PostgreSQL Source Code  git master
pgbench.c
Go to the documentation of this file.
1 /*
2  * pgbench.c
3  *
4  * A simple benchmark program for PostgreSQL
5  * Originally written by Tatsuo Ishii and enhanced by many contributors.
6  *
7  * src/bin/pgbench/pgbench.c
8  * Copyright (c) 2000-2020, PostgreSQL Global Development Group
9  * ALL RIGHTS RESERVED;
10  *
11  * Permission to use, copy, modify, and distribute this software and its
12  * documentation for any purpose, without fee, and without a written agreement
13  * is hereby granted, provided that the above copyright notice and this
14  * paragraph and the following two paragraphs appear in all copies.
15  *
16  * IN NO EVENT SHALL THE AUTHOR OR DISTRIBUTORS BE LIABLE TO ANY PARTY FOR
17  * DIRECT, INDIRECT, SPECIAL, INCIDENTAL, OR CONSEQUENTIAL DAMAGES, INCLUDING
18  * LOST PROFITS, ARISING OUT OF THE USE OF THIS SOFTWARE AND ITS
19  * DOCUMENTATION, EVEN IF THE AUTHOR OR DISTRIBUTORS HAVE BEEN ADVISED OF THE
20  * POSSIBILITY OF SUCH DAMAGE.
21  *
22  * THE AUTHOR AND DISTRIBUTORS SPECIFICALLY DISCLAIMS ANY WARRANTIES,
23  * INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY
24  * AND FITNESS FOR A PARTICULAR PURPOSE. THE SOFTWARE PROVIDED HEREUNDER IS
25  * ON AN "AS IS" BASIS, AND THE AUTHOR AND DISTRIBUTORS HAS NO OBLIGATIONS TO
26  * PROVIDE MAINTENANCE, SUPPORT, UPDATES, ENHANCEMENTS, OR MODIFICATIONS.
27  *
28  */
29 
30 #ifdef WIN32
31 #define FD_SETSIZE 1024 /* must set before winsock2.h is included */
32 #endif
33 
34 #include "postgres_fe.h"
35 
36 #include <ctype.h>
37 #include <float.h>
38 #include <limits.h>
39 #include <math.h>
40 #include <signal.h>
41 #include <time.h>
42 #include <sys/time.h>
43 #ifdef HAVE_SYS_RESOURCE_H
44 #include <sys/resource.h> /* for getrlimit */
45 #endif
46 
47 /* For testing, PGBENCH_USE_SELECT can be defined to force use of that code */
48 #if defined(HAVE_PPOLL) && !defined(PGBENCH_USE_SELECT)
49 #define POLL_USING_PPOLL
50 #ifdef HAVE_POLL_H
51 #include <poll.h>
52 #endif
53 #else /* no ppoll(), so use select() */
54 #define POLL_USING_SELECT
55 #ifdef HAVE_SYS_SELECT_H
56 #include <sys/select.h>
57 #endif
58 #endif
59 
60 #include "common/int.h"
61 #include "common/logging.h"
62 #include "fe_utils/cancel.h"
63 #include "fe_utils/conditional.h"
64 #include "getopt_long.h"
65 #include "libpq-fe.h"
66 #include "pgbench.h"
67 #include "portability/instr_time.h"
68 
69 #ifndef M_PI
70 #define M_PI 3.14159265358979323846
71 #endif
72 
73 #define ERRCODE_UNDEFINED_TABLE "42P01"
74 
75 /*
76  * Hashing constants
77  */
78 #define FNV_PRIME UINT64CONST(0x100000001b3)
79 #define FNV_OFFSET_BASIS UINT64CONST(0xcbf29ce484222325)
80 #define MM2_MUL UINT64CONST(0xc6a4a7935bd1e995)
81 #define MM2_MUL_TIMES_8 UINT64CONST(0x35253c9ade8f4ca8)
82 #define MM2_ROT 47
83 
84 /*
85  * Multi-platform socket set implementations
86  */
87 
88 #ifdef POLL_USING_PPOLL
89 #define SOCKET_WAIT_METHOD "ppoll"
90 
91 typedef struct socket_set
92 {
93  int maxfds; /* allocated length of pollfds[] array */
94  int curfds; /* number currently in use */
95  struct pollfd pollfds[FLEXIBLE_ARRAY_MEMBER];
96 } socket_set;
97 
98 #endif /* POLL_USING_PPOLL */
99 
100 #ifdef POLL_USING_SELECT
101 #define SOCKET_WAIT_METHOD "select"
102 
103 typedef struct socket_set
104 {
105  int maxfd; /* largest FD currently set in fds */
106  fd_set fds;
107 } socket_set;
108 
109 #endif /* POLL_USING_SELECT */
110 
111 /*
112  * Multi-platform pthread implementations
113  */
114 
115 #ifdef WIN32
116 /* Use native win32 threads on Windows */
117 typedef struct win32_pthread *pthread_t;
118 typedef int pthread_attr_t;
119 
120 static int pthread_create(pthread_t *thread, pthread_attr_t *attr, void *(*start_routine) (void *), void *arg);
121 static int pthread_join(pthread_t th, void **thread_return);
122 #elif defined(ENABLE_THREAD_SAFETY)
123 /* Use platform-dependent pthread capability */
124 #include <pthread.h>
125 #else
126 /* No threads implementation, use none (-j 1) */
127 #define pthread_t void *
128 #endif
129 
130 
131 /********************************************************************
132  * some configurable parameters */
133 
134 #define DEFAULT_INIT_STEPS "dtgvp" /* default -I setting */
135 #define ALL_INIT_STEPS "dtgGvpf" /* all possible steps */
136 
137 #define LOG_STEP_SECONDS 5 /* seconds between log messages */
138 #define DEFAULT_NXACTS 10 /* default nxacts */
139 
140 #define MIN_GAUSSIAN_PARAM 2.0 /* minimum parameter for gauss */
141 
142 #define MIN_ZIPFIAN_PARAM 1.001 /* minimum parameter for zipfian */
143 #define MAX_ZIPFIAN_PARAM 1000.0 /* maximum parameter for zipfian */
144 
145 int nxacts = 0; /* number of transactions per client */
146 int duration = 0; /* duration in seconds */
147 int64 end_time = 0; /* when to stop in micro seconds, under -T */
148 
149 /*
150  * scaling factor. for example, scale = 10 will make 1000000 tuples in
151  * pgbench_accounts table.
152  */
153 int scale = 1;
154 
155 /*
156  * fillfactor. for example, fillfactor = 90 will use only 90 percent
157  * space during inserts and leave 10 percent free.
158  */
159 int fillfactor = 100;
160 
161 /*
162  * use unlogged tables?
163  */
164 bool unlogged_tables = false;
165 
166 /*
167  * log sampling rate (1.0 = log everything, 0.0 = option not given)
168  */
169 double sample_rate = 0.0;
170 
171 /*
172  * When threads are throttled to a given rate limit, this is the target delay
173  * to reach that rate in usec. 0 is the default and means no throttling.
174  */
175 double throttle_delay = 0;
176 
177 /*
178  * Transactions which take longer than this limit (in usec) are counted as
179  * late, and reported as such, although they are completed anyway. When
180  * throttling is enabled, execution time slots that are more than this late
181  * are skipped altogether, and counted separately.
182  */
183 int64 latency_limit = 0;
184 
185 /*
186  * tablespace selection
187  */
188 char *tablespace = NULL;
189 char *index_tablespace = NULL;
190 
191 /*
192  * Number of "pgbench_accounts" partitions. 0 is the default and means no
193  * partitioning.
194  */
195 static int partitions = 0;
196 
197 /* partitioning strategy for "pgbench_accounts" */
198 typedef enum
199 {
200  PART_NONE, /* no partitioning */
201  PART_RANGE, /* range partitioning */
202  PART_HASH /* hash partitioning */
204 
205 static partition_method_t partition_method = PART_NONE;
206 static const char *PARTITION_METHOD[] = {"none", "range", "hash"};
207 
208 /* random seed used to initialize base_random_sequence */
209 int64 random_seed = -1;
210 
211 /*
212  * end of configurable parameters
213  *********************************************************************/
214 
215 #define nbranches 1 /* Makes little sense to change this. Change
216  * -s instead */
217 #define ntellers 10
218 #define naccounts 100000
219 
220 /*
221  * The scale factor at/beyond which 32bit integers are incapable of storing
222  * 64bit values.
223  *
224  * Although the actual threshold is 21474, we use 20000 because it is easier to
225  * document and remember, and isn't that far away from the real threshold.
226  */
227 #define SCALE_32BIT_THRESHOLD 20000
228 
229 bool use_log; /* log transaction latencies to a file */
230 bool use_quiet; /* quiet logging onto stderr */
231 int agg_interval; /* log aggregates instead of individual
232  * transactions */
233 bool per_script_stats = false; /* whether to collect stats per script */
234 int progress = 0; /* thread progress report every this seconds */
235 bool progress_timestamp = false; /* progress report with Unix time */
236 int nclients = 1; /* number of clients */
237 int nthreads = 1; /* number of threads */
238 bool is_connect; /* establish connection for each transaction */
239 bool report_per_command; /* report per-command latencies */
240 int main_pid; /* main process id used in log filename */
241 
242 char *pghost = "";
243 char *pgport = "";
244 char *login = NULL;
245 char *dbName;
246 char *logfile_prefix = NULL;
247 const char *progname;
248 
249 #define WSEP '@' /* weight separator */
250 
251 volatile bool timer_exceeded = false; /* flag from signal handler */
252 
253 /*
254  * Variable definitions.
255  *
256  * If a variable only has a string value, "svalue" is that value, and value is
257  * "not set". If the value is known, "value" contains the value (in any
258  * variant).
259  *
260  * In this case "svalue" contains the string equivalent of the value, if we've
261  * had occasion to compute that, or NULL if we haven't.
262  */
263 typedef struct
264 {
265  char *name; /* variable's name */
266  char *svalue; /* its value in string form, if known */
267  PgBenchValue value; /* actual variable's value */
268 } Variable;
269 
270 #define MAX_SCRIPTS 128 /* max number of SQL scripts allowed */
271 #define SHELL_COMMAND_SIZE 256 /* maximum size allowed for shell command */
272 
273 /*
274  * Simple data structure to keep stats about something.
275  *
276  * XXX probably the first value should be kept and used as an offset for
277  * better numerical stability...
278  */
279 typedef struct SimpleStats
280 {
281  int64 count; /* how many values were encountered */
282  double min; /* the minimum seen */
283  double max; /* the maximum seen */
284  double sum; /* sum of values */
285  double sum2; /* sum of squared values */
286 } SimpleStats;
287 
288 /*
289  * Data structure to hold various statistics: per-thread and per-script stats
290  * are maintained and merged together.
291  */
292 typedef struct StatsData
293 {
294  time_t start_time; /* interval start time, for aggregates */
295  int64 cnt; /* number of transactions, including skipped */
296  int64 skipped; /* number of transactions skipped under --rate
297  * and --latency-limit */
300 } StatsData;
301 
302 /*
303  * Struct to keep random state.
304  */
305 typedef struct RandomState
306 {
307  unsigned short xseed[3];
308 } RandomState;
309 
310 /* Various random sequences are initialized from this one. */
312 
313 /*
314  * Connection state machine states.
315  */
316 typedef enum
317 {
318  /*
319  * The client must first choose a script to execute. Once chosen, it can
320  * either be throttled (state CSTATE_PREPARE_THROTTLE under --rate), start
321  * right away (state CSTATE_START_TX) or not start at all if the timer was
322  * exceeded (state CSTATE_FINISHED).
323  */
325 
326  /*
327  * CSTATE_START_TX performs start-of-transaction processing. Establishes
328  * a new connection for the transaction in --connect mode, records the
329  * transaction start time, and proceed to the first command.
330  *
331  * Note: once a script is started, it will either error or run till its
332  * end, where it may be interrupted. It is not interrupted while running,
333  * so pgbench --time is to be understood as tx are allowed to start in
334  * that time, and will finish when their work is completed.
335  */
337 
338  /*
339  * In CSTATE_PREPARE_THROTTLE state, we calculate when to begin the next
340  * transaction, and advance to CSTATE_THROTTLE. CSTATE_THROTTLE state
341  * sleeps until that moment, then advances to CSTATE_START_TX, or
342  * CSTATE_FINISHED if the next transaction would start beyond the end of
343  * the run.
344  */
347 
348  /*
349  * We loop through these states, to process each command in the script:
350  *
351  * CSTATE_START_COMMAND starts the execution of a command. On a SQL
352  * command, the command is sent to the server, and we move to
353  * CSTATE_WAIT_RESULT state. On a \sleep meta-command, the timer is set,
354  * and we enter the CSTATE_SLEEP state to wait for it to expire. Other
355  * meta-commands are executed immediately. If the command about to start
356  * is actually beyond the end of the script, advance to CSTATE_END_TX.
357  *
358  * CSTATE_WAIT_RESULT waits until we get a result set back from the server
359  * for the current command.
360  *
361  * CSTATE_SLEEP waits until the end of \sleep.
362  *
363  * CSTATE_END_COMMAND records the end-of-command timestamp, increments the
364  * command counter, and loops back to CSTATE_START_COMMAND state.
365  *
366  * CSTATE_SKIP_COMMAND is used by conditional branches which are not
367  * executed. It quickly skip commands that do not need any evaluation.
368  * This state can move forward several commands, till there is something
369  * to do or the end of the script.
370  */
376 
377  /*
378  * CSTATE_END_TX performs end-of-transaction processing. It calculates
379  * latency, and logs the transaction. In --connect mode, it closes the
380  * current connection.
381  *
382  * Then either starts over in CSTATE_CHOOSE_SCRIPT, or enters
383  * CSTATE_FINISHED if we have no more work to do.
384  */
386 
387  /*
388  * Final states. CSTATE_ABORTED means that the script execution was
389  * aborted because a command failed, CSTATE_FINISHED means success.
390  */
394 
395 /*
396  * Connection state.
397  */
398 typedef struct
399 {
400  PGconn *con; /* connection handle to DB */
401  int id; /* client No. */
402  ConnectionStateEnum state; /* state machine's current state. */
403  ConditionalStack cstack; /* enclosing conditionals state */
404 
405  /*
406  * Separate randomness for each client. This is used for random functions
407  * PGBENCH_RANDOM_* during the execution of the script.
408  */
410 
411  int use_file; /* index in sql_script for this client */
412  int command; /* command number in script */
413 
414  /* client variables */
415  Variable *variables; /* array of variable definitions */
416  int nvariables; /* number of variables */
417  bool vars_sorted; /* are variables sorted by name? */
418 
419  /* various times about current transaction */
420  int64 txn_scheduled; /* scheduled start time of transaction (usec) */
421  int64 sleep_until; /* scheduled start time of next cmd (usec) */
422  instr_time txn_begin; /* used for measuring schedule lag times */
423  instr_time stmt_begin; /* used for measuring statement latencies */
424 
425  bool prepared[MAX_SCRIPTS]; /* whether client prepared the script */
426 
427  /* per client collected stats */
428  int64 cnt; /* client transaction count, for -t */
429  int ecnt; /* error count */
430 } CState;
431 
432 /*
433  * Thread state
434  */
435 typedef struct
436 {
437  int tid; /* thread id */
438  pthread_t thread; /* thread handle */
439  CState *state; /* array of CState */
440  int nstate; /* length of state[] */
441 
442  /*
443  * Separate randomness for each thread. Each thread option uses its own
444  * random state to make all of them independent of each other and
445  * therefore deterministic at the thread level.
446  */
447  RandomState ts_choose_rs; /* random state for selecting a script */
448  RandomState ts_throttle_rs; /* random state for transaction throttling */
449  RandomState ts_sample_rs; /* random state for log sampling */
450 
451  int64 throttle_trigger; /* previous/next throttling (us) */
452  FILE *logfile; /* where to log, or NULL */
453 
454  /* per thread collected stats */
455  instr_time start_time; /* thread start time */
458  int64 latency_late; /* executed but late transactions */
459 } TState;
460 
461 #define INVALID_THREAD ((pthread_t) 0)
462 
463 /*
464  * queries read from files
465  */
466 #define SQL_COMMAND 1
467 #define META_COMMAND 2
468 
469 /*
470  * max number of backslash command arguments or SQL variables,
471  * including the command or SQL statement itself
472  */
473 #define MAX_ARGS 256
474 
475 typedef enum MetaCommand
476 {
477  META_NONE, /* not a known meta-command */
478  META_SET, /* \set */
479  META_SETSHELL, /* \setshell */
480  META_SHELL, /* \shell */
481  META_SLEEP, /* \sleep */
482  META_GSET, /* \gset */
483  META_ASET, /* \aset */
484  META_IF, /* \if */
485  META_ELIF, /* \elif */
486  META_ELSE, /* \else */
487  META_ENDIF /* \endif */
488 } MetaCommand;
489 
490 typedef enum QueryMode
491 {
492  QUERY_SIMPLE, /* simple query */
493  QUERY_EXTENDED, /* extended query */
494  QUERY_PREPARED, /* extended query with prepared statements */
496 } QueryMode;
497 
498 static QueryMode querymode = QUERY_SIMPLE;
499 static const char *QUERYMODE[] = {"simple", "extended", "prepared"};
500 
501 /*
502  * struct Command represents one command in a script.
503  *
504  * lines The raw, possibly multi-line command text. Variable substitution
505  * not applied.
506  * first_line A short, single-line extract of 'lines', for error reporting.
507  * type SQL_COMMAND or META_COMMAND
508  * meta The type of meta-command, with META_NONE/GSET/ASET if command
509  * is SQL.
510  * argc Number of arguments of the command, 0 if not yet processed.
511  * argv Command arguments, the first of which is the command or SQL
512  * string itself. For SQL commands, after post-processing
513  * argv[0] is the same as 'lines' with variables substituted.
514  * varprefix SQL commands terminated with \gset or \aset have this set
515  * to a non NULL value. If nonempty, it's used to prefix the
516  * variable name that receives the value.
517  * aset do gset on all possible queries of a combined query (\;).
518  * expr Parsed expression, if needed.
519  * stats Time spent in this command.
520  */
521 typedef struct Command
522 {
524  char *first_line;
525  int type;
526  MetaCommand meta;
527  int argc;
528  char *argv[MAX_ARGS];
529  char *varprefix;
532 } Command;
533 
534 typedef struct ParsedScript
535 {
536  const char *desc; /* script descriptor (eg, file name) */
537  int weight; /* selection weight */
538  Command **commands; /* NULL-terminated array of Commands */
539  StatsData stats; /* total time spent in script */
540 } ParsedScript;
541 
542 static ParsedScript sql_script[MAX_SCRIPTS]; /* SQL script files */
543 static int num_scripts; /* number of scripts in sql_script[] */
544 static int64 total_weight = 0;
545 
546 /* Builtin test scripts */
547 typedef struct BuiltinScript
548 {
549  const char *name; /* very short name for -b ... */
550  const char *desc; /* short description */
551  const char *script; /* actual pgbench script */
552 } BuiltinScript;
553 
555 {
556  {
557  "tpcb-like",
558  "<builtin: TPC-B (sort of)>",
559  "\\set aid random(1, " CppAsString2(naccounts) " * :scale)\n"
560  "\\set bid random(1, " CppAsString2(nbranches) " * :scale)\n"
561  "\\set tid random(1, " CppAsString2(ntellers) " * :scale)\n"
562  "\\set delta random(-5000, 5000)\n"
563  "BEGIN;\n"
564  "UPDATE pgbench_accounts SET abalance = abalance + :delta WHERE aid = :aid;\n"
565  "SELECT abalance FROM pgbench_accounts WHERE aid = :aid;\n"
566  "UPDATE pgbench_tellers SET tbalance = tbalance + :delta WHERE tid = :tid;\n"
567  "UPDATE pgbench_branches SET bbalance = bbalance + :delta WHERE bid = :bid;\n"
568  "INSERT INTO pgbench_history (tid, bid, aid, delta, mtime) VALUES (:tid, :bid, :aid, :delta, CURRENT_TIMESTAMP);\n"
569  "END;\n"
570  },
571  {
572  "simple-update",
573  "<builtin: simple update>",
574  "\\set aid random(1, " CppAsString2(naccounts) " * :scale)\n"
575  "\\set bid random(1, " CppAsString2(nbranches) " * :scale)\n"
576  "\\set tid random(1, " CppAsString2(ntellers) " * :scale)\n"
577  "\\set delta random(-5000, 5000)\n"
578  "BEGIN;\n"
579  "UPDATE pgbench_accounts SET abalance = abalance + :delta WHERE aid = :aid;\n"
580  "SELECT abalance FROM pgbench_accounts WHERE aid = :aid;\n"
581  "INSERT INTO pgbench_history (tid, bid, aid, delta, mtime) VALUES (:tid, :bid, :aid, :delta, CURRENT_TIMESTAMP);\n"
582  "END;\n"
583  },
584  {
585  "select-only",
586  "<builtin: select only>",
587  "\\set aid random(1, " CppAsString2(naccounts) " * :scale)\n"
588  "SELECT abalance FROM pgbench_accounts WHERE aid = :aid;\n"
589  }
590 };
591 
592 
593 /* Function prototypes */
594 static void setNullValue(PgBenchValue *pv);
595 static void setBoolValue(PgBenchValue *pv, bool bval);
596 static void setIntValue(PgBenchValue *pv, int64 ival);
597 static void setDoubleValue(PgBenchValue *pv, double dval);
598 static bool evaluateExpr(CState *st, PgBenchExpr *expr,
599  PgBenchValue *retval);
600 static ConnectionStateEnum executeMetaCommand(CState *st, instr_time *now);
601 static void doLog(TState *thread, CState *st,
602  StatsData *agg, bool skipped, double latency, double lag);
603 static void processXactStats(TState *thread, CState *st, instr_time *now,
604  bool skipped, StatsData *agg);
605 static void append_fillfactor(char *opts, int len);
606 static void addScript(ParsedScript script);
607 static void *threadRun(void *arg);
608 static void finishCon(CState *st);
609 static void setalarm(int seconds);
610 static socket_set *alloc_socket_set(int count);
611 static void free_socket_set(socket_set *sa);
612 static void clear_socket_set(socket_set *sa);
613 static void add_socket_to_set(socket_set *sa, int fd, int idx);
614 static int wait_on_socket_set(socket_set *sa, int64 usecs);
615 static bool socket_has_input(socket_set *sa, int fd, int idx);
616 
617 
618 /* callback functions for our flex lexer */
620  NULL, /* don't need get_variable functionality */
621 };
622 
623 
624 static void
625 usage(void)
626 {
627  printf("%s is a benchmarking tool for PostgreSQL.\n\n"
628  "Usage:\n"
629  " %s [OPTION]... [DBNAME]\n"
630  "\nInitialization options:\n"
631  " -i, --initialize invokes initialization mode\n"
632  " -I, --init-steps=[" ALL_INIT_STEPS "]+ (default \"" DEFAULT_INIT_STEPS "\")\n"
633  " run selected initialization steps\n"
634  " -F, --fillfactor=NUM set fill factor\n"
635  " -n, --no-vacuum do not run VACUUM during initialization\n"
636  " -q, --quiet quiet logging (one message each 5 seconds)\n"
637  " -s, --scale=NUM scaling factor\n"
638  " --foreign-keys create foreign key constraints between tables\n"
639  " --index-tablespace=TABLESPACE\n"
640  " create indexes in the specified tablespace\n"
641  " --partition-method=(range|hash)\n"
642  " partition pgbench_accounts with this method (default: range)\n"
643  " --partitions=NUM partition pgbench_accounts into NUM parts (default: 0)\n"
644  " --tablespace=TABLESPACE create tables in the specified tablespace\n"
645  " --unlogged-tables create tables as unlogged tables\n"
646  "\nOptions to select what to run:\n"
647  " -b, --builtin=NAME[@W] add builtin script NAME weighted at W (default: 1)\n"
648  " (use \"-b list\" to list available scripts)\n"
649  " -f, --file=FILENAME[@W] add script FILENAME weighted at W (default: 1)\n"
650  " -N, --skip-some-updates skip updates of pgbench_tellers and pgbench_branches\n"
651  " (same as \"-b simple-update\")\n"
652  " -S, --select-only perform SELECT-only transactions\n"
653  " (same as \"-b select-only\")\n"
654  "\nBenchmarking options:\n"
655  " -c, --client=NUM number of concurrent database clients (default: 1)\n"
656  " -C, --connect establish new connection for each transaction\n"
657  " -D, --define=VARNAME=VALUE\n"
658  " define variable for use by custom script\n"
659  " -j, --jobs=NUM number of threads (default: 1)\n"
660  " -l, --log write transaction times to log file\n"
661  " -L, --latency-limit=NUM count transactions lasting more than NUM ms as late\n"
662  " -M, --protocol=simple|extended|prepared\n"
663  " protocol for submitting queries (default: simple)\n"
664  " -n, --no-vacuum do not run VACUUM before tests\n"
665  " -P, --progress=NUM show thread progress report every NUM seconds\n"
666  " -r, --report-latencies report average latency per command\n"
667  " -R, --rate=NUM target rate in transactions per second\n"
668  " -s, --scale=NUM report this scale factor in output\n"
669  " -t, --transactions=NUM number of transactions each client runs (default: 10)\n"
670  " -T, --time=NUM duration of benchmark test in seconds\n"
671  " -v, --vacuum-all vacuum all four standard tables before tests\n"
672  " --aggregate-interval=NUM aggregate data over NUM seconds\n"
673  " --log-prefix=PREFIX prefix for transaction time log file\n"
674  " (default: \"pgbench_log\")\n"
675  " --progress-timestamp use Unix epoch timestamps for progress\n"
676  " --random-seed=SEED set random seed (\"time\", \"rand\", integer)\n"
677  " --sampling-rate=NUM fraction of transactions to log (e.g., 0.01 for 1%%)\n"
678  " --show-script=NAME show builtin script code, then exit\n"
679  "\nCommon options:\n"
680  " -d, --debug print debugging output\n"
681  " -h, --host=HOSTNAME database server host or socket directory\n"
682  " -p, --port=PORT database server port number\n"
683  " -U, --username=USERNAME connect as specified database user\n"
684  " -V, --version output version information, then exit\n"
685  " -?, --help show this help, then exit\n"
686  "\n"
687  "Report bugs to <%s>.\n"
688  "%s home page: <%s>\n",
689  progname, progname, PACKAGE_BUGREPORT, PACKAGE_NAME, PACKAGE_URL);
690 }
691 
692 /* return whether str matches "^\s*[-+]?[0-9]+$" */
693 static bool
694 is_an_int(const char *str)
695 {
696  const char *ptr = str;
697 
698  /* skip leading spaces; cast is consistent with strtoint64 */
699  while (*ptr && isspace((unsigned char) *ptr))
700  ptr++;
701 
702  /* skip sign */
703  if (*ptr == '+' || *ptr == '-')
704  ptr++;
705 
706  /* at least one digit */
707  if (*ptr && !isdigit((unsigned char) *ptr))
708  return false;
709 
710  /* eat all digits */
711  while (*ptr && isdigit((unsigned char) *ptr))
712  ptr++;
713 
714  /* must have reached end of string */
715  return *ptr == '\0';
716 }
717 
718 
719 /*
720  * strtoint64 -- convert a string to 64-bit integer
721  *
722  * This function is a slightly modified version of scanint8() from
723  * src/backend/utils/adt/int8.c.
