How to write SQL injection proof PL/SQL

ORACLE WHITEPAPER | MAY 2017

 How to write SQL injection proof PL/SQL

CONTENTS

Abstract ......................................................... 1
Introduction ...................................................... 2
Definition of SQL injection .......................................... 4
Introducing a new notion: SQL syntax template . . . . . . . . . . . . . . . . . . . . . . . 4
Distinguishing between compile-time-fixed SQL statement text and
run-time-created SQL statement text . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
Distinguishing between a static SQL syntax template and a dynamic
SQL syntax template . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
Definition of static SQL syntax template . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
Definition of dynamic SQL syntax template . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
SQL injection (finally) defined . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
How can SQL injection happen? .................................... 13
Example 1: user-supplied column-comparison value . . . . . . . . . . . . . . . . . . 13
Example 2: user-supplied table name . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
Counter-example 3: user-supplied where clause . . . . . . . . . . . . . . . . . . . . . . 18
Counter-example 4: SQL syntax template with a questionable intent . . . . . 18
Ensuring the safety of a SQL literal or a simple SQL name ............... 20
Ensuring the safety of a SQL literal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
Ensuring the safety of a SQL text literal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
Ensuring the safety of a SQL datetime literal . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
Ensuring the safety of a SQL numeric literal . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
Ensuring the safety of a simple SQL name . . . . . . . . . . . . . . . . . . . . . . . . . . 26
Rules for cost-effective, guaranteed prevention of SQL injection . . . . . . . . . . 29
Expose the database to clients only via a PL/SQL API . . . . . . . . . . . . . . . . 29
Use compile-time-fixed SQL statement text unless you cannot . . . . . . . . . . 29
Use a static SQL syntax template for run-time-created SQL statement text
unless you cannot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
Replacement of a value placeholder in a SQL syntax template . . . . . . . . . . . . . . 31
Replacement of a simple SQL name placeholder in a SQL syntax template . . . . 33
Don’t confuse the need to use a dynamic SQL syntax template with the need
for dynamic text . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
Formal sufficient prescription for guaranteed safety . . . . . . . . . . . . . . . . . . . 34
Static text . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
Dynamic text . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
Safe dynamic text . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
Safe SQL statement text . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 Establish the safety of run-time-created SQL statement text in the code that
immediately precedes its execution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
Scenarios ....................................................... 40
Make a like predicate by adding leading and trailing % characters . . . . . . . . 40
In list with number of elements not known until run-time . . . . . . . . . . . . . . 41
Query by example form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
Callback . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
Analysis and hardening of extant code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
Appendix A:
Definitions of new terms of art introduced by this paper ............... 52
common SQL name . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
exotic SQL name . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
compile-time-fixed SQL statement text . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
run-time-created SQL statement text . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
SQL syntax template . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
value placeholder . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
simple SQL name placeholder . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
static SQL syntax template . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
dynamic SQL syntax template . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
static text . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
dynamic text . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
safe dynamic text . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
safe SQL statement text . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
top level PL/SQL block . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
Appendix B:
Summary of SQL injection prevention rules . . . . . . . . . . . . . . . . . . . . . . . . . . 56
Appendix C:
Additional Oracle-supplied subprograms that implement dynamic SQL ... 59
DBMS_Utility.Exec_DDL_Statement() . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
DBMS_DDL.Create_Wrapped() . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
DBMS_HS_Passthrough . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
DBMS_HS_Passthrough.Execute_Immediate(). . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
DBMS_HS_Passthrough.Parse() . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
OWA_Util . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
OWA_Util.Bind_Variables() . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
OWA_Util.ListPrint() . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
OWA_Util.TablePrint() . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
Appendix D:
Self-contained code to illustrate implementing callback
using dynamic polymorphism. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 How to write SQL injection proof PL/SQL

ABSTRACT

An internet search for “SQL injection” gets about 4 million hits. The topic
excites interest and superstitious fear. This whitepaper dymystifies the topic and
explains a straightforward approach to writing database PL/SQL programs that
provably guarantees their immunity to SQL injection.
Only when a PL/SQL subprogram executes SQL that it creates at run time is
there a risk of SQL injection; and you’ll see that it’s easier than you might think
to freeze the SQL at PL/SQL compile time. Then you’ll understand that you
need the rules which prevent the risk only for the rare scenarios that do require
run-time-created SQL. It turns out that these rules are simple to state and easy to
follow.

1 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 INTRODUCTION

Make sure that you’re reading the At the time of writing, An internet search for “SQL injection” gets about 4
latest copy of this paper. Check the million hits. The topic excites interest and superstitious fear. This whitepaper
URL given at the top of each page. dymystifies the topic and explains a straightforward approach to writing database
PL/SQL programs that provably guarantees their immunity to SQL injection.
The scope of the discussion is strictly limited to PL/SQL units that are stored in
the database. Similar principles apply in the discussion of languages, such as C or
Java, used to implement client-side programs, but it is very much harder to
control access to such programs. It is even harder to ensure that access to the
database is made only using such client-side programs.
In order best to understand the discussion, the reader must have sound
understanding of the various ways that SQL may be executed from a database
PL/SQL unit. The whitepaper Doing SQL from PL/SQL: Best and Worst Practices1,
addresses this topic in detail. Its study is, therefore, recommended as a
prerequisite for the study of this whitepaper. In particular, the Doing SQL from
PL/SQL whitepaper argues for the strategy that bans direct SQL access to the
database and exposes it to the client only via a strictly minimal PL/SQL API2. If
this strategy is adopted, then the proofing against SQL injection is the sole
responsibility of database PL/SQL; and a sufficient solution is possible in this
regime.
Of course, one cannot avoid what one cannot define — and so we start with the
section “Definition of SQL injection” on page 4. We use this definition, in the
section “How can SQL injection happen?” on page 13, to examine some famous
examples of code that is vulnerable. We also examine some counter-examples in
order to prove our definition of SQL injection.
The discussion in these two sections leads to the understanding that
SQL injection is possible only when a PL/SQL subprogram executes a
SQL statement whose text it has created at run time using what, here, we can
loosely call unchecked user input3. Clearly, then, the best way to avoid
SQL injection is to execute only SQL statements whose text derives entirely
from the source code of the PL/SQL program that executes it.
However, when the watertight approach will not meet the requirements, it is,
after all, necessary to handle user input — and to do so safely. A careful study of
this topic is presented in the section “Ensuring the safety of a SQL literal or a simple
SQL name” on page 20.
The material in these first three sections supports the rationale for, and the
understanding of, what follows in the section “Rules for cost-effective, guaranteed
prevention of SQL injection” on page 29. Indeed, if this paper makes an original
contribution, it is in the development of the conceptual framework, and the
associated terms of art, that then allow the rules to be stated compactly and

1. Doing SQL from PL/SQL: Best and Worst Practices is published on the Oracle Technology
Network website. You can find it easily with Internet search.
2. This is discussed in the section “Expose the database to clients only via a PL/SQL API” on page 29.
3. This notion will be formally defined in the section “Dynamic text” on page 36.

2 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 precisely. The new terms of art introduced and defined at appropriate points in
the discussion. But it proved very hard to find an order of exposition that didn’t
rely on forward reference to terms not yet defined. Therefore, Appendix A:
Definitions of new terms of art introduced by this paper on page 52 lists these terms,
defines each briefly, and cross-references to the section of the paper where it is
introduced.
Next, the section “Scenarios” on page 40 discusses some requirements scenarios
against which the concepts and the rules developed in the preceding sections can
be tested. They seem (to the beginner) to require the use of SQL statements built
from user input. However, such scenarios are very much fewer than many
programmers think; many can be implemented satisfactorily using
SQL statements whose text derives entirely from the source code. Several such
scenarios are illustrated.
This paper focuses unashamedly on writing injection-proof de novo PL/SQL
code. Finally, in the section “Analysis and hardening of extant code” on page 49, it
turns briefly to the topic of extant, and possibly vulnerable, code.
The rules that this paper explains, and insists on, are reproduced4 for quick
reference, in Appendix B: Summary of SQL injection prevention rules on page 56.
They guarantee proof against injection and yet are surprisingly easy to follow.
Moreover, as a bonus, they ensure semantic correctness in edge cases that
programmers often overlook.

4. This paper was prepared using Adobe Framemaker 8.0. Its cross-reference feature allows the
text of a source paragraph to be included by reference at the destination. The reader can be
certain, therefore, that the wording of each rule in the quick-reference summary is identical to
the wording where is stated. (Sadly, the mechanism does not preserve font nuances.)

3 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 DEFINITION OF SQL INJECTION

The definition of SQL injection requires an informal understanding of the

syntax of the SQL language and of how it is parsed.
Consider the two putative SQL statements shown in Code_1
-- Code_1
select c1 from t where c2 = 'Smith'

and Code_2.
-- Code_2
select c1 from t wear c2 = 'Smith'

It is clear, without connecting to an Oracle Database, that Code_1 is syntactically

correct but Code_2 has a syntax error5. Consider next the syntactically correct
SQL statements shown in Code_3
-- Code_3
select a1 from r where a2 = 'Jones'

and Code_4.
-- Code_4
select b1 from s where b2 = :1

The token :1 is a placeholder. If the table s, with columns b1 and b2, is accessible,
then Code_4 will parse without error6.

Introducing a new notion: SQL syntax template

The SQL statements in Code_1, Code_3, and Code_4 are different instances of the
same SQL syntax template; Template_1 shows it7.
-- Template_1
select &&1 from &&2 where &&3 = &4

The SQL syntax template notion and the notation used in Template_1 are
invented for the purposes of this whitepaper.
The notion belongs in the domain of discourse of the Design Specification
document. This document would list, with the notation used in Template_1, the
SQL syntax template or templates that are prescribed for a particular purpose.
Then the implementer would ensure, using the programming techniques that this
paper explains, that only those SQL statements that were instances of the

5. The attempt to execute Code_2 will always cause ORA-00933: SQL command not properly ended
while the attempt to execute Code_1 might cause ORA-00942: table or view does not exist.
ORA-00933 is a syntax error and ORA-00942 is a semantic error. Of course, if table
t(c1 varchar2(30), c2 varchar2(30)) is accessible to the current user, then Code_1 will be parsed
without error.
6. This is confirmed by executing this PL/SQL statement:
DBMS_Sql.Parse(Cur, 'select b1 from s where b2 = :1', DBMS_Sql.Native);
in an appropriately written PL/SQL anonymous block.
7. To emphasize the difference between an ordinary SQL statement and a SQL syntax template,
the latter will be rendered using a proportionally spaced italic font.

4 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 prescribed SQL syntax templates could be issued at the call site that implements
that part of the design.
The & syntax device denotes what we will call a value placeholder; and the &&
syntax device8 denotes what we will call a simple SQL name placeholder. Notice that
the value placeholder in a SQL syntax template is not the same notion as a
regular placeholder in an ordinary SQL statement. A value placeholder in a
SQL syntax template stands for either a well-formed SQL literal or a regular
placeholder in a SQL statement.
A particular SQL syntax template is a specific sequence of intermixed specific
keywords, specific operators, simple SQL name placeholders, and
value placeholders. The simple SQL name placeholders, and value placeholders
imply the possibility of various specializations of the template and so we extend
the SQL syntax template notion to include examples where concrete identifiers,
literals, and regular placeholders are in use. This is illustrated by the
SQL syntax templates shown in Template_2; each is a specialization of the most
generic form shown in Template_1.
-- Template_2
select c1 from &&1 where c2 = &1

select &&1 from t where &&2 = 99

select c1 from &&1 where c2 = :1

select c1 from t where c2 = :1

Notice that the only freedoms available when composing a particular

SQL statement as an instance of a particular SQL syntax template are the textual
substitutions9 of value placeholders and simple SQL name placeholders.

8. Don’t confuse the use of & here with its use in the SQL*Plus scripting language. The choice
of & was, however, made in homage to SQL*Plus’s use. In both cases, the syntax implies early
textual substitution before the “real” processing happens.
9. Notice that by saying “textual substitution” we don’t mean that the SQL syntax template is
represented as such by a value in PL/SQL source code — so that substitution is done
programatically with, for example, Replace(). Rather, we mean that a human who reads an
example of an actual SQL statement that the PL/SQL program uses at run time will be able to
see that it is an instantiation of the SQL syntax template that the Design Specification
document prescribes, achieved by the textual substitution we discuss.
With respect to the auditing of this substitution, we consider whitespace to be insignificant.
The Design Specification document can lay out a SQL syntax template as it pleases. And the
program is free to use a different layout. Ordinary comments of both styles are just a special
case of whitespace. However, when the Design Specification document prescribes the special
comment that starts with */+ and ends with /*, and that expresses a SQL hint, this must be
reproduced faithfully in the actual SQL statements that the program instantiates for this
SQL syntax template. Just as is the case for, for example, a keyword, a SQL hint is definitely
not a candidate for replacement at instantiation time.

5 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 • A value placeholder may be replaced with a regular placeholder10 or a
well-formed SQL literal whose datatype is either text, datetime, or numeric,
according to the specified syntax rules11.
• A simple SQL name placeholder may be replaced only with a simple
SQL name12.
• It is not allowed to substitute any other elements.
In particular, and by definition of the notion that this paper introduces, the . and
@ characters that are used to form qualified SQL names must be an explicit part
of the SQL syntax template; a qualified SQL name may not be used to replace a
simple SQL name placeholder13.

Distinguishing between compile-time-fixed SQL statement text and

run-time-created SQL statement text
We define the term compile-time-fixed SQL statement text to mean the text of a
SQL statement that cannot change at run time and that can be confidently
determined by reading the source code. More precisely, it is the text of a
SQL statement that is a PL/SQL static varchar2 expression14. The value of a
PL/SQL static varchar2 expression cannot change at run time and could be pre-
computed at compile time.
The SQL statement text for embedded SQL is composed by the PL/SQL
compiler and cannot change at run time. Therefore, embedded SQL definitely
executes only compile-time-fixed SQL statement text15.
However, it can easily be arranged that any of PL/SQL’s methods for executing
dynamic SQL will, at a particular call site, execute only compile-time-fixed SQL

10. The freedom to replace a value placeholder with either a regular placeholder or a SQL literal
is only of formal interest; it helps to clarify thought. It is very unlikely indeed that a real
program would be specified to implement the replacement of value placeholder with a regular
placeholder at run time. Such a design would give cause for concern.
11. These rules are specified in the SQL Language Reference book.
12. It is simplest to define a simple SQL name as that which the DBMS_Assert.Simple_Sql_Name()
function will return without raising an exception. Examples are SCOTT (which will be treated
the same as, for example, Scott) and “My Table”. SCOTT is an example of what this paper calls
a common SQL name; and My Table is an example of what it calls an exotic SQL name. These
notions are discussed in the section “Example 2: user-supplied table name” on page 15.
DBMS_Assert.Simple_Sql_Name() is discussed in the section “Ensuring the safety of a simple
SQL name” on page 26.
13. We shall defend this strict approach in the section “Ensuring the safety of a simple SQL name” on
page 26.
14. The term PL/SQL static varchar2 expression is defined in the PL/SQL Language Reference
book. This also defines a PL/SQL static varchar2 constant as a variable that is declared using the
constant keyword and that is initialized using a PL/SQL static varchar2 expression. The
definition is recursive: a PL/SQL static varchar2 constant can be used in the composition of a
PL/SQL static varchar2 expression. We shall return to this notion in the section “Static text” on
page 35.

6 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 statement text; it is necessary only to establish the statement text as a PL/SQL
static varchar2 expression.
We define the term run-time-created SQL statement text to mean the text of a
SQL statement that is not compile-time-fixed SQL statement text.
As we shall see, it is the distinction between compile-time-fixed SQL statement
text and run-time-created SQL statement text that is significant in the discussion
of SQL injection. It is more important to focus on the property of the SQL text
than on the method used for its execution. To (we hope!) avoid confusion, the
execution methods will be referred to as embedded SQL16, native dynamic SQL and
the DBMS_Sql API17.
Embedded SQL always executes compile-time-fixed SQL statement text.
Dynamic SQL may execute compile-time-fixed SQL statement text or may
execute run-time-created SQL statement text.

