0
0
mirror of https://github.com/sqlite/sqlite.git synced 2024-11-28 07:51:10 +01:00
sqlite/test/malloc3.test
dan b8fff29c68 In SQLITE_ENABLE_BATCH_ATOMIC_WRITE builds on F2FS file-systems, invoke
SQLITE_FCNTL_ROLLBACK_ATOMIC_WRITE if an SQLITE_FCNTL_COMMIT_ATOMIC_WRITE call
fails. Also, do not use an atomic transaction to create the initial database.
This is because if an error occurs while writing to the db file, any changes
to the file-size do not seem to be rolled back automatically. The only time
this matters is when the file was 0 bytes in size to start with.

FossilOrigin-Name: b3122db1545aeb48b7c28d480534b4b0fe04e83d5336225714c3cad926e5960e
2018-01-23 14:01:51 +00:00

704 lines
21 KiB
Plaintext

# 2005 November 30
#
# The author disclaims copyright to this source code. In place of
# a legal notice, here is a blessing:
#
# May you do good and not evil.
# May you find forgiveness for yourself and forgive others.
# May you share freely, never taking more than you give.
#
#***********************************************************************
#
# This file contains tests to ensure that the library handles malloc() failures
# correctly. The emphasis of these tests are the _prepare(), _step() and
# _finalize() calls.
#
# $Id: malloc3.test,v 1.24 2008/10/14 15:54:08 drh Exp $
set testdir [file dirname $argv0]
source $testdir/tester.tcl
source $testdir/malloc_common.tcl
# Only run these tests if memory debugging is turned on.
#
if {!$MEMDEBUG} {
puts "Skipping malloc3 tests: not compiled with -DSQLITE_MEMDEBUG..."
finish_test
return
}
# Do not run these tests if F2FS batch writes are supported. In this case,
# it is possible for a single DML statement in an implicit transaction
# to fail with SQLITE_NOMEM, but for the transaction to still end up
# committed to disk. Which confuses the tests in this module.
#
if {[atomic_batch_write test.db]} {
puts "Skipping malloc3 tests: atomic-batch support"
