        int start_offset,
                                         bool chomp);
-extern int    expr_scanner_get_lineno(PsqlScanState state, int offset);

 extern void syntax_error(const char *source, int lineno, const char *line,
                          const char *command, const char *msg,
diff --git a/src/fe_utils/psqlscan.l b/src/fe_utils/psqlscan.l
index b3c6b88e9ca..f2bcaa4cc3a 100644
--- a/src/fe_utils/psqlscan.l
+++ b/src/fe_utils/psqlscan.l
@@ -1079,6 +1079,10 @@ psql_scan_setup(PsqlScanState state,
     /* Set lookaside data in case we have to map unsafe encoding */
     state->curline = state->scanbuf;
     state->refline = state->scanline;
+
+    /* Initialize state for psql_scan_get_location() */
+    state->cur_line_no = 0;        /* yylex not called yet */
+    state->cur_line_start = state->scanbuf;
 }

 /*
@@ -1136,6 +1140,10 @@ psql_scan(PsqlScanState state,
     /* And lex. */
     lexresult = yylex(NULL, state->scanner);

+    /* Notify psql_scan_get_location() that a yylex call has been made. */
+    if (state->cur_line_no == 0)
+        state->cur_line_no = 1;
+
     /*
      * Check termination state and return appropriate result info.
      */
@@ -1311,6 +1319,52 @@ psql_scan_in_quote(PsqlScanState state)
         state->start_state != xqs;
 }

+/*
+ * Return the current scanning location (end+1 of last scanned token),
+ * as a line number counted from 1 and an offset from string start.
+ *
+ * This considers only the outermost input string, and therefore is of
+ * limited use for programs that use psqlscan_push_new_buffer().
+ *
+ * It would be a bit easier probably to use "%option yylineno" to count
+ * lines, but the flex manual says that has a performance cost, and only
+ * a minority of programs using psqlscan have need for this functionality.
+ * So we implement it ourselves without adding overhead to the lexer itself.
+ */
+void
+psql_scan_get_location(PsqlScanState state,
+                       int *lineno, int *offset)
+{
+    const char *line_end;
+
+    /*
+     * We rely on flex's having stored a NUL after the current token in
+     * scanbuf.  Therefore we must specially handle the state before yylex()
+     * has been called, when obviously that won't have happened yet.
+     */
+    if (state->cur_line_no == 0)
+    {
+        *lineno = 1;
+        *offset = 0;
+        return;
+    }
+
+    /*
+     * Advance cur_line_no/cur_line_start if any full lines have been read.
+     * Doing this prevents the operation from being O(N^2) for long inputs.
+     */
+    while ((line_end = strchr(state->cur_line_start, '\n')) != NULL)
+    {
+        state->cur_line_no++;
+        state->cur_line_start = line_end + 1;
+    }
+
+    /* Report current location. */
+    *lineno = state->cur_line_no;
+    *offset = (state->cur_line_start - state->scanbuf) +
+        strlen(state->cur_line_start);
+}
+
 /*
  * Push the given string onto the stack of stuff to scan.
  *
diff --git a/src/include/fe_utils/psqlscan.h b/src/include/fe_utils/psqlscan.h
index 81f792b1733..39d2065fe98 100644
--- a/src/include/fe_utils/psqlscan.h
+++ b/src/include/fe_utils/psqlscan.h
@@ -87,4 +87,7 @@ extern void psql_scan_reselect_sql_lexer(PsqlScanState state);

 extern bool psql_scan_in_quote(PsqlScanState state);

+extern void psql_scan_get_location(PsqlScanState state,
+                                   int *lineno, int *offset);
+
 #endif                            /* PSQLSCAN_H */
diff --git a/src/include/fe_utils/psqlscan_int.h b/src/include/fe_utils/psqlscan_int.h
index 37798114873..453735c9dc8 100644
--- a/src/include/fe_utils/psqlscan_int.h
+++ b/src/include/fe_utils/psqlscan_int.h
@@ -104,6 +104,10 @@ typedef struct PsqlScanStateData
     const char *curline;        /* actual flex input string for cur buf */
     const char *refline;        /* original data for cur buffer */

+    /* status for psql_scan_get_location() */
+    int            cur_line_no;    /* current line#, or 0 if no yylex done */
+    const char *cur_line_start; /* start of cur_line_no'th line in scanbuf */
+
     /*
      * All this state lives across successive input lines, until explicitly
      * reset by psql_scan_reset.  start_state is adopted by yylex() on entry,
--
2.43.5