Distinguishing between a static SQL syntax template and a dynamic

SQL syntax template
Sometimes a PL/SQL subprogram executes a SQL statement whose
SQL syntax template is frozen at compile time. But sometimes the
SQL statement must be composed at run time in such a way that the set of
SQL syntax templates to which these must conform is too large be written down
explicitly in the Design Specification document. This section addresses this
important distinction.
Definition of static SQL syntax template

Consider the definer’s rights18 function f() as shown in Code_5.

-- Code_5
function f(PK in t.PK%type, Wait_Time in pls_integer)
return t.c1%type
authid Definer
is
c1 t.c1%type;
Stmt constant varchar2(32767) :=
'select c1 from t where PK = :b for update wait '
|| Wait_Time;
begin
execute immediate Stmt into c1 using PK;
return c1;
end f;

15. The fact that this text is characterized in a different domain of discourse (the text of the
compiled PL/SQL unit) than text which is a PL/SQL static varchar2 expression (here, the
domain is the source text of the unit) is unimportant with respect to establishing the notion of
compile-time-fixed SQL statement text.
16. We prefer the term embedded SQL to the more familiar static SQL. As we shall see, dynamic SQL
may be used to execute static text. The term static is heavily overloaded.
17. Procedural APIs other than the DBMS_Sql API support the execution of complete SQL
statements of restricted kinds or accept fragments of SQL text and concatenate these
unchecked to compose and then execute run-time-created SQL statement text. These are listed
in Appendix C: Additional Oracle-supplied subprograms that implement dynamic SQL on page 59. The
rules for using these safely are identical to those for using native dynamic SQL and the
DBMS_Sql API safely. Therefore, no further mention will be made of these other APIs in the
body of this paper.

7 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 Its purpose is to return the value of column c1 from the table t for a particular
value of the primary key, PK, and to lock the row for future update. The function
should not fail if another session presently has locked the row — and nor should
it wait indefinitely in such a case. Rather, it should allow the caller to specify the
maximum wait time. As it happens, SQL syntax does not allow a placeholder for
the value that determines the timeout period19.
Manual inspection is sufficient to show us that, while f() does not execute a
compile-time-fixed SQL statement text, it does execute a SQL statement whose
SQL syntax template is frozen at compile time as that shown in Template_320.
-- Template_3
select c1 from t where PK = :b for update wait &1

We shall call such a SQL syntax template a static SQL syntax template.
Notice that, as a consequence of the definitions, compile-time-fixed SQL
statement text always conforms to a static SQL syntax template. But run-time-
created SQL statement text may, or may not, conform to a static
SQL syntax template.
Definition of dynamic SQL syntax template
Now consider the procedure x.p() as shown in Code_6.
-- Code_6
package x is
type cw is varray(20) of boolean;
procedure p(PK in "My Table".PK%type, Wanted in cw);
end x;

Its purpose is to report the values of an arbitrary subset of all the columns in the
table with the exotic SQL name My Table21 for a particular value of the primary
key, PK. The author of code that will call x.p() knows the purpose of My Table,
and the names and significance of all its columns. In particular, he knows the
order in which these are listed in its external documentation. The nth element of
the in formal parameter Wanted determines if the nth column of My Table is to be
included in the report.
Let’s say that Functional Specification document requires that the order of the
columns in the report be the same as that in the external documentation of the

18. A belief seems to have arisen that an invoker’s rights unit is safe and that a definer’s rights unit
is risky. This viewpoint is naïve. The proper choice depends on the purpose of the subprogram.
Here, the purpose of f() is to select from a specific table using the privileges of the table’s
owner. It’s quite likely that Execute on f() will be granted to a user other than its owner, who
has no direct privileges on the table, specifically to give the grantee tightly controlled access to
the table. Invoker’s rights, on the other hand, is appropriate when the purpose of the
subprogram is to perform a parameterized, but powerful, operation using the privileges of the
user that invokes it. A risk would, of course, arise when such a subprogram is definer’s rights
and is owner by, for example, Sys!
19. This code is used to make a teaching point. It would be an unacceptable design in a real world
application. Each executed select statement would probably be textually different from any that
had been seen before; and this would cause a hard parse explosion. A compromise would be
to let the caller chose between, say, four values for the wait time: zero, short, long, and infinite.
The corresponding formal parameter would express the choice. And the select statement would
be build using one of four PL/SQL static varchar2 expressions for the wait time.

8 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 table. If the table has N columns, then the number of distinct select lists is
therefore the sum of the number of ways of choosing 1 item from N, then 2
items from N, through to N-1 items from N, and finally choosing all the items.
This22 is famously 2N-1. For 3 columns, there are 7 distinct select lists, for 10
columns there are 1023 distinct select lists, and for 20 columns, there are just over
one million distinct select lists. The number grows exponentially23 with the number
of columns.
Procedure x.p() models the implementation of a very common requirement: the
customizable report24. And an underlying table with many more than 20
columns is in no way unusual. It is therefore not feasible to provide each possible
statement as compile-time-fixed SQL statement text. Rather, the required
SQL statement must be composed programmatically.
The implementation of p() in the body of package x is shown in Code_7. The
Col_List() inner function composes the select list. In this design, column list is
perhaps not the best mnemonic; as Code_8 shows, the select list is in fact a single
item built by concatenating the right-blank-padded values of the chosen
columns. This approach allows the simple use of execute immediate rather than the

20. This relies on the fact that the intended timeout value is represented by a pls_integer and that
when a value of this type is concatenated to a varchar2, then it is implicitly converted to that
datatype using the single argument overload of To_Char(). The result of this conversion is
guaranteed to be well-formed SQL numeric literal. Because it’s an integral quantity, there’s never
decimal separator. The second actual (that determines the format model) cannot be influenced
through the environment; its default requests a decimal separator but no group separator. And
the third actual, that determines the characters used for the decimal separator and the group
separator, and that can be set using this statement:
alter session set NLS_Numeric_Characters
therefore has no effect. This is a subtle but crucial point. Had the datatype of the in formal
parameter Wait_Time been number, then the SQL syntax template of Stmt would have been
unpredictable. We shall return to it in the section “Ensuring the safety of a SQL numeric literal” on
page 24.
21. We use the exotic SQL name My Table just as a subliminal reminder that such names can occur.
We need do no more in the code discussed in this section than surround its use, as is shown,
with double quote characters. The uses occur only in ordinary PL/SQL source code (in the
declaration of the formal parameter PK) and in a PL/SQL static varchar2 expression. We shall
return to this point in the section “Query by example form” on page 42.
22. The internet searching required to confirm this is left as an exercise for the reader.
23. The expression to grow exponentially is often used metaphorically; here, it is used literally and
correctly!

9 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 more complicated use of the DBMS_Sql API for a select list whose composition is
unknown until run time.
-- Code_7
procedure p(PK in "My Table".PK%type, Wanted in cw) is

function Col_List return varchar2;

Stmt constant varchar2(32767) :=

'select '
|| Col_List()
|| ' Report from "My Table" where PK = :b';

Report varchar2(32767);

function Col_List return varchar2 is ... end Col_List;

begin
execute immediate Stmt into Report using PK;
DBMS_Output.Put_Line(Report);
end p;

Notice that Stmt is declared using the constant keyword and that this means that it
must be initialized as part of the declaration. This is not essential, but it is a
highly recommended approach.
Rule_1
When composing a SQL statement When it is necessary to compose a SQL statement programmatically, the code
programmatically, help the code usually needs, or at least benefits from the use of, variables for intermediate
reviewer by declaring variables used
results. Aim to declare these as constant, assigning the values in the declarations.
for intermediate results as constant.
This sometimes requires the use of nested block-statements or
forward-declared functions. This technique makes code review easier because
the reader can be sure that the value of a variable cannot change between its
initial assignment and its use.
For completeness, the implementation of the Col_List() function is shown in
Code_8.
-- Code_8
function Col_List return varchar2 is
type cn is varray(20) of varchar2(30);
Col_Names constant cn :=
cn('c1', 'c2', 'c3', 'c4', ..., 'c20');
Seen_One boolean := false;
List varchar2(32767);
begin
for j in 1..Wanted.Count() loop
if Wanted(j) then
List :=
List
|| case Seen_One when true then '||'
else ''
end
|| 'Rpad('||Col_Names(j)||', 10)';
Seen_One := true;
end if;
end loop;
return List;
end Col_List;

24. Another common requirement is to let the user specify comparison criteria for any subset of
the columns (the so-called query-by-example paradigm). We shall return to this in the section
“Don’t confuse the need to use a dynamic SQL syntax template with the need for dynamic text” on page 34
and then in the section “Query by example form” on page 42.

10 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 Notice that the elements from which List is composed are all PL/SQL
static varchar2 expressions25.
Suppose that, as the use of the type constructor varray(20) suggests, the table
My Table has 20 columns. The set of the more than one million distinct
SQL statements that x.p() might execute is too numerous to allow each one to be
inspected. Rather, the programmer (and auditor) must reason, from the
self-evident predictability of the possible return values of the Col_List function,
what the possible members are — and from that deduce what the possible
SQL syntax templates are26. In this example, there are 20 distinct
SQL syntax templates. Template_4 shows some.
-- Template_4
select Rpad(&&1, 10) Report
from "My Table" where PK = :b

select Rpad(&&1, 10)||Rpad(&&2, 10)||Rpad(&&3, 10) Report

from "My Table" where PK = :b

select Rpad(&&1, 10)||Rpad(&&2, 10)|| ... ||Rpad(&&20, 10) Report

from "My Table" where PK = :b

We shall call each such a SQL syntax template a dynamic SQL syntax template.
A static SQL syntax template can, by trivial inspection of the PL/SQL source
code that composes it, be written down by the auditor with certainty. A dynamic
SQL syntax template is one of a large set of such templates that are composed
for execution at a particular call site where the set is too large to allow each to be
written down but where, nevertheless, the set can be described with certainty.
Rule_2
Understand what is meant by the term Understand what is meant by the term SQL syntax template. Apply this
SQL syntax template and the understanding to the design of any code that constructs run-time-created SQL
difference between a static
statement text. Understand the difference between a static
SQL syntax template and a dynamic
SQL syntax template and a dynamic SQL syntax template.
SQL syntax template.

SQL injection (finally) defined

We define SQL injection, as it might occur from a PL/SQL subprogram, to
mean the execution, by that subprogram, of a SQL statement with a
SQL syntax template which differs, at a particular call site, from that which the
subprogram’s author intended for that call site.
This dangerous result occurs when the attacker supplies text for what the author
intends will replace a value placeholder or a simple SQL name placeholder in the
SQL syntax template. The attacker’s text subverts the quoting syntax and so is
parsed as a fragment of SQL syntax. This is the origin of the use of the term
“injection”: the attacker’s fragment has been injected into the programmer’s
intended statement. The attacker’s purpose is served when the resulting non-
conforming SQL is legal and so executes without detection to produce an

25. The term PL/SQL static varchar2 expression is defined in the PL/SQL Language Reference
book and its significance is discussed in the section “Static text” on page 35.
26. It would be possible to describe the universe of possible SQL statements and possible
SQL syntax templates using the regular expression syntax.

11 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 unintended result. The next section, How can SQL injection happen?, gives several
examples of the attacks to which badly written code is vulnerable.
Self-evidently, SQL injection cannot occur from a call site that uses
embedded SQL. Nor can it occur from a call site that uses dynamic SQL to
execute compile-time-fixed SQL statement text.
The injected text may come directly via one of the subprogram’s formal
parameters (a so-called first order attack); or it may come indirectly, for example
via a table which the subprogram reads, and trusts, to obtain a component of the
SQL statement that it is building, but into which the attacker has contrived to
insert a rogue value (a so-called second order attack).
Rule_3
Understand that SQL injection is the Understand how to define the term SQL injection as the execution of a
execution of a SQL statement with an SQL statement with an unintended SQL syntax template. Know that only
unintended SQL syntax template and
run-time-created SQL statement text that, therefore, must be executed using
that the risk can occur only when
dynamic SQL is potentially vulnerable to SQL injection.
run-time-created SQL statement text
is executed using dynamic SQL.

12 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 HOW CAN SQL INJECTION HAPPEN?

This is best demonstrated by examples. These will also enable us to tighten the
definition of SQL injection. In this section, we will not ask why the programmer
does not use compile-time-fixed SQL statement text27; we shall see later that
there are indeed use cases that do require run-time-created SQL statement text.
The fragments shown here, though not plausible, illustrate the SQL injection
techniques.

Example 1: user-supplied column-comparison value

Consider the PL/SQL fragment shown in Code_9.
-- Code_9
...
q constant varchar2(1) := '''';
SQL_VC2_Literal constant varchar2(32767) :=
q||Raw_User_Input||q;
begin
Stmt :=
'select c2 from t where c1 = '||SQL_VC2_Literal;
execute immediate Stmt bulk collect into v;
...

Presumably Raw_User_Input holds a PL/SQL text value that the programmer

expects to a possible value for the column t.c2. Is this vulnerable to
SQL injection? It is tempting to reply “no” because the SQL syntax template
seems to be fixed at compile time to this:
-- Template_5
select c1 from t where c2 = &1
If Raw_User_Input has this value:
-- Value_1
Smith

then Stmt will now be this:

-- Value_2
select c1 from t where c2 = 'Smith'

All seems to be well. But suppose that Raw_User_Input has this value:
-- Value_3
O'Brien

The value of the PL/SQL variable Stmt will now be this:

-- Value_4
select c1 from t where c2 = 'O'Brien'

27. Of course, the intent of Template_5 can be achieved using embedded SQL (which, under the
covers, generates at compile time and uses at run time a compile-time-fixed SQL statement text
that includes a placeholder). Code_9 therefore violates the principles advocated in the section
“Use compile-time-fixed SQL statement text unless you cannot” on page 29. However, we have seen
code where native dynamic SQL has been used where embedded SQL was sufficient.
Sometimes the explanation is nothing more than what ordinarily explains poorly conceived
and written programs; sometimes, it might be the work of a disgruntled and malicious
employee.

13 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 While this is unremarkable as a PL/SQL text value, it is syntactically incorrect as
a SQL statement (the single quote characters don’t balance) and will cause
ORA-00933: SQL command not properly ended when it is parsed. The programmer
intended to construct a SQL text literal from a PL/SQL text value by
surrounding it with an opening and a closing single quote character28. But he
forgot that O'Brien is a perfectly plausible value for a text column and that,
therefore, the PL/SQL text value might legally contain interior single quote
characters. The rules for constructing a SQL text literal from a PL/SQL text
value require that each interior single quote character be escaped by doubling it
before surrounding it with an opening and a closing single quote character29.
Because the programmer forgot this, the singleton interior single quote character
closes the SQL text literal, and Brien' is parsed as two SQL tokens, Brien and '.
This might seem to be an ordinary bug — and, of course, it should be avoided
for ordinary reasons. However, if present, it could go undetected indefinitely;
and it has a shocking manifestation when an ingeniously, and maliciously,
contrived value for Raw_User_Input is provided.
Suppose that Raw_User_Input has this value:
-- Value_5
'
union
select Username c1 from All_Users --

Stmt will now be this:

-- Value_6
select c1 from t where c2 = ''
union
select Username c1 from All_Users --'

Notice that the would-be closing single quote character that the programmer
adds, and that would cause a syntax error if it were unbalanced, is disarmed
because Value_5 finishes with the -- token that starts a single line comment.
The result is a legal SQL statement. Because, in Oracle Database, the empty
string is the same as null, and because an equality comparison with null always
results in null, no rows are selected from t; so the SQL statement reduces to this:
-- Value_7
select Username c1 from All_Users

This SQL statement an instance of a very different SQL syntax template than
Template_5 shows — and is surely not what the programmer intended. We have

28. The point here, of course, is that the programmer is writing the source code of a PL/SQL
program whose purpose is to create, at run-time, the source text of another program (a
SQL statement) and then to execute that.
29. Oracle Database 10g introduced an alternative quoting syntax for both SQL and PL/SQL —
the so-called alternative quoting mechanism (sometimes know as the user-defined quoting
mechanism, or the q-quote syntax). It aims to increase usability when the value itself contains
single quote characters. Here is an example in PL/SQL:
v varchar2(80) := q'{You can't do that}';
The opening q' and the closing ' are non-negotiable; The user chooses the inner parentheses,
in this example { and }.We will see in the section “Ensuring the safety of a SQL text literal” on
page 20 that this mechanism should be avoided.

