Commit 7f898e8b authored by John Esmet's avatar John Esmet Committed by Yoni Fogel

fixes #5183 fix stress openclose to use the standard query/update operations...

fixes #5183 fix stress openclose to use the standard query/update operations so that recovery can verify the correctness of the dbs upon recovery. this means we can have a recovery version of the test, which simply sets crash_at_end = true before calling stress table. put all of the common stuff in a header (dirty but better than duplicating the code)


git-svn-id: file:///svn/toku/tokudb@45224 c7de825b-a66e-492c-adef-691d508d4ae1
parent 9c57b1b5
...@@ -117,6 +117,7 @@ set(recover_stress_tests ...@@ -117,6 +117,7 @@ set(recover_stress_tests
ydb/recover-test_stress1.abortrecover ydb/recover-test_stress1.abortrecover
ydb/recover-test_stress2.abortrecover ydb/recover-test_stress2.abortrecover
ydb/recover-test_stress3.abortrecover ydb/recover-test_stress3.abortrecover
ydb/recover-test_stress_openclose.abortrecover
) )
## we run stress tests separately, only run them if asked to ## we run stress tests separately, only run them if asked to
......
/* -*- mode: C; c-basic-offset: 4; indent-tabs-mode: nil -*- */
// vim: expandtab:ts=8:sw=4:softtabstop=4:
#ident "Copyright (c) 2007 Tokutek Inc. All rights reserved."
#ident "$Id: recover-test_stress_openclose.c 44373 2012-06-08 20:33:40Z esmet $"
#include "stress_openclose.h"
int
test_main(int argc, char *const argv[]) {
struct cli_args args = get_default_args();
parse_stress_test_args(argc, argv, &args);
// checkpointing is a part of the ref count, so do it often
args.env_args.checkpointing_period = 5;
// very small dbs, so verification scans are short and sweet
args.num_elements = 1000;
// it's okay for update to get DB_LOCK_NOTGRANTED, etc.
args.crash_on_operation_failure = false;
// set crash at end to true for the recovery version
// then run the test or run recovery, depending on args
stress_openclose_crash_at_end = true;
if (args.do_test_and_crash) {
stress_test_main(&args);
}
if (args.do_recover) {
stress_recover(&args);
}
return 0;
}
/* -*- mode: C; c-basic-offset: 4; indent-tabs-mode: nil -*- */
// vim: expandtab:ts=8:sw=4:softtabstop=4:
#ident "Copyright (c) 2007 Tokutek Inc. All rights reserved."
#ident "$Id: ft-refcount-stress.c 44373 2012-06-08 20:33:40Z esmet $"
#include <toku_pthread.h>
#include "test.h"
#include "threaded_stress_test_helpers.h"
// set this to true for the recovery version of this stress test
// the way this works is to include this header and set
// crash_at_end = true;
static bool stress_openclose_crash_at_end;
//
// Stress test ft reference counting
//
// Three things keep a fractal tree in memory by holding a reference:
// - open ft handle
// - live txn that did a write op
// - checkpoint
//
// To stress reference counting, we would like threads which:
// - take checkpoints at random intervals
// - update random values, random point queries for auditing
// * sometimes close handle before commit.
// - close random dictionaries
//
// Here's how we can do it:
//
// A bunch of threads randomly choose from N buckets. Each bucket
// has a DB, an is_open bit, and a lock.
// - in a single txn, each thread will do some small number of
// queries or updates on randomb uckets, opening the dbs if
// they were closed and possibly closing afterwards.
// - this should stress both open db handles and various txns
// references dbs simultaneously.
//
// and all while this is happening, throw in scanners, updaters,
// and query threads that all assert the contents of these dbs
// is correct, even after recovery.