finish_test
return
}
# Do not run these tests with an in-memory journal.
#
# In the pager layer, if an IO or OOM error occurs during a ROLLBACK, or
# when flushing a page to disk due to cache-stress, the pager enters an
# "error state". The only way out of the error state is to unlock the
# database file and end the transaction, leaving whatever journal and
# database files happen to be on disk in place. The next time the current
# (or any other) connection opens a read transaction, hot-journal rollback
# is performed if necessary.
#
# Of course, this doesn't work with an in-memory journal.
#
if {[permutation]=="inmemory_journal"} {
finish_test
return
}
#--------------------------------------------------------------------------
# NOTES ON RECOVERING FROM A MALLOC FAILURE
#
# The tests in this file test the behaviours described in the following
# paragraphs. These tests test the behaviour of the system when malloc() fails
# inside of a call to _prepare(), _step(), _finalize() or _reset(). The
# handling of malloc() failures within ancillary procedures is tested
# elsewhere.
#
# Overview:
#
# Executing a statement is done in three stages (prepare, step and finalize). A
# malloc() failure may occur within any stage. If a memory allocation fails
# during statement preparation, no statement handle is returned. From the users
# point of view the system state is as if _prepare() had never been called.
#
# If the memory allocation fails during the _step() or _finalize() calls, then
# the database may be left in one of two states (after finalize() has been
# called):
#
# * As if the neither _step() nor _finalize() had ever been called on
# the statement handle (i.e. any changes made by the statement are
# rolled back).
# * The current transaction may be rolled back. In this case a hot-journal
# may or may not actually be present in the filesystem.
#
# The caller can tell the difference between these two scenarios by invoking
# _get_autocommit().
#
#
# Handling of sqlite3_reset():
#
# If a malloc() fails while executing an sqlite3_reset() call, this is handled
# in the same way as a failure within _finalize(). The statement handle
# is not deleted and must be passed to _finalize() for resource deallocation.
# Attempting to _step() or _reset() the statement after a failed _reset() will
# always return SQLITE_NOMEM.
#
#
# Other active SQL statements:
#
# The effect of a malloc failure on concurrently executing SQL statements,
# particularly when the statement is executing with READ_UNCOMMITTED set and
# the malloc() failure mandates statement rollback only. Currently, if
# transaction rollback is required, all other vdbe's are aborted.
#
# Non-transient mallocs in btree.c:
# * The Btree structure itself
# * Each BtCursor structure
#
# Mallocs in pager.c:
# readMasterJournal() - Space to read the master journal name
# pager_delmaster() - Space for the entire master journal file
#
# sqlite3pager_open() - The pager structure itself
# sqlite3_pagerget() - Space for a new page
# pager_open_journal() - Pager.aInJournal[] bitmap
# sqlite3pager_write() - For in-memory databases only: history page and
# statement history page.
# pager_stmt_begin() - Pager.aInStmt[] bitmap
#
# None of the above are a huge problem. The most troublesome failures are the
# transient malloc() calls in btree.c, which can occur during the tree-balance
# operation. This means the tree being balanced will be internally inconsistent
# after the malloc() fails. To avoid the corrupt tree being read by a
# READ_UNCOMMITTED query, we have to make sure the transaction or statement
# rollback occurs before sqlite3_step() returns, not during a subsequent
# sqlite3_finalize().
#--------------------------------------------------------------------------
#--------------------------------------------------------------------------
# NOTES ON TEST IMPLEMENTATION
#
# The tests in this file are implemented differently from those in other
# files. Instead, tests are specified using three primitives: SQL, PREP and
# TEST. Each primitive has a single argument. Primitives are processed in
# the order they are specified in the file.
#
# A TEST primitive specifies a TCL script as its argument. When a TEST
# directive is encountered the Tcl script is evaluated. Usually, this Tcl
# script contains one or more calls to [do_test].
#
# A PREP primitive specifies an SQL script as its argument. When a PREP
# directive is encountered the SQL is evaluated using database connection
# [db].
#
# The SQL primitives are where the action happens. An SQL primitive must
# contain a single, valid SQL statement as its argument. When an SQL
# primitive is encountered, it is evaluated one or more times to test the
# behaviour of the system when malloc() fails during preparation or
# execution of said statement. The Nth time the statement is executed,
# the Nth malloc is said to fail. The statement is executed until it
# succeeds, i.e. (M+1) times, where M is the number of mallocs() required
# to prepare and execute the statement.
#
# Each time an SQL statement fails, the driver program (see proc [run_test]
# below) figures out if a transaction has been automatically rolled back.
# If not, it executes any TEST block immediately proceeding the SQL
# statement, then reexecutes the SQL statement with the next value of N.
#
# If a transaction has been automatically rolled back, then the driver
# program executes all the SQL specified as part of SQL or PREP primitives
# between the current SQL statement and the most recent "BEGIN". Any
# TEST block immediately proceeding the SQL statement is evaluated, and
# then the SQL statement reexecuted with the incremented N value.
#
# That make any sense? If not, read the code in [run_test] and it might.
#
# Extra restriction imposed by the implementation:
#
# * If a PREP block starts a transaction, it must finish it.
# * A PREP block may not close a transaction it did not start.
#
#--------------------------------------------------------------------------
# These procs are used to build up a "program" in global variable
# ::run_test_script. At the end of this file, the proc [run_test] is used
# to execute the program (and all test cases contained therein).
#
set ::run_test_sql_id 0
set ::run_test_script [list]