14 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 seen, therefore, that the subprogram that executes Stmt is indeed vulnerable to
SQL injection.
This is the canonical example of SQL injection. The superstitious fear is that any
program that issues run-time-created SQL statement text is vulnerable to such
an attack. However, as the cartoon30 suggests, such attacks can always be
prevented by following simple immunization practices.

The proper practice in this example is clear: when a PL/SQL text value is to be
converted to a SQL text literal, then the result must be a string that starts and
ends with a single quote character; and between these, there must be no
singleton single quote characters and no runs of an odd number of single quote
characters. This is discussed formally in the section “Ensuring the safety of a
SQL literal” on page 20.

Example 2: user-supplied table name

Consider the PL/SQL fragment shown in Code_10.
-- Code_10
...
Stmt :=
'select c1 from '||Raw_User_Input||' where c2 = ''Smith''';

execute immediate Stmt bulk collect into v;

...

Presumably Raw_User_Input holds a value that the programmer expects to be the

name of a table or view. Is this vulnerable to SQL injection? As with Example 1, it
is tempting to reply “no” because the SQL syntax template seems to be fixed at
compile time to this:
-- Template_6
select c1 from &&1 where c2 = 'Smith'
If Raw_User_Input has this value:
-- Value_8
t

then Stmt will now be this:

-- Value_9
select c1 from t where c2 = 'Smith'

30. The original of this cartoon is to be found at xkcd.com/327/. A footnote states “...you're free
to copy and share these comics (but not to sell them)”.

15 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 Again, all seems to be well. But recall that, though few programmers take
advantage of it31, the following SQL*Plus script runs without error.
-- Code_11
drop table "a /"
/
create table "a /"("?" number, "a'b" number, " " number)
/
insert into "a /"("?", "a'b", " ") values (1, 2, 3)
/
select * from "a /" where "?" = 1
/
select '['||Table_Name||']' x from User_Tables
union
select '['||Column_Name||']' x from User_Tab_Cols
where Table_Name = 'a /'
order by x
/

A SQL identifier can consist of any sequence of one or more characters in the
database character set provided that the representation needs no more than
30 bytes32. The final select in Code_10 produces an output that includes these
rows:
[ ]
[?]
[a /]
[a'b]

Notice how the values end up in metadata in the catalog views; and notice how,
when these views are queried, the rules are no different from those for any other
tables and views. Especially, because an identifier may legally contain singleton
single quote characters, the same care must be used when constructing a SQL
text literal to represent one in a metadata query.
It helps to think that SQL ordinarily requires that every identifier is surrounded
with double quote characters. Only if the identifier happens to start with an
alphabetic character, and happens thereafter to contain only alphanumeric
characters or underscore, # or $, then (as a usability bonus) the double quote
characters may be omitted. And if they are omitted, then the SQL parser upper-
cases the identifier.
There seem to be no established terms of art to capture this distinction. We will
use the term common SQL name to denote one that starts with an upper case
alphabetic character in the range A..Z and then has only upper case
alphanumeric characters in the range A..Z, or underscore, # or $. A common
SQL name doesn’t need to be surrounded by double quote characters in a
SQL statement; and if it is not so surrounded, the case with which it is written
doesn’t matter. It may, however, be so surrounded. If it is, then because the SQL

31. GUI tools for the DBA and the developer are becoming increasingly popular. These support
object creation without typing the SQL directly. It is therefore more likely these days to
encounter exotic SQL names (like Line Items) than it used to be.
32. There is one exception: this DDL statement:
create table "a""b"(n number)
fails with ORA-03001: unimplemented feature. An identifier may not contain the double quote
character. Oracle Database no longer allows this. The text of the message might seem to
suggest a possibility that a later version of Oracle Database might, again, allow this. This is
misleading. There is no intention to do this.

16 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 parser preserves its case, it must be written in all upper case. We will use the term
exotic SQL name to denote one that violates the rules for a common SQL name
and that must, therefore, be surrounded by double quote characters in a
SQL statement.
Suppose, then, that the value for the simple SQL name placeholder in Template_6
is a / rather than t; Raw_User_Input will have this value:
-- Value_10
a /

Stmt will now be this:

-- Value_11
select c1 from a / where c2 = 'Smith'

This is syntactically incorrect and will cause ORA-00933: SQL command not
properly ended when it is parsed.
The programmer intended to construct a SQL identifier from a PL/SQL text
value. But he forgot about the possibility of getting a exotic SQL name. The
rules for constructing a SQL identifier from a PL/SQL text value need to take
account of this. This is discussed formally in the section “Ensuring the safety of a
simple SQL name” on page 26.
Again, this might seem to be an ordinary bug; and again it could go undetected
indefinitely. (The use of exotic SQL names is relatively rare.) But again the bug
has a shocking manifestation when an ingeniously, and maliciously, contrived
value for Raw_User_Input is provided.
Suppose that Raw_User_Input has this value:
-- Value_12
t where 1=2
union
select Username c1 from All_Users --

Stmt will now be this:

-- Value_13
select c1 from t where 1=2
union
select Username c1 from All_Users -- where c2 = 'Smith'

The code expects only a common SQL name. But the caller has supplied an
exotic SQL name. It has been cunningly designed so that the result is a legal
SQL statement. As in Example 1, again the SQL statement reduces to this:
-- Value_14
select Username c1 from All_Users

Again, this is an instance of a very different SQL syntax template than

Template_6 shows and again it is surely not what the programmer intended. Of
course, then, the subprogram that executes Stmt is vulnerable to SQL injection.
The proper practice in this example, too, is clear: when a PL/SQL text value is to
be used to compose a SQL identifier, then the conversion must acknowledge the
difference between a common SQL name and an exotic SQL name, must handle
case accordingly, and must surround the final result with double quote
characters.

17 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 Counter-example 3: user-supplied where clause
Consider next the PL/SQL fragment shown in Code_12.
-- Code_12
...
Stmt := 'select c1 from t where '||Raw_User_Input;
execute immediate Stmt bulk collect into v;
...

Assuming that the PL/SQL identifier is appropriately named, then the intended
SQL syntax template is simply unspecified. Formally, therefore, it falls outside
the scope of the discussion of SQL injection — just as does SQL*Plus, which is
specified to be able to execute any arbitrary SQL statement.
The programmer probably did have an informal notion of the set of possible
SQL syntax templates that Stmt may instantiate, for example an equality
comparison with a literal for each of any arbitrary combination of columns from
table t. However, unlike checking that a string is a well-formed SQL literal or
SQL identifier, it is very difficult to check that a putative complete where clause is
an instantiation of one of a set of intended SQL syntax templates.
A Functional Specification document for a PL/SQL program that offers the
powerful flexibility implied by Code_12 must be rejected unless it can be
guaranteed that it can be executed only by a user who anyway could connect
directly to the database and execute arbitrary select statements against the same
table.
We shall return to this point in the section “Use a static SQL syntax template for
run-time-created SQL statement text unless you cannot” on page 31.

Counter-example 4: SQL syntax template with a questionable intent

Consider next the procedure shown in Code_13.
-- Code_13
procedure Make_DBA(Raw_User_Input in varchar2)
authid Definer -- Current_User
is
Double_Quote_Test constant char(3) := '%"%';
begin
if Raw_User_Input like Double_Quote_Test then
Raise_Application_Error(-20000, 'Interior " is illegal');
end if;

declare
Username constant varchar2(32767) :=
'"'||Raw_User_Input||'"';
Stmt constant varchar2(32767) :=
'grant DBA to '
|| Username
|| ' identified by x with Admin Option';
begin
DBMS_Output.Put_Line(Stmt);
execute immediate Stmt;
end;
end Make_DBA;

18 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 Because a value of Raw_User_Input that has an interior double quote character33
causes an error, and because a value that does not is then surrounded by double
quotes34, there is no risk that the SQL syntax template will be anything other
than this:
-- Template_7
grant DBA to &&1 identified by x

There is, therefore, no risk of SQL injection35, and further discussion of it is

beyond the scope of this whitepaper.
Not to leave it unsaid, though, the intent of this SQL syntax template does signal
the need for some serious analysis36. Because Make_DBA() is an invoker’s rights
procedure, then it can be executed without error only by a user that already has
the DBA role with Admin Option. Nominally, then, it should be safe to grant
Execute on this to public. There is, however, a subtle secondary point to consider.
If there exists a subprogram in the database, whose owner has the DBA role
with Admin Option, and that is vulnerable to injection, then it might be possible to
inject an invocation of this procedure37 where it would not be possible to inject
the SQL that it executes. Clearly, then, the existence of a procedure like
Make_DBA() makes a security audit harder.

33. A simple SQL name must not contain a double quote character. An attempt to use such a
name in a SQL statement causes ORA-03001: unimplemented feature.
34. This crude programming device is used in this example in order not to pre-empt the discussion,
in the section “Ensuring the safety of a simple SQL name” on page 26, of the proper approach. We
shall see there that the proper approach is to use DBMS_Assert.Simple_Sql_Name().
35. You might question the usability of the interpretation of Raw_User_Input. If you wanted to
specify a user whose name was the common SQL name SCOTT, then you would have to be
careful to spell it in all upper case. This is not the convention in SQL. However, the usability
is not interesting in this example. The example is designed to illuminate the discussion of
SQL injection.
36. To keep the code simple, the password is specified as an explicit identifier. But the point of
this example would be unchanged even if the password were specified by user input that not
only was surrounded with double quotes but also was required to pass a stringent password
strength test.
37. This would require wrapping it in an invoker’s rights function that was defined with the
Autonomous_Transaction pragma.

19 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 ENSURING THE SAFETY OF A SQL LITERAL OR A SIMPLE SQL NAME

Only two kinds of element in a SQL syntax template may be replaced with
dynamic text38: a value placeholder may be replaced with a SQL literal and a
simple SQL name placeholder may be replaced with a simple SQL name.
However, as we have seen in the section “Example 1: user-supplied column-comparison
value” on page 13 and the section “Example 2: user-supplied table name” on page 15,
it is precisely in this replacement that the risk of SQL injection occurs. This
section, therefore, prescribes the approaches that guarantee immunity to that
risk.

Ensuring the safety of a SQL literal

There are three kinds of SQL literal: text, datetime, and numeric39. Each deserves
separate attention.
Ensuring the safety of a SQL text literal
If the value O’Brien is to be presented as a SQL text literal, then it can be written
either using the default mechanism shown in Code_14 or using the so-called
alternative quoting mechanism40 shown in Code_15.
-- Code_14
...where Last_Name = 'O''Brien'

-- Code_15
...where Last_Name = q'{O'Brien}'

While the alternative quoting mechanism can improve readability for values that
have interior singleton occurrences of the single quote character, its use must be
banned for the purpose of composing a safe SQL text literal41.
The DBMS_Assert package42 exposes the function Enquote_Literal(). It has a
single formal parameter, Str, of datatype varchar2 and mode in.; and its return
datatype is varchar2. When the input is a well-formed SQL text literal, then the
output is identical to the input; but when the input is not a well-formed SQL text
literal, then the predefined exception Standard.Value_Error is raised43.

38. The term dynamic text is formally defined in the section “Dynamic text” on page 36. For now, it
is sufficient to assume the intuitive meaning: text whose composition cannot be traced back to
only PL/SQL static varchar2 expressions.
39. This syntax definition for these items is given in the SQL Language Reference book.
40. Support for the alternative quoting mechanism was introduced in Oracle Database 10g.
41. The reason is simple: the DBMS_Assert package does not provide a function to assert the safety
of this syntax.
42. The DBMS_Assert package was first documented in Oracle Database 11g.
43. There is one exception: when the input has only properly doubled interior single quote
characters, but neither starts nor ends with a single quote character, then Enquote_Literal() will
quietly add these. There are those (for example, Bryn Llewellyn!) who feel that this was an
unfortunate design and that the function would have been better conceived of as a pure
asserter. However, it is what it is; and backwards compatibility considerations prohibit
changing it. This paper encourages the approach, which is always possible, that uses it as a pure
asserter.

20 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 These, then, are the rules for composing a safe SQL text literal from a PL/SQL
text value:
• Replace each singleton occurrence, within the PL/SQL text value, of the single
quote character with two consecutive single quote characters.
• Concatenate one single quote character before the start of the value and one
single quote character after the end of the value.

• Assert that the result is safe with DBMS_Assert.Enquote_Literal()44.

Notice that the mandate in the third bullet is the crucial one. It is this one that
guarantees immunity to injection; the first mandate prevents annoying run-time
errors.
Ensuring the safety of a SQL datetime literal
A SQL datetime literal is a SQL text literal which must conform to the additional
condition that it must be capable of conversion to a value of a datetime datatype.
Obviously, then, the safety of a SQL datetime literal must be asserted with
DBMS_Assert.Enquote_Literal() in exactly the same way as a SQL text literal.
It is interesting to see what can happen if this simple rule is forgotten, and if the
documented behaviors of the single argument To_Char(d), where d has a datetime
datatype, and the single argument To_Date(t), where t has a text datatype, are
forgotten too. Consider the contrived procedure whose first few lines are shown
in Code_16.
-- Code_16
procedure p is
q constant varchar2(1) := '''';
d constant date :=
To_Date('2008-09-22 17:30:00', 'yyyy-mm-dd hh24:mi:ss');
Stmt constant varchar2(32767) :=
'select t.PK from t where t.d > ' || q||d||q;
...
begin
execute immediate Stmt bulk collect into Results;
...

The author might understand that, with the code as presented, the single
argument To_Char() is implicitly invoked by PL/SQL when d is concatenated
into Stmt; and that, next, the single argument To_Date() is implicitly invoked by
SQL when Stmt is executed. He might even remember that both the output of
the single argument To_Char() and that of the single argument To_Date() are
affected by the environment setting of the NLS_Date_Format parameter — and
might reason that the second conversion is the antidote to the first, and that
there is, therefore, no cause for concern.
However, this understanding is naïve. At the very least, he has programmed an
ordinary bug that can lead to unreliable query results because of a loss of date

44. In all the code examples in this paper, the name of this package is always qualified with its
owner, as Sys.DBMS_Assert, rather than using the bare public synonym. This measure is
essential for safety. It ensures that the simple name cannot be captured by a private synonym
or package in the same schema as the PL/SQL unit which refers to what is intended to be the
Oracle-supplied DBMS_Assert package.

21 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 precision. Suppose that table t had been populated by the SQL*Plus script
shown in Code_17.
-- Code_17
alter session set NLS_Date_Format = 'yyyy-mm-dd hh24:mi:ss'
/
begin
insert into t(PK, d) values(1, '2008-09-23 17:30:00');
insert into t(PK, d) values(2, '2008-09-21 17:30:00');
commit;
end;
/

And suppose, now, that p() is executed using the SQL*Plus script shown in
Code_18.
-- Code_18
alter session set NLS_Date_Format = 'dd-Mon-yy hh24:mi:ss'
/
begin p(); end;
/
alter session set NLS_Date_Format = 'yyyy'
/
begin p(); end;
/

The first execution of p(), though it runs with a different setting of

NLS_Date_Format than was current when table t was populated, still produces
the expected result: only the row with PK=1 is selected. (This is because it
honors the precision with which the dates were stored.) However, the second
execution produces what is almost certainly an unexpected result: both the row
with PK=1 and the row with PK=2 are selected. The reason, of course, is that the
conversion of the value in the PL/SQL variable d, after lossy conversion to text
and then conversion back to a datetime datatype, has ended up as (what would
display as) 2008-01-01 00:00:00. This no longer honors the intended precision.
One who has followed the discussion of the case presented in the section
“Example 1: user-supplied column-comparison value” on page 13), will not be surprised
to learn that procedure p() is not only ordinarily buggy; it is also vulnerable to
SQL injection. Suppose that p() is executed using the SQL*Plus script shown in
Code_19.
-- Code_19
alter session set
NLS_Date_Format = '"'' and Scott.Evil()=1--"'
/
begin p(); end;
/

The value shown in Value_15 is assigned to Stmt.