724  *
725  * The function returns whether the conversion worked, and if so
726  * "*result" is set to the result.
727  *
728  * If not errorOK, an error message is also printed out on errors.
729  */
730 bool
731 strtoint64(const char *str, bool errorOK, int64 *result)
732 {
733  const char *ptr = str;
734  int64 tmp = 0;
735  bool neg = false;
736 
737  /*
738  * Do our own scan, rather than relying on sscanf which might be broken
739  * for long long.
740  *
741  * As INT64_MIN can't be stored as a positive 64 bit integer, accumulate
742  * value as a negative number.
743  */
744 
745  /* skip leading spaces */
746  while (*ptr && isspace((unsigned char) *ptr))
747  ptr++;
748 
749  /* handle sign */
750  if (*ptr == '-')
751  {
752  ptr++;
753  neg = true;
754  }
755  else if (*ptr == '+')
756  ptr++;
757 
758  /* require at least one digit */
759  if (unlikely(!isdigit((unsigned char) *ptr)))
760  goto invalid_syntax;
761 
762  /* process digits */
763  while (*ptr && isdigit((unsigned char) *ptr))
764  {
765  int8 digit = (*ptr++ - '0');
766 
767  if (unlikely(pg_mul_s64_overflow(tmp, 10, &tmp)) ||
768  unlikely(pg_sub_s64_overflow(tmp, digit, &tmp)))
769  goto out_of_range;
770  }
771 
772  /* allow trailing whitespace, but not other trailing chars */
773  while (*ptr != '\0' && isspace((unsigned char) *ptr))
774  ptr++;
775 
776  if (unlikely(*ptr != '\0'))
777  goto invalid_syntax;
778 
779  if (!neg)
780  {
781  if (unlikely(tmp == PG_INT64_MIN))
782  goto out_of_range;
783  tmp = -tmp;
784  }
785 
786  *result = tmp;
787  return true;
788 
789 out_of_range:
790  if (!errorOK)
791  pg_log_error("value \"%s\" is out of range for type bigint", str);
792  return false;
793 
794 invalid_syntax:
795  if (!errorOK)
796  pg_log_error("invalid input syntax for type bigint: \"%s\"", str);
797  return false;
798 }
799 
800 /* convert string to double, detecting overflows/underflows */
801 bool
802 strtodouble(const char *str, bool errorOK, double *dv)
803 {
804  char *end;
805 
806  errno = 0;
807  *dv = strtod(str, &end);
808 
809  if (unlikely(errno != 0))
810  {
811  if (!errorOK)
812  pg_log_error("value \"%s\" is out of range for type double", str);
813  return false;
814  }
815 
816  if (unlikely(end == str || *end != '\0'))
817  {
818  if (!errorOK)
819  pg_log_error("invalid input syntax for type double: \"%s\"", str);
820  return false;
821  }
822  return true;
823 }
824 
825 /*
826  * Initialize a random state struct.
827  *
828  * We derive the seed from base_random_sequence, which must be set up already.
829  */
830 static void
832 {
833  random_state->xseed[0] = (unsigned short)
834  (pg_jrand48(base_random_sequence.xseed) & 0xFFFF);
835  random_state->xseed[1] = (unsigned short)
836  (pg_jrand48(base_random_sequence.xseed) & 0xFFFF);
837  random_state->xseed[2] = (unsigned short)
838  (pg_jrand48(base_random_sequence.xseed) & 0xFFFF);
839 }
840 
841 /*
842  * Random number generator: uniform distribution from min to max inclusive.
843  *
844  * Although the limits are expressed as int64, you can't generate the full
845  * int64 range in one call, because the difference of the limits mustn't
846  * overflow int64. In practice it's unwise to ask for more than an int32
847  * range, because of the limited precision of pg_erand48().
848  */
849 static int64
850 getrand(RandomState *random_state, int64 min, int64 max)
851 {
852  /*
853  * Odd coding is so that min and max have approximately the same chance of
854  * being selected as do numbers between them.
855  *
856  * pg_erand48() is thread-safe and concurrent, which is why we use it
857  * rather than random(), which in glibc is non-reentrant, and therefore
858  * protected by a mutex, and therefore a bottleneck on machines with many
859  * CPUs.
860  */
861  return min + (int64) ((max - min + 1) * pg_erand48(random_state->xseed));
862 }
863 
864 /*
865  * random number generator: exponential distribution from min to max inclusive.
866  * the parameter is so that the density of probability for the last cut-off max
867  * value is exp(-parameter).
868  */
869 static int64
870 getExponentialRand(RandomState *random_state, int64 min, int64 max,
871  double parameter)
872 {
873  double cut,
874  uniform,
875  rand;
876 
877  /* abort if wrong parameter, but must really be checked beforehand */
878  Assert(parameter > 0.0);
879  cut = exp(-parameter);
880  /* erand in [0, 1), uniform in (0, 1] */
881  uniform = 1.0 - pg_erand48(random_state->xseed);
882 
883  /*
884  * inner expression in (cut, 1] (if parameter > 0), rand in [0, 1)
885  */
886  Assert((1.0 - cut) != 0.0);
887  rand = -log(cut + (1.0 - cut) * uniform) / parameter;
888  /* return int64 random number within between min and max */
889  return min + (int64) ((max - min + 1) * rand);
890 }
891 
892 /* random number generator: gaussian distribution from min to max inclusive */
893 static int64
894 getGaussianRand(RandomState *random_state, int64 min, int64 max,
895  double parameter)
896 {
897  double stdev;
898  double rand;
899 
900  /* abort if parameter is too low, but must really be checked beforehand */
901  Assert(parameter >= MIN_GAUSSIAN_PARAM);
902 
903  /*
904  * Get user specified random number from this loop, with -parameter <
905  * stdev <= parameter
906  *
907  * This loop is executed until the number is in the expected range.
908  *
909  * As the minimum parameter is 2.0, the probability of looping is low:
910  * sqrt(-2 ln(r)) <= 2 => r >= e^{-2} ~ 0.135, then when taking the
911  * average sinus multiplier as 2/pi, we have a 8.6% looping probability in
912  * the worst case. For a parameter value of 5.0, the looping probability
913  * is about e^{-5} * 2 / pi ~ 0.43%.
914  */
915  do
916  {
917  /*
918  * pg_erand48 generates [0,1), but for the basic version of the
919  * Box-Muller transform the two uniformly distributed random numbers
920  * are expected in (0, 1] (see
921  * https://en.wikipedia.org/wiki/Box-Muller_transform)
922  */
923  double rand1 = 1.0 - pg_erand48(random_state->xseed);
924  double rand2 = 1.0 - pg_erand48(random_state->xseed);
925 
926  /* Box-Muller basic form transform */
927  double var_sqrt = sqrt(-2.0 * log(rand1));
928 
929  stdev = var_sqrt * sin(2.0 * M_PI * rand2);
930 
931  /*
932  * we may try with cos, but there may be a bias induced if the
933  * previous value fails the test. To be on the safe side, let us try
934  * over.
935  */
936  }
937  while (stdev < -parameter || stdev >= parameter);
938 
939  /* stdev is in [-parameter, parameter), normalization to [0,1) */
940  rand = (stdev + parameter) / (parameter * 2.0);
941 
942  /* return int64 random number within between min and max */
943  return min + (int64) ((max - min + 1) * rand);
944 }
945 
946 /*
947  * random number generator: generate a value, such that the series of values
948  * will approximate a Poisson distribution centered on the given value.
949  *
950  * Individual results are rounded to integers, though the center value need
951  * not be one.
952  */
953 static int64
954 getPoissonRand(RandomState *random_state, double center)
955 {
956  /*
957  * Use inverse transform sampling to generate a value > 0, such that the
958  * expected (i.e. average) value is the given argument.
959  */
960  double uniform;
961 
962  /* erand in [0, 1), uniform in (0, 1] */
963  uniform = 1.0 - pg_erand48(random_state->xseed);
964 
965  return (int64) (-log(uniform) * center + 0.5);
966 }
967 
968 /*
969  * Computing zipfian using rejection method, based on
970  * "Non-Uniform Random Variate Generation",
971  * Luc Devroye, p. 550-551, Springer 1986.
972  *
973  * This works for s > 1.0, but may perform badly for s very close to 1.0.
974  */
975 static int64
976 computeIterativeZipfian(RandomState *random_state, int64 n, double s)
977 {
978  double b = pow(2.0, s - 1.0);
979  double x,
980  t,
981  u,
982  v;
983 
984  /* Ensure n is sane */
985  if (n <= 1)
986  return 1;
987 
988  while (true)
989  {
990  /* random variates */
991  u = pg_erand48(random_state->xseed);
992  v = pg_erand48(random_state->xseed);
993 
994  x = floor(pow(u, -1.0 / (s - 1.0)));
995 
996  t = pow(1.0 + 1.0 / x, s - 1.0);
997  /* reject if too large or out of bound */
998  if (v * x * (t - 1.0) / (b - 1.0) <= t / b && x <= n)
999  break;
1000  }
1001  return (int64) x;
1002 }
1003 
1004 /* random number generator: zipfian distribution from min to max inclusive */
1005 static int64
1006 getZipfianRand(RandomState *random_state, int64 min, int64 max, double s)
1007 {
1008  int64 n = max - min + 1;
1009 
1010  /* abort if parameter is invalid */
1012 
1013  return min - 1 + computeIterativeZipfian(random_state, n, s);
1014 }
1015 
1016 /*
1017  * FNV-1a hash function
1018  */
1019 static int64
1020 getHashFnv1a(int64 val, uint64 seed)
1021 {
1022  int64 result;
1023  int i;
1024 
1025  result = FNV_OFFSET_BASIS ^ seed;
1026  for (i = 0; i < 8; ++i)
1027  {
1028  int32 octet = val & 0xff;
1029 
1030  val = val >> 8;
1031  result = result ^ octet;
1032  result = result * FNV_PRIME;
1033  }
1034 
1035  return result;
1036 }
1037 
1038 /*
1039  * Murmur2 hash function
1040  *
1041  * Based on original work of Austin Appleby
1042  * https://github.com/aappleby/smhasher/blob/master/src/MurmurHash2.cpp
1043  */
1044 static int64
1045 getHashMurmur2(int64 val, uint64 seed)
1046 {
1047  uint64 result = seed ^ MM2_MUL_TIMES_8; /* sizeof(int64) */
1048  uint64 k = (uint64) val;
1049 
1050  k *= MM2_MUL;
1051  k ^= k >> MM2_ROT;
1052  k *= MM2_MUL;
1053 
1054  result ^= k;
1055  result *= MM2_MUL;
1056 
1057  result ^= result >> MM2_ROT;
1058  result *= MM2_MUL;
1059  result ^= result >> MM2_ROT;
1060 
1061  return (int64) result;
1062 }
1063 
1064 /*
1065  * Initialize the given SimpleStats struct to all zeroes
1066  */
1067 static void
1069 {
1070  memset(ss, 0, sizeof(SimpleStats));
1071 }
1072 
1073 /*
1074  * Accumulate one value into a SimpleStats struct.
1075  */
1076 static void
1078 {
1079  if (ss->count == 0 || val < ss->min)
1080  ss->min = val;
1081  if (ss->count == 0 || val > ss->max)
1082  ss->max = val;
1083  ss->count++;
1084  ss->sum += val;
1085  ss->sum2 += val * val;
1086 }
1087 
1088 /*
1089  * Merge two SimpleStats objects
1090  */
1091 static void
1093 {
1094  if (acc->count == 0 || ss->min < acc->min)
1095  acc->min = ss->min;
1096  if (acc->count == 0 || ss->max > acc->max)
1097  acc->max = ss->max;
1098  acc->count += ss->count;
1099  acc->sum += ss->sum;
1100  acc->sum2 += ss->sum2;
1101 }
1102 
1103 /*
1104  * Initialize a StatsData struct to mostly zeroes, with its start time set to
1105  * the given value.
1106  */
1107 static void
1109 {
1110  sd->start_time = start_time;
1111  sd->cnt = 0;
1112  sd->skipped = 0;
1113  initSimpleStats(&sd->latency);
1114  initSimpleStats(&sd->lag);
1115 }
1116 
1117 /*
1118  * Accumulate one additional item into the given stats object.
1119  */
1120 static void
1121 accumStats(StatsData *stats, bool skipped, double lat, double lag)
1122 {
1123  stats->cnt++;
1124 
1125  if (skipped)
1126  {
1127  /* no latency to record on skipped transactions */
1128  stats->skipped++;
1129  }
1130  else
1131  {
1132  addToSimpleStats(&stats->latency, lat);
1133 
1134  /* and possibly the same for schedule lag */
1135  if (throttle_delay)
1136  addToSimpleStats(&stats->lag, lag);
1137  }
1138 }
1139 
1140 /* call PQexec() and exit() on failure */
1141 static void
1142 executeStatement(PGconn *con, const char *sql)
1143 {
1144  PGresult *res;
1145 
1146  res = PQexec(con, sql);
1147  if (PQresultStatus(res) != PGRES_COMMAND_OK)
1148  {
1149  pg_log_fatal("query failed: %s", PQerrorMessage(con));
1150  pg_log_info("query was: %s", sql);
1151  exit(1);
1152  }
1153  PQclear(res);
1154 }
1155 
1156 /* call PQexec() and complain, but without exiting, on failure */
1157 static void
1158 tryExecuteStatement(PGconn *con, const char *sql)
1159 {
1160  PGresult *res;
1161 
1162  res = PQexec(con, sql);
1163  if (PQresultStatus(res) != PGRES_COMMAND_OK)
1164  {
1165  pg_log_error("%s", PQerrorMessage(con));
1166  pg_log_info("(ignoring this error and continuing anyway)");
1167  }
1168  PQclear(res);
1169 }
1170 
1171 /* set up a connection to the backend */
1172 static PGconn *
1174 {
1175  PGconn *conn;
1176  bool new_pass;
1177  static bool have_password = false;
1178  static char password[100];
1179 
1180  /*
1181  * Start the connection. Loop until we have a password if requested by
1182  * backend.
1183  */
1184  do
1185  {
1186 #define PARAMS_ARRAY_SIZE 7
1187 
1188  const char *keywords[PARAMS_ARRAY_SIZE];
1189  const char *values[PARAMS_ARRAY_SIZE];
1190 
1191  keywords[0] = "host";
1192  values[0] = pghost;
1193  keywords[1] = "port";
1194  values[1] = pgport;
1195  keywords[2] = "user";
1196  values[2] = login;
1197  keywords[3] = "password";
1198  values[3] = have_password ? password : NULL;
1199  keywords[4] = "dbname";
1200  values[4] = dbName;
1201  keywords[5] = "fallback_application_name";
1202  values[5] = progname;
1203  keywords[6] = NULL;
1204  values[6] = NULL;
1205 
1206  new_pass = false;
1207 
1208  conn = PQconnectdbParams(keywords, values, true);
1209 
1210  if (!conn)
1211  {
1212  pg_log_error("connection to database \"%s\" failed", dbName);
1213  return NULL;
1214  }
1215 
1216  if (PQstatus(conn) == CONNECTION_BAD &&
1217  PQconnectionNeedsPassword(conn) &&
1218  !have_password)
1219  {
1220  PQfinish(conn);
1221  simple_prompt("Password: ", password, sizeof(password), false);
1222  have_password = true;
1223  new_pass = true;
1224  }
1225  } while (new_pass);
1226 
1227  /* check to see that the backend connection was successfully made */
1228  if (PQstatus(conn) == CONNECTION_BAD)
1229  {
1230  pg_log_error("connection to database \"%s\" failed: %s",
1231  dbName, PQerrorMessage(conn));
1232  PQfinish(conn);
1233  return NULL;
1234  }
1235 
1236  return conn;
1237 }
1238 
1239 /* qsort comparator for Variable array */
1240 static int
1241 compareVariableNames(const void *v1, const void *v2)
1242 {
1243  return strcmp(((const Variable *) v1)->name,
1244  ((const Variable *) v2)->name);
1245 }
1246 
1247 /* Locate a variable by name; returns NULL if unknown */
1248 static Variable *
1250 {
1251  Variable key;
1252 
1253  /* On some versions of Solaris, bsearch of zero items dumps core */
1254  if (st->nvariables <= 0)
1255  return NULL;
1256 
1257  /* Sort if we have to */
1258  if (!st->vars_sorted)
1259  {
1260  qsort((void *) st->variables, st->nvariables, sizeof(Variable),
1262  st->vars_sorted = true;
1263  }
1264 
1265  /* Now we can search */
1266  key.name = name;
1267  return (Variable *) bsearch((void *) &key,
1268  (void *) st->variables,
1269  st->nvariables,
1270  sizeof(Variable),
1272 }
1273 
1274 /* Get the value of a variable, in string form; returns NULL if unknown */
1275 static char *
1277 {
1278  Variable *var;
1279  char stringform[64];
1280 
1281  var = lookupVariable(st, name);
1282  if (var == NULL)
1283  return NULL; /* not found */
1284 
1285  if (var->svalue)
1286  return var->svalue; /* we have it in string form */
1287 
1288  /* We need to produce a string equivalent of the value */
1289  Assert(var->value.type != PGBT_NO_VALUE);
1290  if (var->value.type == PGBT_NULL)
1291  snprintf(stringform, sizeof(stringform), "NULL");
1292  else if (var->value.type == PGBT_BOOLEAN)
1293  snprintf(stringform, sizeof(stringform),
1294  "%s", var->value.u.bval ? "true" : "false");
1295  else if (var->value.type == PGBT_INT)
1296  snprintf(stringform, sizeof(stringform),
1297  INT64_FORMAT, var->value.u.ival);
1298  else if (var->value.type == PGBT_DOUBLE)
1299  snprintf(stringform, sizeof(stringform),
1300  "%.*g", DBL_DIG, var->value.u.dval);
1301  else /* internal error, unexpected type */
1302  Assert(0);
1303  var->svalue = pg_strdup(stringform);
1304  return var->svalue;
1305 }
1306 
1307 /* Try to convert variable to a value; return false on failure */
1308 static bool
1310 {
1311  size_t slen;
1312 
1313  if (var->value.type != PGBT_NO_VALUE)
1314  return true; /* no work */
1315 
1316  slen = strlen(var->svalue);
1317 
1318  if (slen == 0)
1319  /* what should it do on ""? */
1320  return false;
1321 
1322  if (pg_strcasecmp(var->svalue, "null") == 0)
1323  {
1324  setNullValue(&var->value);
1325  }
1326 
1327  /*
1328  * accept prefixes such as y, ye, n, no... but not for "o". 0/1 are
1329  * recognized later as an int, which is converted to bool if needed.
1330  */
1331  else if (pg_strncasecmp(var->svalue, "true", slen) == 0 ||
1332  pg_strncasecmp(var->svalue, "yes", slen) == 0 ||
1333  pg_strcasecmp(var->svalue, "on") == 0)
1334  {
1335  setBoolValue(&var->value, true);
1336  }
1337  else if (pg_strncasecmp(var->svalue, "false", slen) == 0 ||
1338  pg_strncasecmp(var->svalue, "no", slen) == 0 ||
1339  pg_strcasecmp(var->svalue, "off") == 0 ||
1340  pg_strcasecmp(var->svalue, "of") == 0)
1341  {
1342  setBoolValue(&var->value, false);
1343  }
1344  else if (is_an_int(var->svalue))
1345  {
1346  /* if it looks like an int, it must be an int without overflow */
1347  int64 iv;
1348 
1349  if (!strtoint64(var->svalue, false, &iv))
1350  return false;
1351 
1352  setIntValue(&var->value, iv);
1353  }
1354  else /* type should be double */
1355  {
1356  double dv;
1357 
1358  if (!strtodouble(var->svalue, true, &dv))
1359  {
1360  pg_log_error("malformed variable \"%s\" value: \"%s\"",
1361  var->name, var->svalue);
1362  return false;
1363  }
1364  setDoubleValue(&var->value, dv);
1365  }
1366  return true;
1367 }
1368 
1369 /*
1370  * Check whether a variable's name is allowed.