#define verbose_printf(...) \
do { \
if (verbose) { \
printf(__VA_ARGS__); \
fflush(stdout); \
} \
} while (0)
// a bunch of buckets with dbs, a lock, and an is_open bit
// threads will choose buckets randomly for update, query,
// and then maybe open/close the bucket's db.
struct db_bucket {
DB_ENV *env;
DB *db;
bool is_open;
toku_mutex_t mutex;
};
static struct db_bucket *buckets;
static int num_buckets;
// each operation can do at most this many operations in one txn
static int
choose_random_iteration_count(ARG arg) {
const int max_iteration_count = 8;
int k = myrandom_r(arg->random_data) % max_iteration_count;
return k + 1;
}
// open the ith db in the array, asserting success
static void
open_ith_db(DB_ENV *env, DB **db, int i) {
char name[30];
memset(name, 0, sizeof(name));
get_ith_table_name(name, sizeof(name), i);
int r = db_create(db, env, 0);
CKERR(r);
r = (*db)->open(*db, null_txn, name, NULL, DB_BTREE, 0, 0666);
CKERR(r);
}
// debugging counter to maintain the invariant that open_buckets <= num_buckets
static int open_buckets;
// choose and lock a random bucket, possibly opening a db
static struct db_bucket *
lock_and_maybe_open_some_db(ARG arg) {
int k = myrandom_r(arg->random_data) % num_buckets;
struct db_bucket *bucket = &buckets[k];
toku_mutex_lock(&bucket->mutex);
if (!bucket->is_open) {
// choose a random DB from 0..k-1 to associate with this bucket
// then, mark the bucket as open.
int i = myrandom_r(arg->random_data) % num_buckets;
open_ith_db(bucket->env, &bucket->db, i);
bucket->is_open = true;
assert(__sync_fetch_and_add(&open_buckets, 1) < num_buckets);
verbose_printf("opened db %d in bucket %d\n", i, k);
}
return bucket;
}
// release the lock on a bucket, possibly closing its db
static void
unlock_and_maybe_close_db(struct db_bucket *bucket, ARG arg) {
static const int p = 5;
int k = ((unsigned) myrandom_r(arg->random_data)) % 100;
// we should close with probability approximately p / 100
assert(bucket->is_open);
if (false && k <= p) {
DB *db = bucket->db;
int r = db->close(db, 0);
CKERR(r);
bucket->is_open = false;
assert(__sync_fetch_and_add(&open_buckets, -1) > 0);
verbose_printf("decided to close a bucket's db before unlocking\n");
}
toku_mutex_unlock(&bucket->mutex);
}
// scan some dbs, verifying the correct sum.
static int
scan_some_dbs(DB_TXN *txn, ARG arg, void* operation_extra, void *UU(stats_extra)) {
int r = 0;
verbose_printf("scanning some dbs\n");
struct scan_op_extra* extra = operation_extra;
// scan every db, one by one, and verify that the contents are correct
for (int i = 0; r == 0 && run_test && i < arg->cli->num_DBs; i++) {
struct db_bucket *bucket = lock_and_maybe_open_some_db(arg);
const bool check_sum = true;
r = scan_op_and_maybe_check_sum(bucket->db, txn, extra, check_sum);
invariant(r == 0 || r == DB_LOCK_NOTGRANTED);
unlock_and_maybe_close_db(bucket, arg);
}
return r;
}
// update a couple of dbs in some buckets with a txn
static int
update_some_dbs(DB_TXN *txn, ARG arg, void *op_extra, void *stats_extra) {
int r = 0;
verbose_printf("updating some dbs\n");
const int iterations = choose_random_iteration_count(arg);
for (int i = 0; r == 0 && run_test && i < iterations; i++) {
struct db_bucket *bucket = lock_and_maybe_open_some_db(arg);
// does an update operation on this bucket's db
r = update_op_db(bucket->db, txn, arg, op_extra, stats_extra);
invariant(r == 0 || r == DB_LOCK_NOTGRANTED);
unlock_and_maybe_close_db(bucket, arg);
}
return r;
}
// point query a couple of dbs in some buckets with a txn
static int