-- Value_15
select t.PK from t where t.d > '' and Scott.Evil()=1--'

This is surely not an instance of the intended SQL syntax template. Therefore,
SQL injection has occurred. The explanation is a twist on the usual one: a
singleton occurrence of the single quote character has been injected into the
run-time composed SQL statement, but this time from the side45 — via an NLS
environment parameter — rather than directly. Of course, all bets are off now.
The invocation of the malicious function is injected, and the would-be closing

22 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 single quote character that p() adds is disarmed by the -- token that starts a single
line comment. Scott.Evil() is invoked!
While the argument just presented might seem tortuous, the conclusion is simple
to understand and simple to implement. These, then, are the rules for composing
a safe SQL datetime literal from a PL/SQL datetime value:
• Use the two-parameter overload, for an input of datatype date, To_Char(d, Fmt),
to compose a SQL datetime literal46, t. (This, of course, will be a PL/SQL
varchar2.) Use a value for Fmt that is consistent with what the Functional
Specification document requires for the precision.
• Concatenate one single quote character before the start of this value and one
single quote character after its end.
• Assert that the result is safe with DBMS_Assert.Enquote_Literal().
• Compose the date predicate in the SQL statement using the two-parameter
overload for To_Date(t, Fmt) and using the identical value for Fmt as was used
to compose t.
Notice that the mandate in the third bullet is the crucial one. It is this one that
guarantees immunity to injection; the first two and the fourth mandates prevent
annoying run-time errors.
The procedure p_Safe(), whose first few lines are shown in Code_20, implements
this approach.
-- Code_20
procedure p_Safe(d in date) is
q constant varchar2(1) := '''';

-- Choose precision according to purpose.

Fmt constant varchar2(32767) := 'J hh24:mi:ss';

Safe_Date_Literal constant varchar2(32767) :=

Sys.DBMS_Assert.Enquote_Literal(q||To_Char(d, Fmt)||q);

Fmt_Literal constant varchar2(32767) := q||Fmt||q;

Safe_Stmt constant varchar2(32767) :=

' insert into t(d) values(To_Date('
|| Safe_Date_Literal
|| ', '
|| Fmt_Literal
|| '))';
begin
execute immediate Safe_Stmt;
...

45. This form of attack has been referred to as lateral SQL injection by David Litchfield. The article
is available on the internet here:
www.databasesecurity.com/dbsec/lateral-sql-injection.pdf
46. Of course, date is not the only datetime datatype. The same reasoning applies for, for example,
a timestamp literal.

23 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 When p_Safe() is invoked using Sysdate(), a value like that shown in Value_16 47 is
assigned to Safe_Stmt.
-- Value_16
insert into t(d) values(To_Date('2454723 18:01:05', 'J hh24:mi:ss'))

The composition of Safe_Stmt is immune to the effect of changes to the

NLS_Date_Format parameter.
Ensuring the safety of a SQL numeric literal
Unlike a SQL datetime literal, which is a specialized kind of SQL text literal, a SQL
numeric literal has specific syntax. It always uses a dot as the decimal character and
never contains a group separator. Here are some examples (notice that there are
no surrounding single quote characters)48:
42
-1
+6.34
0.5
-123.4567
25e-03
25f
+6.34F
0.5d
-1D

The SQL syntax does not recognize national differences. However, the output of
the overload of the To_Char() function for a numeric datatype does. This function
has three sub-overloads with one, two, and three formal parameters.
For the overload with three formal parameters, the second formal parameter,
called Fmt, specifies the format model and the third formal parameter, called
NLSparam, specifies the actual characters that will be used for notions like the
decimal character, the group separator, the currency symbol, and so on.
For the overload with two formal parameters, the second formal parameter
specifies the format model. Notably, if this overload is used, the value for (the
components of) NLSparam is determined by the NLS_Numeric_Characters,
NLS_Currency, and NLS_ISO_Currency environment parameters. Even more
notably, if the overload with one formal parameter is used, then the value for Fmt
has a fixed default: it cannot be controlled from the environment. But the value
for NLSparam is still determined by the environment and still has an effect.
Suppose that table t has been populated using the insert statement shown in
Code_21.
-- Code_21
insert into t(n) values (123456.789)

47. The Julian date 2454723 is 13-Sep-2008.

48. These are taken from the SQL Language Reference book.

24 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 Then the SQL*Plus script shown in Code_22
-- Code_22
select n from t
/
alter session set NLS_Numeric_Characters = ',.'
/
alter session set NLS_Currency = 'NOK '
/
select To_Char(n, 'L999G999G999D999') n from t
/
select n from t
/
select To_Char(n, 'TM', 'NLS_Numeric_Characters = ''!.''') n
from t
/
alter session set NLS_Numeric_Characters = '''.'
/
select 'c1 = '||n x from t
/

will produce this output:

123456.789

NOK 123.456,789

123456,789

123456!789

c1 = 123456'789

If a SQL numeric literal is not composed carefully, there is a risk of generating a

SQL statement with a syntax error — in other words, a different
SQL syntax template than was intended. The ability to inject a singleton
single quote character is particularly alarming. This is another example of lateral
SQL injection, and the moral is clear.
These, then, are the rules for composing a safe SQL numeric literal from a
PL/SQL numeric value:
• Use explicit conversion with the To_Char() overload with three formal
parameters. This overload requires that a value be supplied for Fmt. Explicitly
provide the value that supplies the default when the overload with one formal
parameter is used. This is 'TM'.49
• Explicitly provide the value that supplies the default for the
NLS_Numeric_Characters parameter when the one of the overloads with one or
two formal parameters is used. This is '.,'.
We shall return to this point in the section “Safe dynamic text” on page 36, where
we introduce the shorthand To_Char(x f, n) for this overload.

49. 'TM' is the so-called text minimum number format model. It returns the smallest number of
characters possible in fixed notation unless the output exceeds 64 characters. In that case, the
number is returned as if the model had been 'TMe'. 'TMe' returns the smallest number of
characters possible in scientific notation. If the Functional Specification document suggests
this, it is safe also to specify the 'TMe' model explicitly. In fact, any format model is safe
provided that it is chosen deliberately, together with the value for the NLSparam argument, in
the light of the requirements for precision given in the Functional Specification document.

25 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 Ensuring the safety of a simple SQL name

Suppose that a user exists with the exotic SQL name50 O’Brien in whose schema
there exists a procedure with the common SQL name PROC and a package with
the common SQL name PKG. Suppose that this package has a procedure with
the exotic SQL name Do it. Finally, suppose there exists, in a different database, a
link to this one with the common SQL name LNK. Here are some examples of
the various qualified SQL names that, in different contexts, might be used in a
SQL statement:
Proc
"O'Brien".Pkg
Pkg."Do it"
Proc@"Lnk"
"O'Brien".Pkg."Do it"
"O'Brien".Pkg@lnk
"O'Brien".Pkg."Do it"@LNK

This is the general form of a qualified SQL name:

a [ .b [ .c ]][ @dblink ]

Each of the items a, b, c, and DBlink is a simple SQL name. A simple SQL name
is either a common SQL name or an exotic SQL name and an exotic SQL name
must be surrounded with double quote characters. Notice that this is just a
prescription for the syntax; a.b might denote the element b in the package a (as in
Proc."Do it") or it might denote the object b owner by the user a (as in
"O’Brien".Pkg).
The ability to compose a qualified SQL name by assembling simple SQL names
with the punctuation characters . and @ delivers a powerful generality in the
context of a SQL statement. However, we have argued in this paper that the
power of expression of SQL should be fully available only to the author of a
subprogram51 that executes it. This logic applies no less to the assembly of a
qualified SQL name. This is the reason that we insist that the . and @
punctuation characters are part of the SQL syntax template and that the simple
SQL name placeholder be exactly that: this placeholder must not be replaced by
a qualified SQL name. This rule will be restated formally in the section
“Safe SQL statement text” on page 36.
It might be argued that the usability of a subprogram which allows the caller to
choose the object it operates on is improved if the familiar syntax of the qualified
SQL name can be expressed in a single actual argument. There is a simple way to
accommodate such a requirement. The Oracle-supplied procedure
DBMS_Utility.Name_Tokenize() decomposes a qualified SQL name into its simple
SQL names. It takes no account of semantics: the denoted object need not exist;

50. The rules for composing a common SQL name and an exotic SQL name are given in the
SQL Language Reference book; however, as was mentioned in the section “Example 2: user-
supplied table name” on page 15, these terms of art were invented for this paper.
51. To be precise, we should use the term top level PL/SQL block, as is defined in the section
“Static text” on page 35, rather than subprogram.

26 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 and it therefore gives no information about whether a.b is the element b in the
package a or the object b owner by the user a. Code_23 shows how to invoke it.
-- Code_23
procedure p(Qualified_SQL_Name in varchar2) is
a varchar2(32767);
b varchar2(32767);
c varchar2(32767);
DBlink varchar2(32767);
Dummy pls_integer;
begin
DBMS_Utility.Name_Tokenize(
Qualified_SQL_Name,
a,
b,
c,
DBlink,
Dummy);

The utility’s first formal parameter (for which the actual Qualified_SQL_Name is
used) has the mode in; and the remaining formal parameters (for which the
actuals a, b, c, DBlink, and Dummy are used) have the mode out. The returned
value of a will never be null. The values of b, c, and DBlink might be null; but c will
never be not null unless b is also not null. Dummy carries no useful information: it is
always equal to Length(Qualified_SQL_Name).
An arguably better API design for a subprogram which allows the caller to
choose the object it operates on is to provide an explicit formal parameter for
each simple SQL name that will be used to compose the qualified SQL name.
However, the choice of approach is unimportant with respect to this paper’s
focus.
Notice that by allowing the replacement of a simple SQL name placeholder in a
SQL syntax template only with a simple SQL name, we force the programmer to
take a deliberate decision in the design about, for example, whether an object in a
remote database may be referenced. This self-evidently improves the safety of
the design52.
These, then, are the rules for composing a safe simple SQL name from a
PL/SQL numeric value:
• Design the API so that the caller provides the putative name, in the PL/SQL
numeric value, using exactly the syntax that would be used in a SQL statement: a
common SQL name may be presented without enclosing double quote
characters (when it will then be treated case-insensitively); an exotic SQL name
must be presented with enclosing double quote characters.
• Ensure the safety of the name with DBMS_Assert.Simple_Sql_Name().
The DBMS_Assert.Simple_Sql_Name() is a pure asserter: either the output is
identical to the input; or the exception DBMS_Assert.Invalid_Sql_Name is raised.

52. The have been some discussions in public internet forums about how a program that uses the
DBMS_Assert functions inappropriately can be vulnerable to injection. See, for example, the
paper Bypassing Oracle DBMS_Assert by Alexander Kornbrust and discussions about it. This
paper insists that the safety of object identification be asserted always and only with
DBMS_Assert.Simple_Sql_Name(). Of course, other functions from the package may be used,
before the final check, for their ordinary utility value.

27 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 If the input is not surrounded by double quote characters, it is taken to be a
common SQL name; here, the exception is raised if it violates the rules for such
a name. If the input is surrounded by double quote characters, it is taken to be an
exotic SQL name; here, the exception is raised only if the name contains a
singleton double quote character or a consecutive run of an odd number of such
characters. However, a name that contains only properly doubled double quote
characters will fail when it is used53.
Finally, in this section, we should note that the DBMS_Assert package has
functions to establish whether a putative first class object or a putative schema
actually exists. While these might be convenient utilities for use in some very
well-controlled scenarios, they have no value with respect to immunizing against
the risk of SQL injection. Oracle Database is famously a multiuser, multiaccess
environment. Therefore, a check for object existence might report a fact which
no longer holds true at the moment an action is taken that should supposedly
depend of the outcome of that check.

53. The attempt will cause ORA-03001: unimplemented feature.

28 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 RULES FOR COST-EFFECTIVE, GUARANTEED PREVENTION OF
SQL INJECTION

Like any rules of best practice, the following do not prescribe the only way to
achieve the goal. These rules are recommended because they are simple to state,
easy to follow, easy to audit in a manual code review54, and — most importantly
— because they guarantee the objective: they make PL/SQL database code that
adheres to them proof against SQL injection. The rules necessarily limit
freedom. But we argue that the limited freedom that remains is sufficient to
allow all reasonable application requirements to be supported.

Expose the database to clients only via a PL/SQL API

Briefly, the paradigm is to establish a database user as the only one to which a
client may connect55. This user may own only private synonyms; and these may
denote only PL/SQL units. Necessarily, by application of this rule, the denoted
PL/SQL units are owned by users as whom the client may not connect. The
Execute privilege on exactly and only these denoted PL/SQL units is granted to
the user as whom the client may connect.
This approach formally defines the API — subject to the caveat that objects that
are accessible via privileges granted to public are therefore part of the API. (If
appropriate, the scheme can be extended to several such users, for access by
different kinds of client end-user, where each user has appropriate privileges.)
The API-defining PL/SQL units may be definer’s rights or invoker’s rights
according to their purpose.
This paradigm locates the responsibility to prevent SQL injection where it
belongs: in the subsystem of the overall application stack that executes the SQL.
And it offers the only approach whose safety can be proved. The paradigm is a
natural extension of the thinking that places all code that is responsible for the
enforcement of data integrity in the database56.
Rule_4
Expose the database to clients Expose the database to clients only via a PL/SQL API. Carefully control
only via a PL/SQL API. privileges so that the client has no direct access to the application’s objects of
other kinds — especially tables and views.

Use compile-time-fixed SQL statement text unless you cannot

If the complete text of a SQL statement, and not just the SQL syntax template,
is fixed at compile time, then it takes no effort to prove that the call site that
issues that statement is proof against SQL injection: because the statement text is
fixed, then so, in turn, is the SQL syntax template. We don’t even need to discuss

54. They also allow the possibility, in a future release of Oracle Database, of at least partial
mechanical auditing.
55. The passwords of all the other users are closely guarded and are not revealed to engineers who
implement client-side code.
56. Notice that, as a side-benefit, the paradigm liberates the designer from requiring to use triggers
to implement data integrity logic. When all data changes are made through PL/SQL
subprograms, these can directly implement the data integrity logic.

29 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 whether dynamic text57 has been made safe, and the auditing task is therefore
trivial (and correspondingly cheap).
Notice that while using embedded SQL guarantees a compile-time-fixed SQL
statement text, it is not the only way to guarantee this. Code_24 shows an obvious
alternative.
-- Code_24
declare
Stmt constant varchar2(80) :=
'alter session
set NLS_Date_Format = ''AD yyyy-mm-dd hh24:mi:ss''';
begin
execute immediate Stmt;
...
end;

Here, we imagine that the Functional Specification document for the application
insists that a report which lists some datetime values do so in a particular way58.
The use of constant with an assignment statement that uses only PL/SQL
static varchar2 expressions59 provably fixes the SQL statement text at compile
time. The auditor need not study whatever expanse of code lies between the
declaration of Stmt and its use as the argument of execute immediate; the PL/SQL
compiler will refuse to compile the unit if it includes code that would change
Stmt.
It is important, therefore, to separate the discussion of the method for executing
the SQL at a particular call site from the discussion of the SQL syntax template
that will be executed at that call site. Embedded SQL supports only these kinds
of statement: select, insert, update, delete, merge, lock table, commit, rollback, savepoint, and
set transaction.
For all other kinds of statement, one of PL/SQL’s methods for dynamic SQL
must be used. Moreover, for these, the execute immediate statement is sufficient in
almost every case60.
There are plausible use cases that require SQL statements of the kinds that
embedded SQL does not support in ordinary application code. But these occur
very much more rarely for ordinary application code than for system code.
Rule_5
Insist that the Design Specification When the Design Specification document for ordinary application code
document for ordinary application proposes to use anything other than embedded SQL, insist on examining the
code defends any propsal to execute
SQL using any method other than
Embedded SQL.
57. The notion of dynamic text is defined formally in the section “Dynamic text” on page 36.
58. For example, the report could be generated in a marked-up format for the specific purpose of
import into a spreadsheet where the convention for representing dates has already been
established.
59. The term PL/SQL static varchar2 expression is defined in the PL/SQL Language Reference
book and its significance is discussed in the section “Static text” on page 35.
60. The cases where execute immediate is not sufficient are very rare indeed, and all need special
consideration with respect to security. One example is when DDL statement must be executed
in a remote database. In such cases, the DBMS_Sql API is needed.