1371  *
1372  * We allow any non-ASCII character, as well as ASCII letters, digits, and
1373  * underscore.
1374  *
1375  * Keep this in sync with the definitions of variable name characters in
1376  * "src/fe_utils/psqlscan.l", "src/bin/psql/psqlscanslash.l" and
1377  * "src/bin/pgbench/exprscan.l". Also see parseVariable(), below.
1378  *
1379  * Note: this static function is copied from "src/bin/psql/variables.c"
1380  */
1381 static bool
1383 {
1384  const unsigned char *ptr = (const unsigned char *) name;
1385 
1386  /* Mustn't be zero-length */
1387  if (*ptr == '\0')
1388  return false;
1389 
1390  while (*ptr)
1391  {
1392  if (IS_HIGHBIT_SET(*ptr) ||
1393  strchr("ABCDEFGHIJKLMNOPQRSTUVWXYZ" "abcdefghijklmnopqrstuvwxyz"
1394  "_0123456789", *ptr) != NULL)
1395  ptr++;
1396  else
1397  return false;
1398  }
1399 
1400  return true;
1401 }
1402 
1403 /*
1404  * Lookup a variable by name, creating it if need be.
1405  * Caller is expected to assign a value to the variable.
1406  * Returns NULL on failure (bad name).
1407  */
1408 static Variable *
1409 lookupCreateVariable(CState *st, const char *context, char *name)
1410 {
1411  Variable *var;
1412 
1413  var = lookupVariable(st, name);
1414  if (var == NULL)
1415  {
1416  Variable *newvars;
1417 
1418  /*
1419  * Check for the name only when declaring a new variable to avoid
1420  * overhead.
1421  */
1422  if (!valid_variable_name(name))
1423  {
1424  pg_log_error("%s: invalid variable name: \"%s\"", context, name);
1425  return NULL;
1426  }
1427 
1428  /* Create variable at the end of the array */
1429  if (st->variables)
1430  newvars = (Variable *) pg_realloc(st->variables,
1431  (st->nvariables + 1) * sizeof(Variable));
1432  else
1433  newvars = (Variable *) pg_malloc(sizeof(Variable));
1434 
1435  st->variables = newvars;
1436 
1437  var = &newvars[st->nvariables];
1438 
1439  var->name = pg_strdup(name);
1440  var->svalue = NULL;
1441  /* caller is expected to initialize remaining fields */
1442 
1443  st->nvariables++;
1444  /* we don't re-sort the array till we have to */
1445  st->vars_sorted = false;
1446  }
1447 
1448  return var;
1449 }
1450 
1451 /* Assign a string value to a variable, creating it if need be */
1452 /* Returns false on failure (bad name) */
1453 static bool
1454 putVariable(CState *st, const char *context, char *name, const char *value)
1455 {
1456  Variable *var;
1457  char *val;
1458 
1459  var = lookupCreateVariable(st, context, name);
1460  if (!var)
1461  return false;
1462 
1463  /* dup then free, in case value is pointing at this variable */
1464  val = pg_strdup(value);
1465 
1466  if (var->svalue)
1467  free(var->svalue);
1468  var->svalue = val;
1469  var->value.type = PGBT_NO_VALUE;
1470 
1471  return true;
1472 }
1473 
1474 /* Assign a value to a variable, creating it if need be */
1475 /* Returns false on failure (bad name) */
1476 static bool
1477 putVariableValue(CState *st, const char *context, char *name,
1478  const PgBenchValue *value)
1479 {
1480  Variable *var;
1481 
1482  var = lookupCreateVariable(st, context, name);
1483  if (!var)
1484  return false;
1485 
1486  if (var->svalue)
1487  free(var->svalue);
1488  var->svalue = NULL;
1489  var->value = *value;
1490 
1491  return true;
1492 }
1493 
1494 /* Assign an integer value to a variable, creating it if need be */
1495 /* Returns false on failure (bad name) */
1496 static bool
1497 putVariableInt(CState *st, const char *context, char *name, int64 value)
1498 {
1499  PgBenchValue val;
1500 
1501  setIntValue(&val, value);
1502  return putVariableValue(st, context, name, &val);
1503 }
1504 
1505 /*
1506  * Parse a possible variable reference (:varname).
1507  *
1508  * "sql" points at a colon. If what follows it looks like a valid
1509  * variable name, return a malloc'd string containing the variable name,
1510  * and set *eaten to the number of characters consumed.
1511  * Otherwise, return NULL.
1512  */
1513 static char *
1514 parseVariable(const char *sql, int *eaten)
1515 {
1516  int i = 0;
1517  char *name;
1518 
1519  do
1520  {
1521  i++;
1522  } while (IS_HIGHBIT_SET(sql[i]) ||
1523  strchr("ABCDEFGHIJKLMNOPQRSTUVWXYZ" "abcdefghijklmnopqrstuvwxyz"
1524  "_0123456789", sql[i]) != NULL);
1525  if (i == 1)
1526  return NULL; /* no valid variable name chars */
1527 
1528  name = pg_malloc(i);
1529  memcpy(name, &sql[1], i - 1);
1530  name[i - 1] = '\0';
1531 
1532  *eaten = i;
1533  return name;
1534 }
1535 
1536 static char *
1537 replaceVariable(char **sql, char *param, int len, char *value)
1538 {
1539  int valueln = strlen(value);
1540 
1541  if (valueln > len)
1542  {
1543  size_t offset = param - *sql;
1544 
1545  *sql = pg_realloc(*sql, strlen(*sql) - len + valueln + 1);
1546  param = *sql + offset;
1547  }
1548 
1549  if (valueln != len)
1550  memmove(param + valueln, param + len, strlen(param + len) + 1);
1551  memcpy(param, value, valueln);
1552 
1553  return param + valueln;
1554 }
1555 
1556 static char *
1557 assignVariables(CState *st, char *sql)
1558 {
1559  char *p,
1560  *name,
1561  *val;
1562 
1563  p = sql;
1564  while ((p = strchr(p, ':')) != NULL)
1565  {
1566  int eaten;
1567 
1568  name = parseVariable(p, &eaten);
1569  if (name == NULL)
1570  {
1571  while (*p == ':')
1572  {
1573  p++;
1574  }
1575  continue;
1576  }
1577 
1578  val = getVariable(st, name);
1579  free(name);
1580  if (val == NULL)
1581  {
1582  p++;
1583  continue;
1584  }
1585 
1586  p = replaceVariable(&sql, p, eaten, val);
1587  }
1588 
1589  return sql;
1590 }
1591 
1592 static void
1593 getQueryParams(CState *st, const Command *command, const char **params)
1594 {
1595  int i;
1596 
1597  for (i = 0; i < command->argc - 1; i++)
1598  params[i] = getVariable(st, command->argv[i + 1]);
1599 }
1600 
1601 static char *
1603 {
1604  if (pval->type == PGBT_NO_VALUE)
1605  return "none";
1606  else if (pval->type == PGBT_NULL)
1607  return "null";
1608  else if (pval->type == PGBT_INT)
1609  return "int";
1610  else if (pval->type == PGBT_DOUBLE)
1611  return "double";
1612  else if (pval->type == PGBT_BOOLEAN)
1613  return "boolean";
1614  else
1615  {
1616  /* internal error, should never get there */
1617  Assert(false);
1618  return NULL;
1619  }
1620 }
1621 
1622 /* get a value as a boolean, or tell if there is a problem */
1623 static bool
1624 coerceToBool(PgBenchValue *pval, bool *bval)
1625 {
1626  if (pval->type == PGBT_BOOLEAN)
1627  {
1628  *bval = pval->u.bval;
1629  return true;
1630  }
1631  else /* NULL, INT or DOUBLE */
1632  {
1633  pg_log_error("cannot coerce %s to boolean", valueTypeName(pval));
1634  *bval = false; /* suppress uninitialized-variable warnings */
1635  return false;
1636  }
1637 }
1638 
1639 /*
1640  * Return true or false from an expression for conditional purposes.
1641  * Non zero numerical values are true, zero and NULL are false.
1642  */
1643 static bool
1645 {
1646  switch (pval->type)
1647  {
1648  case PGBT_NULL:
1649  return false;
1650  case PGBT_BOOLEAN:
1651  return pval->u.bval;
1652  case PGBT_INT:
1653  return pval->u.ival != 0;
1654  case PGBT_DOUBLE:
1655  return pval->u.dval != 0.0;
1656  default:
1657  /* internal error, unexpected type */
1658  Assert(0);
1659  return false;
1660  }
1661 }
1662 
1663 /* get a value as an int, tell if there is a problem */
1664 static bool
1665 coerceToInt(PgBenchValue *pval, int64 *ival)
1666 {
1667  if (pval->type == PGBT_INT)
1668  {
1669  *ival = pval->u.ival;
1670  return true;
1671  }
1672  else if (pval->type == PGBT_DOUBLE)
1673  {
1674  double dval = rint(pval->u.dval);
1675 
1676  if (isnan(dval) || !FLOAT8_FITS_IN_INT64(dval))
1677  {
1678  pg_log_error("double to int overflow for %f", dval);
1679  return false;
1680  }
1681  *ival = (int64) dval;
1682  return true;
1683  }
1684  else /* BOOLEAN or NULL */
1685  {
1686  pg_log_error("cannot coerce %s to int", valueTypeName(pval));
1687  return false;
1688  }
1689 }
1690 
1691 /* get a value as a double, or tell if there is a problem */
1692 static bool
1693 coerceToDouble(PgBenchValue *pval, double *dval)
1694 {
1695  if (pval->type == PGBT_DOUBLE)
1696  {
1697  *dval = pval->u.dval;
1698  return true;
1699  }
1700  else if (pval->type == PGBT_INT)
1701  {
1702  *dval = (double) pval->u.ival;
1703  return true;
1704  }
1705  else /* BOOLEAN or NULL */
1706  {
1707  pg_log_error("cannot coerce %s to double", valueTypeName(pval));
1708  return false;
1709  }
1710 }
1711 
1712 /* assign a null value */
1713 static void
1715 {
1716  pv->type = PGBT_NULL;
1717  pv->u.ival = 0;
1718 }
1719 
1720 /* assign a boolean value */
1721 static void
1723 {
1724  pv->type = PGBT_BOOLEAN;
1725  pv->u.bval = bval;
1726 }
1727 
1728 /* assign an integer value */
1729 static void
1730 setIntValue(PgBenchValue *pv, int64 ival)
1731 {
1732  pv->type = PGBT_INT;
1733  pv->u.ival = ival;
1734 }
1735 
1736 /* assign a double value */
1737 static void
1738 setDoubleValue(PgBenchValue *pv, double dval)
1739 {
1740  pv->type = PGBT_DOUBLE;
1741  pv->u.dval = dval;
1742 }
1743 
1744 static bool
1746 {
1747  return func == PGBENCH_AND || func == PGBENCH_OR || func == PGBENCH_CASE;
1748 }
1749 
1750 /* lazy evaluation of some functions */
1751 static bool
1754 {
1755  PgBenchValue a1,
1756  a2;
1757  bool ba1,
1758  ba2;
1759 
1760  Assert(isLazyFunc(func) && args != NULL && args->next != NULL);
1761 
1762  /* args points to first condition */
1763  if (!evaluateExpr(st, args->expr, &a1))
1764  return false;
1765 
1766  /* second condition for AND/OR and corresponding branch for CASE */
1767  args = args->next;
1768 
1769  switch (func)
1770  {
1771  case PGBENCH_AND:
1772  if (a1.type == PGBT_NULL)
1773  {
1774  setNullValue(retval);
1775  return true;
1776  }
1777 
1778  if (!coerceToBool(&a1, &ba1))
1779  return false;
1780 
1781  if (!ba1)
1782  {
1783  setBoolValue(retval, false);
1784  return true;
1785  }
1786 
1787  if (!evaluateExpr(st, args->expr, &a2))
1788  return false;
1789 
1790  if (a2.type == PGBT_NULL)
1791  {
1792  setNullValue(retval);
1793  return true;
1794  }
1795  else if (!coerceToBool(&a2, &ba2))
1796  return false;
1797  else
1798  {
1799  setBoolValue(retval, ba2);
1800  return true;
1801  }
1802 
1803  return true;
1804 
1805  case PGBENCH_OR:
1806 
1807  if (a1.type == PGBT_NULL)
1808  {
1809  setNullValue(retval);
1810  return true;
1811  }
1812 
1813  if (!coerceToBool(&a1, &ba1))
1814  return false;
1815 
1816  if (ba1)
1817  {
1818  setBoolValue(retval, true);
1819  return true;
1820  }
1821 
1822  if (!evaluateExpr(st, args->expr, &a2))
1823  return false;
1824 
1825  if (a2.type == PGBT_NULL)
1826  {
1827  setNullValue(retval);
1828  return true;
1829  }
1830  else if (!coerceToBool(&a2, &ba2))
1831  return false;
1832  else
1833  {
1834  setBoolValue(retval, ba2);
1835  return true;
1836  }
1837 
1838  case PGBENCH_CASE:
1839  /* when true, execute branch */
1840  if (valueTruth(&a1))
1841  return evaluateExpr(st, args->expr, retval);
1842 
1843  /* now args contains next condition or final else expression */
1844  args = args->next;
1845 
1846  /* final else case? */
1847  if (args->next == NULL)
1848  return evaluateExpr(st, args->expr, retval);
1849 
1850  /* no, another when, proceed */
1851  return evalLazyFunc(st, PGBENCH_CASE, args, retval);
1852 
1853  default:
1854  /* internal error, cannot get here */
1855  Assert(0);
1856  break;
1857  }
1858  return false;
1859 }
1860 
1861 /* maximum number of function arguments */
1862 #define MAX_FARGS 16
1863 
1864 /*
1865  * Recursive evaluation of standard functions,
1866  * which do not require lazy evaluation.
1867  */
1868 static bool
1871  PgBenchValue *retval)
1872 {
1873  /* evaluate all function arguments */
1874  int nargs = 0;
1875  PgBenchValue vargs[MAX_FARGS];
1876  PgBenchExprLink *l = args;
1877  bool has_null = false;
1878 
1879  for (nargs = 0; nargs < MAX_FARGS && l != NULL; nargs++, l = l->next)
1880  {
1881  if (!evaluateExpr(st, l->expr, &vargs[nargs]))
1882  return false;
1883  has_null |= vargs[nargs].type == PGBT_NULL;
1884  }
1885 
1886  if (l != NULL)
1887  {
1888  pg_log_error("too many function arguments, maximum is %d", MAX_FARGS);
1889  return false;
1890  }
1891 
1892  /* NULL arguments */
1893  if (has_null && func != PGBENCH_IS && func != PGBENCH_DEBUG)
1894  {
1895  setNullValue(retval);
1896  return true;
1897  }
1898 
1899  /* then evaluate function */
1900  switch (func)
1901  {
1902  /* overloaded operators */
1903  case PGBENCH_ADD:
1904  case PGBENCH_SUB:
1905  case PGBENCH_MUL:
1906  case PGBENCH_DIV:
1907  case PGBENCH_MOD:
1908  case PGBENCH_EQ:
1909  case PGBENCH_NE:
1910  case PGBENCH_LE:
1911  case PGBENCH_LT:
1912  {
1913  PgBenchValue *lval = &vargs[0],
1914  *rval = &vargs[1];
1915 
1916  Assert(nargs == 2);
1917 
1918  /* overloaded type management, double if some double */
1919  if ((lval->type == PGBT_DOUBLE ||
1920  rval->type == PGBT_DOUBLE) && func != PGBENCH_MOD)
1921  {
1922  double ld,
1923  rd;
1924 
1925  if (!coerceToDouble(lval, &ld) ||
1926  !coerceToDouble(rval, &rd))
1927  return false;
1928 
1929  switch (func)
1930  {
1931  case PGBENCH_ADD:
1932  setDoubleValue(retval, ld + rd);
1933  return true;
1934 
1935  case PGBENCH_SUB:
1936  setDoubleValue(retval, ld - rd);
1937  return true;
1938 
1939  case PGBENCH_MUL:
1940  setDoubleValue(retval, ld * rd);
1941  return true;
1942 
1943  case PGBENCH_DIV:
1944  setDoubleValue(retval, ld / rd);
1945  return true;
1946 
1947  case PGBENCH_EQ:
1948  setBoolValue(retval, ld == rd);
1949  return true;
1950 
1951  case PGBENCH_NE:
1952  setBoolValue(retval, ld != rd);
1953  return true;
1954 
1955  case PGBENCH_LE:
1956  setBoolValue(retval, ld <= rd);
1957  return true;
1958 
1959  case PGBENCH_LT:
1960  setBoolValue(retval, ld < rd);
1961  return true;
1962 
1963  default:
1964  /* cannot get here */
1965  Assert(0);
1966  }
1967  }
1968  else /* we have integer operands, or % */
1969  {
1970  int64 li,
1971  ri,
1972  res;
1973 
1974  if (!coerceToInt(lval, &li) ||
1975  !coerceToInt(rval, &ri))
1976  return false;
1977 
1978  switch (func)
1979  {
1980  case PGBENCH_ADD:
1981  if (pg_add_s64_overflow(li, ri, &res))
1982  {
1983  pg_log_error("bigint add out of range");
1984  return false;
1985  }
1986  setIntValue(retval, res);
1987  return true;
1988 
1989  case PGBENCH_SUB:
1990  if (pg_sub_s64_overflow(li, ri, &res))
1991  {
1992  pg_log_error("bigint sub out of range");
1993  return false;
1994  }
1995  setIntValue(retval, res);
1996  return true;
1997 
1998  case PGBENCH_MUL:
1999  if (pg_mul_s64_overflow(li, ri, &res))
2000  {
2001  pg_log_error("bigint mul out of range");
2002  return false;
2003  }
2004  setIntValue(retval, res);
2005  return true;
2006 
2007  case PGBENCH_EQ:
2008  setBoolValue(retval, li == ri);
2009  return true;
2010 
2011  case PGBENCH_NE:
2012  setBoolValue(retval, li != ri);
2013  return true;
2014 
2015  case PGBENCH_LE:
2016  setBoolValue(retval, li <= ri);
2017  return true;
2018 
2019  case PGBENCH_LT:
2020  setBoolValue(retval, li < ri);
2021  return true;
2022 
2023  case PGBENCH_DIV:
2024  case PGBENCH_MOD:
2025  if (ri == 0)
2026  {
2027  pg_log_error("division by zero");
2028  return false;
2029  }
2030  /* special handling of -1 divisor */
2031  if (ri == -1)
2032  {
2033  if (func == PGBENCH_DIV)
2034  {
2035  /* overflow check (needed for INT64_MIN) */
2036  if (li == PG_INT64_MIN)
2037  {
2038  pg_log_error("bigint div out of range");
2039  return false;
2040  }
2041  else
2042  setIntValue(retval, -li);
2043  }
2044  else
2045  setIntValue(retval, 0);
2046  return true;
2047  }
2048  /* else divisor is not -1 */
2049  if (func == PGBENCH_DIV)
2050  setIntValue(retval, li / ri);
2051  else /* func == PGBENCH_MOD */
2052  setIntValue(retval, li % ri);
2053 
2054  return true;
2055 
2056  default:
2057  /* cannot get here */
2058  Assert(0);
2059  }
2060  }
2061 
2062  Assert(0);
2063  return false; /* NOTREACHED */
2064  }
2065 
2066  /* integer bitwise operators */
2067  case PGBENCH_BITAND:
2068  case PGBENCH_BITOR:
2069  case PGBENCH_BITXOR:
2070  case PGBENCH_LSHIFT:
2071  case PGBENCH_RSHIFT:
2072  {
2073  int64 li,
2074  ri;
2075 
2076  if (!coerceToInt(&vargs[0], &li) || !coerceToInt(&vargs[1], &ri))
2077  return false;
2078 
2079  if (func == PGBENCH_BITAND)
2080  setIntValue(retval, li & ri);
2081  else if (func == PGBENCH_BITOR)
2082  setIntValue(retval, li | ri);
2083  else if (func == PGBENCH_BITXOR)
2084  setIntValue(retval, li ^ ri);
2085  else if (func == PGBENCH_LSHIFT)
2086  setIntValue(retval, li << ri);
2087  else if (func == PGBENCH_RSHIFT)
2088  setIntValue(retval, li >> ri);
2089  else /* cannot get here */
2090  Assert(0);
2091 
2092  return true;
2093  }
2094 
2095  /* logical operators */
2096  case PGBENCH_NOT:
2097  {
2098  bool b;
2099 
2100  if (!coerceToBool(&vargs[0], &b))
2101  return false;
2102 
2103  setBoolValue(retval, !b);
2104  return true;
2105  }
2106 
2107  /* no arguments */
2108  case PGBENCH_PI:
2109  setDoubleValue(retval, M_PI);
2110  return true;
2111 
2112  /* 1 overloaded argument */
2113  case PGBENCH_ABS:
2114  {
2115  PgBenchValue *varg = &vargs[0];
2116 
2117  Assert(nargs == 1);
2118 
2119  if (varg->type == PGBT_INT)
2120  {
2121  int64 i = varg->u.ival;
2122 
2123  setIntValue(retval, i < 0 ? -i : i);
2124  }
2125  else
2126  {
2127  double d = varg->u.dval;
2128 
2129  Assert(varg->type == PGBT_DOUBLE);
2130  setDoubleValue(retval, d < 0.0 ? -d : d);
2131  }
2132 
2133  return true;
2134  }
2135 
2136  case PGBENCH_DEBUG:
2137  {
2138  PgBenchValue *varg = &vargs[0];
2139 
2140  Assert(nargs == 1);
2141 
2142  fprintf(stderr, "debug(script=%d,command=%d): ",
2143  st->use_file, st->command + 1);
2144 
2145  if (varg->type == PGBT_NULL)
2146  fprintf(stderr, "null\n");
2147  else if (varg->type == PGBT_BOOLEAN)
2148  fprintf(stderr, "boolean %s\n", varg->u.bval ? "true" : "false");
2149  else if (varg->type == PGBT_INT)
2150  fprintf(stderr, "int " INT64_FORMAT "\n", varg->u.ival);
2151  else if (varg->type == PGBT_DOUBLE)
2152  fprintf(stderr, "double %.*g\n", DBL_DIG, varg->u.dval);
2153  else /* internal error, unexpected type */
2154  Assert(0);
2155 
2156  *retval = *varg;
2157 
2158  return true;
2159  }
2160 
2161  /* 1 double argument */
2162  case PGBENCH_DOUBLE:
2163  case PGBENCH_SQRT:
2164  case PGBENCH_LN:
2165  case PGBENCH_EXP:
2166  {
2167  double dval;
2168 
2169  Assert(nargs == 1);
2170 
2171  if (!coerceToDouble(&vargs[0], &dval))
2172  return false;
2173 
2174  if (func == PGBENCH_SQRT)
2175  dval = sqrt(dval);
2176  else if (func == PGBENCH_LN)
2177  dval = log(dval);
2178  else if (func == PGBENCH_EXP)
2179  dval = exp(dval);
2180  /* else is cast: do nothing */
2181 
2182  setDoubleValue(retval, dval);
2183  return true;
2184  }
2185 
2186  /* 1 int argument */
2187  case PGBENCH_INT:
2188  {
2189  int64 ival;
2190 
2191  Assert(nargs == 1);
2192 
2193  if (!coerceToInt(&vargs[0], &ival))
2194  return false;
2195 
2196  setIntValue(retval, ival);
2197  return true;
2198  }
2199 
2200  /* variable number of arguments */
2201  case PGBENCH_LEAST:
2202  case PGBENCH_GREATEST:
2203  {
2204  bool havedouble;
2205  int i;
2206 
2207  Assert(nargs >= 1);
2208 
2209  /* need double result if any input is double */
2210  havedouble = false;
2211  for (i = 0; i < nargs; i++)
2212  {
2213  if (vargs[i].type == PGBT_DOUBLE)
2214  {
2215  havedouble = true;
2216  break;
2217  }
2218  }
2219  if (havedouble)
2220  {
2221  double extremum;
2222 
2223  if (!coerceToDouble(&vargs[0], &extremum))
2224  return false;
2225  for (i = 1; i < nargs; i++)
2226  {
2227  double dval;
2228 
2229  if (!coerceToDouble(&vargs[i], &dval))
2230  return false;
2231  if (func == PGBENCH_LEAST)
2232  extremum = Min(extremum, dval);
2233  else
2234  extremum = Max(extremum, dval);
2235  }
2236  setDoubleValue(retval, extremum);
2237  }
2238  else
2239  {
2240  int64 extremum;
2241 
2242  if (!coerceToInt(&vargs[0], &extremum))
2243  return false;
2244  for (i = 1; i < nargs; i++)
2245  {
2246  int64 ival;
2247 
2248  if (!coerceToInt(&vargs[i], &ival))
2249  return false;
2250  if (func == PGBENCH_LEAST)
2251  extremum = Min(extremum, ival);
2252  else
2253  extremum = Max(extremum, ival);
2254  }
2255  setIntValue(retval, extremum);
2256  }
2257  return true;
2258  }
2259 
2260  /* random functions */
2261  case PGBENCH_RANDOM:
2265  {
2266  int64 imin,
2267  imax;
2268 
2269  Assert(nargs >= 2);
2270 
2271  if (!coerceToInt(&vargs[0], &imin) ||
2272  !coerceToInt(&vargs[1], &imax))
2273  return false;
2274 
2275  /* check random range */
2276  if (imin > imax)
2277  {
2278  pg_log_error("empty range given to random");
2279  return false;
2280  }
2281  else if (imax - imin < 0 || (imax - imin) + 1 < 0)
2282  {
2283  /* prevent int overflows in random functions */
2284  pg_log_error("random range is too large");
2285  return false;
2286  }
2287 
2288  if (func == PGBENCH_RANDOM)
2289  {
2290  Assert(nargs == 2);
2291  setIntValue(retval, getrand(&st->cs_func_rs, imin, imax));
2292  }
2293  else /* gaussian & exponential */
2294  {
2295  double param;
2296 
2297  Assert(nargs == 3);
2298 
2299  if (!coerceToDouble(&vargs[2], &param))
2300  return false;
2301 
2302  if (func == PGBENCH_RANDOM_GAUSSIAN)
2303  {
2304  if (param < MIN_GAUSSIAN_PARAM)
2305  {
2306  pg_log_error("gaussian parameter must be at least %f (not %f)",
2307  MIN_GAUSSIAN_PARAM, param);
2308  return false;
2309  }
2310 
2311  setIntValue(retval,
2313  imin, imax, param));
2314  }
2315  else if (func == PGBENCH_RANDOM_ZIPFIAN)
2316  {
2317  if (param < MIN_ZIPFIAN_PARAM || param > MAX_ZIPFIAN_PARAM)
2318  {
2319  pg_log_error("zipfian parameter must be in range [%.3f, %.0f] (not %f)",
2320  MIN_ZIPFIAN_PARAM, MAX_ZIPFIAN_PARAM, param);
2321  return false;
2322  }
2323 
2324  setIntValue(retval,
2325  getZipfianRand(&st->cs_func_rs, imin, imax, param));
2326  }
2327  else /* exponential */
2328  {
2329  if (param <= 0.0)
2330  {
2331  pg_log_error("exponential parameter must be greater than zero (not %f)",
2332  param);
2333  return false;
2334  }
2335 
2336  setIntValue(retval,
2338  imin, imax, param));
2339  }
2340  }
2341 
2342  return true;
2343  }
2344 
2345  case PGBENCH_POW:
2346  {
2347  PgBenchValue *lval = &vargs[0];
2348  PgBenchValue *rval = &vargs[1];
2349  double ld,
2350  rd;
2351 
2352  Assert(nargs == 2);
2353 
2354  if (!coerceToDouble(lval, &ld) ||
2355  !coerceToDouble(rval, &rd))
2356  return false;
2357 
2358  setDoubleValue(retval, pow(ld, rd));
2359 
2360  return true;
2361  }
2362 
2363  case PGBENCH_IS:
2364  {
2365  Assert(nargs == 2);
2366 
2367  /*
2368  * note: this simple implementation is more permissive than
2369  * SQL
2370  */
2371  setBoolValue(retval,
2372  vargs[0].type == vargs[1].type &&
2373  vargs[0].u.bval == vargs[1].u.bval);
2374  return true;
2375  }
2376 
2377  /* hashing */
2378  case PGBENCH_HASH_FNV1A:
2379  case PGBENCH_HASH_MURMUR2:
2380  {
2381  int64 val,
2382  seed;
2383 
2384  Assert(nargs == 2);
2385 
2386  if (!coerceToInt(&vargs[0], &val) ||
2387  !coerceToInt(&vargs[1], &seed))
2388  return false;
2389 
2390  if (func == PGBENCH_HASH_MURMUR2)
2391  setIntValue(retval, getHashMurmur2(val, seed));
2392  else if (func == PGBENCH_HASH_FNV1A)
2393  setIntValue(retval, getHashFnv1a(val, seed));
2394  else
2395  /* cannot get here */
2396  Assert(0);
2397 
2398  return true;
2399  }
2400 
2401  default:
2402  /* cannot get here */
2403  Assert(0);
2404  /* dead code to avoid a compiler warning */
2405  return false;
2406  }
2407 }
2408 
2409 /* evaluate some function */
2410 static bool
2413 {
2414  if (isLazyFunc(func))
2415  return evalLazyFunc(st, func, args, retval);
2416  else
2417  return evalStandardFunc(st, func, args, retval);
2418 }
2419 
2420 /*
2421  * Recursive evaluation of an expression in a pgbench script
2422  * using the current state of variables.