ptquery_some_dbs(DB_TXN *txn, ARG arg, void *UU(op_extra), void *UU(stats_extra)) {
int r = 0;
verbose_printf("querying some dbs\n");
const int iterations = choose_random_iteration_count(arg);
for (int i = 0; r == 0 && run_test && i < iterations; i++) {
struct db_bucket *bucket = lock_and_maybe_open_some_db(arg);
// does a point query on a random key for this bucket's db
const bool check_sum = true;
r = ptquery_and_maybe_check_op(bucket->db, txn, arg, check_sum);
invariant(r == 0 || r == DB_LOCK_NOTGRANTED);
unlock_and_maybe_close_db(bucket, arg);
}
return r;
}
static void
stress_table(DB_ENV *env, DB **dbp, struct cli_args *cli_args) {
const int update_threads = cli_args->num_update_threads;
const int query_threads = cli_args->num_ptquery_threads;
const int total_threads = update_threads + query_threads + 1;
struct arg myargs[total_threads];
for (int i = 0; i < total_threads; i++) {
arg_init(&myargs[i], dbp, env, cli_args);
}
struct scan_op_extra soe[4];
// make the forward fast scanner
soe[0].fast = TRUE;
soe[0].fwd = TRUE;
soe[0].prefetch = FALSE;
myargs[0].operation_extra = &soe[0];
myargs[0].operation = scan_some_dbs;
struct update_op_args uoe = get_update_op_args(cli_args, NULL);
// make the guy that updates the db
for (int i = 1; i < 1 + update_threads; ++i) {
myargs[i].operation_extra = &uoe;
myargs[i].operation = update_some_dbs;
myargs[i].do_prepare = true;
}
// make the guy that does point queries
for (int i = 1 + update_threads; i < total_threads; i++) {
myargs[i].operation = ptquery_some_dbs;
myargs[i].do_prepare = true;
}
num_buckets = cli_args->num_DBs;
open_buckets = num_buckets;
// each thread gets access to this array of db buckets, from
// which they can choose a random db to either touch or query
buckets = toku_malloc(sizeof(struct db_bucket) * num_buckets);
for (int i = 0; i < num_buckets; i++) {
struct db_bucket bucket = {
.is_open = true,
.db = dbp[i],
.env = env,
};
buckets[i] = bucket;
toku_mutex_init(&buckets[i].mutex, NULL);
}
// run all of the query and update workers. they may randomly open
// and close the dbs in each db_bucket to be some random dictionary,
// so when they're done we'll have to clean up the mess so this
// stress test can exit gracefully expecting db[i] = the ith db
//verbose_printf("stressing %d tables using %d update threads, %d query threads\n",
// num_buckets, update_threads, query_threads);
verbose_printf("stressing %d tables using %d update threads\n",
num_buckets, update_threads);
// stress_openclose_crash_at_end should be changed to true or false,
// depending if this test is for recovery or not.
const bool crash_at_end = stress_openclose_crash_at_end;
run_workers(myargs, total_threads, cli_args->time_of_test, crash_at_end, cli_args);
// the stress test is now complete. get ready for shutdown/close.
//
// make sure that every db in the original array is opened
// as it was when it was passed in.
for (int i = 0; i < num_buckets; i++) {
// close whatever is open
if (buckets[i].is_open) {
DB *db = buckets[i].db;
int r = db->close(db, 0);
CKERR(r);
}
// put the correct db back, then save the pointer
// into the dbp array we were given
open_ith_db(env, &buckets[i].db, i);
dbp[i] = buckets[i].db;
}
toku_free(buckets);
}
/* -*- mode: C; c-basic-offset: 4; indent-tabs-mode: nil -*- */ /* -*- mode: C; c-basic-offset: 4; indent-tabs-mode: nil -*- */
// vim: expandtab:ts=8:sw=4:softtabstop=4: // vim: expandtab:ts=8:sw=4:softtabstop=4:
#ident "Copyright (c) 2007 Tokutek Inc. All rights reserved." #ident "Copyright (c) 2007 Tokutek Inc. All rights reserved."