30 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 rationale very carefully when the document is reviewed. The design may well
be defensible. But this defense must be made explicitly.
Notice that the converse is true: when a statement of one of the kinds that
Embedded SQL does support is to be executed, and when the statement text can
be fixed at compile time, then dynamic SQL is never needed. Therefore,
embedded SQL should always be used for these cases. We understand now why
the example in Code_9 is a violation of the rule stated in this section.
There is, however, a practical consideration which forces us to moderate the
statement of this rule. Sometimes, while the universe of possible
SQL syntax templates that may be issued from a particular call site is fixed at
compile time, the number of these is so great (because they are generated
programatically, albeit by assembling components all of which are fixed at
compile time) that they cannot be set up at compile time as embedded SQL
statements. This use case was discussed in the section “Distinguishing between a
static SQL syntax template and a dynamic SQL syntax template” on page 7. And it is
illustrated in the section “Query by example form” on page 42.
Rule_6
Use compile-time-fixed SQL Use compile-time-fixed SQL statement text unless you cannot. Use
statement text unless you cannot. embedded SQL when the SQL statement is of a kind that this supports.
Use Embedded SQL or
Otherwise, use execute immediate with single PL/SQL constant argument
execute immediate with a PL/SQL
composed using only PL/SQL static varchar2 expressions. When you conclude
static varchar2 expression.
that you cannot, review the Functional Specification document and the Design
Specification document carefully with colleagues and then explain in the latter
exactly why compile-time-fixed SQL statement text cannot be used.

Use a static SQL syntax template for run-time-created SQL statement

text unless you cannot
There are only very few requirements scenarios that might be used to justify a
deliberate violation of Rule_6. Each must be examined very carefully when the
Design Specification document is reviewed.
Replacement of a value placeholder in a SQL syntax template
This requirement arises for two different reasons.
• There is no SQL syntax for binding to a regular placeholder for the required
parameterization.
• Pre Oracle Database 11g, a literal value is required to encourage an optimal
execution plan61.
There are three reasons why binding to a placeholder might not be supported.
• While the SQL statement is of the kind that embedded SQL supports, there is
a requirement to parameterize a value like that which determines the timeout
period in a select... for update statement62. SQL does not support the use of a
regular placeholder for this purpose.
• There is a requirement to execute a SQL statement of the kind that requires
Dynamic SQL and to parameterize it in some way63. An example is provided

31 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 by an extension to the requirement that led to the implementation shown in
Code_24. Suppose, now, that NLS_Date_Format must be set at run-time in
response to conditions that are first discovered then.
• The SQL statement (or fragment, like a where clause) is to be executed by one of
the subprograms listed in Appendix C: Additional Oracle-supplied subprograms that
implement dynamic SQL on page 59 and this subprogram does not support
binding to placeholders.
The safety of the dynamic text that is used to replace a value placeholder must be
ensured according to the principles established in the section “Ensuring the safety of
a SQL literal” on page 20. This is stated compactly in Rule_10 on page 36.
Consider the implied Functional Specification document that led to the
injectable implementation shown in Code_12 on page 18. Because the from list
specifies exactly one item, it is likely that the requirement is to let the user
mention some subset of the item’s columns and, for each, to express a
comparison with a value. And it is likely that the combination of the predicates
can be constrained to either uniformly mutual or (any of the predicates is
satisfied) or mutual and (all of the predicates is satisfied). With about a dozen
columns, the number of distinct SQL syntax templates is several thousand, so it
is easy to see why it is tempting to let the user supply the where clause as string.
However, there is no general way to define an algorithm that checks if such a
string is indeed one of the several thousand allowed templates. There is, then,
only one acceptable approach; this discussion is taken up in the section “Don’t
confuse the need to use a dynamic SQL syntax template with the need for dynamic text” on
page 34.
Rule_7
Insist on a justification for a design A Design Specification document that proposes to replace a value placeholder
that proposes to replace a in a SQL syntax template (whether this is a static SQL syntax template or a
value placeholder in a
SQL syntax template.

61. This use case used to occur in the implementation of an application to expose the information
in a datawarehouse. In such cases, the tables are enormous and the queries are complex.
Moreover, the number of concurrent users is relatively low. The classical wisdom (bind to
placeholders rather than using literals to reduce the frequency of the so-called SQL hard parse)
used not to hold. Parse time can be tiny in comparison to execution time, and contention is
unimportant. Here, the use of a literal in a predicate, allowing more specific advantage to be
taken of statistics, used sometimes to be critical for getting acceptably short execution times.

However, Oracle Database 11g brings enhanced bind peeking. Later releases brought more
adaptive methods for calculating the execution plan. There is now never any need to encode
literal values into a SQL statement in order to get the optimal execution plan. Nor is there any
other reason to prefer literals.
62. Code to implement this use case was shown in Code_5 on page 7.
63. It so happens that none of the kinds of SQL statement that embedded SQL does not support
allows the use of a regular placeholder.
However it is relatively rare for ordinary application code to need to execute the kind of SQL
that isn’t supported by embedded SQL. This need occurs only within special kinds of program.
The higher level Functional Specification document should be very carefully reviewed to
ensure that the proposed approach is essential.

32 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 dynamic SQL syntax template) should be regarded with extreme suspicion. It
must present a convincing argument for this approach before it is signed off.

Replacement of a simple SQL name placeholder in a SQL syntax template

The general characterization of the use case is that the application needs to
operate on an object (which is typically a table) whose shape and purpose is given
in the Design Specification document but whose identity is not known until run
time.
While the safe approach for this scenario has been prescribed in the section
“Ensuring the safety of a simple SQL name” on page 26, it is still useful, as each new
Design Specification document is written, to determine if an alternative
approach is possible which uses only compile-time-fixed SQL statement text,
and (for accessing a table) can therefore be implemented with embedded SQL.
Consider the sort of utility that is aimed manipulating data in one of a set of
schemas where each such schema is populated by objects with the same names
and purposes. A naïve approach would encode the value returned by the
Sys_Context() function for Current_Schema in the Unserenv namespace into a
qualified SQL name for each object to be accessed and would execute these
statements with Dynamic SQL. The ideal solution, though, is to encapsulate the
SQL statements that access the objects in an Invoker’s Rights unit. The access
can then be implemented ideally with embedded SQL which uses unqualified
names. The code can be pointed at the intended objects by changing the
Current_Schema — either programatically or by simply letting it follow the
Current_User.
Sometimes an application needs to handle a very large quantity of data (more
than can safely be held in a PL/SQL collection) but does not need to retain it
beyond the lifetime of a session. Historically, applications used a pool of
nominally temporary tables together with a management system to allocate, and
later reclaim, such a table for use by a particular session. This, of course, meant
that the table name was not known until run time. Oracle8i Database introduced
the global temporary table. It is perfect for this purpose because it can be
accessed using embedded SQL but no session sees another session’s data.
However, there are some compelling use cases that non-negotiably require the
replacement of a simple SQL name placeholder in a SQL syntax template with a
simple SQL name. They tend to arise in the design of what is loosely referred to
as system software. A convincing example is given by an implementation of a
domain index that uses the ODCI framework. Each different index will use a set
of tables whose names are typically derived from that of that index. The ODCI
client code will need to derive such names at run time when it is called, by the
Oracle system, with the name of the current index of interest.64
Rule_8
Insist on a justification for a design A Design Specification document that proposes to replace a simple
that proposes to replace a simple SQL name placeholder in a SQL syntax template (whether this is a static
SQL name placeholder in a
SQL syntax template or a dynamic SQL syntax template) should be regarded
SQL syntax template.

64. Oracle Text works this way.

33 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 with some suspicion. It must present a convincing argument for this approach
before it is signed off.

Don’t confuse the need to use a dynamic SQL syntax template with the
need for dynamic text
The kind of Functional Specification document that cannot be implemented
using a static SQL syntax template has already been briefly discussed65 and the
concept for the safe implementation has been illustrated with code66. There, the
requirement was to handle a select list whose composition is not known until run
time. The requirement often goes hand-in-hand with one optionally to specify a
restriction criterion for each column by supplying both the value and the
comparison operator (exact equals, like, less than, and so on). There is also often a
requirement to specify which columns to use, in which order, and with ascending
or descending for each, to sort the results.
The requirement to handle a where clause whose composition is not known until
run time can lead junior programmers to decide that binding to placeholders
when the requirement is not known at compile time is not feasible67. More
mature programmers realize that the DBMS_Sql API provides exactly the
primitives that are needed to implement this. The approach is illustrated in the
section “Query by example form” on page 42.
The coding effort is certainly nontrivial. But the size of that effort is never
enough to justify abandoning binding in favor of encoding literals directly into
the SQL statement.
In other words, the requirement to encode literals directly into the
SQL statement (in this use case only for the reason of improving the execution
plan) is orthogonal to the requirement to implement binding to a set of
placeholders whose composition emerges first at run time.
Rule_9
Don’t confuse the need to use a Don’t confuse the requirement to bind to a set of placeholders whose
dynamic SQL syntax template with composition emerges first at run time, which is fully supported by the
the need to replace
DBMS_Sql API, with a requirement to use directly encoded literals in pursuit
value placeholders in the template
of optimal query execution performance.

Formal sufficient prescription for guaranteed safety

The foregoing exposition has aimed to show how a PL/SQL application that is
vulnerable to SQL injection is also ordinarily buggy. It focussed, therefore, on
how to write bug-free code that has the pleasant bonus of being proof against
SQL injection. This section takes the auditor’s viewpoint whose aim is to certify
proof against SQL injection by inspecting extant code.
First, we need to establish some definitions.

65. See the section “Definition of dynamic SQL syntax template” on page 8.
66. See Code_6 on page 8, Code_7 on page 10, and Code_8 on page 10.
67. You need only to try to write the code to do this using execute immediate to realize that, because
its using clause is fixed at compile time, the task is impossible.

34 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 Static text
Static text is either
• a PL/SQL static varchar2 expression as defined in the
68,
PL/SQL Language Reference book or
• an expression formed by an arbitrary concatenation of static text items, or

• the value of a local variable69 that has been visibly assigned with static text.
This definition is intentionally recursive. A variable can be assigned some
static text and then that variable (strictly speaking, its value70) can be used to
build another yet larger static text. The tower can be built as high as one likes;
but its foundation must consist only of PL/SQL static varchar2 expressions.
Notice that the concatenation may be controlled by tests whose outcome is not
known until run time. But it must be self-evident, following trivial human
inspection, that every possible concatenation will result only in static text
according to the rules stated in the preceding three bullets; this is what we mean
by visibly assigned. (One might say that you must be able to see the foundation
when you are standing on the top of a static text tower.)
Code_25 shows an extract from the declarations in function f() shown in
Code_26 on page 38.
-- Code_25
Tab_1 constant varchar2(32767) := 'Tab_1';
Tab_2 constant varchar2(32767) := 'Tab_2';
...
Tab constant varchar2(32767) :=
case b
when true then Tab_1
else Tab_2
end;

68. This is the list:

— a simple literal, for example, 'abcdef'
— a concatenation of simple literals, for example, 'abc'||'def'
— the literal null
— To_Char(x), where x is a pls_integer static expression
— To_Char(x f, n) where x is a pls_integer static expression
and f and n are PL/SQL static varchar2 expressions
— x||y where each of x and y is either a PL/SQL static varchar2 expression
or a pls_integer static expression.
The list is extended to include a constant declared in the specification of a package that has been
assigned with PL/SQL static varchar2 expressions.
Of course, pls_integer static expression is also defined in the same book.
The value of a PL/SQL static varchar2 expressions is known, and therefore fixed, at compile
time.
69. A local variable is one that is declared within the present top level PL/SQL block A
top level PL/SQL block is one of the following: a schema-level function; a schema-level
procedure; a function or procedure defined at top level within a package body or a type body;
a package’s initialization block; or the implementation of a trigger. Thus a variable that is
declared at top level in a package or package body is, by definition, not a local variable.

35 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 Notice that b is a formal parameter and so the value of Tab will not be known
until run time and will be, in general, different for each invocation of f().
Nevertheless, Tab is a static text according to the definition given above.
Dynamic text
Dynamic text is any text that is not static text. Obvious examples are formal
parameters, variables declared at top level in a package without the constant
keyword, and ordinary variables that are assigned by executing a SQL statement
or as the actual argument for the Buffer formal parameter to Utl_File.Get_Line().
Safe dynamic text
Safe dynamic text is the output of either
• DBMS_Assert.Enquote_Literal(), or
• To_Char(x f, 'NLS_Numeric_Characters = ''.,''') where x is a variable of a
numeric datatype and f is an the explicit format model 'TM', or
• DBMS_Assert.Simple_Sql_Name().
Hereinafter, the shorthand To_Char(x f, n) will be used to denote
To_Char(x f, 'NLS_Numeric_Characters = ''.,'''). The background for this rule was
established in the section “Ensuring the safety of a SQL numeric literal” on page 24.
Notice that it follows from the definition of dynamic text that the safety of
safe dynamic text must be established within the present
top level PL/SQL block.
Safe SQL statement text
Safe SQL statement text is an arbitrary concatenation of static text and
safe dynamic text. (A careful definition would require similar wording to that
used in the definition of static text. Local variables may be used for intermediate
results or for the final result; and it must be self-evident that every possible
concatenation will result only in safe SQL statement text.)
We can now define a necessary prescription for guaranteed safety:
Rule_10
Dynamic SQL may execute only a When a SQL statement represented by a PL/SQL text expression is executed
concatentation of static text and using one of PL/SQL’s APIs for dynamic SQL, then the expression must be
safe dynamic text. Safe dynamic text
safe SQL statement text. Safe SQL statement text is a concatenation of
is the output of one of exactly three
static text and safe dynamic text. Static text is composed only of PL/SQL
Oracle-supplied functions:
Simple_Sql_Name(),
static varchar2 expressions. Dynamic text is anything that isn’t static text.
Enquote_Literal(), and Safe dynamic text is the output of one of exactly three Oracle-supplied
To_Char(x f, n). functions: DBMS_Assert.Simple_Sql_Name(), DBMS_Assert.Enquote_Literal(),
and To_Char(x f, n).

70. The distinction between a variable and its (current) value becomes uninteresting when the
declaration of the variable uses the constant keyword. It is always possible, with appropriate use
of nested block statements, to declare all the variables that are used for the composition of
static text this way. This practice is strongly encouraged because it makes the task of the human
auditor very much easier.

36 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 Notice that the definition of safe SQL statement text pays no attention to the
semantics of the resulting SQL statement and is, therefore, not sufficient.
Nevertheless, this rule is useful because, as it is easy to imagine, the PL/SQL
compiler, in a future version of Oracle Database, could be enhanced to detect a
violation of Rule_10. Application code that violates the rule is definitely suspect
and must be studied.
We must also guarantee that the composed SQL syntax template is only one of
those that the programmer intended. We need to add Rule_11 to Rule_10 to give
the sufficient prescription for guaranteed safety.
Rule_11
Use Simple_Sql_Name(), to ensure The safe dynamic text produced by DBMS_Assert.Enquote_Literal() or by
the safety of a SQL name. Use To_Char(x f, n) should be used only at a spot within the SQL statement where a
Enquote_Literal() to ensure the
SQL literal is intended; and the safe dynamic text produced by
safety of a string or datetime literal.
DBMS_Assert.Simple_Sql_Name() should be used only at a spot within the
Use To_Char(x f, n) to ensure the
safety of a numeric literal.
SQL statement where a simple SQL name is intended.
It is very much harder to imagine how adherence to the stricter Rule_11 could be
mechanically policed. It would require a run-time check, for every single dynamic
execution of SQL, by a subsystem that implemented complete knowledge of
SQL syntax. Anyway, both Rule_10 and Rule_11 must be policed by ordinary
human code review.