2423  * Returns whether the evaluation was ok,
2424  * the value itself is returned through the retval pointer.
2425  */
2426 static bool
2428 {
2429  switch (expr->etype)
2430  {
2431  case ENODE_CONSTANT:
2432  {
2433  *retval = expr->u.constant;
2434  return true;
2435  }
2436 
2437  case ENODE_VARIABLE:
2438  {
2439  Variable *var;
2440 
2441  if ((var = lookupVariable(st, expr->u.variable.varname)) == NULL)
2442  {
2443  pg_log_error("undefined variable \"%s\"", expr->u.variable.varname);
2444  return false;
2445  }
2446 
2447  if (!makeVariableValue(var))
2448  return false;
2449 
2450  *retval = var->value;
2451  return true;
2452  }
2453 
2454  case ENODE_FUNCTION:
2455  return evalFunc(st,
2456  expr->u.function.function,
2457  expr->u.function.args,
2458  retval);
2459 
2460  default:
2461  /* internal error which should never occur */
2462  pg_log_fatal("unexpected enode type in evaluation: %d", expr->etype);
2463  exit(1);
2464  }
2465 }
2466 
2467 /*
2468  * Convert command name to meta-command enum identifier
2469  */
2470 static MetaCommand
2471 getMetaCommand(const char *cmd)
2472 {
2473  MetaCommand mc;
2474 
2475  if (cmd == NULL)
2476  mc = META_NONE;
2477  else if (pg_strcasecmp(cmd, "set") == 0)
2478  mc = META_SET;
2479  else if (pg_strcasecmp(cmd, "setshell") == 0)
2480  mc = META_SETSHELL;
2481  else if (pg_strcasecmp(cmd, "shell") == 0)
2482  mc = META_SHELL;
2483  else if (pg_strcasecmp(cmd, "sleep") == 0)
2484  mc = META_SLEEP;
2485  else if (pg_strcasecmp(cmd, "if") == 0)
2486  mc = META_IF;
2487  else if (pg_strcasecmp(cmd, "elif") == 0)
2488  mc = META_ELIF;
2489  else if (pg_strcasecmp(cmd, "else") == 0)
2490  mc = META_ELSE;
2491  else if (pg_strcasecmp(cmd, "endif") == 0)
2492  mc = META_ENDIF;
2493  else if (pg_strcasecmp(cmd, "gset") == 0)
2494  mc = META_GSET;
2495  else if (pg_strcasecmp(cmd, "aset") == 0)
2496  mc = META_ASET;
2497  else
2498  mc = META_NONE;
2499  return mc;
2500 }
2501 
2502 /*
2503  * Run a shell command. The result is assigned to the variable if not NULL.
2504  * Return true if succeeded, or false on error.
2505  */
2506 static bool
2507 runShellCommand(CState *st, char *variable, char **argv, int argc)
2508 {
2509  char command[SHELL_COMMAND_SIZE];
2510  int i,
2511  len = 0;
2512  FILE *fp;
2513  char res[64];
2514  char *endptr;
2515  int retval;
2516 
2517  /*----------
2518  * Join arguments with whitespace separators. Arguments starting with
2519  * exactly one colon are treated as variables:
2520  * name - append a string "name"
2521  * :var - append a variable named 'var'
2522  * ::name - append a string ":name"
2523  *----------
2524  */
2525  for (i = 0; i < argc; i++)
2526  {
2527  char *arg;
2528  int arglen;
2529 
2530  if (argv[i][0] != ':')
2531  {
2532  arg = argv[i]; /* a string literal */
2533  }
2534  else if (argv[i][1] == ':')
2535  {
2536  arg = argv[i] + 1; /* a string literal starting with colons */
2537  }
2538  else if ((arg = getVariable(st, argv[i] + 1)) == NULL)
2539  {
2540  pg_log_error("%s: undefined variable \"%s\"", argv[0], argv[i]);
2541  return false;
2542  }
2543 
2544  arglen = strlen(arg);
2545  if (len + arglen + (i > 0 ? 1 : 0) >= SHELL_COMMAND_SIZE - 1)
2546  {
2547  pg_log_error("%s: shell command is too long", argv[0]);
2548  return false;
2549  }
2550 
2551  if (i > 0)
2552  command[len++] = ' ';
2553  memcpy(command + len, arg, arglen);
2554  len += arglen;
2555  }
2556 
2557  command[len] = '\0';
2558 
2559  /* Fast path for non-assignment case */
2560  if (variable == NULL)
2561  {
2562  if (system(command))
2563  {
2564  if (!timer_exceeded)
2565  pg_log_error("%s: could not launch shell command", argv[0]);
2566  return false;
2567  }
2568  return true;
2569  }
2570 
2571  /* Execute the command with pipe and read the standard output. */
2572  if ((fp = popen(command, "r")) == NULL)
2573  {
2574  pg_log_error("%s: could not launch shell command", argv[0]);
2575  return false;
2576  }
2577  if (fgets(res, sizeof(res), fp) == NULL)
2578  {
2579  if (!timer_exceeded)
2580  pg_log_error("%s: could not read result of shell command", argv[0]);
2581  (void) pclose(fp);
2582  return false;
2583  }
2584  if (pclose(fp) < 0)
2585  {
2586  pg_log_error("%s: could not close shell command", argv[0]);
2587  return false;
2588  }
2589 
2590  /* Check whether the result is an integer and assign it to the variable */
2591  retval = (int) strtol(res, &endptr, 10);
2592  while (*endptr != '\0' && isspace((unsigned char) *endptr))
2593  endptr++;
2594  if (*res == '\0' || *endptr != '\0')
2595  {
2596  pg_log_error("%s: shell command must return an integer (not \"%s\")", argv[0], res);
2597  return false;
2598  }
2599  if (!putVariableInt(st, "setshell", variable, retval))
2600  return false;
2601 
2602  pg_log_debug("%s: shell parameter name: \"%s\", value: \"%s\"", argv[0], argv[1], res);
2603 
2604  return true;
2605 }
2606 
2607 #define MAX_PREPARE_NAME 32
2608 static void
2609 preparedStatementName(char *buffer, int file, int state)
2610 {
2611  sprintf(buffer, "P%d_%d", file, state);
2612 }
2613 
2614 static void
2615 commandFailed(CState *st, const char *cmd, const char *message)
2616 {
2617  pg_log_error("client %d aborted in command %d (%s) of script %d; %s",
2618  st->id, st->command, cmd, st->use_file, message);
2619 }
2620 
2621 /* return a script number with a weighted choice. */
2622 static int
2624 {
2625  int i = 0;
2626  int64 w;
2627 
2628  if (num_scripts == 1)
2629  return 0;
2630 
2631  w = getrand(&thread->ts_choose_rs, 0, total_weight - 1);
2632  do
2633  {
2634  w -= sql_script[i++].weight;
2635  } while (w >= 0);
2636 
2637  return i - 1;
2638 }
2639 
2640 /* Send a SQL command, using the chosen querymode */
2641 static bool
2643 {
2644  int r;
2645 
2646  if (querymode == QUERY_SIMPLE)
2647  {
2648  char *sql;
2649 
2650  sql = pg_strdup(command->argv[0]);
2651  sql = assignVariables(st, sql);
2652 
2653  pg_log_debug("client %d sending %s", st->id, sql);
2654  r = PQsendQuery(st->con, sql);
2655  free(sql);
2656  }
2657  else if (querymode == QUERY_EXTENDED)
2658  {
2659  const char *sql = command->argv[0];
2660  const char *params[MAX_ARGS];
2661 
2662  getQueryParams(st, command, params);
2663 
2664  pg_log_debug("client %d sending %s", st->id, sql);
2665  r = PQsendQueryParams(st->con, sql, command->argc - 1,
2666  NULL, params, NULL, NULL, 0);
2667  }
2668  else if (querymode == QUERY_PREPARED)
2669  {
2670  char name[MAX_PREPARE_NAME];
2671  const char *params[MAX_ARGS];
2672 
2673  if (!st->prepared[st->use_file])
2674  {
2675  int j;
2676  Command **commands = sql_script[st->use_file].commands;
2677 
2678  for (j = 0; commands[j] != NULL; j++)
2679  {
2680  PGresult *res;
2681  char name[MAX_PREPARE_NAME];
2682 
2683  if (commands[j]->type != SQL_COMMAND)
2684  continue;
2685  preparedStatementName(name, st->use_file, j);
2686  res = PQprepare(st->con, name,
2687  commands[j]->argv[0], commands[j]->argc - 1, NULL);
2688  if (PQresultStatus(res) != PGRES_COMMAND_OK)
2689  pg_log_error("%s", PQerrorMessage(st->con));
2690  PQclear(res);
2691  }
2692  st->prepared[st->use_file] = true;
2693  }
2694 
2695  getQueryParams(st, command, params);
2696  preparedStatementName(name, st->use_file, st->command);
2697 
2698  pg_log_debug("client %d sending %s", st->id, name);
2699  r = PQsendQueryPrepared(st->con, name, command->argc - 1,
2700  params, NULL, NULL, 0);
2701  }
2702  else /* unknown sql mode */
2703  r = 0;
2704 
2705  if (r == 0)
2706  {
2707  pg_log_debug("client %d could not send %s", st->id, command->argv[0]);
2708  st->ecnt++;
2709  return false;
2710  }
2711  else
2712  return true;
2713 }
2714 
2715 /*
2716  * Process query response from the backend.
2717  *
2718  * If varprefix is not NULL, it's the variable name prefix where to store
2719  * the results of the *last* command (META_GSET) or *all* commands
2720  * (META_ASET).
2721  *
2722  * Returns true if everything is A-OK, false if any error occurs.
2723  */
2724 static bool
2725 readCommandResponse(CState *st, MetaCommand meta, char *varprefix)
2726 {
2727  PGresult *res;
2728  PGresult *next_res;
2729  int qrynum = 0;
2730 
2731  /*
2732  * varprefix should be set only with \gset or \aset, and SQL commands do
2733  * not need it.
2734  */
2735  Assert((meta == META_NONE && varprefix == NULL) ||
2736  ((meta == META_GSET || meta == META_ASET) && varprefix != NULL));
2737 
2738  res = PQgetResult(st->con);
2739 
2740  while (res != NULL)
2741  {
2742  bool is_last;
2743 
2744  /* peek at the next result to know whether the current is last */
2745  next_res = PQgetResult(st->con);
2746  is_last = (next_res == NULL);
2747 
2748  switch (PQresultStatus(res))
2749  {
2750  case PGRES_COMMAND_OK: /* non-SELECT commands */
2751  case PGRES_EMPTY_QUERY: /* may be used for testing no-op overhead */
2752  if (is_last && meta == META_GSET)
2753  {
2754  pg_log_error("client %d script %d command %d query %d: expected one row, got %d",
2755  st->id, st->use_file, st->command, qrynum, 0);
2756  goto error;
2757  }
2758  break;
2759 
2760  case PGRES_TUPLES_OK:
2761  if ((is_last && meta == META_GSET) || meta == META_ASET)
2762  {
2763  int ntuples = PQntuples(res);
2764 
2765  if (meta == META_GSET && ntuples != 1)
2766  {
2767  /* under \gset, report the error */
2768  pg_log_error("client %d script %d command %d query %d: expected one row, got %d",
2769  st->id, st->use_file, st->command, qrynum, PQntuples(res));
2770  goto error;
2771  }
2772  else if (meta == META_ASET && ntuples <= 0)
2773  {
2774  /* coldly skip empty result under \aset */
2775  break;
2776  }
2777 
2778  /* store results into variables */
2779  for (int fld = 0; fld < PQnfields(res); fld++)
2780  {
2781  char *varname = PQfname(res, fld);
2782 
2783  /* allocate varname only if necessary, freed below */
2784  if (*varprefix != '\0')
2785  varname = psprintf("%s%s", varprefix, varname);
2786 
2787  /* store last row result as a string */
2788  if (!putVariable(st, meta == META_ASET ? "aset" : "gset", varname,
2789  PQgetvalue(res, ntuples - 1, fld)))
2790  {
2791  /* internal error */
2792  pg_log_error("client %d script %d command %d query %d: error storing into variable %s",
2793  st->id, st->use_file, st->command, qrynum, varname);
2794  goto error;
2795  }
2796 
2797  if (*varprefix != '\0')
2798  pg_free(varname);
2799  }
2800  }
2801  /* otherwise the result is simply thrown away by PQclear below */
2802  break;
2803 
2804  default:
2805  /* anything else is unexpected */
2806  pg_log_error("client %d script %d aborted in command %d query %d: %s",
2807  st->id, st->use_file, st->command, qrynum,
2808  PQerrorMessage(st->con));
2809  goto error;
2810  }
2811 
2812  PQclear(res);
2813  qrynum++;
2814  res = next_res;
2815  }
2816 
2817  if (qrynum == 0)
2818  {
2819  pg_log_error("client %d command %d: no results", st->id, st->command);
2820  st->ecnt++;
2821  return false;
2822  }
2823 
2824  return true;
2825 
2826 error:
2827  st->ecnt++;
2828  PQclear(res);
2829  PQclear(next_res);
2830  do
2831  {
2832  res = PQgetResult(st->con);
2833  PQclear(res);
2834  } while (res);
2835 
2836  return false;
2837 }
2838 
2839 /*
2840  * Parse the argument to a \sleep command, and return the requested amount
2841  * of delay, in microseconds. Returns true on success, false on error.
2842  */
2843 static bool
2844 evaluateSleep(CState *st, int argc, char **argv, int *usecs)
2845 {
2846  char *var;
2847  int usec;
2848 
2849  if (*argv[1] == ':')
2850  {
2851  if ((var = getVariable(st, argv[1] + 1)) == NULL)
2852  {
2853  pg_log_error("%s: undefined variable \"%s\"", argv[0], argv[1] + 1);
2854  return false;
2855  }
2856  usec = atoi(var);
2857  }
2858  else
2859  usec = atoi(argv[1]);
2860 
2861  if (argc > 2)
2862  {
2863  if (pg_strcasecmp(argv[2], "ms") == 0)
2864  usec *= 1000;
2865  else if (pg_strcasecmp(argv[2], "s") == 0)
2866  usec *= 1000000;
2867  }
2868  else
2869  usec *= 1000000;
2870 
2871  *usecs = usec;
2872  return true;
2873 }
2874 
2875 /*
2876  * Advance the state machine of a connection.
2877  */
2878 static void
2880 {
2881  instr_time now;
2882 
2883  /*
2884  * gettimeofday() isn't free, so we get the current timestamp lazily the
2885  * first time it's needed, and reuse the same value throughout this
2886  * function after that. This also ensures that e.g. the calculated
2887  * latency reported in the log file and in the totals are the same. Zero
2888  * means "not set yet". Reset "now" when we execute shell commands or
2889  * expressions, which might take a non-negligible amount of time, though.
2890  */
2891  INSTR_TIME_SET_ZERO(now);
2892 
2893  /*
2894  * Loop in the state machine, until we have to wait for a result from the
2895  * server or have to sleep for throttling or \sleep.
2896  *
2897  * Note: In the switch-statement below, 'break' will loop back here,
2898  * meaning "continue in the state machine". Return is used to return to
2899  * the caller, giving the thread the opportunity to advance another
2900  * client.
2901  */
2902  for (;;)
2903  {
2904  Command *command;
2905 
2906  switch (st->state)
2907  {
2908  /* Select transaction (script) to run. */
2909  case CSTATE_CHOOSE_SCRIPT:
2910  st->use_file = chooseScript(thread);
2912 
2913  pg_log_debug("client %d executing script \"%s\"",
2914  st->id, sql_script[st->use_file].desc);
2915 
2916  /*
2917  * If time is over, we're done; otherwise, get ready to start
2918  * a new transaction, or to get throttled if that's requested.
2919  */
2920  st->state = timer_exceeded ? CSTATE_FINISHED :
2921  throttle_delay > 0 ? CSTATE_PREPARE_THROTTLE : CSTATE_START_TX;
2922  break;
2923 
2924  /* Start new transaction (script) */
2925  case CSTATE_START_TX:
2926 
2927  /* establish connection if needed, i.e. under --connect */
2928  if (st->con == NULL)
2929  {
2930  instr_time start;
2931 
2933  start = now;
2934  if ((st->con = doConnect()) == NULL)
2935  {
2936  pg_log_error("client %d aborted while establishing connection", st->id);
2937  st->state = CSTATE_ABORTED;
2938  break;
2939  }
2941  INSTR_TIME_ACCUM_DIFF(thread->conn_time, now, start);
2942 
2943  /* Reset session-local state */
2944  memset(st->prepared, 0, sizeof(st->prepared));
2945  }
2946 
2947  /* record transaction start time */
2949  st->txn_begin = now;
2950 
2951  /*
2952  * When not throttling, this is also the transaction's
2953  * scheduled start time.
2954  */
2955  if (!throttle_delay)
2957 
2958  /* Begin with the first command */
2960  st->command = 0;
2961  break;
2962 
2963  /*
2964  * Handle throttling once per transaction by sleeping.
2965  */
2967 
2968  /*
2969  * Generate a delay such that the series of delays will
2970  * approximate a Poisson distribution centered on the
2971  * throttle_delay time.
2972  *
2973  * If transactions are too slow or a given wait is shorter
2974  * than a transaction, the next transaction will start right
2975  * away.
2976  */
2977  Assert(throttle_delay > 0);
2978 
2979  thread->throttle_trigger +=
2980  getPoissonRand(&thread->ts_throttle_rs, throttle_delay);
2981  st->txn_scheduled = thread->throttle_trigger;
2982 
2983  /*
2984  * If --latency-limit is used, and this slot is already late
2985  * so that the transaction will miss the latency limit even if
2986  * it completed immediately, skip this time slot and schedule
2987  * to continue running on the next slot that isn't late yet.
2988  * But don't iterate beyond the -t limit, if one is given.
2989  */
2990  if (latency_limit)
2991  {
2992  int64 now_us;
2993 
2995  now_us = INSTR_TIME_GET_MICROSEC(now);
2996 
2997  while (thread->throttle_trigger < now_us - latency_limit &&
2998  (nxacts <= 0 || st->cnt < nxacts))
2999  {
3000  processXactStats(thread, st, &now, true, agg);
3001  /* next rendez-vous */
3002  thread->throttle_trigger +=
3003  getPoissonRand(&thread->ts_throttle_rs, throttle_delay);
3004  st->txn_scheduled = thread->throttle_trigger;
3005  }
3006 
3007  /*
3008  * stop client if -t was exceeded in the previous skip
3009  * loop
3010  */
3011  if (nxacts > 0 && st->cnt >= nxacts)
3012  {
3013  st->state = CSTATE_FINISHED;
3014  break;
3015  }
3016  }
3017 
3018  /*
3019  * stop client if next transaction is beyond pgbench end of
3020  * execution; otherwise, throttle it.
3021  */
3022  st->state = end_time > 0 && st->txn_scheduled > end_time ?
3024  break;
3025 
3026  /*
3027  * Wait until it's time to start next transaction.
3028  */
3029  case CSTATE_THROTTLE:
3031 
3032  if (INSTR_TIME_GET_MICROSEC(now) < st->txn_scheduled)
3033  return; /* still sleeping, nothing to do here */
3034 
3035  /* done sleeping, but don't start transaction if we're done */
3036  st->state = timer_exceeded ? CSTATE_FINISHED : CSTATE_START_TX;
3037  break;
3038 
3039  /*
3040  * Send a command to server (or execute a meta-command)
3041  */
3042  case CSTATE_START_COMMAND:
3043  command = sql_script[st->use_file].commands[st->command];
3044 
3045  /* Transition to script end processing if done */
3046  if (command == NULL)
3047  {
3048  st->state = CSTATE_END_TX;
3049  break;
3050  }
3051 
3052  /* record begin time of next command, and initiate it */
3053  if (report_per_command)
3054  {
3056  st->stmt_begin = now;
3057  }
3058 
3059  /* Execute the command */
3060  if (command->type == SQL_COMMAND)
3061  {
3062  if (!sendCommand(st, command))
3063  {
3064  commandFailed(st, "SQL", "SQL command send failed");
3065  st->state = CSTATE_ABORTED;
3066  }
3067  else
3068  st->state = CSTATE_WAIT_RESULT;
3069  }
3070  else if (command->type == META_COMMAND)
3071  {
3072  /*-----
3073  * Possible state changes when executing meta commands:
3074  * - on errors CSTATE_ABORTED
3075  * - on sleep CSTATE_SLEEP
3076  * - else CSTATE_END_COMMAND
3077  */
3078  st->state = executeMetaCommand(st, &now);
3079  }
3080 
3081  /*
3082  * We're now waiting for an SQL command to complete, or
3083  * finished processing a metacommand, or need to sleep, or
3084  * something bad happened.
3085  */
3086  Assert(st->state == CSTATE_WAIT_RESULT ||
3087  st->state == CSTATE_END_COMMAND ||
3088  st->state == CSTATE_SLEEP ||
3089  st->state == CSTATE_ABORTED);
3090  break;
3091 
3092  /*
3093  * non executed conditional branch
3094  */
3095  case CSTATE_SKIP_COMMAND:
3097  /* quickly skip commands until something to do... */
3098  while (true)
3099  {
3100  Command *command;
3101 
3102  command = sql_script[st->use_file].commands[st->command];
3103 
3104  /* cannot reach end of script in that state */
3105  Assert(command != NULL);
3106 
3107  /*
3108  * if this is conditional related, update conditional
3109  * state
3110  */
3111  if (command->type == META_COMMAND &&
3112  (command->meta == META_IF ||
3113  command->meta == META_ELIF ||
3114  command->meta == META_ELSE ||
3115  command->meta == META_ENDIF))
3116  {
3117  switch (conditional_stack_peek(st->cstack))
3118  {
3119  case IFSTATE_FALSE:
3120  if (command->meta == META_IF ||
3121  command->meta == META_ELIF)
3122  {
3123  /* we must evaluate the condition */
3125  }
3126  else if (command->meta == META_ELSE)
3127  {
3128  /* we must execute next command */
3132  st->command++;
3133  }
3134  else if (command->meta == META_ENDIF)
3135  {
3138  if (conditional_active(st->cstack))
3140 
3141  /*
3142  * else state remains in
3143  * CSTATE_SKIP_COMMAND
3144  */
3145  st->command++;
3146  }
3147  break;
3148 
3149  case IFSTATE_IGNORED:
3150  case IFSTATE_ELSE_FALSE:
3151  if (command->meta == META_IF)
3153  IFSTATE_IGNORED);
3154  else if (command->meta == META_ENDIF)
3155  {
3158  if (conditional_active(st->cstack))
3160  }
3161  /* could detect "else" & "elif" after "else" */
3162  st->command++;
3163  break;
3164 
3165  case IFSTATE_NONE:
3166  case IFSTATE_TRUE:
3167  case IFSTATE_ELSE_TRUE:
3168  default:
3169 
3170  /*
3171  * inconsistent if inactive, unreachable dead
3172  * code
3173  */
3174  Assert(false);
3175  }
3176  }
3177  else
3178  {
3179  /* skip and consider next */
3180  st->command++;
3181  }
3182 
3183  if (st->state != CSTATE_SKIP_COMMAND)
3184  /* out of quick skip command loop */
3185  break;
3186  }
3187  break;
3188 
3189  /*
3190  * Wait for the current SQL command to complete
3191  */
3192  case CSTATE_WAIT_RESULT:
3193  pg_log_debug("client %d receiving", st->id);
3194  if (!PQconsumeInput(st->con))
3195  {
3196  /* there's something wrong */
3197  commandFailed(st, "SQL", "perhaps the backend died while processing");
3198  st->state = CSTATE_ABORTED;
3199  break;
3200  }
3201  if (PQisBusy(st->con))
3202  return; /* don't have the whole result yet */
3203 
3204  /* store or discard the query results */
3205  if (readCommandResponse(st,
3206  sql_script[st->use_file].commands[st->command]->meta,
3207  sql_script[st->use_file].commands[st->command]->varprefix))
3208  st->state = CSTATE_END_COMMAND;
3209  else
3210  st->state = CSTATE_ABORTED;
3211  break;
3212 
3213  /*
3214  * Wait until sleep is done. This state is entered after a
3215  * \sleep metacommand. The behavior is similar to
3216  * CSTATE_THROTTLE, but proceeds to CSTATE_START_COMMAND
3217  * instead of CSTATE_START_TX.
3218  */
3219  case CSTATE_SLEEP:
3221  if (INSTR_TIME_GET_MICROSEC(now) < st->sleep_until)
3222  return; /* still sleeping, nothing to do here */
3223  /* Else done sleeping. */
3224  st->state = CSTATE_END_COMMAND;
3225  break;
3226 
3227  /*
3228  * End of command: record stats and proceed to next command.
3229  */
3230  case CSTATE_END_COMMAND:
3231 
3232  /*
3233  * command completed: accumulate per-command execution times
3234  * in thread-local data structure, if per-command latencies
3235  * are requested.
3236  */
3237  if (report_per_command)
3238  {
3239  Command *command;
3240 
3242 
3243  command = sql_script[st->use_file].commands[st->command];
3244  /* XXX could use a mutex here, but we choose not to */
3245  addToSimpleStats(&command->stats,
3246  INSTR_TIME_GET_DOUBLE(now) -
3248  }
3249 
3250  /* Go ahead with next command, to be executed or skipped */
3251  st->command++;
3252  st->state = conditional_active(st->cstack) ?
3254  break;
3255 
3256  /*
3257  * End of transaction (end of script, really).
3258  */
3259  case CSTATE_END_TX:
3260 
3261  /* transaction finished: calculate latency and do log */
3262  processXactStats(thread, st, &now, false, agg);
3263 
3264  /*
3265  * missing \endif... cannot happen if CheckConditional was
3266  * okay
3267  */
3269 
3270  if (is_connect)
3271  {
3272  finishCon(st);
3273  INSTR_TIME_SET_ZERO(now);
3274  }
3275 
3276  if ((st->cnt >= nxacts && duration <= 0) || timer_exceeded)
3277  {
3278  /* script completed */
3279  st->state = CSTATE_FINISHED;
3280  break;
3281  }
3282 
3283  /* next transaction (script) */
3285 
3286  /*
3287  * Ensure that we always return on this point, so as to avoid
3288  * an infinite loop if the script only contains meta commands.
3289  */
3290  return;
3291 
3292  /*
3293  * Final states. Close the connection if it's still open.
3294  */
3295  case CSTATE_ABORTED:
3296  case CSTATE_FINISHED:
3297  finishCon(st);
3298  return;
3299  }
3300  }
3301 }
3302 
3303 /*
3304  * Subroutine for advanceConnectionState -- initiate or execute the current
3305  * meta command, and return the next state to set.
3306  *
3307  * *now is updated to the current time, unless the command is expected to
3308  * take no time to execute.
3309  */
3310 static ConnectionStateEnum
3312 {
3313  Command *command = sql_script[st->use_file].commands[st->command];
3314  int argc;
3315  char **argv;
3316 
3317  Assert(command != NULL && command->type == META_COMMAND);
3318 
3319  argc = command->argc;
3320  argv = command->argv;
3321 
3323  {
3325 
3326  initPQExpBuffer(&buf);
3327 
3328  printfPQExpBuffer(&buf, "client %d executing \\%s", st->id, argv[0]);
3329  for (int i = 1; i < argc; i++)
3330  appendPQExpBuffer(&buf, " %s", argv[i]);
3331 
3332  pg_log_debug("%s", buf.data);
3333 
3334  termPQExpBuffer(&buf);
3335  }
3336 
3337  if (command->meta == META_SLEEP)
3338  {
3339  int usec;
3340 
3341  /*
3342  * A \sleep doesn't execute anything, we just get the delay from the
3343  * argument, and enter the CSTATE_SLEEP state. (The per-command
3344  * latency will be recorded in CSTATE_SLEEP state, not here, after the
3345  * delay has elapsed.)
3346  */
3347  if (!evaluateSleep(st, argc, argv, &usec))
3348  {
3349  commandFailed(st, "sleep", "execution of meta-command failed");
3350  return CSTATE_ABORTED;
3351  }
3352 
3354  st->sleep_until = INSTR_TIME_GET_MICROSEC(*now) + usec;
3355  return CSTATE_SLEEP;
3356  }
3357  else if (command->meta == META_SET)
3358  {
3359  PgBenchExpr *expr = command->expr;
3360  PgBenchValue result;
3361 
3362  if (!evaluateExpr(st, expr, &result))
3363  {
3364  commandFailed(st, argv[0], "evaluation of meta-command failed");
3365  return CSTATE_ABORTED;
3366  }
3367 
3368  if (!putVariableValue(st, argv[0], argv[1], &result))
3369  {
3370  commandFailed(st, "set", "assignment of meta-command failed");
3371  return CSTATE_ABORTED;
3372  }
3373  }
3374  else if (command->meta == META_IF)
3375  {
3376  /* backslash commands with an expression to evaluate */
3377  PgBenchExpr *expr = command->expr;
3378  PgBenchValue result;
3379  bool cond;
3380 
3381  if (!evaluateExpr(st, expr, &result))
3382  {
3383  commandFailed(st, argv[0], "evaluation of meta-command failed");
3384  return CSTATE_ABORTED;
3385  }
3386 
3387  cond = valueTruth(&result);
3389  }
3390  else if (command->meta == META_ELIF)
3391  {
3392  /* backslash commands with an expression to evaluate */
3393  PgBenchExpr *expr = command->expr;
3394  PgBenchValue result;
3395  bool cond;
3396 
3398  {
3399  /* elif after executed block, skip eval and wait for endif. */
3401  return CSTATE_END_COMMAND;
3402  }
3403 
3404  if (!evaluateExpr(st, expr, &result))
3405  {
3406  commandFailed(st, argv[0], "evaluation of meta-command failed");
3407  return CSTATE_ABORTED;
3408  }
3409 
3410  cond = valueTruth(&result);
3413  }
3414  else if (command->meta == META_ELSE)
3415  {
3416  switch (conditional_stack_peek(st->cstack))
3417  {
3418  case IFSTATE_TRUE:
3420  break;
3421  case IFSTATE_FALSE: /* inconsistent if active */
3422  case IFSTATE_IGNORED: /* inconsistent if active */
3423  case IFSTATE_NONE: /* else without if */
3424  case IFSTATE_ELSE_TRUE: /* else after else */
3425  case IFSTATE_ELSE_FALSE: /* else after else */
3426  default:
3427  /* dead code if conditional check is ok */
3428  Assert(false);
3429  }
3430  }
3431  else if (command->meta == META_ENDIF)
3432  {
3435  }
3436  else if (command->meta == META_SETSHELL)
3437  {
3438  if (!runShellCommand(st, argv[1], argv + 2, argc - 2))
3439  {
3440  commandFailed(st, "setshell", "execution of meta-command failed");
3441  return CSTATE_ABORTED;
3442  }
3443  }
3444  else if (command->meta == META_SHELL)
3445  {
3446  if (!runShellCommand(st, NULL, argv + 1, argc - 1))
3447  {
3448  commandFailed(st, "shell", "execution of meta-command failed");
3449  return CSTATE_ABORTED;
3450  }
3451  }
3452 
3453  /*
3454  * executing the expression or shell command might have taken a
3455  * non-negligible amount of time, so reset 'now'
3456  */
3457  INSTR_TIME_SET_ZERO(*now);
3458 
3459  return CSTATE_END_COMMAND;
3460 }
3461 
3462 /*
3463  * Print log entry after completing one transaction.
3464  *
3465  * We print Unix-epoch timestamps in the log, so that entries can be
3466  * correlated against other logs. On some platforms this could be obtained
3467  * from the instr_time reading the caller has, but rather than get entangled
3468  * with that, we just eat the cost of an extra syscall in all cases.
3469  */
3470 static void
3471 doLog(TState *thread, CState *st,
3472  StatsData *agg, bool skipped, double latency, double lag)
3473 {
3474  FILE *logfile = thread->logfile;
3475 
3476  Assert(use_log);
3477 
3478  /*
3479  * Skip the log entry if sampling is enabled and this row doesn't belong
3480  * to the random sample.
3481  */
3482  if (sample_rate != 0.0 &&
3483  pg_erand48(thread->ts_sample_rs.xseed) > sample_rate)
3484  return;
3485 
3486  /* should we aggregate the results or not? */
3487  if (agg_interval > 0)
3488  {
3489  /*
3490  * Loop until we reach the interval of the current moment, and print
3491  * any empty intervals in between (this may happen with very low tps,
3492  * e.g. --rate=0.1).
3493  */
3494  time_t now = time(NULL);
3495 
3496  while (agg->start_time + agg_interval <= now)
3497  {
3498  /* print aggregated report to logfile */
3499  fprintf(logfile, "%ld " INT64_FORMAT " %.0f %.0f %.0f %.0f",
3500  (long) agg->start_time,
3501  agg->cnt,
3502  agg->latency.sum,
3503  agg->latency.sum2,
3504  agg->latency.min,
3505  agg->latency.max);
3506  if (throttle_delay)
3507  {
3508  fprintf(logfile, " %.0f %.0f %.0f %.0f",
3509  agg->lag.sum,
3510  agg->lag.sum2,
3511  agg->lag.min,
3512  agg->lag.max);
3513  if (latency_limit)
3514  fprintf(logfile, " " INT64_FORMAT, agg->skipped);
3515  }
3516  fputc('\n', logfile);
3517 
3518  /* reset data and move to next interval */
3519  initStats(agg, agg->start_time + agg_interval);
3520  }
3521 
3522  /* accumulate the current transaction */
3523  accumStats(agg, skipped, latency, lag);
3524  }
3525  else
3526  {
3527  /* no, print raw transactions */
3528  struct timeval tv;
3529 
3530  gettimeofday(&tv, NULL);
3531  if (skipped)
3532  fprintf(logfile, "%d " INT64_FORMAT " skipped %d %ld %ld",
3533  st->id, st->cnt, st->use_file,
3534  (long) tv.tv_sec, (long) tv.tv_usec);
3535  else
3536  fprintf(logfile, "%d " INT64_FORMAT " %.0f %d %ld %ld",
3537  st->id, st->cnt, latency, st->use_file,
3538  (long) tv.tv_sec, (long) tv.tv_usec);
3539  if (throttle_delay)
3540  fprintf(logfile, " %.0f", lag);
3541  fputc('\n', logfile);
3542  }
3543 }
3544 
3545 /*
3546  * Accumulate and report statistics at end of a transaction.
3547  *
3548  * (This is also called when a transaction is late and thus skipped.
3549  * Note that even skipped transactions are counted in the "cnt" fields.)
3550  */
3551 static void
3553  bool skipped, StatsData *agg)
3554 {
3555  double latency = 0.0,
3556  lag = 0.0;
3557  bool thread_details = progress || throttle_delay || latency_limit,
3558  detailed = thread_details || use_log || per_script_stats;
3559 
3560  if (detailed && !skipped)
3561  {
3563 
3564  /* compute latency & lag */
3565  latency = INSTR_TIME_GET_MICROSEC(*now) - st->txn_scheduled;
3567  }
3568 
3569  if (thread_details)
3570  {
3571  /* keep detailed thread stats */
3572  accumStats(&thread->stats, skipped, latency, lag);
3573 
3574  /* count transactions over the latency limit, if needed */
3575  if (latency_limit && latency > latency_limit)
3576  thread->latency_late++;
3577  }
3578  else
3579  {
3580  /* no detailed stats, just count */
3581  thread->stats.cnt++;
3582  }
3583 
3584  /* client stat is just counting */
3585  st->cnt++;
3586 
3587  if (use_log)
3588  doLog(thread, st, agg, skipped, latency, lag);
3589 
3590  /* XXX could use a mutex here, but we choose not to */
3591  if (per_script_stats)
3592  accumStats(&sql_script[st->use_file].stats, skipped, latency, lag);
3593 }
3594 
3595 
3596 /* discard connections */
3597 static void
3599 {
3600  int i;
3601 
3602  for (i = 0; i < length; i++)
3603  finishCon(&state[i]);
3604 }
3605 
3606 /*
3607  * Remove old pgbench tables, if any exist
3608  */
3609 static void
3611 {
3612  fprintf(stderr, "dropping old tables...\n");
3613 
3614  /*
3615  * We drop all the tables in one command, so that whether there are
3616  * foreign key dependencies or not doesn't matter.
3617  */
3618  executeStatement(con, "drop table if exists "
3619  "pgbench_accounts, "
3620  "pgbench_branches, "
3621  "pgbench_history, "
3622  "pgbench_tellers");
3623 }
3624 
3625 /*
3626  * Create "pgbench_accounts" partitions if needed.
3627  *
3628  * This is the larger table of pgbench default tpc-b like schema
3629  * with a known size, so we choose to partition it.
3630  */
3631 static void
3633 {
3634  char ff[64];
3635 
3636  ff[0] = '\0';
3637 
3638  /*
3639  * Per ddlinfo in initCreateTables, fillfactor is needed on table
3640  * pgbench_accounts.
3641  */
3642  append_fillfactor(ff, sizeof(ff));
3643 
3644  /* we must have to create some partitions */
3645  Assert(partitions > 0);
3646 
3647  fprintf(stderr, "creating %d partitions...\n", partitions);
3648 
3649  for (int p = 1; p <= partitions; p++)
3650  {
3651  char query[256];
3652 
3653  if (partition_method == PART_RANGE)
3654  {
3655  int64 part_size = (naccounts * (int64) scale + partitions - 1) / partitions;
3656  char minvalue[32],
3657  maxvalue[32];
3658 
3659  /*
3660  * For RANGE, we use open-ended partitions at the beginning and
3661  * end to allow any valid value for the primary key. Although the
3662  * actual minimum and maximum values can be derived from the
3663  * scale, it is more generic and the performance is better.
3664  */
3665  if (p == 1)
3666  sprintf(minvalue, "minvalue");
3667  else
3668  sprintf(minvalue, INT64_FORMAT, (p - 1) * part_size + 1);
3669 
3670  if (p < partitions)
3671  sprintf(maxvalue, INT64_FORMAT, p * part_size + 1);
3672  else
3673  sprintf(maxvalue, "maxvalue");
3674 
3675  snprintf(query, sizeof(query),
3676  "create%s table pgbench_accounts_%d\n"
3677  " partition of pgbench_accounts\n"
3678  " for values from (%s) to (%s)%s\n",
3679  unlogged_tables ? " unlogged" : "", p,
3680  minvalue, maxvalue, ff);
3681  }
3682  else if (partition_method == PART_HASH)
3683  snprintf(query, sizeof(query),
3684  "create%s table pgbench_accounts_%d\n"
3685  " partition of pgbench_accounts\n"
3686  " for values with (modulus %d, remainder %d)%s\n",
3687  unlogged_tables ? " unlogged" : "", p,
3688  partitions, p - 1, ff);
3689  else /* cannot get there */
3690  Assert(0);
3691 
3692  executeStatement(con, query);
3693  }
3694 }
3695 
3696 /*
3697  * Create pgbench's standard tables
3698  */
3699 static void
3701 {
3702  /*
3703  * Note: TPC-B requires at least 100 bytes per row, and the "filler"
3704  * fields in these table declarations were intended to comply with that.
3705  * The pgbench_accounts table complies with that because the "filler"
3706  * column is set to blank-padded empty string. But for all other tables
3707  * the columns default to NULL and so don't actually take any space. We
3708  * could fix that by giving them non-null default values. However, that
3709  * would completely break comparability of pgbench results with prior
3710  * versions. Since pgbench has never pretended to be fully TPC-B compliant
3711  * anyway, we stick with the historical behavior.
3712  */
3713  struct ddlinfo
3714  {
3715  const char *table; /* table name */
3716  const char *smcols; /* column decls if accountIDs are 32 bits */
3717  const char *bigcols; /* column decls if accountIDs are 64 bits */
3718  int declare_fillfactor;
3719  };
3720  static const struct ddlinfo DDLs[] = {
3721  {
3722  "pgbench_history",
3723  "tid int,bid int,aid int,delta int,mtime timestamp,filler char(22)",
3724  "tid int,bid int,aid bigint,delta int,mtime timestamp,filler char(22)",
3725  0
3726  },
3727  {
3728  "pgbench_tellers",
3729  "tid int not null,bid int,tbalance int,filler char(84)",
3730  "tid int not null,bid int,tbalance int,filler char(84)",
3731  1
3732  },
3733  {
3734  "pgbench_accounts",
3735  "aid int not null,bid int,abalance int,filler char(84)",
3736  "aid bigint not null,bid int,abalance int,filler char(84)",
3737  1
3738  },
3739  {
3740  "pgbench_branches",
3741  "bid int not null,bbalance int,filler char(88)",
3742  "bid int not null,bbalance int,filler char(88)",
3743  1
3744  }
3745  };
3746  int i;
3747 
3748  fprintf(stderr, "creating tables...\n");
3749 
3750  for (i = 0; i < lengthof(DDLs); i++)
3751  {
3752  char opts[256];
3753  char buffer[256];
3754  const struct ddlinfo *ddl = &DDLs[i];
3755  const char *cols;
3756 
3757  /* Construct new create table statement. */
3758  opts[0] = '\0';
3759 
3760  /* Partition pgbench_accounts table */
3761  if (partition_method != PART_NONE && strcmp(ddl->table, "pgbench_accounts") == 0)
3762  snprintf(opts + strlen(opts), sizeof(opts) - strlen(opts),
3763  " partition by %s (aid)", PARTITION_METHOD[partition_method]);
3764  else if (ddl->declare_fillfactor)
3765  /* fillfactor is only expected on actual tables */
3766  append_fillfactor(opts, sizeof(opts));
3767 
3768  if (tablespace != NULL)
3769  {
3770  char *escape_tablespace;
3771 
3772  escape_tablespace = PQescapeIdentifier(con, tablespace,
3773  strlen(tablespace));
3774  snprintf(opts + strlen(opts), sizeof(opts) - strlen(opts),
3775  " tablespace %s", escape_tablespace);
3776  PQfreemem(escape_tablespace);
3777  }
3778 
3779  cols = (scale >= SCALE_32BIT_THRESHOLD) ? ddl->bigcols : ddl->smcols;
3780 
3781  snprintf(buffer, sizeof(buffer), "create%s table %s(%s)%s",
3782  unlogged_tables ? " unlogged" : "",
3783  ddl->table, cols, opts);
3784 
3785  executeStatement(con, buffer);
3786  }
3787 
3788  if (partition_method != PART_NONE)
3789  createPartitions(con);
3790 }
3791 
3792 /*
3793  * add fillfactor percent option.
3794  *
3795  * XXX - As default is 100, it could be removed in this case.
3796  */
3797 static void
3798 append_fillfactor(char *opts, int len)
3799 {
3800  snprintf(opts + strlen(opts), len - strlen(opts),
3801  " with (fillfactor=%d)", fillfactor);
3802 }
3803 
3804 /*
3805  * Truncate away any old data, in one command in case there are foreign keys
3806  */
3807 static void
3809 {
3810  executeStatement(con, "truncate table "
3811  "pgbench_accounts, "
3812  "pgbench_branches, "
3813  "pgbench_history, "
3814  "pgbench_tellers");
3815 }
3816 
3817 /*
3818  * Fill the standard tables with some data generated and sent from the client
3819  */
3820 static void
3822 {
3823  char sql[256];
3824  PGresult *res;
3825  int i;
3826  int64 k;
3827 
3828  /* used to track elapsed time and estimate of the remaining time */
3829  instr_time start,
3830  diff;
3831  double elapsed_sec,
3832  remaining_sec;
3833  int log_interval = 1;
3834 
3835  /* Stay on the same line if reporting to a terminal */
3836  char eol = isatty(fileno(stderr)) ? '\r' : '\n';
3837 
3838  fprintf(stderr, "generating data (client-side)...\n");
3839 
3840  /*
3841  * we do all of this in one transaction to enable the backend's
3842  * data-loading optimizations
3843  */
3844  executeStatement(con, "begin");
3845 
3846  /* truncate away any old data */
3847  initTruncateTables(con);
3848 
3849  /*
3850  * fill branches, tellers, accounts in that order in case foreign keys
3851  * already exist
3852  */
3853  for (i = 0; i < nbranches * scale; i++)
3854  {
3855  /* "filler" column defaults to NULL */
3856  snprintf(sql, sizeof(sql),
3857  "insert into pgbench_branches(bid,bbalance) values(%d,0)",
3858  i + 1);
3859  executeStatement(con, sql);
3860  }
3861 
3862  for (i = 0; i < ntellers * scale; i++)
3863  {
3864  /* "filler" column defaults to NULL */
3865  snprintf(sql, sizeof(sql),
3866  "insert into pgbench_tellers(tid,bid,tbalance) values (%d,%d,0)",
3867  i + 1, i / ntellers + 1);
3868  executeStatement(con, sql);
3869  }
3870 
3871  /*
3872  * accounts is big enough to be worth using COPY and tracking runtime
3873  */
3874  res = PQexec(con, "copy pgbench_accounts from stdin");
3875  if (PQresultStatus(res) != PGRES_COPY_IN)
3876  {
3877  pg_log_fatal("unexpected copy in result: %s", PQerrorMessage(con));
3878  exit(1);
3879  }
3880  PQclear(res);
3881 
3882  INSTR_TIME_SET_CURRENT(start);
3883 
3884  for (k = 0; k < (int64) naccounts * scale; k++)
3885  {
3886  int64 j = k + 1;
3887 
3888  /* "filler" column defaults to blank padded empty string */
3889  snprintf(sql, sizeof(sql),
3890  INT64_FORMAT "\t" INT64_FORMAT "\t%d\t\n",
3891  j, k / naccounts + 1, 0);
3892  if (PQputline(con, sql))
3893  {
3894  pg_log_fatal("PQputline failed");
3895  exit(1);
3896  }
3897 
3898  if (CancelRequested)
3899  break;
3900 
3901  /*
3902  * If we want to stick with the original logging, print a message each
3903  * 100k inserted rows.
3904  */
3905  if ((!use_quiet) && (j % 100000 == 0))
3906  {
3907  INSTR_TIME_SET_CURRENT(diff);
3908  INSTR_TIME_SUBTRACT(diff, start);
3909 
3910  elapsed_sec = INSTR_TIME_GET_DOUBLE(diff);
3911  remaining_sec = ((double) scale * naccounts - j) * elapsed_sec / j;
3912 
3913  fprintf(stderr, INT64_FORMAT " of " INT64_FORMAT " tuples (%d%%) done (elapsed %.2f s, remaining %.2f s)%c",
3914  j, (int64) naccounts * scale,
3915  (int) (((int64) j * 100) / (naccounts * (int64) scale)),
3916  elapsed_sec, remaining_sec, eol);
3917  }
3918  /* let's not call the timing for each row, but only each 100 rows */
3919  else if (use_quiet && (j % 100 == 0))
3920  {
3921  INSTR_TIME_SET_CURRENT(diff);
3922  INSTR_TIME_SUBTRACT(diff, start);
3923 
3924  elapsed_sec = INSTR_TIME_GET_DOUBLE(diff);
3925  remaining_sec = ((double) scale * naccounts - j) * elapsed_sec / j;
3926 
3927  /* have we reached the next interval (or end)? */
3928  if ((j == scale * naccounts) || (elapsed_sec >= log_interval * LOG_STEP_SECONDS))
3929  {
3930  fprintf(stderr, INT64_FORMAT " of " INT64_FORMAT " tuples (%d%%) done (elapsed %.2f s, remaining %.2f s)%c",
3931  j, (int64) naccounts * scale,
3932  (int) (((int64) j * 100) / (naccounts * (int64) scale)), elapsed_sec, remaining_sec, eol);
3933 
3934  /* skip to the next interval */
3935  log_interval = (int) ceil(elapsed_sec / LOG_STEP_SECONDS);
3936  }
3937  }
3938  }
3939 
3940  if (eol != '\n')
3941  fputc('\n', stderr); /* Need to move to next line */
3942 
3943  if (PQputline(con, "\\.\n"))
3944  {
3945  pg_log_fatal("very last PQputline failed");
3946  exit(1);
3947  }
3948  if (PQendcopy(con))
3949  {
3950  pg_log_fatal("PQendcopy failed");
3951  exit(1);
3952  }
3953 
3954  executeStatement(con, "commit");
3955 }
3956 
3957 /*
3958  * Fill the standard tables with some data generated on the server
3959  *
3960  * As already the case with the client-side data generation, the filler
3961  * column defaults to NULL in pgbench_branches and pgbench_tellers,
3962  * and is a blank-padded string in pgbench_accounts.
3963  */
3964 static void
3966 {
3967  char sql[256];
3968 
3969  fprintf(stderr, "generating data (server-side)...\n");
3970 
3971  /*
3972  * we do all of this in one transaction to enable the backend's
3973  * data-loading optimizations
3974  */
3975  executeStatement(con, "begin");
3976 
3977  /* truncate away any old data */
3978  initTruncateTables(con);
3979 
3980  snprintf(sql, sizeof(sql),
3981  "insert into pgbench_branches(bid,bbalance) "
3982  "select bid, 0 "
3983  "from generate_series(1, %d) as bid", nbranches * scale);
3984  executeStatement(con, sql);
3985 
3986  snprintf(sql, sizeof(sql),
3987  "insert into pgbench_tellers(tid,bid,tbalance) "
3988  "select tid, (tid - 1) / %d + 1, 0 "
3989  "from generate_series(1, %d) as tid", ntellers, ntellers * scale);
3990  executeStatement(con, sql);
3991 
3992  snprintf(sql, sizeof(sql),
3993  "insert into pgbench_accounts(aid,bid,abalance,filler) "
3994  "select aid, (aid - 1) / %d + 1, 0, '' "
3995  "from generate_series(1, " INT64_FORMAT ") as aid",
3996  naccounts, (int64) naccounts * scale);
3997  executeStatement(con, sql);
3998 
3999  executeStatement(con, "commit");
4000 }
4001 
4002 /*
4003  * Invoke vacuum on the standard tables
4004  */
4005 static void
4007 {
4008  fprintf(stderr, "vacuuming...\n");
4009  executeStatement(con, "vacuum analyze pgbench_branches");
4010  executeStatement(con, "vacuum analyze pgbench_tellers");
4011  executeStatement(con, "vacuum analyze pgbench_accounts");
4012  executeStatement(con, "vacuum analyze pgbench_history");
4013 }
4014 
4015 /*
4016  * Create primary keys on the standard tables
4017  */
4018 static void
4020 {
4021  static const char *const DDLINDEXes[] = {
4022  "alter table pgbench_branches add primary key (bid)",
4023  "alter table pgbench_tellers add primary key (tid)",
4024  "alter table pgbench_accounts add primary key (aid)"
4025  };
4026  int i;
4027 
4028  fprintf(stderr, "creating primary keys...\n");
4029  for (i = 0; i < lengthof(DDLINDEXes); i++)
4030  {
4031  char buffer[256];
4032 
4033  strlcpy(buffer, DDLINDEXes[i], sizeof(buffer));
4034 
4035  if (index_tablespace != NULL)
4036  {
4037  char *escape_tablespace;
4038 
4039  escape_tablespace = PQescapeIdentifier(con, index_tablespace,
4040  strlen(index_tablespace));
4041  snprintf(buffer + strlen(buffer), sizeof(buffer) - strlen(buffer),
4042  " using index tablespace %s", escape_tablespace);
4043  PQfreemem(escape_tablespace);
4044  }
4045 
4046  executeStatement(con, buffer);
4047  }
4048 }
4049 
4050 /*
4051  * Create foreign key constraints between the standard tables
4052  */
4053 static void
4055 {
4056  static const char *const DDLKEYs[] = {
4057  "alter table pgbench_tellers add constraint pgbench_tellers_bid_fkey foreign key (bid) references pgbench_branches",
4058  "alter table pgbench_accounts add constraint pgbench_accounts_bid_fkey foreign key (bid) references pgbench_branches",
4059  "alter table pgbench_history add constraint pgbench_history_bid_fkey foreign key (bid) references pgbench_branches",
4060  "alter table pgbench_history add constraint pgbench_history_tid_fkey foreign key (tid) references pgbench_tellers",
4061  "alter table pgbench_history add constraint pgbench_history_aid_fkey foreign key (aid) references pgbench_accounts"
4062  };
4063  int i;
4064 
4065  fprintf(stderr, "creating foreign keys...\n");
4066  for (i = 0; i < lengthof(DDLKEYs); i++)
4067  {
4068  executeStatement(con, DDLKEYs[i]);
4069  }
4070 }
4071 
4072 /*
4073  * Validate an initialization-steps string
4074  *
4075  * (We could just leave it to runInitSteps() to fail if there are wrong
4076  * characters, but since initialization can take awhile, it seems friendlier
4077  * to check during option parsing.)
4078  */
4079 static void
4080 checkInitSteps(const char *initialize_steps)
4081 {
4082  if (initialize_steps[0] == '\0')
4083  {
4084  pg_log_fatal("no initialization steps specified");
4085  exit(1);
4086  }
4087 
4088  for (const char *step = initialize_steps; *step != '\0'; step++)
4089  {
4090  if (strchr(ALL_INIT_STEPS " ", *step) == NULL)
4091  {
4092  pg_log_fatal("unrecognized initialization step \"%c\"", *step);
4093  pg_log_info("Allowed step characters are: \"" ALL_INIT_STEPS "\".");
4094  exit(1);
4095  }
4096  }
4097 }
4098 
4099 /*
4100  * Invoke each initialization step in the given string
4101  */
4102 static void
4103 runInitSteps(const char *initialize_steps)
4104 {
4105  PQExpBufferData stats;
4106  PGconn *con;
4107  const char *step;
4108  double run_time = 0.0;
4109  bool first = true;
4110 
4111  initPQExpBuffer(&stats);
4112 
4113  if ((con = doConnect()) == NULL)
4114  exit(1);
4115 
4116  setup_cancel_handler(NULL);
4117  SetCancelConn(con);
4118 
4119  for (step = initialize_steps; *step != '\0'; step++)
4120  {
4121  instr_time start;
4122  char *op = NULL;
4123 
4124  INSTR_TIME_SET_CURRENT(start);
4125 
4126  switch (*step)
4127  {
4128  case 'd':
4129  op = "drop tables";
4130  initDropTables(con);
4131  break;
4132  case 't':
4133  op = "create tables";
4134  initCreateTables(con);
4135  break;
4136  case 'g':
4137  op = "client-side generate";
4139  break;
4140  case 'G':
4141  op = "server-side generate";
4143  break;
4144  case 'v':
4145  op = "vacuum";
4146  initVacuum(con);
4147  break;
4148  case 'p':
4149  op = "primary keys";
4150  initCreatePKeys(con);
4151  break;
4152  case 'f':
4153  op = "foreign keys";
4154  initCreateFKeys(con);
4155  break;
4156  case ' ':
4157  break; /* ignore */
4158  default:
4159  pg_log_fatal("unrecognized initialization step \"%c\"", *step);
4160  PQfinish(con);
4161  exit(1);
4162  }
4163 
4164  if (op != NULL)
4165  {
4166  instr_time diff;
4167  double elapsed_sec;
4168 
4169  INSTR_TIME_SET_CURRENT(diff);
4170  INSTR_TIME_SUBTRACT(diff, start);
4171  elapsed_sec = INSTR_TIME_GET_DOUBLE(diff);
4172 
4173  if (!first)
4174  appendPQExpBufferStr(&stats, ", ");
4175  else
4176  first = false;
4177 
4178  appendPQExpBuffer(&stats, "%s %.2f s", op, elapsed_sec);
4179 
4180  run_time += elapsed_sec;
4181  }
4182  }
4183 
4184  fprintf(stderr, "done in %.2f s (%s).\n", run_time, stats.data);
4185  ResetCancelConn();
4186  PQfinish(con);
4187  termPQExpBuffer(&stats);
4188 }
4189 
4190 /*
4191  * Extract pgbench table informations into global variables scale,
4192  * partition_method and partitions.
4193  */
4194 static void
4195 GetTableInfo(PGconn *con, bool scale_given)
4196 {
4197  PGresult *res;
4198 
4199  /*
4200  * get the scaling factor that should be same as count(*) from
4201  * pgbench_branches if this is not a custom query
4202  */
4203  res = PQexec(con, "select count(*) from pgbench_branches");
4204  if (PQresultStatus(res) != PGRES_TUPLES_OK)
4205  {
4206  char *sqlState = PQresultErrorField(res, PG_DIAG_SQLSTATE);
4207 
4208  pg_log_fatal("could not count number of branches: %s", PQerrorMessage(con));
4209 
4210  if (sqlState && strcmp(sqlState, ERRCODE_UNDEFINED_TABLE) == 0)
4211  pg_log_info("Perhaps you need to do initialization (\"pgbench -i\") in database \"%s\"",
4212  PQdb(con));
4213 
4214  exit(1);
4215  }
4216  scale = atoi(PQgetvalue(res, 0, 0));
4217  if (scale < 0)
4218  {
4219  pg_log_fatal("invalid count(*) from pgbench_branches: \"%s\"",
4220  PQgetvalue(res, 0, 0));
4221  exit(1);
4222  }
4223  PQclear(res);
4224 
4225  /* warn if we override user-given -s switch */
4226  if (scale_given)
4227  pg_log_warning("scale option ignored, using count from pgbench_branches table (%d)",
4228  scale);
4229 
4230  /*
4231  * Get the partition information for the first "pgbench_accounts" table
4232  * found in search_path.
4233  *
4234  * The result is empty if no "pgbench_accounts" is found.
4235  *
4236  * Otherwise, it always returns one row even if the table is not
4237  * partitioned (in which case the partition strategy is NULL).
4238  *
4239  * The number of partitions can be 0 even for partitioned tables, if no
4240  * partition is attached.
4241  *
4242  * We assume no partitioning on any failure, so as to avoid failing on an
4243  * old version without "pg_partitioned_table".
4244  */
4245  res = PQexec(con,
4246  "select o.n, p.partstrat, pg_catalog.count(i.inhparent) "
4247  "from pg_catalog.pg_class as c "
4248  "join pg_catalog.pg_namespace as n on (n.oid = c.relnamespace) "
4249  "cross join lateral (select pg_catalog.array_position(pg_catalog.current_schemas(true), n.nspname)) as o(n) "
4250  "left join pg_catalog.pg_partitioned_table as p on (p.partrelid = c.oid) "
4251  "left join pg_catalog.pg_inherits as i on (c.oid = i.inhparent) "
4252  "where c.relname = 'pgbench_accounts' and o.n is not null "
4253  "group by 1, 2 "
4254  "order by 1 asc "
4255  "limit 1");
4256 
4257  if (PQresultStatus(res) != PGRES_TUPLES_OK)
4258  {
4259  /* probably an older version, coldly assume no partitioning */
4260  partition_method = PART_NONE;
4261  partitions = 0;
4262  }
4263  else if (PQntuples(res) == 0)
4264  {
4265  /*
4266  * This case is unlikely as pgbench already found "pgbench_branches"
4267  * above to compute the scale.
4268  */
4269  pg_log_fatal("no pgbench_accounts table found in search_path");
4270  pg_log_info("Perhaps you need to do initialization (\"pgbench -i\") in database \"%s\".", PQdb(con));
4271  exit(1);
4272  }
4273  else /* PQntupes(res) == 1 */
4274  {
4275  /* normal case, extract partition information */
4276  if (PQgetisnull(res, 0, 1))
4277  partition_method = PART_NONE;
4278  else
4279  {
4280  char *ps = PQgetvalue(res, 0, 1);
4281 
4282  /* column must be there */
4283  Assert(ps != NULL);
4284 
4285  if (strcmp(ps, "r") == 0)
4286  partition_method = PART_RANGE;
4287  else if (strcmp(ps, "h") == 0)
4288  partition_method = PART_HASH;
4289  else
4290  {
4291  /* possibly a newer version with new partition method */
4292  pg_log_fatal("unexpected partition method: \"%s\"", ps);
4293  exit(1);
4294  }
4295  }
4296 
4297  partitions = atoi(PQgetvalue(res, 0, 2));
4298  }
4299 
4300  PQclear(res);
4301 }
4302 
4303 /*
4304  * Replace :param with $n throughout the command's SQL text, which
4305  * is a modifiable string in cmd->lines.
4306  */
4307 static bool
4309 {
4310  char *sql,
4311  *p;
4312 
4313  cmd->argc = 1;
4314 
4315  p = sql = pg_strdup(cmd->lines.data);
4316  while ((p = strchr(p, ':')) != NULL)
4317  {
4318  char var[13];
4319  char *name;
4320  int eaten;
4321 
4322  name = parseVariable(p, &eaten);
4323  if (name == NULL)
4324  {
4325  while (*p == ':')
4326  {
4327  p++;
4328  }
4329  continue;
4330  }
4331 
4332  /*
4333  * cmd->argv[0] is the SQL statement itself, so the max number of
4334  * arguments is one less than MAX_ARGS
4335  */
4336  if (cmd->argc >= MAX_ARGS)
4337  {
4338  pg_log_error("statement has too many arguments (maximum is %d): %s",
4339  MAX_ARGS - 1, cmd->lines.data);
4340  pg_free(name);
4341  return false;
4342  }
4343 
4344  sprintf(var, "$%d", cmd->argc);
4345  p = replaceVariable(&sql, p, eaten, var);
4346 
4347  cmd->argv[cmd->argc] = name;
4348  cmd->argc++;
4349  }
4350 
4351  Assert(cmd->argv[0] == NULL);
4352  cmd->argv[0] = sql;
4353  return true;
4354 }
4355 
4356 /*
4357  * syntax error while parsing a script (in practice, while parsing a
4358  * backslash command, because we don't detect syntax errors in SQL)
4359  *
4360  * source: source of script (filename or builtin-script ID)
4361  * lineno: line number within script (count from 1)
4362  * line: whole line of backslash command, if available
4363  * command: backslash command name, if available
4364  * msg: the actual error message
4365  * more: optional extra message
4366  * column: zero-based column number, or -1 if unknown
4367  */
4368 void
4369 syntax_error(const char *source, int lineno,
4370  const char *line, const char *command,
4371  const char *msg, const char *more, int column)
4372 {
4374 
4375  initPQExpBuffer(&buf);
4376 
4377  printfPQExpBuffer(&buf, "%s:%d: %s", source, lineno, msg);
4378  if (more != NULL)
4379  appendPQExpBuffer(&buf, " (%s)", more);
4380  if (column >= 0 && line == NULL)
4381  appendPQExpBuffer(&buf, " at column %d", column + 1);
4382  if (command != NULL)
4383  appendPQExpBuffer(&buf, " in command \"%s\"", command);
4384 
4385  pg_log_fatal("%s", buf.data);
4386 
4387  termPQExpBuffer(&buf);
4388 
4389  if (line != NULL)
4390  {
4391  fprintf(stderr, "%s\n", line);
4392  if (column >= 0)
4393  fprintf(stderr, "%*c error found here\n", column + 1, '^');
4394  }
4395 
4396  exit(1);
4397 }
4398 
4399 /*
4400  * Return a pointer to the start of the SQL command, after skipping over
4401  * whitespace and "--" comments.
4402  * If the end of the string is reached, return NULL.
4403  */
4404 static char *
4405 skip_sql_comments(char *sql_command)
4406 {
4407  char *p = sql_command;
4408 
4409  /* Skip any leading whitespace, as well as "--" style comments */
4410  for (;;)
4411  {
4412  if (isspace((unsigned char) *p))
4413  p++;
4414  else if (strncmp(p, "--", 2) == 0)
4415  {
4416  p = strchr(p, '\n');
4417  if (p == NULL)
4418  return NULL;
4419  p++;
4420  }
4421  else
4422  break;
4423  }
4424 
4425  /* NULL if there's nothing but whitespace and comments */
4426  if (*p == '\0')
4427  return NULL;
4428 
4429  return p;
4430 }
4431 
4432 /*
4433  * Parse a SQL command; return a Command struct, or NULL if it's a comment
4434  *
4435  * On entry, psqlscan.l has collected the command into "buf", so we don't
4436  * really need to do much here except check for comments and set up a Command
4437  * struct.
4438  */
4439 static Command *
4440 create_sql_command(PQExpBuffer buf, const char *source)
4441 {
4442  Command *my_command;
4443  char *p = skip_sql_comments(buf->data);
4444 
4445  if (p == NULL)
4446  return NULL;
4447 
4448  /* Allocate and initialize Command structure */
4449  my_command = (Command *) pg_malloc(sizeof(Command));
4450  initPQExpBuffer(&my_command->lines);
4451  appendPQExpBufferStr(&my_command->lines, p);
4452  my_command->first_line = NULL; /* this is set later */
4453  my_command->type = SQL_COMMAND;
4454  my_command->meta = META_NONE;
4455  my_command->argc = 0;
4456  memset(my_command->argv, 0, sizeof(my_command->argv));
4457  my_command->varprefix = NULL; /* allocated later, if needed */
4458  my_command->expr = NULL;
4459  initSimpleStats(&my_command->stats);
4460 
4461  return my_command;
4462 }
4463 
4464 /* Free a Command structure and associated data */
4465 static void
4467 {
4468  termPQExpBuffer(&command->lines);
4469  if (command->first_line)
4470  pg_free(command->first_line);
4471  for (int i = 0; i < command->argc; i++)
4472  pg_free(command->argv[i]);
4473  if (command->varprefix)
4474  pg_free(command->varprefix);
4475 
4476  /*
4477  * It should also free expr recursively, but this is currently not needed
4478  * as only gset commands (which do not have an expression) are freed.
4479  */
4480  pg_free(command);
4481 }
4482 
4483 /*
4484  * Once an SQL command is fully parsed, possibly by accumulating several
4485  * parts, complete other fields of the Command structure.
4486  */
4487 static void
4489 {
4490  char buffer[128];
4491 
4492  Assert(my_command->type == SQL_COMMAND);
4493 
4494  /* Save the first line for error display. */
4495  strlcpy(buffer, my_command->lines.data, sizeof(buffer));
4496  buffer[strcspn(buffer, "\n\r")] = '\0';
4497  my_command->first_line = pg_strdup(buffer);
4498 
4499  /* parse query if necessary */
4500  switch (querymode)
4501  {
4502  case QUERY_SIMPLE:
4503  my_command->argv[0] = my_command->lines.data;
4504  my_command->argc++;
4505  break;
4506  case QUERY_EXTENDED:
4507  case QUERY_PREPARED:
4508  if (!parseQuery(my_command))
4509  exit(1);
4510  break;
4511  default:
4512  exit(1);
4513  }
4514 }
4515 
4516 /*
4517  * Parse a backslash command; return a Command struct, or NULL if comment
4518  *
4519  * At call, we have scanned only the initial backslash.
4520  */
4521 static Command *
4522 process_backslash_command(PsqlScanState sstate, const char *source)
4523 {
4524  Command *my_command;
4525  PQExpBufferData word_buf;
4526  int word_offset;
4527  int offsets[MAX_ARGS]; /* offsets of argument words */
4528  int start_offset;
4529  int lineno;
4530  int j;
4531 
4532  initPQExpBuffer(&word_buf);
4533 
4534  /* Remember location of the backslash */
4535  start_offset = expr_scanner_offset(sstate) - 1;
4536  lineno = expr_scanner_get_lineno(sstate, start_offset);
4537 
4538  /* Collect first word of command */
4539  if (!expr_lex_one_word(sstate, &word_buf, &word_offset))
4540  {
4541  termPQExpBuffer(&word_buf);
4542  return NULL;
4543  }
4544 
4545  /* Allocate and initialize Command structure */
4546  my_command = (Command *) pg_malloc0(sizeof(Command));
4547  my_command->type = META_COMMAND;
4548  my_command->argc = 0;
4549  initSimpleStats(&my_command->stats);
4550 
4551  /* Save first word (command name) */
4552  j = 0;
4553  offsets[j] = word_offset;
4554  my_command->argv[j++] = pg_strdup(word_buf.data);
4555  my_command->argc++;
4556 
4557  /* ... and convert it to enum form */
4558  my_command->meta = getMetaCommand(my_command->argv[0]);
4559 
4560  if (my_command->meta == META_SET ||
4561  my_command->meta == META_IF ||
4562  my_command->meta == META_ELIF)
4563  {
4565 
4566  /* For \set, collect var name */
4567  if (my_command->meta == META_SET)
4568  {
4569  if (!expr_lex_one_word(sstate, &word_buf, &word_offset))
4570  syntax_error(source, lineno, my_command->first_line, my_command->argv[0],
4571  "missing argument", NULL, -1);
4572 
4573  offsets[j] = word_offset;
4574  my_command->argv[j++] = pg_strdup(word_buf.data);
4575  my_command->argc++;
4576  }
4577 
4578  /* then for all parse the expression */
4579  yyscanner = expr_scanner_init(sstate, source, lineno, start_offset,
4580  my_command->argv[0]);
4581 
4582  if (expr_yyparse(yyscanner) != 0)
4583  {
4584  /* dead code: exit done from syntax_error called by yyerror */
4585  exit(1);
4586  }
4587 
4588  my_command->expr = expr_parse_result;
4589 
4590  /* Save line, trimming any trailing newline */
4591  my_command->first_line =
4593  start_offset,
4594  expr_scanner_offset(sstate),
4595  true);
4596 
4597  expr_scanner_finish(yyscanner);
4598 
4599  termPQExpBuffer(&word_buf);
4600 
4601  return my_command;
4602  }
4603 
4604  /* For all other commands, collect remaining words. */
4605  while (expr_lex_one_word(sstate, &word_buf, &word_offset))
4606  {
4607  /*
4608  * my_command->argv[0] is the command itself, so the max number of
4609  * arguments is one less than MAX_ARGS
4610  */
4611  if (j >= MAX_ARGS)
4612  syntax_error(source, lineno, my_command->first_line, my_command->argv[0],
4613  "too many arguments", NULL, -1);
4614 
4615  offsets[j] = word_offset;
4616  my_command->argv[j++] = pg_strdup(word_buf.data);
4617  my_command->argc++;
4618  }
4619 
4620  /* Save line, trimming any trailing newline */
4621  my_command->first_line =
4623  start_offset,
4624  expr_scanner_offset(sstate),
4625  true);
4626 
4627  if (my_command->meta == META_SLEEP)
4628  {
4629  if (my_command->argc < 2)
4630  syntax_error(source, lineno, my_command->first_line, my_command->argv[0],
4631  "missing argument", NULL, -1);
4632 
4633  if (my_command->argc > 3)
4634  syntax_error(source, lineno, my_command->first_line, my_command->argv[0],
4635  "too many arguments", NULL,
4636  offsets[3] - start_offset);
4637 
4638  /*
4639  * Split argument into number and unit to allow "sleep 1ms" etc. We
4640  * don't have to terminate the number argument with null because it
4641  * will be parsed with atoi, which ignores trailing non-digit
4642  * characters.
4643  */
4644  if (my_command->argc == 2 && my_command->argv[1][0] != ':')
4645  {
4646  char *c = my_command->argv[1];
4647 
4648  while (isdigit((unsigned char) *c))
4649  c++;
4650  if (*c)
4651  {
4652  my_command->argv[2] = c;
4653  offsets[2] = offsets[1] + (c - my_command->argv[1]);
4654  my_command->argc = 3;
4655  }
4656  }
4657 
4658  if (my_command->argc == 3)
4659  {
4660  if (pg_strcasecmp(my_command->argv[2], "us") != 0 &&
4661  pg_strcasecmp(my_command->argv[2], "ms") != 0 &&
4662  pg_strcasecmp(my_command->argv[2], "s") != 0)
4663  syntax_error(source, lineno, my_command->first_line, my_command->argv[0],
4664  "unrecognized time unit, must be us, ms or s",
4665  my_command->argv[2], offsets[2] - start_offset);
4666  }
4667  }
4668  else if (my_command->meta == META_SETSHELL)
4669  {
4670  if (my_command->argc < 3)
4671  syntax_error(source, lineno, my_command->first_line, my_command->argv[0],
4672  "missing argument", NULL, -1);
4673  }
4674  else if (my_command->meta == META_SHELL)
4675  {
4676  if (my_command->argc < 2)
4677  syntax_error(source, lineno, my_command->first_line, my_command->argv[0],
4678  "missing command", NULL, -1);
4679  }
4680  else if (my_command->meta == META_ELSE || my_command->meta == META_ENDIF)
4681  {
4682  if (my_command->argc != 1)
4683  syntax_error(source, lineno, my_command->first_line, my_command->argv[0],
4684  "unexpected argument", NULL, -1);
4685  }
4686  else if (my_command->meta == META_GSET || my_command->meta == META_ASET)
4687  {
4688  if (my_command->argc > 2)
4689  syntax_error(source, lineno, my_command->first_line, my_command->argv[0],
4690  "too many arguments", NULL, -1);
4691  }
4692  else
4693  {
4694  /* my_command->meta == META_NONE */
4695  syntax_error(source, lineno, my_command->first_line, my_command->argv[0],
4696  "invalid command", NULL, -1);
4697  }
4698 
4699  termPQExpBuffer(&word_buf);
4700 
4701  return my_command;
4702 }
4703 
4704 static void
4705 ConditionError(const char *desc, int cmdn, const char *msg)
4706 {
4707  pg_log_fatal("condition error in script \"%s\" command %d: %s",
4708  desc, cmdn, msg);
4709  exit(1);
4710 }
4711 
4712 /*
4713  * Partial evaluation of conditionals before recording and running the script.
4714  */
4715 static void
4717 {
4718  /* statically check conditional structure */
4720  int i;
4721 
4722  for (i = 0; ps.commands[i] != NULL; i++)
4723  {
4724  Command *cmd = ps.commands[i];
4725 
4726  if (cmd->type == META_COMMAND)
4727  {
4728  switch (cmd->meta)
4729  {
4730  case META_IF:
4732  break;
4733  case META_ELIF:
4734  if (conditional_stack_empty(cs))
4735  ConditionError(ps.desc, i + 1, "\\elif without matching \\if");
4737  ConditionError(ps.desc, i + 1, "\\elif after \\else");
4738  break;
4739  case META_ELSE:
4740  if (conditional_stack_empty(cs))
4741  ConditionError(ps.desc, i + 1, "\\else without matching \\if");
4743  ConditionError(ps.desc, i + 1, "\\else after \\else");
4745  break;
4746  case META_ENDIF:
4747  if (!conditional_stack_pop(cs))
4748  ConditionError(ps.desc, i + 1, "\\endif without matching \\if");
4749  break;
4750  default:
4751  /* ignore anything else... */
4752  break;
4753  }
4754  }
4755  }
4756  if (!conditional_stack_empty(cs))
4757  ConditionError(ps.desc, i + 1, "\\if without matching \\endif");
4759 }
4760 
4761 /*
4762  * Parse a script (either the contents of a file, or a built-in script)
4763  * and add it to the list of scripts.
4764  */
4765 static void
4766 ParseScript(const char *script, const char *desc, int weight)
4767 {
4768  ParsedScript ps;
4769  PsqlScanState sstate;
4770  PQExpBufferData line_buf;
4771  int alloc_num;
4772  int index;
4773  int lineno;
4774  int start_offset;
4775 
4776 #define COMMANDS_ALLOC_NUM 128
4777  alloc_num = COMMANDS_ALLOC_NUM;
4778 
4779  /* Initialize all fields of ps */
4780  ps.desc = desc;
4781  ps.weight = weight;
4782  ps.commands = (Command **) pg_malloc(sizeof(Command *) * alloc_num);
4783  initStats(&ps.stats, 0);
4784 
4785  /* Prepare to parse script */
4786  sstate = psql_scan_create(&pgbench_callbacks);
4787 
4788  /*
4789  * Ideally, we'd scan scripts using the encoding and stdstrings settings
4790  * we get from a DB connection. However, without major rearrangement of
4791  * pgbench's argument parsing, we can't have a DB connection at the time
4792  * we parse scripts. Using SQL_ASCII (encoding 0) should work well enough
4793  * with any backend-safe encoding, though conceivably we could be fooled
4794  * if a script file uses a client-only encoding. We also assume that
4795  * stdstrings should be true, which is a bit riskier.
4796  */
4797  psql_scan_setup(sstate, script, strlen(script), 0, true);
4798  start_offset = expr_scanner_offset(sstate) - 1;
4799 
4800  initPQExpBuffer(&line_buf);
4801 
4802  index = 0;
4803 
4804  for (;;)
4805  {
4806  PsqlScanResult sr;
4807  promptStatus_t prompt;
4808  Command *command = NULL;
4809 
4810  resetPQExpBuffer(&line_buf);
4811  lineno = expr_scanner_get_lineno(sstate, start_offset);
4812 
4813  sr = psql_scan(sstate, &line_buf, &prompt);
4814 
4815  /* If we collected a new SQL command, process that */
4816  command = create_sql_command(&line_buf, desc);
4817 
4818  /* store new command */
4819  if (command)
4820  ps.commands[index++] = command;
4821 
4822  /* If we reached a backslash, process that */
4823  if (sr == PSCAN_BACKSLASH)
4824  {
4825  command = process_backslash_command(sstate, desc);
4826 
4827  if (command)
4828  {
4829  /*
4830  * If this is gset or aset, merge into the preceding command.
4831  * (We don't use a command slot in this case).
4832  */
4833  if (command->meta == META_GSET || command->meta == META_ASET)
4834  {
4835  Command *cmd;
4836 
4837  if (index == 0)
4838  syntax_error(desc, lineno, NULL, NULL,
4839  "\\gset must follow a SQL command",
4840  NULL, -1);
4841 
4842  cmd = ps.commands[index - 1];
4843 
4844  if (cmd->type != SQL_COMMAND ||
4845  cmd->varprefix != NULL)
4846  syntax_error(desc, lineno, NULL, NULL,
4847  "\\gset must follow a SQL command",
4848  cmd->first_line, -1);
4849 
4850  /* get variable prefix */
4851  if (command->argc <= 1 || command->argv[1][0] == '\0')
4852  cmd->varprefix = pg_strdup("");
4853  else
4854  cmd->varprefix = pg_strdup(command->argv[1]);
4855 
4856  /* update the sql command meta */
4857  cmd->meta = command->meta;
4858 
4859  /* cleanup unused command */
4860  free_command(command);
4861 
4862  continue;
4863  }
4864 
4865  /* Attach any other backslash command as a new command */
4866  ps.commands[index++] = command;
4867  }
4868  }
4869 
4870  /*
4871  * Since we used a command slot, allocate more if needed. Note we
4872  * always allocate one more in order to accommodate the NULL
4873  * terminator below.
4874  */
4875  if (index >= alloc_num)
4876  {
4877  alloc_num += COMMANDS_ALLOC_NUM;
4878  ps.commands = (Command **)
4879  pg_realloc(ps.commands, sizeof(Command *) * alloc_num);
4880  }
4881 
4882  /* Done if we reached EOF */
4883  if (sr == PSCAN_INCOMPLETE || sr == PSCAN_EOL)
4884  break;
4885  }
4886 
4887  ps.commands[index] = NULL;
4888 
4889  addScript(ps);
4890 
4891  termPQExpBuffer(&line_buf);
4892  psql_scan_finish(sstate);
4893  psql_scan_destroy(sstate);
4894 }
4895 
4896 /*
4897  * Read the entire contents of file fd, and return it in a malloc'd buffer.
4898  *
4899  * The buffer will typically be larger than necessary, but we don't care
4900  * in this program, because we'll free it as soon as we've parsed the script.
4901  */
4902 static char *
4904 {
4905  char *buf;
4906  size_t buflen = BUFSIZ;
4907  size_t used = 0;
4908 
4909  buf = (char *) pg_malloc(buflen);
4910 
4911  for (;;)
4912  {
4913  size_t nread;
4914 
4915  nread = fread(buf + used, 1, BUFSIZ, fd);
4916  used += nread;
4917  /* If fread() read less than requested, must be EOF or error */
4918  if (nread < BUFSIZ)
4919  break;
4920  /* Enlarge buf so we can read some more */
4921  buflen += BUFSIZ;
4922  buf = (char *) pg_realloc(buf, buflen);
4923  }
4924  /* There is surely room for a terminator */
4925  buf[used] = '\0';
4926 
4927  return buf;
4928 }
4929 
4930 /*
4931  * Given a file name, read it and add its script to the list.
4932  * "-" means to read stdin.
4933  * NB: filename must be storage that won't disappear.
4934  */
4935 static void
4936 process_file(const char *filename, int weight)
4937 {
4938  FILE *fd;
4939  char *buf;
4940 
4941  /* Slurp the file contents into "buf" */
4942  if (strcmp(filename, "-") == 0)
4943  fd = stdin;
4944  else if ((fd = fopen(filename, "r")) == NULL)
4945  {
4946  pg_log_fatal("could not open file \"%s\": %m", filename);
4947  exit(1);
4948  }
4949 
4950  buf = read_file_contents(fd);
4951 
4952  if (ferror(fd))
4953  {
4954  pg_log_fatal("could not read file \"%s\": %m", filename);
4955  exit(1);
4956  }
4957 
4958  if (fd != stdin)
4959  fclose(fd);
4960 
4961  ParseScript(buf, filename, weight);
4962 
4963  free(buf);
4964 }
4965 
4966 /* Parse the given builtin script and add it to the list. */
4967 static void
4968 process_builtin(const BuiltinScript *bi, int weight)
4969 {
4970  ParseScript(bi->script, bi->desc, weight);
4971 }
4972 
4973 /* show available builtin scripts */
4974 static void
4976 {
4977  int i;
4978 
4979  fprintf(stderr, "Available builtin scripts:\n");
4980  for (i = 0; i < lengthof(builtin_script); i++)
4981  fprintf(stderr, " %13s: %s\n", builtin_script[i].name, builtin_script[i].desc);
4982  fprintf(stderr, "\n");
4983 }
4984 
4985 /* return builtin script "name" if unambiguous, fails if not found */
4986 static const BuiltinScript *
4987 findBuiltin(const char *name)
4988 {
4989  int i,
4990  found = 0,
4991  len = strlen(name);
4992  const BuiltinScript *result = NULL;
4993 
4994  for (i = 0; i < lengthof(builtin_script); i++)
4995  {
4996  if (strncmp(builtin_script[i].name, name, len) == 0)
4997  {
4998  result = &builtin_script[i];
4999  found++;
5000  }
5001  }
5002 
5003  /* ok, unambiguous result */
5004  if (found == 1)
5005  return result;
5006 
5007  /* error cases */
5008  if (found == 0)
5009  pg_log_fatal("no builtin script found for name \"%s\"", name);
5010  else /* found > 1 */
5011  pg_log_fatal("ambiguous builtin name: %d builtin scripts found for prefix \"%s\"", found, name);
5012 
5014  exit(1);
5015 }
5016 
5017 /*
5018  * Determine the weight specification from a script option (-b, -f), if any,
5019  * and return it as an integer (1 is returned if there's no weight). The
5020  * script name is returned in *script as a malloc'd string.
5021  */
5022 static int
5023 parseScriptWeight(const char *option, char **script)
5024 {
5025  char *sep;
5026  int weight;
5027 
5028  if ((sep = strrchr(option, WSEP)))
5029  {
5030  int namelen = sep - option;
5031  long wtmp;
5032  char *badp;
5033 
5034  /* generate the script name */
5035  *script = pg_malloc(namelen + 1);
5036  strncpy(*script, option, namelen);
5037  (*script)[namelen] = '\0';
5038 
5039  /* process digits of the weight spec */
5040  errno = 0;
5041  wtmp = strtol(sep + 1, &badp, 10);
5042  if (errno != 0 || badp == sep + 1 || *badp != '\0')
5043  {
5044  pg_log_fatal("invalid weight specification: %s", sep);
5045  exit(1);
5046  }
5047  if (wtmp > INT_MAX || wtmp < 0)
5048  {
5049  pg_log_fatal("weight specification out of range (0 .. %u): " INT64_FORMAT,
5050  INT_MAX, (int64) wtmp);
5051  exit(1);
5052  }
5053  weight = wtmp;
5054  }
5055  else
5056  {
5057  *script = pg_strdup(option);
5058  weight = 1;
5059  }
5060 
5061  return weight;
5062 }
5063 
5064 /* append a script to the list of scripts to process */
5065 static void
5067 {
5068  if (script.commands == NULL || script.commands[0] == NULL)
5069  {
5070  pg_log_fatal("empty command list for script \"%s\"", script.desc);
5071  exit(1);
5072  }
5073 
5074  if (num_scripts >= MAX_SCRIPTS)
5075  {
5076  pg_log_fatal("at most %d SQL scripts are allowed", MAX_SCRIPTS);
5077  exit(1);
5078  }
5079 
5080  CheckConditional(script);
5081 
5082  sql_script[num_scripts] = script;
5083  num_scripts++;
5084 }
5085 
5086 /*
5087  * Print progress report.
5088  *
5089  * On entry, *last and *last_report contain the statistics and time of last
5090  * progress report. On exit, they are updated with the new stats.
5091  */
5092 static void
5093 printProgressReport(TState *threads, int64 test_start, int64 now,
5094  StatsData *last, int64 *last_report)
5095 {
5096  /* generate and show report */
5097  int64 run = now - *last_report,
5098  ntx;
5099  double tps,
5100  total_run,
5101  latency,
5102  sqlat,
5103  lag,
5104  stdev;
5105  char tbuf[315];
5106  StatsData cur;
5107 
5108  /*
5109  * Add up the statistics of all threads.
5110  *
5111  * XXX: No locking. There is no guarantee that we get an atomic snapshot
5112  * of the transaction count and latencies, so these figures can well be
5113  * off by a small amount. The progress report's purpose is to give a
5114  * quick overview of how the test is going, so that shouldn't matter too
5115  * much. (If a read from a 64-bit integer is not atomic, you might get a
5116  * "torn" read and completely bogus latencies though!)
5117  */
5118  initStats(&cur, 0);
5119  for (int i = 0; i < nthreads; i++)
5120  {
5121  mergeSimpleStats(&cur.latency, &threads[i].stats.latency);
5122  mergeSimpleStats(&cur.lag, &threads[i].stats.lag);
5123  cur.cnt += threads[i].stats.cnt;
5124  cur.skipped += threads[i].stats.skipped;
5125  }
5126 
5127  /* we count only actually executed transactions */
5128  ntx = (cur.cnt - cur.skipped) - (last->cnt - last->skipped);
5129  total_run = (now - test_start) / 1000000.0;
5130  tps = 1000000.0 * ntx / run;
5131  if (ntx > 0)
5132  {
5133  latency = 0.001 * (cur.latency.sum - last->latency.sum) / ntx;
5134  sqlat = 1.0 * (cur.latency.sum2 - last->latency.sum2) / ntx;
5135  stdev = 0.001 * sqrt(sqlat - 1000000.0 * latency * latency);
5136  lag = 0.001 * (cur.lag.sum - last->lag.sum) / ntx;
5137  }
5138  else
5139  {
5140  latency = sqlat = stdev = lag = 0;
5141  }
5142 
5143  if (progress_timestamp)
5144  {
5145  /*
5146  * On some platforms the current system timestamp is available in
5147  * now_time, but rather than get entangled with that, we just eat the
5148  * cost of an extra syscall in all cases.
5149  */
5150  struct timeval tv;
5151 
5152  gettimeofday(&tv, NULL);
5153  snprintf(tbuf, sizeof(tbuf), "%ld.%03ld s",
5154  (long) tv.tv_sec, (long) (tv.tv_usec / 1000));
5155  }
5156  else
5157  {
5158  /* round seconds are expected, but the thread may be late */
5159  snprintf(tbuf, sizeof(tbuf), "%.1f s", total_run);
5160  }
5161 
5162  fprintf(stderr,
5163  "progress: %s, %.1f tps, lat %.3f ms stddev %.3f",
5164  tbuf, tps, latency, stdev);
5165 
5166  if (throttle_delay)
5167  {
5168  fprintf(stderr, ", lag %.3f ms", lag);
5169  if (latency_limit)
5170  fprintf(stderr, ", " INT64_FORMAT " skipped",
5171  cur.skipped - last->skipped);
5172  }
5173  fprintf(stderr, "\n");
5174 
5175  *last = cur;
5176  *last_report = now;
5177 }
5178 
5179 static void
5180 printSimpleStats(const char *prefix, SimpleStats *ss)
5181 {
5182  if (ss->count > 0)
5183  {
5184  double latency = ss->sum / ss->count;
5185  double stddev = sqrt(ss->sum2 / ss->count - latency * latency);
5186 
5187  printf("%s average = %.3f ms\n", prefix, 0.001 * latency);
5188  printf("%s stddev = %.3f ms\n", prefix, 0.001 * stddev);
5189  }
5190 }
5191 
5192 /* print out results */
5193 static void
5194 printResults(StatsData *total, instr_time total_time,
5195  instr_time conn_total_time, int64 latency_late)
5196 {
5197  double time_include,
5198  tps_include,
5199  tps_exclude;
5200  int64 ntx = total->cnt - total->skipped;
5201 
5202  time_include = INSTR_TIME_GET_DOUBLE(total_time);
5203 
5204  /* tps is about actually executed transactions */
5205  tps_include = ntx / time_include;
5206  tps_exclude = ntx /
5207  (time_include - (INSTR_TIME_GET_DOUBLE(conn_total_time) / nclients));
5208 
5209  /* Report test parameters. */
5210  printf("transaction type: %s\n",
5211  num_scripts == 1 ? sql_script[0].desc : "multiple scripts");
5212  printf("scaling factor: %d\n", scale);
5213  /* only print partitioning information if some partitioning was detected */
5214  if (partition_method != PART_NONE)
5215  printf("partition method: %s\npartitions: %d\n",
5216  PARTITION_METHOD[partition_method], partitions);
5217  printf("query mode: %s\n", QUERYMODE[querymode]);
5218  printf("number of clients: %d\n", nclients);
5219  printf("number of threads: %d\n", nthreads);
5220  if (duration <= 0)
5221  {
5222  printf("number of transactions per client: %d\n", nxacts);
5223  printf("number of transactions actually processed: " INT64_FORMAT "/%d\n",
5224  ntx, nxacts * nclients);
5225  }
5226  else
5227  {
5228  printf("duration: %d s\n", duration);
5229  printf("number of transactions actually processed: " INT64_FORMAT "\n",
5230  ntx);
5231  }
5232 
5233  /* Remaining stats are nonsensical if we failed to execute any xacts */
5234  if (total->cnt <= 0)
5235  return;
5236 
5237  if (throttle_delay && latency_limit)
5238  printf("number of transactions skipped: " INT64_FORMAT " (%.3f %%)\n",
5239  total->skipped,
5240  100.0 * total->skipped / total->cnt);
5241 
5242  if (latency_limit)
5243  printf("number of transactions above the %.1f ms latency limit: " INT64_FORMAT "/" INT64_FORMAT " (%.3f %%)\n",
5244  latency_limit / 1000.0, latency_late, ntx,
5245  (ntx > 0) ? 100.0 * latency_late / ntx : 0.0);
5246 
5247  if (throttle_delay || progress || latency_limit)
5248  printSimpleStats("latency", &total->latency);
5249  else
5250  {
5251  /* no measurement, show average latency computed from run time */
5252  printf("latency average = %.3f ms\n",
5253  1000.0 * time_include * nclients / total->cnt);
5254  }
5255 
5256  if (throttle_delay)
5257  {
5258  /*
5259  * Report average transaction lag under rate limit throttling. This
5260  * is the delay between scheduled and actual start times for the
5261  * transaction. The measured lag may be caused by thread/client load,
5262  * the database load, or the Poisson throttling process.
5263  */
5264  printf("rate limit schedule lag: avg %.3f (max %.3f) ms\n",
5265  0.001 * total->lag.sum / total->cnt, 0.001 * total->lag.max);
5266  }
5267 
5268  printf("tps = %f (including connections establishing)\n", tps_include);
5269  printf("tps = %f (excluding connections establishing)\n", tps_exclude);
5270 
5271  /* Report per-script/command statistics */
5272  if (per_script_stats || report_per_command)
5273  {
5274  int i;
5275 
5276  for (i = 0; i < num_scripts; i++)
5277  {
5278  if (per_script_stats)
5279  {
5280  StatsData *sstats = &sql_script[i].stats;
5281 
5282  printf("SQL script %d: %s\n"
5283  " - weight: %d (targets %.1f%% of total)\n"
5284  " - " INT64_FORMAT " transactions (%.1f%% of total, tps = %f)\n",
5285  i + 1, sql_script[i].desc,
5286  sql_script[i].weight,
5287  100.0 * sql_script[i].weight / total_weight,
5288  sstats->cnt,
5289  100.0 * sstats->cnt / total->cnt,
5290  (sstats->cnt - sstats->skipped) / time_include);
5291 
5292  if (throttle_delay && latency_limit && sstats->cnt > 0)
5293  printf(" - number of transactions skipped: " INT64_FORMAT " (%.3f%%)\n",
5294  sstats->skipped,
5295  100.0 * sstats->skipped / sstats->cnt);
5296 
5297  printSimpleStats(" - latency", &sstats->latency);
5298  }
5299 
5300  /* Report per-command latencies */
5301  if (report_per_command)
5302  {
5303  Command **commands;
5304 
5305  if (per_script_stats)
5306  printf(" - statement latencies in milliseconds:\n");
5307  else
5308  printf("statement latencies in milliseconds:\n");
5309 
5310  for (commands = sql_script[i].commands;
5311  *commands != NULL;
5312  commands++)
5313  {
5314  SimpleStats *cstats = &(*commands)->stats;
5315 
5316  printf(" %11.3f %s\n",
5317  (cstats->count > 0) ?
5318  1000.0 * cstats->sum / cstats->count : 0.0,
5319  (*commands)->first_line);
5320  }
5321  }
5322  }
5323  }
5324 }
5325 
5326 /*
5327  * Set up a random seed according to seed parameter (NULL means default),
5328  * and initialize base_random_sequence for use in initializing other sequences.
5329  */
5330 static bool
5331 set_random_seed(const char *seed)
5332 {
5333  uint64 iseed;
5334 
5335  if (seed == NULL || strcmp(seed, "time") == 0)
5336  {
5337  /* rely on current time */
5338  instr_time now;
5339 
5341  iseed = (uint64) INSTR_TIME_GET_MICROSEC(now);
5342  }
5343  else if (strcmp(seed, "rand") == 0)
5344  {
5345  /* use some "strong" random source */
5346  if (!pg_strong_random(&iseed, sizeof(iseed)))
5347  {
5348  pg_log_error("could not generate random seed");
5349  return false;
5350  }
5351  }
5352  else
5353  {
5354  /* parse unsigned-int seed value */
5355  unsigned long ulseed;
5356  char garbage;
5357 
5358  /* Don't try to use UINT64_FORMAT here; it might not work for sscanf */
5359  if (sscanf(seed, "%lu%c", &ulseed, &garbage) != 1)
5360  {
5361  pg_log_error("unrecognized random seed option \"%s\"", seed);
5362  pg_log_info("Expecting an unsigned integer, \"time\" or \"rand\"");
5363  return false;
5364  }
5365  iseed = (uint64) ulseed;
5366  }
5367 
5368  if (seed != NULL)
5369  pg_log_info("setting random seed to " UINT64_FORMAT, iseed);
5370  random_seed = iseed;
5371 
5372  /* Fill base_random_sequence with low-order bits of seed */
5373  base_random_sequence.xseed[0] = iseed & 0xFFFF;
5374  base_random_sequence.xseed[1] = (iseed >> 16) & 0xFFFF;
5375  base_random_sequence.xseed[2] = (iseed >> 32) & 0xFFFF;
5376 
5377  return true;
5378 }
5379 
5380 int
5381 main(int argc, char **argv)
5382 {
5383  static struct option long_options[] = {
5384  /* systematic long/short named options */
5385  {"builtin", required_argument, NULL, 'b'},
5386  {"client", required_argument, NULL, 'c'},
5387  {"connect", no_argument, NULL, 'C'},
5388  {"debug", no_argument, NULL, 'd'},
5389  {"define", required_argument, NULL, 'D'},
5390  {"file", required_argument, NULL, 'f'},
5391  {"fillfactor", required_argument, NULL, 'F'},
5392  {"host", required_argument, NULL, 'h'},
5393  {"initialize", no_argument, NULL, 'i'},
5394  {"init-steps", required_argument, NULL, 'I'},
5395  {"jobs", required_argument, NULL, 'j'},
5396  {"log", no_argument, NULL, 'l'},
5397  {"latency-limit", required_argument, NULL, 'L'},
5398  {"no-vacuum", no_argument, NULL, 'n'},
5399  {"port", required_argument, NULL, 'p'},
5400  {"progress", required_argument, NULL, 'P'},
5401  {"protocol", required_argument, NULL, 'M'},
5402  {"quiet", no_argument, NULL, 'q'},
5403  {"report-latencies", no_argument, NULL, 'r'},
5404  {"rate", required_argument, NULL, 'R'},
5405  {"scale", required_argument, NULL, 's'},
5406  {"select-only", no_argument, NULL, 'S'},
5407  {"skip-some-updates", no_argument, NULL, 'N'},
5408  {"time", required_argument, NULL, 'T'},
5409  {"transactions", required_argument, NULL, 't'},
5410  {"username", required_argument, NULL, 'U'},
5411  {"vacuum-all", no_argument, NULL, 'v'},
5412  /* long-named only options */
5413  {"unlogged-tables", no_argument, NULL, 1},
5414  {"tablespace", required_argument, NULL, 2},
5415  {"index-tablespace", required_argument, NULL, 3},
5416  {"sampling-rate", required_argument, NULL, 4},
5417  {"aggregate-interval", required_argument, NULL, 5},
5418  {"progress-timestamp", no_argument, NULL, 6},
5419  {"log-prefix", required_argument, NULL, 7},
5420  {"foreign-keys", no_argument, NULL, 8},
5421  {"random-seed", required_argument, NULL, 9},
5422  {"show-script", required_argument, NULL, 10},
5423  {"partitions", required_argument, NULL, 11},
5424  {"partition-method", required_argument, NULL, 12},
5425  {NULL, 0, NULL, 0}
5426  };
5427 
5428  int c;
5429  bool is_init_mode = false; /* initialize mode? */
5430  char *initialize_steps = NULL;
5431  bool foreign_keys = false;
5432  bool is_no_vacuum = false;
5433  bool do_vacuum_accounts = false; /* vacuum accounts table? */
5434  int optindex;
5435  bool scale_given = false;
5436 
5437  bool benchmarking_option_set = false;
5438  bool initialization_option_set = false;
5439  bool internal_script_used = false;
5440 
5441  CState *state; /* status of clients */
5442  TState *threads; /* array of thread */
5443 
5444  instr_time start_time; /* start up time */
5445  instr_time total_time;
5446  instr_time conn_total_time;
5447  int64 latency_late = 0;
5448  StatsData stats;
5449  int weight;
5450 
5451  int i;
5452  int nclients_dealt;
5453 
5454 #ifdef HAVE_GETRLIMIT
5455  struct rlimit rlim;
5456 #endif
5457 
5458  PGconn *con;
5459  char *env;
5460 
5461  int exit_code = 0;
5462 
5463  pg_logging_init(argv[0]);
5464  progname = get_progname(argv[0]);
5465 
5466  if (argc > 1)
5467  {
5468  if (strcmp(argv[1], "--help") == 0 || strcmp(argv[1], "-?") == 0)
5469  {
5470  usage();
5471  exit(0);
5472  }
5473  if (strcmp(argv[1], "--version") == 0 || strcmp(argv[1], "-V") == 0)
5474  {
5475  puts("pgbench (PostgreSQL) " PG_VERSION);
5476  exit(0);
5477  }
5478  }
5479 
5480  if ((env = getenv("PGHOST")) != NULL && *env != '\0')
5481  pghost = env;
5482  if ((env = getenv("PGPORT")) != NULL && *env != '\0')
5483  pgport = env;
5484  else if ((env = getenv("PGUSER")) != NULL && *env != '\0')
5485  login = env;
5486 
5487  state = (CState *) pg_malloc0(sizeof(CState));
5488 
5489  /* set random seed early, because it may be used while parsing scripts. */
5490  if (!set_random_seed(getenv("PGBENCH_RANDOM_SEED")))
5491  {
5492  pg_log_fatal("error while setting random seed from PGBENCH_RANDOM_SEED environment variable");
5493  exit(1);
5494  }
5495 
5496  while ((c = getopt_long(argc, argv, "iI:h:nvp:dqb:SNc:j:Crs:t:T:U:lf:D:F:M:P:R:L:", long_options, &optindex)) != -1)
5497  {
5498  char *script;
5499 
5500  switch (c)
5501  {
5502  case 'i':
5503  is_init_mode = true;
5504  break;
5505  case 'I':
5506  if (initialize_steps)
5507  pg_free(initialize_steps);
5508  initialize_steps = pg_strdup(optarg);
5509  checkInitSteps(initialize_steps);
5510  initialization_option_set = true;
5511  break;
5512  case 'h':
5513  pghost = pg_strdup(optarg);
5514  break;
5515  case 'n':
5516  is_no_vacuum = true;
5517  break;
5518  case 'v':
5519  benchmarking_option_set = true;
5520  do_vacuum_accounts = true;
5521  break;
5522  case 'p':
5523  pgport = pg_strdup(optarg);
5524  break;
5525  case 'd':
5527  break;
5528  case 'c':
5529  benchmarking_option_set = true;
5530  nclients = atoi(optarg);
5531  if (nclients <= 0)
5532  {
5533  pg_log_fatal("invalid number of clients: \"%s\"", optarg);
5534  exit(1);
5535  }
5536 #ifdef HAVE_GETRLIMIT
5537 #ifdef RLIMIT_NOFILE /* most platforms use RLIMIT_NOFILE */
5538  if (getrlimit(RLIMIT_NOFILE, &rlim) == -1)
5539 #else /* but BSD doesn't ... */
5540  if (getrlimit(RLIMIT_OFILE, &rlim) == -1)
5541 #endif /* RLIMIT_NOFILE */
5542  {
5543  pg_log_fatal("getrlimit failed: %m");
5544  exit(1);
5545  }
5546  if (rlim.rlim_cur < nclients + 3)
5547  {
5548  pg_log_fatal("need at least %d open files, but system limit is %ld",
5549  nclients + 3, (long) rlim.rlim_cur);
5550  pg_log_info("Reduce number of clients, or use limit/ulimit to increase the system limit.");
5551  exit(1);
5552  }
5553 #endif /* HAVE_GETRLIMIT */
5554  break;
5555  case 'j': /* jobs */
5556  benchmarking_option_set = true;
5557  nthreads = atoi(optarg);
5558  if (nthreads <= 0)
5559  {
5560  pg_log_fatal("invalid number of threads: \"%s\"", optarg);
5561  exit(1);
5562  }
5563 #ifndef ENABLE_THREAD_SAFETY
5564  if (nthreads != 1)
5565  {
5566  pg_log_fatal("threads are not supported on this platform; use -j1");
5567  exit(1);
5568  }
5569 #endif /* !ENABLE_THREAD_SAFETY */
5570  break;
5571  case 'C':
5572  benchmarking_option_set = true;
5573  is_connect = true;
5574  break;
5575  case 'r':
5576  benchmarking_option_set = true;
5577  report_per_command = true;
5578  break;
5579  case 's':
5580  scale_given = true;
5581  scale = atoi(optarg);
5582  if (scale <= 0)
5583  {
5584  pg_log_fatal("invalid scaling factor: \"%s\"", optarg);
5585  exit(1);
5586  }
5587  break;
5588  case 't':
5589  benchmarking_option_set = true;
5590  nxacts = atoi(optarg);
5591  if (nxacts <= 0)
5592  {
5593  pg_log_fatal("invalid number of transactions: \"%s\"", optarg);
5594  exit(1);
5595  }
5596  break;
5597  case 'T':
5598  benchmarking_option_set = true;
5599  duration = atoi(optarg);
5600  if (duration <= 0)
5601  {
5602  pg_log_fatal("invalid duration: \"%s\"", optarg);
5603  exit(1);
5604  }
5605  break;
5606  case 'U':
5607  login = pg_strdup(optarg);
5608  break;
5609  case 'l':
5610  benchmarking_option_set = true;
5611  use_log = true;
5612  break;
5613  case 'q':
5614  initialization_option_set = true;
5615  use_quiet = true;
5616  break;
5617  case 'b':
5618  if (strcmp(optarg, "list") == 0)
5619  {
5621  exit(0);
5622  }
5623  weight = parseScriptWeight(optarg, &script);
5624  process_builtin(findBuiltin(script), weight);
5625  benchmarking_option_set = true;
5626  internal_script_used = true;
5627  break;
5628  case 'S':
5629  process_builtin(findBuiltin("select-only"), 1);
5630  benchmarking_option_set = true;
5631  internal_script_used = true;
5632  break;
5633  case 'N':
5634  process_builtin(findBuiltin("simple-update"), 1);
5635  benchmarking_option_set = true;
5636  internal_script_used = true;
5637  break;
5638  case 'f':
5639  weight = parseScriptWeight(optarg, &script);
5640  process_file(script, weight);
5641  benchmarking_option_set = true;
5642  break;
5643  case 'D':
5644  {
5645  char *p;
5646 
5647  benchmarking_option_set = true;
5648 
5649  if ((p = strchr(optarg, '=')) == NULL || p == optarg || *(p + 1) == '\0')
5650  {
5651  pg_log_fatal("invalid variable definition: \"%s\"", optarg);
5652  exit(1);
5653  }
5654 
5655  *p++ = '\0';
5656  if (!putVariable(&state[0], "option", optarg, p))
5657  exit(1);
5658  }
5659  break;
5660  case 'F':
5661  initialization_option_set = true;
5662  fillfactor = atoi(optarg);
5663  if (fillfactor < 10 || fillfactor > 100)
5664  {
5665  pg_log_fatal("invalid fillfactor: \"%s\"", optarg);
5666  exit(1);
5667  }
5668  break;
5669  case 'M':
5670  benchmarking_option_set = true;
5671  for (querymode = 0; querymode < NUM_QUERYMODE; querymode++)
5672  if (strcmp(optarg, QUERYMODE[querymode]) == 0)
5673  break;
5674  if (querymode >= NUM_QUERYMODE)
5675  {
5676  pg_log_fatal("invalid query mode (-M): \"%s\"", optarg);
5677  exit(1);
5678  }
5679  break;
5680  case 'P':
5681  benchmarking_option_set = true;
5682  progress = atoi(optarg);
5683  if (progress <= 0)
5684  {
5685  pg_log_fatal("invalid thread progress delay: \"%s\"", optarg);
5686  exit(1);
5687  }
5688  break;
5689  case 'R':
5690  {
5691  /* get a double from the beginning of option value */
5692  double throttle_value = atof(optarg);
5693 
5694  benchmarking_option_set = true;
5695 
5696  if (throttle_value <= 0.0)
5697  {
5698  pg_log_fatal("invalid rate limit: \"%s\"", optarg);
5699  exit(1);
5700  }
5701  /* Invert rate limit into per-transaction delay in usec */
5702  throttle_delay = 1000000.0 / throttle_value;
5703  }
5704  break;
5705  case 'L':
5706  {
5707  double limit_ms = atof(optarg);
5708 
5709  if (limit_ms <= 0.0)
5710  {
5711  pg_log_fatal("invalid latency limit: \"%s\"", optarg);
5712  exit(1);
5713  }
5714  benchmarking_option_set = true;
5715  latency_limit = (int64) (limit_ms * 1000);
5716  }
5717  break;
5718  case 1: /* unlogged-tables */
5719  initialization_option_set = true;
5720  unlogged_tables = true;
5721  break;
5722  case 2: /* tablespace */
5723  initialization_option_set = true;
5724  tablespace = pg_strdup(optarg);
5725  break;
5726  case 3: /* index-tablespace */
5727  initialization_option_set = true;
5728  index_tablespace = pg_strdup(optarg);
5729  break;
5730  case 4: /* sampling-rate */
5731  benchmarking_option_set = true;
5732  sample_rate = atof(optarg);
5733  if (sample_rate <= 0.0 || sample_rate > 1.0)
5734  {
5735  pg_log_fatal("invalid sampling rate: \"%s\"", optarg);
5736  exit(1);
5737  }
5738  break;
5739  case 5: /* aggregate-interval */
5740  benchmarking_option_set = true;
5741  agg_interval = atoi(optarg);
5742  if (agg_interval <= 0)
5743  {
5744  pg_log_fatal("invalid number of seconds for aggregation: \"%s\"", optarg);
5745  exit(1);
5746  }
5747  break;
5748  case 6: /* progress-timestamp */
5749  progress_timestamp = true;
5750  benchmarking_option_set = true;
5751  break;
5752  case 7: /* log-prefix */
5753  benchmarking_option_set = true;
5754  logfile_prefix = pg_strdup(optarg);
5755  break;
5756  case 8: /* foreign-keys */
5757  initialization_option_set = true;
5758  foreign_keys = true;
5759  break;
5760  case 9: /* random-seed */
5761  benchmarking_option_set = true;
5762  if (!set_random_seed(optarg))
5763  {
5764  pg_log_fatal("error while setting random seed from --random-seed option");
5765  exit(1);
5766  }
5767  break;
5768  case 10: /* list */
5769  {
5770  const BuiltinScript *s = findBuiltin(optarg);
5771 
5772  fprintf(stderr, "-- %s: %s\n%s\n", s->name, s->desc, s->script);
5773  exit(0);
5774  }
5775  break;
5776  case 11: /* partitions */
5777  initialization_option_set = true;
5778  partitions = atoi(optarg);
5779  if (partitions < 0)
5780  {
5781  pg_log_fatal("invalid number of partitions: \"%s\"", optarg);
5782  exit(1);
5783  }
5784  break;
5785  case 12: /* partition-method */
5786  initialization_option_set = true;
5787  if (pg_strcasecmp(optarg, "range") == 0)
5788  partition_method = PART_RANGE;
5789  else if (pg_strcasecmp(optarg, "hash") == 0)
5790  partition_method = PART_HASH;
5791  else
5792  {
5793  pg_log_fatal("invalid partition method, expecting \"range\" or \"hash\", got: \"%s\"",
5794  optarg);
5795  exit(1);
5796  }
5797  break;
5798  default:
5799  fprintf(stderr, _("Try \"%s --help\" for more information.\n"), progname);
5800  exit(1);
5801  break;
5802  }
5803  }
5804 
5805  /* set default script if none */
5806  if (num_scripts == 0 && !is_init_mode)
5807  {
5808  process_builtin(findBuiltin("tpcb-like"), 1);
5809  benchmarking_option_set = true;
5810  internal_script_used = true;
5811  }
5812 
5813  /* complete SQL command initialization and compute total weight */
5814  for (i = 0; i < num_scripts; i++)
5815  {
5816  Command **commands = sql_script[i].commands;
5817 
5818  for (int j = 0; commands[j] != NULL; j++)
5819  if (commands[j]->type == SQL_COMMAND)
5820  postprocess_sql_command(commands[j]);
5821 
5822  /* cannot overflow: weight is 32b, total_weight 64b */
5823  total_weight += sql_script[i].weight;
5824  }
5825 
5826  if (total_weight == 0 && !is_init_mode)
5827  {
5828  pg_log_fatal("total script weight must not be zero");
5829  exit(1);
5830  }
5831 
5832  /* show per script stats if several scripts are used */
5833  if (num_scripts > 1)
5834  per_script_stats = true;
5835 
5836  /*
5837  * Don't need more threads than there are clients. (This is not merely an
5838  * optimization; throttle_delay is calculated incorrectly below if some
5839  * threads have no clients assigned to them.)
5840  */
5841  if (nthreads > nclients)
5842  nthreads = nclients;
5843 
5844  /*
5845  * Convert throttle_delay to a per-thread delay time. Note that this
5846  * might be a fractional number of usec, but that's OK, since it's just
5847  * the center of a Poisson distribution of delays.
5848  */
5849  throttle_delay *= nthreads;
5850 
5851  if (argc > optind)
5852  dbName = argv[optind++];
5853  else
5854  {
5855  if ((env = getenv("PGDATABASE")) != NULL && *env != '\0')
5856  dbName = env;
5857