#ident "$Id: ft-refcount-stress.c 44373 2012-06-08 20:33:40Z esmet $" #ident "$Id: test_stress_openclose.c 44373 2012-06-08 20:33:40Z esmet $"
#include <toku_pthread.h> #include "stress_openclose.h"
#include "test.h"
#include "threaded_stress_test_helpers.h"
//
// Stress test ft reference counting
//
// Three things keep a fractal tree in memory by holding a reference:
// - open ft handle
// - live txn that did a write op
// - checkpoint
//
// To stress reference counting, we would like threads which:
// - take checkpoints at random intervals
// - update random values, random point queries for auditing
// * sometimes close handle before commit.
// - close random dictionaries
//
// Here's how we can do it:
//
// A bunch of threads randomly choose from N buckets. Each bucket
// has a DB, an is_open bit, and a lock.
// - in a single txn, each thread will do some small number of
// queries or updates on randomb uckets, opening the dbs if
// they were closed and possibly closing afterwards.
// - this should stress both open db handles and v arious txns
// references dbs simultaneously.
//
// update_thread {
// for i in random() % small_int
// db = lock_and_maybe_open_random_db()
// db->update(db, touch);
// maybe_close(db);
// unlock(db);
// }
// query_thread {
// for i in random() % small_int
// db = lock_and_maybe_open_random_db()
// db->get(db, val);
// maybe_close(db);
// unlock(db);
// }
// a bunch of buckets with dbs, a lock, and an is_open bit
// threads will choose buckets randomly for update, query,
// and then maybe open/close the bucket's db.
struct db_bucket {
DB_ENV *env;
DB *db;
bool is_open;
toku_mutex_t mutex;
};
static struct db_bucket *buckets;
static int num_buckets;
static void
init_dbt(DBT *dbt, void *data, int len) {
memset(dbt, 0, sizeof(DBT));
dbt->data = data;
dbt->size = len;
dbt->ulen = len;
dbt->flags = DB_DBT_USERMEM;
}
// each operation can do at most this many operations in one txn
static int
choose_random_iteration_count(ARG arg) {
const int max_iteration_count = 8;
int k = myrandom_r(arg->random_data) % max_iteration_count;
return k + 1;
}
// open the ith db in the array, asserting success
static void
open_ith_db(DB_ENV *env, DB **db, int i) {
char name[30];
memset(name, 0, sizeof(name));
get_ith_table_name(name, sizeof(name), i);
int r = db_create(db, env, 0);
CKERR(r);
r = (*db)->open(*db, null_txn, name, NULL, DB_BTREE, 0, 0666);
CKERR(r);
}
static int open_buckets;
// choose and lock a random bucket, possibly opening a db
static struct db_bucket *
lock_and_maybe_open_some_db(ARG arg) {
int k = myrandom_r(arg->random_data) % num_buckets;
struct db_bucket *bucket = &buckets[k];
toku_mutex_lock(&bucket->mutex);
if (!bucket->is_open) {
// choose a random DB from 0..k-1 to associate with this bucket
// then, mark the bucket as open.