37 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 Code_26 shows an example71 that honors Rule_10 and Rule_11. Function f() is
guaranteed always to compose a SQL statement that conforms to the
SQL syntax template shown in Template_8.
-- Template_8
select PK from &&1 where VC2 = &2
-- Code_26
function f(b in boolean, VC2 in varchar2) return number is
Quote constant varchar2(1) := '''';
Doubled_Quote constant varchar2(2) := Quote||Quote;

Start_ constant varchar2(32767) := 'select PK from ';

Tab_1 constant varchar2(32767) := 'Tab_1';
Tab_2 constant varchar2(32767) := 'Tab_2';
Where_ constant varchar2(32767) := ' where VC2 = ';

Tab constant varchar2(32767) :=

case b
when true then Tab_1
else Tab_2
end;

SQL_Text_Literal constant varchar2(32767) :=

Quote||Replace(VC2, Quote, Doubled_Quote)||Quote;

Stmt constant varchar2(32767) :=

Start_||
Tab||
Where_||
Sys.DBMS_Assert.Enquote_Literal(SQL_Text_Literal);

PK number;
begin
execute immediate Stmt into PK;
return PK;
end f;

When f() is invoked with actual arguments with values true for b and O’Brien for
VC2, then the safe SQL statement text shown in Value_17 is composed.
-- Value_17
select PK from Tab_1 where VC2 = 'O''Brien'

Establish the safety of run-time-created SQL statement text in the code

that immediately precedes its execution

The definition established for safe dynamic text72 insists that it be composed in
the same top level PL/SQL block that executes the SQL statement, into which it
is concatenated, using dynamic SQL. This is a sufficient condition to allow the
future possibility of a mechanical check. However, the task of the human auditor
is made very much easier if all calls to DBMS_Assert.Simple_Sql_Name(),
DBMS_Assert.Enquote_Literal(), or to To_Char(x f, n) are made in code that
composes the SQL statement immediately before the PL/SQL statement that
executes it.

71. Because this example replaces a value placeholder in a SQL syntax template, we must assume
that this approach was deliberately chosen in preference to binding to a regular placeholder in
order to increase the probability of a good execution plan as was discussed in the section
“Replacement of a value placeholder in a SQL syntax template” on page 31.
72. See the section “Safe dynamic text” on page 36.

38 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 Code_26 honors this principle.
Rule_12
Keep the code that invokes Make the job of the auditor easy by establishing the safety of run-time-created
Simple_Sql_Name(), SQL statement text in the code that immediately precedes the PL/SQL
Enquote_Literal(), or To_Char(x f, n)
statement that executes it.
very close to the code that executes
the SQL statement that these calls
make safe.

39 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 SCENARIOS

Each of the following requirements scenarios can be implemented very well

using embedded SQL. Yet (incredibly, in the case of the first scenario) we have
seen real production code where dynamic SQL has been used unsafely because
the author didn’t know how to program the embedded SQL solution.

Make a like predicate by adding leading and trailing % characters

Code_27 shows an embedded SQL statement that implements a like predicate.
-- Code_27
select t.PK, t.c1 bulk
collect into Results
from t where c1 like '%'||x||'%';

It’s impossible to speculate on the origin of the strange myth that says that this
needs run-time-created SQL statement text.
It might help to see the SQL statement that the PL/SQL compiler constructs
from this source code. The simplest way to observe this is, on one’s own private
development database, to encapsulate the code shown in Code_27 in a procedure,
to execute it, and to query the v$Sql view using a restriction that is selective
enough to filter out irrelevant results73. Code_28 shows such a query.
-- Code_28
select Sql_Text
from v$Sql
where Lower(Sql_Text) not like '%v$sql%'
and Lower(Sql_Text) like 'select%t.pk%t.c1%'

This is the output of Code_28:

SELECT T.PK, T.C1 FROM T WHERE C1 LIKE '%'||:B1 ||'%'

Notice that the regular placeholder :B1 has been established in the position
where the variable x was used in the embedded SQL statement. The PL/SQL
compiler generates appropriate code to do the binding.
As the companion paper, Doing SQL from PL/SQL: Best and Worst Practices,
explains at some length, PL/SQL’s embedded SQL and PL/SQL’s dynamic SQL
are both processed the same way at run time: the text of the SQL statement is
submitted by the PL/SQL run-time system to the SQL subsystem for execution.
The difference is that the PL/SQL compiler generates the text of the
SQL statement from an embedded SQL statement at compile time and stores it
with the unit’s compiled code; but the SQL statement executed by dynamic SQL
is composed at run time when the unit executes. Of course, as this paper has
advocated, that composition might be no more demanding than the initialization
of a PL/SQL constant with a PL/SQL static varchar2 expression74.
A programmer who thought that embedded SQL could not support a like
predicate must have thought, if pushed to explain his thoughts, that the
expression '%'||:B1||'%' was not legal in SQL; it most certainly is.

73. This is easy to arrange by inventing an unusual name for the test table. An alternative is simply
to flush the shared pool before executing the test procedure.
74. It turns out that the optimizing PL/SQL compiler can do this at compile time.

40 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 In list with number of elements not known until run-time
This use case is described in the companion paper, Doing SQL from PL/SQL: Best
and Worst Practices. For convenience, the account is reproduced here — slightly
reworded to suit the needs of this paper’s pedagogy.
If, for some relatively improbable reasons, you need a query whose where clause
uses an in list with, say, exactly five items, then you can write the embedded SQL
statement without noticing the challenge that will present itself later when the
requirements change to reflect the more probable use case. Code_29 shows this
unlikely statement.
-- Code_29
select a.PK, a.v1
bulk collect into b.Results
from t a
where a.v1 in (b.p1, b.p2, b.p3, b.p4, b.p5);

Code_30 expresses the intention of the far more likely statement.

-- Code_30
select a.PK, a.v1
bulk collect into b.Results
from t a
where a.v1 in (b.ps(1), b.ps(2), b.ps(3),
b.ps(4), b.ps(5), b.ps(6),
b.ps(7), b.ps(8), b.ps(9),
... );

The problem is immediately apparent: we cannot bear to write an explicit

reference to each element in a collection using a literal for the index value — and
even if we could, the text would become unmanageably voluminous. Rather, we
need a syntax that expresses the notion “all the elements in this collection,
however many that might be”. Such a syntax exists and is supported in
embedded SQL; Code_31 shows it.
-- Code_31
ps Strings_t;
begin
select a.PK, a.v1
bulk collect into b.Results
from t a
where a.v1 in (select Column_Value
from table(b.ps));

However, this seems to be relatively little known, possibly because it uses the
table operator75. The datatype of ps must be declared at schema level. Code_32
shows the SQL*Plus script that creates it.
-- Code_32
create type Strings_t is table of varchar2(30)
/

It is not uncommon for programmers who don’t know about the table operator
to satisfy the functionality requirement by building the text of the SQL statement
at run time. At best, this implementation is cumbersome and inefficient; and at
worst, it will be vulnerable to SQL injection.

75. The use of the table operator is explained in the section Manipulating Individual Collection Elements
with SQL in the Object-Relational Developer’s Guide. The PL/SQL Language Reference
mentions it, with no explanation, only in the PL/SQL Language Elements section.

41 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 Query by example form
This scenario complements the one discussed in the section “Definition of dynamic
SQL syntax template” on page 8 76. There, the Functional Specification document
required a customizable report format for a fixed restriction condition (primary
key identity). In this section, the Functional Specification document requires a
fixed report format but a customizable restriction condition. This leads to the
need for a design which, if it is to avoid replacing value placeholders in the
dynamic SQL syntax template with SQL literals, needs to bind to placeholders,
the number of which, and the datatypes of what they stand for, are unknown
until run time.
Neither the requirement to handle a select list whose composition is not known
until run time nor the requirement to handle an order by clause whose
composition is not known until run time leads to design choices that impact the
discussion of exposure to the risk of SQL injection. Therefore, to keep the code
illustrations as short as possible, this section shows only how to code when the
where clause is not known at compile time.
This paper needs to do only two things: to show how to build a SQL statement
as an instance of a dynamic SQL syntax template but using only static text; and
to show how to implement the binding that this implies77.
consider the procedure x.p() as shown in Code_33.
-- Code_33
package x is
type t1 is Record(
Val varchar2(4000),
Exact boolean := false);
type t2 is varray(20) of t1;
procedure p(Cols in t2);
end x;

Its purpose is to list the values of the primary key, PK, for rows that satisfy the
customizable where clause. Just as with Code_6 on page 8, the author of code that
will call x.p() knows the purpose of My Table78, and the names and significance of
all its columns. All this, and especially the names, is fixed in the Design
Specification document. In particular, the programmer knows the order in which
the columns are listed in its external documentation. The nth element of the in
formal parameter Columns determines if, and how, the nth column of My Table is
to be used in the where clause. When Val is not null, the nth column is to be

76. The code that implements this scenario is shown in Code_6 on page 8, Code_7 on page 10, and
Code_8 on page 10.
77. The need to compose the order by clause at run time impacts only the composition of the
SQL statement; it has no impact on binding nor on how the results are fetched. The need to
compose the select list at run time, while it has no impact on binding, does potentially affect how
the results are fetched. The approach shown in Code_7 on page 10 sidesteps this by
concatenating all the required columns into a single text item. Sometimes, though, each select list
item must be fetched into its own variable, and the datatypes of the columns have to be taken
into account. The DBMS_Sql API is well able to support this; but the techniques have no
consequence for the subject of this paper.
78. The exotic name My Table was chosen deliberately to make a teaching point in this example.

42 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 included; and, in this case, Exact specifies whether an equality comparison or a
like comparison is to be used.
The implementation of p() in the body of package x is shown in Code_34. The
Where_Clause() inner function composes the where clause.
-- Code_34
procedure p(Cols in t2) is
type t3 is varray(200) of "My Table".PK%type;
Results t3;
type cn is varray(20) of varchar2(30);
Col_Names constant cn := cn('c1', 'c2', 'c3', 'c4');

function Where_Clause return varchar2;

Stmt constant varchar2(32767) :=

'select PK from "My Table"'
|| Where_Clause()
|| ' order by PK';
function Where_Clause return varchar2 is ... end Where_Clause;

begin
DBMS_Output.Put_Line(Stmt);

declare
nc integer := DBMS_Sql.Open_Cursor(Security_Level=>2);
rc Sys_Refcursor;
Dummy number;
begin
DBMS_Sql.Parse(nc, Stmt, DBMS_Sql.Native);

for j in 1..Cols.Count() loop

if Cols(j).Val is not null then
DBMS_Sql.Bind_Variable(nc, ':b'||j, Cols(j).Val);
end if;
end loop;

Dummy := DBMS_Sql.Execute(nc);

rc := DBMS_Sql.To_Refcursor(nc);
fetch rc bulk collect into Results;
for j in 1..Results.Count() loop
DBMS_Output.Put_Line(Results(j));
end loop;
close rc;
end;
end p;

Notice that, following this paper’s recommendation in Rule_1 on page 10, Stmt is
declared using the constant keyword and that this means that it must be initialized
as part of the declaration. Notice too that the assignment to Stmt starts like this:
Stmt constant varchar2(32767) :=
'select PK from "My Table"' || ...

The exotic SQL name My Table is enclosed with double quote characters but its
safety is not ensured with DBMS_Assert.Simple_Sql_Name(). This is entirely safe.
The reason is that it occurs within a PL/SQL static varchar2 expression and
therefore its spelling is established with certainty by the PL/SQL source text79.
The same logic applies to the list ('c1', 'c2', 'c3', 'c4') that initializes the Col_Names

79. It would not be harmful to use DBMS_Assert.Simple_Sql_Name() in the initialization of Stmt.

However, because this is manifestly unnecessary, to do so might confuse subsequent
maintainers of the code.

43 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 static constant. The names of the columns are given in the Design Specification
document; and x.p() is specified to work against My Table with the columns it has.
Because the columns all have common SQL names, there is no need to surround
each with double quote characters80.
The key feature in this example, with respect to the binding challenge, is that
DBMS_Sql.Bind_Variable(), as a procedure, can be invoked in a loop as many
times as emerges at run time to be necessary and with actual arguments whose
values are known only then. The execute immediate statement and the open rc for
statement, because the elements to be bound are written in the source code of
the PL/SQL unit, support only a binding requirement that is fixed at compile
time.
The design capitalizes on the fact that the select list is known at compile time. It
uses DBMS_Sql.To_Refcursor()81 to move from the regime of the DBMS_Sql API
to that of native dynamic SQL to fetch the results in a single fetch... bulk collect
statement. Notice that the call to DBMS_Sql.Open_Cursor() uses an actual value 2
for the formal Security_Level82. Using this ensures that every operation attempted
using a cursor opened this way must be done by the same user that opened it
with the same, or greater, set of enabled roles83.

80. Writing the list as ('“C1”', '“C2”', '“C3”' '“C4”') would have no consequence for correctness of
for immunity to SQL injection. (Curly double quote characters are used to write this because
the straight double quote character that source code demands is, in the font used, visually
indistinguishable from two straight single quote characters.)
81. This function, and its counterpart DBMS_Sql.To_Cursor_Number(), were introduced in
Oracle Database 11g.
82. The new overload for DBMS_Sql.Open_Cursor(), that has the Security_Level formal, was
introduced in Oracle Database 11g.
83. This eliminates the a rather different kind of vulnerability named by David Litchfield as Cursor
Injection and described his article available on the internet here:
www.databasesecurity.com/dbsec/cursor-injection.pdf.

44 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 For completeness, the implementation of the Where_Clause() function is shown in
Code_35.
-- Code_35
function Where_Clause return varchar2 is
Clause varchar2(32767);
Seen_One boolean := false;
begin
for j in 1..Cols.Count() loop
if Cols(j).Val is not null then
Clause :=
Clause
|| case Seen_One
when true then ' and '
else ' where '
end
|| Col_Names(j)
|| case Cols(j).Exact
when true then ' = :b'||j
else ' like ''%''||:b'||j||'||''%'''
end;
Seen_One := true;
end if;
end loop;
return Clause;
end Where_Clause;

Notice that the design is very similar to that for the Column_List() function
shown in Code_8 on page 10 and that elements from which Clause is composed,
here too, are all PL/SQL static varchar2 expressions.
Again as with the Column_List() function, the set of the more than one million
distinct SQL statements that x.p() might execute is too numerous to allow each
one to be inspected. Rather, the programmer (and auditor) must reason what the
possible members are — and from that deduce what the possible
SQL syntax templates are. In this example, there are very many distinct
SQL syntax templates. Template_9 shows some.
-- Template_9
select PK from "My Table"
where c1 like '%'||:b1||'%' order by PK

select PK from "My Table"

where c1 = :b1 order by PK

select PK from "My Table"

where c1 like '%'||:b1||'%' and c3 = :b3 and c4 = :b4 order by PK

Because these templates use no value placeholders or simple

SQL name placeholders, there is no need for dynamic text and, tautologically, no
need to ensure its safety. This approach is self-evidently not vulnerable to
SQL injection. Should, in a bungled maintenance cycle, the name of the table or
any of its columns be changed without correspondingly updating the
implementation of x.p(), then it would simply fail at run time with an ordinary
semantic error.

Callback
The requirement is that a compiled PL/SQL unit must, without recompiling it,
be somehow instructed to call a subprogram whose identity isn’t known until run
time. Though the identity is unknown, the signature of formal parameters and
their meanings is known84. This scenario is relatively common in ISV

45 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 applications that allow the customer to supply code to perform a generically
characterized operation in a customer-specific fashion.
The challenge can be illustrated by specifying this simple shape for the to-be-
called subprogram:
function Callback(Input in integer) return integer

The “obvious” design, shown in Code_36, uses dynamic SQL.

-- Code_36
procedure p(Name in varchar2) is
...
Safe_Name constant varchar2(32767) :=
Sys.DBMS_Assert.Simple_Sql_Name(Name);
Stmt constant varchar2(32767) :=
'begin :x := '||Safe_Name||'(:n); end;';
begin
execute immediate Stmt using out x, in n;

Here, p is the unit that must not be recompiled. The actual for P.Name would be
obtained at run-time by, for example, selecting from a schema-level table85.
The approach does follow the rules that this paper sets out and is proof against
SQL injection. However, as we shall see, a better approach is available; it takes
advantage of dynamic polymorphism — sometimes known as virtual dispatch86.
This is one of the cornerstones of object-oriented programming and it goes
hand-in-hand with subtyping. The to-be-called subprogram is described first as
the not final member method of an abstract datatype. Code_37 shows this.
-- Code_37
type Hook authid Current_User is object(
Dummy number,
not instantiable member function Callback(
self in Hook,
Input in integer ) return integer
)
not final
not instantiable

Other syntactic elements are needed in the definition of the type.