proc TEST {id t} {lappend ::run_test_script -test [list $id $t]}
proc PREP {p} {lappend ::run_test_script -prep [string trim $p]}
proc DEBUG {s} {lappend ::run_test_script -debug $s}
# SQL --
#
# SQL ?-norollback? <sql-text>
#
# Add an 'SQL' primitive to the program (see notes above). If the -norollback
# switch is present, then the statement is not allowed to automatically roll
# back any active transaction if malloc() fails. It must rollback the statement
# transaction only.
#
proc SQL {a1 {a2 ""}} {
# An SQL primitive parameter is a list of three elements, an id, a boolean
# value indicating if the statement may cause transaction rollback when
# malloc() fails, and the sql statement itself.
set id [incr ::run_test_sql_id]
if {$a2 == ""} {
lappend ::run_test_script -sql [list $id true [string trim $a1]]
} else {
lappend ::run_test_script -sql [list $id false [string trim $a2]]
}
}
# TEST_AUTOCOMMIT --
#
# A shorthand test to see if a transaction is active or not. The first
# argument - $id - is the integer number of the test case. The second
# argument is either 1 or 0, the expected value of the auto-commit flag.
#
proc TEST_AUTOCOMMIT {id a} {
TEST $id "do_test \$testid { sqlite3_get_autocommit \$::DB } {$a}"
}
#--------------------------------------------------------------------------
# Start of test program declaration
#
# Warm body test. A malloc() fails in the middle of a CREATE TABLE statement
# in a single-statement transaction on an empty database. Not too much can go
# wrong here.
#
TEST 1 {
do_test $testid {
execsql {SELECT tbl_name FROM sqlite_master;}
} {}
}
SQL {
CREATE TABLE IF NOT EXISTS abc(a, b, c);
}
TEST 2 {
do_test $testid.1 {
execsql {SELECT tbl_name FROM sqlite_master;}
} {abc}
}
# Insert a couple of rows into the table. each insert is in its own
# transaction. test that the table is unpopulated before running the inserts
# (and hence after each failure of the first insert), and that it has been
# populated correctly after the final insert succeeds.
#
TEST 3 {
do_test $testid.2 {
execsql {SELECT * FROM abc}
} {}
}
SQL {INSERT INTO abc VALUES(1, 2, 3);}
SQL {INSERT INTO abc VALUES(4, 5, 6);}
SQL {INSERT INTO abc VALUES(7, 8, 9);}
TEST 4 {
do_test $testid {
execsql {SELECT * FROM abc}
} {1 2 3 4 5 6 7 8 9}
}
# Test a CREATE INDEX statement. Because the table 'abc' is so small, the index
# will all fit on a single page, so this doesn't test too much that the CREATE
# TABLE statement didn't test. A few of the transient malloc()s in btree.c
# perhaps.
#
SQL {CREATE INDEX abc_i ON abc(a, b, c);}
TEST 4 {
do_test $testid {
execsql {
SELECT * FROM abc ORDER BY a DESC;
}
} {7 8 9 4 5 6 1 2 3}
}
# Test a DELETE statement. Also create a trigger and a view, just to make sure
# these statements don't have any obvious malloc() related bugs in them. Note
# that the test above will be executed each time the DELETE fails, so we're
# also testing rollback of a DELETE from a table with an index on it.
#
SQL {DELETE FROM abc WHERE a > 2;}
SQL {CREATE TRIGGER abc_t AFTER INSERT ON abc BEGIN SELECT 'trigger!'; END;}
SQL {CREATE VIEW abc_v AS SELECT * FROM abc;}
TEST 5 {
do_test $testid {
execsql {
SELECT name, tbl_name FROM sqlite_master ORDER BY name;
SELECT * FROM abc;
}
} {abc abc abc_i abc abc_t abc abc_v abc_v 1 2 3}
}
set sql {
BEGIN;DELETE FROM abc;
}
for {set i 1} {$i < 100} {incr i} {
set a $i
set b "String value $i"
set c [string repeat X $i]
append sql "INSERT INTO abc VALUES ($a, '$b', '$c');"
}
append sql {COMMIT;}
PREP $sql
SQL {
DELETE FROM abc WHERE oid IN (SELECT oid FROM abc ORDER BY random() LIMIT 5);
}
TEST 6 {
do_test $testid.1 {
execsql {SELECT count(*) FROM abc}
} {94}
do_test $testid.2 {
execsql {
SELECT min(
(oid == a) AND 'String value ' || a == b AND a == length(c)
) FROM abc;
}
} {1}
}
SQL {
DELETE FROM abc WHERE oid IN (SELECT oid FROM abc ORDER BY random() LIMIT 5);
}
TEST 7 {
do_test $testid {
execsql {SELECT count(*) FROM abc}
} {89}
do_test $testid {
execsql {
SELECT min(
(oid == a) AND 'String value ' || a == b AND a == length(c)
) FROM abc;
}
} {1}
}
SQL {
DELETE FROM abc WHERE oid IN (SELECT oid FROM abc ORDER BY random() LIMIT 5);
}
TEST 9 {
do_test $testid {
execsql {SELECT count(*) FROM abc}
} {84}
do_test $testid {
execsql {
SELECT min(
(oid == a) AND 'String value ' || a == b AND a == length(c)
) FROM abc;
}
} {1}
}
set padding [string repeat X 500]
PREP [subst {
DROP TABLE abc;
CREATE TABLE abc(a PRIMARY KEY, padding, b, c);
INSERT INTO abc VALUES(0, '$padding', 2, 2);
INSERT INTO abc VALUES(3, '$padding', 5, 5);
INSERT INTO abc VALUES(6, '$padding', 8, 8);
}]
TEST 10 {
do_test $testid {
execsql {SELECT a, b, c FROM abc}
} {0 2 2 3 5 5 6 8 8}
}
SQL {BEGIN;}
SQL {INSERT INTO abc VALUES(9, 'XXXXX', 11, 12);}
TEST_AUTOCOMMIT 11 0
SQL -norollback {UPDATE abc SET a = a + 1, c = c + 1;}
TEST_AUTOCOMMIT 12 0
SQL {DELETE FROM abc WHERE a = 10;}
TEST_AUTOCOMMIT 13 0
SQL {COMMIT;}
TEST 14 {
do_test $testid.1 {
sqlite3_get_autocommit $::DB
} {1}
do_test $testid.2 {
execsql {SELECT a, b, c FROM abc}
} {1 2 3 4 5 6 7 8 9}
}
PREP [subst {
DROP TABLE abc;
CREATE TABLE abc(a, padding, b, c);
INSERT INTO abc VALUES(1, '$padding', 2, 3);
INSERT INTO abc VALUES(4, '$padding', 5, 6);
INSERT INTO abc VALUES(7, '$padding', 8, 9);
CREATE INDEX abc_i ON abc(a, padding, b, c);
}]
TEST 15 {
db eval {PRAGMA cache_size = 10}
}
SQL {BEGIN;}
SQL -norllbck {INSERT INTO abc (oid, a, padding, b, c) SELECT NULL, * FROM abc}
TEST 16 {
do_test $testid {
execsql {SELECT a, count(*) FROM abc GROUP BY a;}
} {1 2 4 2 7 2}
}
SQL -norllbck {INSERT INTO abc (oid, a, padding, b, c) SELECT NULL, * FROM abc}
TEST 17 {
do_test $testid {
execsql {SELECT a, count(*) FROM abc GROUP BY a;}
} {1 4 4 4 7 4}
}
SQL -norllbck {INSERT INTO abc (oid, a, padding, b, c) SELECT NULL, * FROM abc}
TEST 18 {
do_test $testid {
execsql {SELECT a, count(*) FROM abc GROUP BY a;}
} {1 8 4 8 7 8}
}
SQL -norllbck {INSERT INTO abc (oid, a, padding, b, c) SELECT NULL, * FROM abc}
TEST 19 {
do_test $testid {
execsql {SELECT a, count(*) FROM abc GROUP BY a;}
} {1 16 4 16 7 16}
}
SQL {COMMIT;}
TEST 21 {
do_test $testid {
execsql {SELECT a, count(*) FROM abc GROUP BY a;}
} {1 16 4 16 7 16}
}
SQL {BEGIN;}
SQL {DELETE FROM abc WHERE oid %2}
TEST 22 {
do_test $testid {
execsql {SELECT a, count(*) FROM abc GROUP BY a;}
} {1 8 4 8 7 8}
}
SQL {DELETE FROM abc}
TEST 23 {
do_test $testid {
execsql {SELECT * FROM abc}
} {}
}
SQL {ROLLBACK;}
TEST 24 {
do_test $testid {
execsql {SELECT a, count(*) FROM abc GROUP BY a;}
} {1 16 4 16 7 16}
}
# Test some schema modifications inside of a transaction. These should all
# cause transaction rollback if they fail. Also query a view, to cover a bit
# more code.
#
PREP {DROP VIEW abc_v;}
TEST 25 {
do_test $testid {
execsql {
SELECT name, tbl_name FROM sqlite_master;
}
} {abc abc abc_i abc}
}
SQL {BEGIN;}
SQL {CREATE TABLE def(d, e, f);}
SQL {CREATE TABLE ghi(g, h, i);}
TEST 26 {
do_test $testid {
execsql {
SELECT name, tbl_name FROM sqlite_master;
}
} {abc abc abc_i abc def def ghi ghi}
}
SQL {CREATE VIEW v1 AS SELECT * FROM def, ghi}
SQL {CREATE UNIQUE INDEX ghi_i1 ON ghi(g);}
TEST 27 {
do_test $testid {
execsql {
SELECT name, tbl_name FROM sqlite_master;
}
} {abc abc abc_i abc def def ghi ghi v1 v1 ghi_i1 ghi}
}
SQL {INSERT INTO def VALUES('a', 'b', 'c')}
SQL {INSERT INTO def VALUES(1, 2, 3)}
SQL -norollback {INSERT INTO ghi SELECT * FROM def}
TEST 28 {
do_test $testid {
execsql {
SELECT * FROM def, ghi WHERE d = g;
}
} {a b c a b c 1 2 3 1 2 3}
}
SQL {COMMIT}
TEST 29 {
do_test $testid {
execsql {
SELECT * FROM v1 WHERE d = g;
}
} {a b c a b c 1 2 3 1 2 3}
}
# Test a simple multi-file transaction
#
forcedelete test2.db
ifcapable attach {
SQL {ATTACH 'test2.db' AS aux;}
SQL {BEGIN}
SQL {CREATE TABLE aux.tbl2(x, y, z)}
SQL {INSERT INTO tbl2 VALUES(1, 2, 3)}
SQL {INSERT INTO def VALUES(4, 5, 6)}
TEST 30 {
do_test $testid {
execsql {
SELECT * FROM tbl2, def WHERE d = x;
}
} {1 2 3 1 2 3}
}
SQL {COMMIT}
TEST 31 {
do_test $testid {
execsql {
SELECT * FROM tbl2, def WHERE d = x;
}
} {1 2 3 1 2 3}
}
}
# Test what happens when a malloc() fails while there are other active
# statements. This changes the way sqlite3VdbeHalt() works.
TEST 32 {
if {![info exists ::STMT32]} {
set sql "SELECT name FROM sqlite_master"
set ::STMT32 [sqlite3_prepare $::DB $sql -1 DUMMY]
do_test $testid {
sqlite3_step $::STMT32
} {SQLITE_ROW}
}
}
SQL BEGIN
TEST 33 {
do_test $testid {
execsql {SELECT * FROM ghi}
} {a b c 1 2 3}
}
SQL -norollback {
-- There is a unique index on ghi(g), so this statement may not cause
-- an automatic ROLLBACK. Hence the "-norollback" switch.
INSERT INTO ghi SELECT '2'||g, h, i FROM ghi;
}
TEST 34 {
if {[info exists ::STMT32]} {
do_test $testid {
sqlite3_finalize $::STMT32
} {SQLITE_OK}
unset ::STMT32
}
}
SQL COMMIT
#
# End of test program declaration
#--------------------------------------------------------------------------
proc run_test {arglist iRepeat {pcstart 0} {iFailStart 1}} {
if {[llength $arglist] %2} {
error "Uneven number of arguments to TEST"
}
for {set i 0} {$i < $pcstart} {incr i} {
set k2 [lindex $arglist [expr {2 * $i}]]
set v2 [lindex $arglist [expr {2 * $i + 1}]]
set ac [sqlite3_get_autocommit $::DB] ;# Auto-Commit
switch -- $k2 {
-sql {db eval [lindex $v2 2]}
-prep {db eval $v2}
-debug {eval $v2}
}
set nac [sqlite3_get_autocommit $::DB] ;# New Auto-Commit
if {$ac && !$nac} {set begin_pc $i}
}
db rollback_hook [list incr ::rollback_hook_count]
set iFail $iFailStart
set pc $pcstart
while {$pc*2 < [llength $arglist]} {
# Fetch the current instruction type and payload.
set k [lindex $arglist [expr {2 * $pc}]]
set v [lindex $arglist [expr {2 * $pc + 1}]]
# Id of this iteration:
set iterid "pc=$pc.iFail=$iFail$k"
switch -- $k {
-test {
foreach {id script} $v {}
set testid "malloc3-(test $id).$iterid"
eval $script
incr pc
}
-sql {
set ::rollback_hook_count 0
set id [lindex $v 0]
set testid "malloc3-(integrity $id).$iterid"
set ac [sqlite3_get_autocommit $::DB] ;# Auto-Commit
sqlite3_memdebug_fail $iFail -repeat 0
set rc [catch {db eval [lindex $v 2]} msg] ;# True error occurs
set nac [sqlite3_get_autocommit $::DB] ;# New Auto-Commit
if {$rc != 0 && $nac && !$ac} {
# Before [db eval] the auto-commit flag was clear. Now it
# is set. Since an error occurred we assume this was not a
# commit - therefore a rollback occurred. Check that the
# rollback-hook was invoked.
do_test malloc3-rollback_hook_count.$iterid {
set ::rollback_hook_count
} {1}
}
set nFail [sqlite3_memdebug_fail -1 -benigncnt nBenign]
if {$rc == 0} {
# Successful execution of sql. The number of failed malloc()
# calls should be equal to the number of benign failures.
# Otherwise a malloc() failed and the error was not reported.
#
set expr {$nFail!=$nBenign}
if {[expr $expr]} {
error "Unreported malloc() failure, test \"$testid\", $expr"
}
if {$ac && !$nac} {
# Before the [db eval] the auto-commit flag was set, now it
# is clear. We can deduce that a "BEGIN" statement has just
# been successfully executed.
set begin_pc $pc
}
incr pc
set iFail 1
integrity_check $testid
} elseif {[regexp {.*out of memory} $msg] || [db errorcode] == 3082} {
# Out of memory error, as expected.
#
integrity_check $testid
incr iFail
if {$nac && !$ac} {
if {![lindex $v 1] && [db errorcode] != 3082} {
# error "Statement \"[lindex $v 2]\" caused a rollback"
}
for {set i $begin_pc} {$i < $pc} {incr i} {
set k2 [lindex $arglist [expr {2 * $i}]]
set v2 [lindex $arglist [expr {2 * $i + 1}]]
set catchupsql ""
switch -- $k2 {
-sql {set catchupsql [lindex $v2 2]}
-prep {set catchupsql $v2}
}
db eval $catchupsql
}
}
} else {
error $msg
}
# back up to the previous "-test" block.
while {[lindex $arglist [expr {2 * ($pc - 1)}]] == "-test"} {
incr pc -1
}
}
-prep {
db eval $v
incr pc
}
-debug {
eval $v
incr pc
}
default { error "Unknown switch: $k" }
}
}
}
# Turn off the Tcl interface's prepared statement caching facility. Then
# run the tests with "persistent" malloc failures.
sqlite3_extended_result_codes db 1
db cache size 0
run_test $::run_test_script 1
# Close and reopen the db.
db close
forcedelete test.db test.db-journal test2.db test2.db-journal
sqlite3 db test.db
sqlite3_extended_result_codes db 1
set ::DB [sqlite3_connection_pointer db]
# Turn off the Tcl interface's prepared statement caching facility in
# the new connnection. Then run the tests with "transient" malloc failures.
db cache size 0
run_test $::run_test_script 0
sqlite3_memdebug_fail -1
finish_test