int i = myrandom_r(arg->random_data) % num_buckets;
open_ith_db(bucket->env, &bucket->db, i);
bucket->is_open = true;
printf("opened db %d in bucket %d\n", i, k);
}
assert(bucket->is_open);
assert(__sync_fetch_and_add(&open_buckets, 1) < num_buckets);
return bucket;
}
// release the lock on a bucket, possibly closing its db
static void
unlock_and_maybe_close_db(struct db_bucket *bucket, ARG arg) {
static const int p = 5;
int k = ((unsigned) myrandom_r(arg->random_data)) % 100;
// we should close with probability approximately p / 100
assert(bucket->is_open);
if (k <= p) {
DB *db = bucket->db;
int r = db->close(db, 0);
CKERR(r);
bucket->is_open = false;
printf("decided to close a bucket's db before unlocking\n");
}
assert(__sync_fetch_and_add(&open_buckets, -1) > 0);
toku_mutex_unlock(&bucket->mutex);
}
// put a value of 1 for the key or increment what's there
static int
increment(DB *db, const DBT *key, const DBT *old_val, const DBT *extra,
void (*set_val)(const DBT *new_val, void *set_extra), void *set_extra) {
assert(db);
assert(key);
assert(extra && extra->size == 0);
DBT new_val;
int k = 1;
if (old_val) {
assert(old_val->size == sizeof(int));
k = *(int *) old_val->data;
k++;
}
init_dbt(&new_val, &k, sizeof(int));
set_val(&new_val, set_extra);
return 0;
}
static const int max_key = 10000;
// read the one and only row for a db with the bucket.
// should be holding the bucket's lock.
static int
read_row(DB_TXN *txn, struct db_bucket *bucket, ARG arg) {
int k, v;
DBT key, value;
k = myrandom_r(arg->random_data) % max_key;
init_dbt(&key, &k, sizeof(int));
init_dbt(&value, &v, sizeof(int));
DB *db = bucket->db;
int r = db->get(db, txn, &key, &value, 0);
invariant(r == 0 || r == DB_NOTFOUND || r == DB_LOCK_NOTGRANTED);
return k;
}
// update a random row, should be holding the bucket's lock
static void
update_row(DB_TXN *txn, struct db_bucket *bucket, ARG arg) {
DBT key, extra;
int k = myrandom_r(arg->random_data) % max_key;
init_dbt(&key, &k, sizeof(int));
init_dbt(&extra, NULL, 0);
DB *db = bucket->db;
int r = db->update(db, txn, &key, &extra, DB_TXN_SNAPSHOT);
invariant(r == 0 || r == DB_LOCK_NOTGRANTED);
}
// touch a couple of dbs in some buckets with a txn
static int
touch_some_dbs(DB_TXN *txn, ARG arg, void *UU(op_extra), void *UU(stats_extra)) {
const int iterations = choose_random_iteration_count(arg);
for (int i = 0; i < iterations; i++) {
struct db_bucket *bucket = lock_and_maybe_open_some_db(arg);
update_row(txn, bucket, arg);
unlock_and_maybe_close_db(bucket, arg);
}
return 0;
}
// query a couple of dbs, asserting that the value read is equal
// to the value in the dbs bucket, while holding the bucket lock
static int
query_some_dbs(DB_TXN *txn, ARG arg, void *UU(op_extra), void *UU(stats_extra)) {
printf("querying some dbs\n");
const int iterations = choose_random_iteration_count(arg);
for (int i = 0; i < iterations; i++) {
struct db_bucket *bucket = lock_and_maybe_open_some_db(arg);
int k = read_row(txn, bucket, arg);
assert(k >= 0);
unlock_and_maybe_close_db(bucket, arg);
}
return 0;
}
static void
stress_table(DB_ENV *env, DB **dbp, struct cli_args *cli_args) {
const int update_threads = cli_args->num_update_threads;
const int query_threads = cli_args->num_ptquery_threads;
const int total_threads = update_threads + query_threads;
num_buckets = cli_args->num_DBs;
printf("stressing %d tables using %d update threads, %d query threads\n",
num_buckets, update_threads, query_threads);
// each thread gets access to this array of db buckets, from
// which they can choose a random db to either touch or query
buckets = toku_malloc(sizeof(struct db_bucket) * num_buckets);
for (int i = 0; i < num_buckets; i++) {
struct db_bucket bucket = {
.is_open = true,
.db = dbp[i],
.env = env,
};
buckets[i] = bucket;
toku_mutex_init(&buckets[i].mutex, NULL);
}
struct arg myargs[total_threads];
for (int i = 0; i < total_threads; i++) {
arg_init(&myargs[i], dbp, env, cli_args);
// first i threads are updaters, rest are verifiers
if (i < update_threads) {
myargs[i].operation = touch_some_dbs;
} else {
myargs[i].operation = query_some_dbs;
}
}
// run all of the query and update workers. they may randomly open
// and close the dbs in each db_bucket to be some random dictionary,
// so when they're done we'll have to clean up the mess so this
// stress test can exit gracefully expecting db[i] = the ith db
run_workers(myargs, total_threads, cli_args->time_of_test, false, cli_args);
// make sure that every db in the original array is opened
// as it was when it was passed in.