• A type has at least one data attribute. It is declared as “Dummy number” to
emphasize that it has no significance for the implementation.
• The element not final allows a type to be created under Hook.
• The Callback() method is declared as not instantiable because no implementation
will be provided at this level in the type hierarchy. Hook will be the supertype of
a subtype which will describe the actual implementation. Because it has a not
instantiable method, Hook itself must also be declared as not instantiable.

84. This is analogous to the scenario where a table’s identity is unknown until run time but the
name and purpose of each column is.
85. The ISV’s instructions would explain that the customer should create a callback function with
the required shape and with any name. To simplify this example, we assume that this must be
done in a specified schema. Then the customer should insert a row into a specified table to
state what name he used for the callback.
86. The complete listing, from which the code in this section is extracted, is shown in Appendix D:
Self-contained code to illustrate implementing callback using dynamic polymorphism. on page 62.

46 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 • Every member method of a type has the implicit first formal self which by
default is of in out mode. It may alternatively be declared as in, but not as out. In
out is the default because member methods are typically used to set data
attributes. In out formals are passed by value unless the nocopy hint is used. For
this particular use case, where the attribute Dummy is never referenced, the best
practice is to establish self explicitly as in; formals with mode in are passed by
reference.
• The qualifier authid current_user — to establish the supertype with invoker’s
rights — is used in accordance with probable best practice. The choice
between invoker’s rights and definer’s rights would be taken in a real project in
the light of the overall requirements.
As mentioned, the implementation of the to-be-called subprogram is described
in a subtype of Hook :
-- Code_38
type My_Implementation under Hook(
overriding member function Callback(
...
input in integer) return integer
)

type body My_Implementation is

overriding member function Callback(
...
input in integer) return integer
is
begin
...
return ...;
end Callback;
end;

The element overriding member establishes the method as the actual implemen-
tation for the virtual declaration in the supertype.
The magic of the above construct is that a variable of the supertype Hook — say,
Obj — can be given a value at run-time which is an instance of the subtype
My_Implementation; and an invocation of what is apparently the supertype’s virtual
method — Obj.Callback() — will in that case denote the overriding match
declared in the subtype. This gives you what we could loosely call a dynamic case
statement whose legs can be determined after having compiled the statement,
and without recompiling it, simply by creating new compilation units. It’s as if
you could make p(), in Code_39, acquire an new leg to invoke Callback3() in its case

47 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 expression by merely by establishing Callback3() in a new compilation unit —
without needing to recompile p().
-- Code_39
function Callback1(input in integer) return integer is
begin return input*2; end Callback1;

function Callback2(input in integer) return integer is

begin return input*2; end Callback2;

procedure P(Which pls_integer) is

...
begin
x := case Which
when 1 then Callback1(n)
when 2 then Callback2(n)
...
end;
...

Code_40 shows how to implement intent of Code_36 using dynamic

polymorhism.
-- Code_40
procedure p(Obj in Hook) is
...
begin
...
x := Obj.Callback(n);
...

The actual for p.Obj would be obtained at run-time by, for example, selecting
from a schema-level table. This time, instead of using a varchar2 column to hold
the name of the implementation, the column would have datatype Hook and
would be populated with an instance of the subtype My_Implementation.
The dynamic SQL approach, shown in Code_36, uses a static
SQL syntax template and safely replaces a simple SQL name placeholder with a
simple SQL name. It would take the auditor some effort to prove that this code
was safe. The approach that uses dynamic polymorhism, shown in Code_40,
doesn’t use SQL at all and so is self-evidently safe.
As a bonus, the approach that uses dynamic polymorhism has two additional
advantages.
• Tests show that the dynamic polymorhism approach is on the order of 10
times faster than the dynamic SQL approach when, as is typical, the body of
the callback executes relatively quickly compared to the cost of the PL/SQL to
SQL to PL/SQL context switch.
• When the dynamic polymorhism approach is used, Callback()’s formal
parameters may have any PL/SQL datatype but when the dynamic SQL
approach is used, these datatypes are limited to only those that SQL
understands.

48 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 ANALYSIS AND HARDENING OF EXTANT CODE

This must be a mainly manual task. Notice that Oracle Database 12c Release 2
brought enhancements to PL/Scope to list every source code location where
dynamic SQL is is used. Ad hoc methods must be used to find those
top level PL/SQL blocks that do dynamic SQL.
It is easy to find all of the call sites where a procedural API is used87.
PL/Scope88 can be used to identify each call site with no risk of omission or of
false positives. However, there is no corresponding method to find the call sites
that issue native dynamic SQL89. There is no alternative except to search the
source code manually. Any PL/SQL unit with such a call sites will contain either
both words execute and immediate or both words open and for. Therefore, a
mechanical search in User_Source, seen from the viewpoint of the relevant users,
might be able to prune away some units.
Once the interesting units have been identified then it remains, for each relevant
top level PL/SQL block, to inspect the source code manually to determine if the
rules that this paper has advocated have been followed.
When violations are found, they will be where a SQL statement is executed using
a run-time-created SQL statement text composed with dynamic text that has not
been made safe according to the principles that this paper has described90.
In a fortunate case, the auditor will be able to establish quickly that a static
SQL syntax template was intended but that dynamic text is being used unsafely.
Here, the addition of calls to DBMS_Assert.Simple_Sql_Name(),
DBMS_Assert.Enquote_Literal(), or To_Char(x f, n) just before the PL/SQL
statement that executes the run-time-created SQL statement text will eliminate
the vulnerability.
In a less fortunate case, where the design relies on the assumption that SQL
keywords and operators will be supplied as dynamic text, the solution can only
be radical re-design. But, at least, the vulnerability will have been identified.

87. The relevant subprograms are listed in Appendix C: Additional Oracle-supplied subprograms that
implement dynamic SQL on page 59.
88. PL/Scope was introduced by Oracle Database 11g. It is described in the chapter
Using PL/Scope in the Oracle Database Development Guide.
89. Enhancement request 6913337 asks to provide a way to do this.
90. See especially the section “Ensuring the safety of a SQL literal or a simple SQL name” on page 20
and the section “Formal sufficient prescription for guaranteed safety” on page 34.

49 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 CONCLUSION

This paper has carefully defined, and named, some notions91 that are essential to
support a proper discussion of SQL injection and its avoidance92:
• common SQL name and exotic SQL name
• compile-time-fixed SQL statement text and run-time-created SQL statement text
• value placeholder, simple SQL name placeholder
• SQL syntax template, static SQL syntax template, and dynamic SQL syntax template
• static text, dynamic text, safe dynamic text, and safe SQL statement text
• top level PL/SQL block
And it has relied on careful use of some established notions and terminology:
• regular placeholder, simple SQL name, qualified SQL name
• embedded SQL, native dynamic SQL, and the DBMS_Sql API
• PL/SQL numeric value, PL/SQL datetime value, and PL/SQL text value
• SQL numeric literal, SQL datetime literal, and SQL text literal
• PL/SQL static varchar2 expression and PL/SQL static varchar2 constant
This allowed us to establish a simple, brief definition of the paper’s topic, from
its asserted perspective: SQL injection occurs at a particular call site in a
PL/SQL subprogram when a SQL statement us executed whose
SQL syntax template differs from what the subprogram’s author intended. It
also allowed us to see that SQL injection is possible only when run-time-created
SQL statement text is executed.
This gave the basis for understanding how most cost-effectively to ensure that
PL/SQL code is not vulnerable to SQL injection.
• Aim to use execute only compile-time-fixed SQL statement text using, as
appropriate, either embedded SQL, execute immediate, or open for with a PL/SQL
static varchar2 expression.
• Only after explaining carefully in the Design Specification document why
compile-time-fixed SQL statement text is insufficient to meet the requirements
in the Functional Specification document, aim to execute run-time-created
SQL statement text that is entirely static text.
• Only after explaining carefully why neither compile-time-fixed SQL statement
text nor run-time-created SQL statement text that is entirely static text is
sufficient, use run-time-created SQL statement text that is
safe SQL statement text.
• When composing this safe SQL statement text

91. See Appendix A: Definitions of new terms of art introduced by this paper on page 52.
92. This seems never to have been done before.

50 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 - ensure the safety of dynamic text that replaces a value placeholder in the
SQL syntax template with DBMS_Assert.Enquote_Literal() or with
To_Char(x f, n). For this design, explain explicitly why binding to a regular
placeholder is insufficient. Use the former for a SQL text literal or a SQL
datetime literal. Additionally, when composing a SQL datetime literal, use
To_Char() with a format model appropriate for the required precision and
encode To_Date() with the same model into the safe SQL statement text.
Use To_Char(x f, n) for a SQL numeric literal
- ensure the safety of dynamic text that replaces a simple
SQL name placeholder in the SQL syntax template with
DBMS_Assert.Simple_Sql_Name().
Clearly, then, the expense of implementing a design that is proof against
SQL injection increases with each successively more sophisticated approach.
However, this expense must never be used to justify taking shortcuts in the
implementation. Only by honoring the principles that this paper has set out can
you be sure that you are safe from the threat of SQL injection.

Mark Fallon, [email protected]

Bryn Llewellyn, [email protected]
Howard Smith, [email protected]
Oracle Headquarters
10-May-2017

51 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 APPENDIX A:
DEFINITIONS OF NEW TERMS OF ART INTRODUCED BY THIS PAPER

This appendix collects the definitions of the new terms of art that this paper has
introduced.

common SQL name

A common SQL name starts with an upper case alphabetic character in the range
A..Z and then has only upper case alphanumeric characters in the range A..Z, or
underscore, # or $. A common SQL name doesn’t need to be surrounded by
double quote characters in a SQL statement; and if it is not so surrounded, the
case with which it is written doesn’t matter. It may, however, be so surrounded. If
it is, then because the SQL parser preserves its case, it must be written in all
upper case.
See Example 2: user-supplied table name on page 15.

exotic SQL name

An exotic SQL name is one that violates the rules for a common SQL name and
that must, therefore, be surrounded by double quote characters in a
SQL statement.
See Example 2: user-supplied table name on page 15.

compile-time-fixed SQL statement text

Compile-time-fixed SQL statement text is the text of a SQL statement that cannot
change at run time and that can be confidently determined by reading the source
code of the PL/SQL unit that executes the SQL statement in question. For
dynamic SQL, it is the text of a SQL statement that is a PL/SQL
static varchar2 expression. The value of a PL/SQL static varchar2 expression
cannot change at run time and could be pre-computed at compile time. Because
the PL/SQL compiler composes the text of the SQL statement that implements
a PL/SQL embedded SQL statement, we regard this, too, as compile-time-fixed
SQL statement text.
See Distinguishing between compile-time-fixed SQL statement text and run-time-created
SQL statement text on page 6.

run-time-created SQL statement text

Run-time-created SQL statement text is the text of a SQL statement that is not
compile-time-fixed SQL statement text.
See Distinguishing between compile-time-fixed SQL statement text and run-time-created
SQL statement text on page 6.

SQL syntax template

A SQL syntax template looks something like a regular SQL statement but the
notion belongs in the domain of discourse of the Design Specification
document. Here is an example:
select &&1 from &&2 where &&3 = &4

52 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 It is a documentation device to prescribe the set of regular SQL statements that
are instances of the template. Instantiation, by a PL/SQL program, is achieved
by using regular SQL elements in place of the elements that start with && or &
in the template. All the other non- whitespace elements in the template must be
reproduced faithfully in the instantiation as a regular SQL statement. This
includes a SQL hint — but excludes ordinary comments, which can be treated as
whitespace.
See Introducing a new notion: SQL syntax template on page 4.

value placeholder
A value placeholder is the element in a SQL syntax template that starts with &. It
stands for either a well-formed SQL literal or a regular placeholder in a regular
SQL statement.
See Introducing a new notion: SQL syntax template on page 4.

simple SQL name placeholder

A simple SQL name placeholder is the element in a SQL syntax template that starts
with &&. It stands for a simple SQL name in a regular SQL statement.
See Introducing a new notion: SQL syntax template on page 4.

static SQL syntax template

A static SQL syntax template is a SQL syntax template that can be written down
explicitly in the Design Specification document. The implementation that
instantiates and executes a static SQL syntax template can always be written like
this, and should be:
<<Some_Inner_Block>>declare
Stmt constant varchar2(32767) :=
'select c1 from '
|| Sys.DBMS_Assert.Simple_Sql_Name(Dynamic_Name)
|| ' where PK = :b for update wait '
|| Sys.DBMS_Assert.Enquote_Literal(Dynamic_Value);
begin
execute immediate Stmt...
...
end Some_Inner_Block;

The assignment is a single concatenation of one or several PL/SQL

static varchar2 expressions with one or several invocations of any of
• DBMS_Assert.Enquote_Literal(), or
• To_Char(x f, 'NLS_Numeric_Characters = ''.,'''), or
• DBMS_Assert.Simple_Sql_Name().
A static SQL syntax template can be used, for example, when it would have been
possible to use embedded SQL were it not for the fact that the name(s) of the
from list items are not known until run time.
By definition, compile-time-fixed SQL statement text conforms to a static
SQL syntax template.
See Definition of static SQL syntax template on page 7.

53 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 dynamic SQL syntax template
A dynamic SQL syntax template is one of a large set of such templates that are
designed to prescribe the regular SQL statements for execution at a particular
call site, where the set is too large to allow each to be written down but where,
nevertheless, the set can be described with certainty. The description could be
written in the Design Specification document using, for example, regular
expression syntax.
A set of dynamic SQL syntax templates is needed, for example, to implement a
query by example interface of the kind illustrated in the section “Query by example
form” on page 42. A dynamic SQL syntax template needs to use neither simple
SQL name placeholders nor value placeholders in order to be useful — and its
safety is easier to determine in a code review when it does not. In such a use,
every regular SQL statement that conforms to the set of dynamic
SQL syntax templates for a particular call site will be composed using only
PL/SQL static varchar2 expressions.
Run-time-created SQL statement text might conform to a static
SQL syntax template or might conform to a dynamic SQL syntax template.
See Definition of dynamic SQL syntax template on page 8.

static text
Static text is either
• a PL/SQL static varchar2 expression as defined in the
PL/SQL Language Reference book, or
• an expression formed by an arbitrary concatenation of static text items, or
• the value of a local variable that has been visibly assigned with static text. (A
local variable is one that is declared within the present
top level PL/SQL block.)
This definition is intentionally recursive. The concatenation may be controlled by
tests whose outcome is not known until run time. But it must be self-evident,
following trivial human inspection, that every possible concatenation will result
only in static text whose ultimate source is only PL/SQL
static varchar2 expressions.
See Static text on page 35.

dynamic text
Dynamic text is any text that is not static text. Obvious examples are formal
parameters, variables declared at top level in a package without the constant
keyword, and ordinary variables that are assigned by executing a SQL statement
or as the actual argument for the Buffer formal parameter to Utl_File.Get_Line().
See Dynamic text on page 36.

safe dynamic text

Safe dynamic text is the output of either
• DBMS_Assert.Enquote_Literal(), or

54 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 • To_Char(x f, 'NLS_Numeric_Characters = ''.,''') where x is a variable of a
numeric datatype and f is an the explicit format model 'TM', or
• DBMS_Assert.Simple_Sql_Name().
See Safe dynamic text on page 36.

safe SQL statement text

Safe SQL statement text is an arbitrary concatenation of static text and
safe dynamic text. Local variables may be used for intermediate results or for the
final result.
See Safe SQL statement text on page 36.

top level PL/SQL block

A top level PL/SQL block is one of the following: a schema-level function; a
schema-level procedure; a function or procedure defined at top level within a
package body or a type body; a package’s initialization block; or the
implementation of a trigger. Thus a variable that is declared at top level in a
package or package body is, by definition, not a local variable.
See Static text on page 35.

55 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 APPENDIX B:
SUMMARY OF SQL INJECTION PREVENTION RULES

This appendix collects the rules for immunization against SQL injection that
have been advocated in this paper.

When composing a SQL statement  Rule_1: When it is necessary to compose a SQL statement programmatically,
programmatically, help the code the code usually needs, or at least benefits from the use of, variables for
reviewer by declaring variables used
intermediate results. Aim to declare these as constant, assigning the values in
for intermediate results as constant.
the declarations. This sometimes requires the use of nested block-statements
or forward-declared functions. This technique makes code review easier
because the reader can be sure that the value of a variable cannot change
between its initial assignment and its use. (page 10)

Understand what is meant by the term  Rule_2: Understand what is meant by the term SQL syntax template. Apply
SQL syntax template and the this understanding to the design of any code that constructs run-time-
difference between a static
created SQL statement text. Understand the difference between a static
SQL syntax template and a dynamic
SQL syntax template and a dynamic SQL syntax template. (page 11)
SQL syntax template.

Understand that SQL injection is the  Rule_3: Understand how to define the term SQL injection as the execution
execution of a SQL statement with an of a SQL statement with an unintended SQL syntax template. Know that
unintended SQL syntax template and
only run-time-created SQL statement text that, therefore, must be executed
that the risk can occur only when
using dynamic SQL is potentially vulnerable to SQL injection. (page 12)
run-time-created SQL statement text
is executed using dynamic SQL.

Expose the database to clients only  Rule_4: Expose the database to clients only via a PL/SQL API. Carefully
via a PL/SQL API. control privileges so that the client has no direct access to the application’s
objects of other kinds — especially tables and views. (page 29)

Insist that the Design Specification  Rule_5: When the Design Specification document for ordinary application
document for ordinary application code proposes to use anything other than embedded SQL, insist on
code defends any propsal to execute
examining the rationale very carefully when the document is reviewed. The
SQL using any method other than
design may well be defensible. But this defense must be made explicitly.
Embedded SQL.
(page 30)

56 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

Use compile-time-fixed SQL  Rule_6: Use compile-time-fixed SQL statement text unless you cannot. Use
statement text unless you cannot. embedded SQL when the SQL statement is of a kind that this supports.
Use Embedded SQL or
Otherwise, use execute immediate with single PL/SQL constant argument
execute immediate with a PL/SQL
composed using only PL/SQL static varchar2 expressions. When you
static varchar2 expression.
conclude that you cannot, review the Functional Specification document and
the Design Specification document carefully with colleagues and then
explain in the latter exactly why compile-time-fixed SQL statement text
cannot be used. (page 31)

Insist on a justification for a design  Rule_7: A Design Specification document that proposes to replace a
that proposes to replace a value placeholder in a SQL syntax template (whether this is a static
value placeholder in a
SQL syntax template or a dynamic SQL syntax template) should be regarded
SQL syntax template.
with extreme suspicion. It must present a convincing argument for this
approach before it is signed off. (page 32)

Insist on a justification for a design  Rule_8: A Design Specification document that proposes to replace a simple
that proposes to replace a simple SQL name placeholder in a SQL syntax template (whether this is a static
SQL name placeholder in a
SQL syntax template or a dynamic SQL syntax template) should be regarded
SQL syntax template.
with some suspicion. It must present a convincing argument for this
approach before it is signed off. (page 33)

Don’t confuse the need to use a  Rule_9: Don’t confuse the requirement to bind to a set of placeholders
dynamic SQL syntax template with whose composition emerges first at run time, which is fully supported by the
the need to replace
DBMS_Sql API, with a requirement to use directly encoded literals in
value placeholders in the template
pursuit of optimal query execution performance. (page 34)
with SQL literals.

Dynamic SQL may execute only a  Rule_10: When a SQL statement represented by a PL/SQL text expression
concatentation of static text and is executed using one of PL/SQL’s APIs for dynamic SQL, then the
safe dynamic text. Safe dynamic text
expression must be safe SQL statement text. Safe SQL statement text is a
is the output of one of exactly three
concatenation of static text and safe dynamic text. Static text is composed
Oracle-supplied functions:
Simple_Sql_Name(),
only of PL/SQL static varchar2 expressions. Dynamic text is anything that
Enquote_Literal(), and isn’t static text. Safe dynamic text is the output of one of exactly three
To_Char(x f, n). Oracle-supplied functions: DBMS_Assert.Simple_Sql_Name(),
DBMS_Assert.Enquote_Literal(), and To_Char(x f, n). (page 36)

57 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

Use Simple_Sql_Name(), to ensure  Rule_11: The safe dynamic text produced by
the safety of a SQL name. Use DBMS_Assert.Enquote_Literal() or by To_Char(x f, n) should be used only
Enquote_Literal() to ensure the safety
at a spot within the SQL statement where a SQL literal is intended; and the
of a string or datetime literal. Use
safe dynamic text produced by DBMS_Assert.Simple_Sql_Name() should
To_Char(x f, n) to ensure the safety of
a numeric literal.
be used only at a spot within the SQL statement where a simple SQL name
is intended. (page 37)

Keep the code that invokes  Rule_12: Make the job of the auditor easy by establishing the safety of
Simple_Sql_Name(), run-time-created SQL statement text in the code that immediately precedes
Enquote_Literal(), or To_Char(x f, n)
the PL/SQL statement that executes it. (page 39)
very close to the code that executes
the SQL statement that these calls
make safe.

Ensure the safety of the run-time-  Rule_13: Learn the full list of Oracle-supplied APIs that are designed to
created SQL statement text in the execute run-time-created SQL statement text. Ensure the safety such text by
same way no matter to which of the
using exactly the same rules as you use to ensure the safety of text that is
several Oracle-supplied APIs for
used with native dynamic SQL and the DBMS_Sql API. The approach is
executing dynamic SQL you submit it.
identical whether the text represents a complete SQL statement or a
component of one. (page 59)

Don’t use  Rule_14: There is never a good reason to use

DBMS_Utility.Exec_DDL_Statement() DBMS_Utility.Exec_DDL_Statement() in new code. Avoid it. (page 60)
in new code.

58 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 APPENDIX C:
ADDITIONAL ORACLE-SUPPLIED SUBPROGRAMS THAT IMPLEMENT
DYNAMIC SQL

PL/SQL programmers are used to thinking that there are just two ways to
execute a SQL statement that is presented as a PL/SQL text value: native
dynamic SQL and the DBMS_Sql API. This is not the case. Others exist, and
they are described in this appendix.
Some are designed to accept and execute a complete SQL statement but insist
that it must be of a certain kind (for example, a DDL statement of any kind or a
DDL statement to create a PL/SQL unit). Others are designed to accept a
component of a SQL statement (for example. a where clause) and concatenate this
with other components that they compose programatically to compose a
SQL statement and then execute it.
All these APIs share the property that they were designed specifically on the
assumption that it is the caller’s responsibility to ensure immunity to
SQL injection.
Rule_13
Ensure the safety of the run-time- Learn the full list of Oracle-supplied APIs that are designed to execute
created SQL statement text in the run-time-created SQL statement text. Ensure the safety such text by using
same way no matter to which of the
exactly the same rules as you use to ensure the safety of text that is used with
several Oracle-supplied APIs for
native dynamic SQL and the DBMS_Sql API. The approach is identical
executing dynamic SQL you submit it.
whether the text represents a complete SQL statement or a component of one.

DBMS_Utility.Exec_DDL_Statement()
Code_41 shows the important part of the implementation of
DBMS_Utility.Exec_DDL_Statement().
-- Code_41
Cur := DBMS_Sql.Open_Cursor();
DBMS_Sql.Parse(Cur, Stmt, DBMS_Sql.Native);
DBMS_Sql.Close_Cursor(Cur);

It turns out that when Stmt is a DDL statement93, then this code is sufficient to
execute it94. But for other kinds of statement that are not supported by
embedded SQL, like for, for example, alter session, it is necessary to call
DBMS_Sql.Execute(). This fact has the consequence that
DBMS_Utility.Exec_DDL_Statement() will quietly ignore anything but a
DDL statement.
It’s hard to imagine a Design Specification document that would prescribe this as
the intended behavior. Therefore, in new code, you should use execute immediate to
execute a DDL statement. And you should use native dynamic SQL or the
DBMS_Sql API to execute other kinds of statement when embedded SQL
cannot be used. The purpose of the calling code is known; and the
SQL statement text will have been composed specifically for that purpose.

93. The list of the kinds of statement that are classified as DDL statements is given in the
SQL Language Reference book.
94. This fact is documented in the PL/SQL Packages and Types Reference book.

59 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 Rule_14
Don’t use There is never a good reason to use DBMS_Utility.Exec_DDL_Statement() in
DBMS_Utility.Exec_DDL_Statement() new code. Avoid it.
in new code.

DBMS_DDL.Create_Wrapped()
The purpose of this procedure is to create a PL/SQL unit and to store the code
in the catalog using the obfuscated representation. The input is the plain text of
the appropriate create or replace statement. If it was not required to obfuscate the
source, you would simply use execute immediate.
Use cases do arise where it is appropriate generate a PL/SQL unit
programmatically. One example is given by code that implements
post-installation steps when ISV code is installed at a customer site. The logic of
particular subprograms needs to reflect specific properties of the installation
environment. Notice that Rule_10 on page 36 and Rule_11 on page 37 limit the
freedom with which PL/SQL can be programmatically composed. The huge
bulk of such source must be composed as static text; only identifiers and text
literals within the source may derive from dynamic text. Huge caution must be
used with respect to text literals95, even when, as this paper requires,
safe dynamic text is used. For example, it would clearly be unsafe to allow
safe dynamic text as the argument for execute immediate. Probably the only sensible
use of anything other than static text is for externally visible names, like the name
of the unit itself or of the subprograms it exposes.

DBMS_HS_Passthrough
The pass-through facility provides a mechanism for developers to execute a
SQL statement on a non-Oracle system without being interpreted by Oracle
Database.
DBMS_HS_Passthrough.Execute_Immediate().
This function runs a SQL statement immediately. Any valid SQL statement
except a select statement will be run immediately. Internally, the SQL statement is
run using the passthrough SQL protocol sequence of Open_Cursor(), Parse(),
Execute_Non_Query(), Close_Cursor(). The SQL statement cannot contain
placeholders.
DBMS_HS_Passthrough.Parse()
This function parses the provided SQL statement at the non-Oracle system. The
execution of the statement occurs at the subsequent
DBMS_HS_Passthrough.Execute_Non_Query() or in the case of a select statement
DBMS_HS_Passthrough.Fetch(). The SQL statement cannot contain placeholders.

95. While composing the source code of a PL/SQL unit using safe numeric and
datetime literals isn’t intrinsically risky, it’s very hard to see how a sensible
Design Specification document could call for this.

60 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 OWA_Util
The OWA_Util package provides, among other things, the ability to generate
HTML pages or page fragments using the results of a select statement presented
as a PL/SQL text value.
OWA_Util.Bind_Variables()
This function is intended to prepare a select statement for subsequent use by
other OWA_Util subprograms. However, it is nothing other than a wrapper for
calls to DBMS_Sql.Open_Cursor(), DBMS_Sql.Parse(), and a series of calls to
DBMS_Sql.Bind_Variable(). It returns an ordinary DBMS_Sql numeric cursor.
OWA_Util.Bind_Variables() has 25 pairs of formal parameters with names like
bv17Name, bv17Value, all with datatype varchar2, and with default value null so the
they can be omitted according to purpose. Each not null pair causes a
corresponding call to DBMS_Sql.Bind_Variable().
OWA_Util.ListPrint()
This procedure generates an HTML form user-interface element (a selection list)
from the output of a select statement. It has two versions: the first takes a
complete select statement as a PL/SQL text value; and the second takes a
DBMS_Sql numeric cursor prepared, for example, by a preceding call to
OWA_Util.Bind_Variables().
OWA_Util.TablePrint()
This function generates an HTML fragment to represent the contents of a
database table. It does not take full select statement; rather, the caller supplies the
table’s name, a comma separated list of the required columns, and specifies the
where clause and order by clause. The the where clause cannot contain placeholders.

61 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 APPENDIX D:
SELF-CONTAINED CODE TO ILLUSTRATE IMPLEMENTING CALLBACK
USING DYNAMIC POLYMORPHISM.

This code can be run as any ordinary user. To save space, it is written as if none
of the objects it creates already exist.
To prove that the implementation that uses dynamic polymorphism gives the
same results as the naïve implementation with execute immediate, the test harness
— procedure p() at the end — uses each method in turn to compute a checksum
over many invocations with different arguments. You can easily add some timing
code — use DBMS_Utility.Get_CPU_Time() — and greatly increase the range of
input values over which the checksum is computed.
create function Callback(Input in integer) return integer is
begin
return Input*Input;
end Callback;
/

create type Hook authid Current_User is object(

Dummy number,
not instantiable member function Callback(
self in Hook,
Input in integer ) return integer
)
not final
not instantiable
/

create type My_Implementation under Hook(

overriding member function Callback(
self in My_Implementation,
Input in integer) return integer
)
/

create type body My_Implementation is

overriding member function Callback(
self in My_Implementation,
Input in integer) return integer
is
begin
return Input*Input;
end Callback;
end;
/

create package Which_Callback is

function The_Callback return Hook;
end Which_Callback;
/

62 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 create package body Which_Callback is
-- Implements some logic to chose an instance of
-- the appropriate subtype of Hook
function The_Callback return Hook is
Which constant Hook := My_Implementation(0);
begin
return Which;
end;
end Which_Callback;
/

create procedure p(Name in varchar2, Obj in Hook) is

No_Of_Repeats constant pls_integer := 100;
Checksum integer;
n integer;
x integer;

Safe_Name constant varchar2(32767) :=

Sys.DBMS_Assert.Simple_Sql_Name(Name);
Stmt constant varchar2(32767) :=
'begin :x := '||Safe_Name||'(:n); end;';

procedure Check_It is
begin
if(Checksum is null or Checksum <> 338350) then
Raise_Application_Error(-20000,
'Callback: unexpected Checksum: ' || Checksum);
end if;
end Check_It;
begin
Checksum := 0;
n := 0;
for r in 1..No_Of_Repeats loop
n := n + 1;
execute immediate Stmt using out x, in n;
Checksum := Checksum + x;
end loop;
Check_It();

Checksum := 0;
n := 0;
for r in 1..No_Of_Repeats loop
n := n + 1;
Checksum := Checksum + Obj.Callback(n);
end loop;
Check_It();
end p;
/

begin p('Callback', Which_Callback.The_Callback()); end;

/

63 | HOW TO WRITE SQL INJECTION PROOF PL/SQL 10-May-2017

 Oracle Corporation, World Headquarters Worldwide Inquiries
500 Oracle Parkway Phone: +1.650.506.7000
Redwood Shores, CA 94065, USA Fax: +1.650.506.7200

CONNECT WITH US Integrated Cloud Applications & Platform Services

Copyright © 2017, Oracle and/or its affiliates. All rights reserved. This document is provided for information purposes only, and the
contents hereof are subject to change without notice. This document is not warranted to be error-free, nor subject to any other
blogs.oracle.com/oracle warranties or conditions, whether expressed orally or implied in law, including implied warranties and conditions of merchantability or
fitness for a particular purpose. We specifically disclaim any liability with respect to this document, and no contractual obligations are
formed either directly or indirectly by this document. This document may not be reproduced or transmitted in any form or by any
facebook.com/oracle means, electronic or mechanical, for any purpose, without our prior written permission.

twitter.com/oracle Oracle and Java are registered trademarks of Oracle and/or its affiliates. Other names may be trademarks of their respective owners.

Intel and Intel Xeon are trademarks or registered trademarks of Intel Corporation. All SPARC trademarks are used under license and
oracle.com
are trademarks or registered trademarks of SPARC International, Inc. AMD, Opteron, the AMD logo, and the AMD Opteron logo are
trademarks or registered trademarks of Advanced Micro Devices. UNIX is a registered trademark of The Open Group. 1014

How to write SQL injection proof PL/SQL

May 2017
Authors:
Mark Fallon, Architect, Oracle Headquarters
Bryn Llewellyn, PL/SQL Product Manager, Oracle Headquarters
Howard Smith, Senior Director, Global Product Security, Oracle UK