for (int i = 0; i < num_buckets; i++) {
// close whatever is open
if (buckets[i].is_open) {
DB *db = buckets[i].db;
int r = db->close(db, 0);
CKERR(r);
}
// put the correct db back, then save the pointer
// into the dbp array we were given
open_ith_db(env, &buckets[i].db, i);
dbp[i] = buckets[i].db;
}
toku_free(buckets);
}
int int
test_main(int argc, char *const argv[]) { test_main(int argc, char *const argv[]) {
struct cli_args args = get_default_args(); struct cli_args args = get_default_args();
parse_stress_test_args(argc, argv, &args); parse_stress_test_args(argc, argv, &args);
args.env_args.update_function = increment;
// checkpointing is a part of the ref count, so do it often // checkpointing is a part of the ref count, so do it often
args.env_args.checkpointing_period = 5; args.env_args.checkpointing_period = 5;
// very small dbs, so verification scans are short and sweet
args.num_elements = 1000;
// it's okay for update to get DB_LOCK_NOTGRANTED, etc.
args.crash_on_operation_failure = false;
// just run the stress test, no crashing and recovery test
stress_openclose_crash_at_end = false;
stress_test_main(&args); stress_test_main(&args);
return 0; return 0;
} }
...@@ -1118,10 +1118,10 @@ static int UU()update_op2(DB_TXN* txn, ARG arg, void* UU(operation_extra), void ...@@ -1118,10 +1118,10 @@ static int UU()update_op2(DB_TXN* txn, ARG arg, void* UU(operation_extra), void
return r; return r;
} }
static int UU()update_op(DB_TXN *txn, ARG arg, void* operation_extra, void *UU(stats_extra)) { // take the given db and do an update on it
static int
UU() update_op_db(DB *db, DB_TXN *txn, ARG arg, void* operation_extra, void *UU(stats_extra)) {
int r; int r;
int db_index = myrandom_r(arg->random_data)%arg->cli->num_DBs;
DB* db = arg->dbp[db_index];
int curr_val_sum = 0; int curr_val_sum = 0;
DBT key, val; DBT key, val;
int rand_key; int rand_key;
...@@ -1180,6 +1180,14 @@ static int UU()update_op(DB_TXN *txn, ARG arg, void* operation_extra, void *UU(s ...@@ -1180,6 +1180,14 @@ static int UU()update_op(DB_TXN *txn, ARG arg, void* operation_extra, void *UU(s
return r; return r;
} }
// choose a random DB and do an update on it
static int
UU() update_op(DB_TXN *txn, ARG arg, void* operation_extra, void *stats_extra) {
int db_index = myrandom_r(arg->random_data) % arg->cli->num_DBs;
DB *db = arg->dbp[db_index];
return update_op_db(db, txn, arg, operation_extra, stats_extra);
}
static int UU() update_with_history_op(DB_TXN *txn, ARG arg, void* operation_extra, void *UU(stats_extra)) { static int UU() update_with_history_op(DB_TXN *txn, ARG arg, void* operation_extra, void *UU(stats_extra)) {
struct update_op_args* op_args = operation_extra; struct update_op_args* op_args = operation_extra;
assert(arg->bounded_element_range); assert(arg->bounded_element_range);
......
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment