Note: The Oracle documentation uses two terms: defining (for instructing the library where the output data should go) and binding (for the input data and input/output PL/SQL parameters). For the sake of simplicity, SOCI uses the term binding for both of these.
The into
expression is used to add binding information to
the statement:
int count; sql << "select count(*) from person", into(count); string name; sql << "select name from person where id = 7", into(name);
In the above examples, some data is retrieved from the database and transmitted into the given local variable.
There should be as many into
elements as there are
expected columns in the result (see dynamic
resultset binding for the exception to this rule).
The use
expression associates the SQL placeholder (written with colon) with the local data:
int val = 7; sql << "insert into numbers(val) values(:val)", use(val);
In the above statement, the first "val" is a column name (assuming
that
there is appropriate table numbers
with this column), the
second "val" (with colon) is a
placeholder and its name is ignored here, and the third "val" is a name
of local variable.
To better understand the meaning of each "val" above, consider also:
int number = 7; sql << "insert into numbers(val) values(:blabla)", use(number);
Both examples above will insert the value of some local variable into
the table numbers
- we say that the local variable is used in the SQL statement.
There should be as many use
elements as there are
parameters used in the SQL query.
Portability note:
Older versions of the PostgreSQL client API do not allow to use input
parameters at all. In order to compile SOCI with those old client
libraries, define the SOCI_PGSQL_NOPARAMS
preprocessor
name.
If there is more output or input "holes" in the single statement, it
is possible to use many into
and use
expressions, separated by commas, where each expression will be
responsible for the consecutive "hole" in the statement:
string firstName = "John", lastName = "Smith"; int personId = 7; sql << "insert into person(id, firstname, lastname) values(:id, :fn, :ln)", use(personId), use(firstName), use(lastName); sql << "select firstname, lastname from person where id = :id", into(firstName), into(lastName), use(personId);
In the code above, the order of "holes" in the SQL statement and the
order of into
and use
expression should
match.
The SQL placeholders that have their names (with colon) can be bound by name to clearly associate the local variable with the given placehorder.
This explicit naming allows to use different order of elements:
string firstName = "John", lastName = "Smith"; int personId = 7; sql << "insert into person(id, firstname, lastname) values(:id, :fn, :ln)", use(firstName, "fn"), use(lastName, "ln"), use(personId, "id");
or bind the same local data to many "holes" at the same time:
string addr = "..."; sql << "update person" " set mainaddress = :addr, contactaddress = :addr" " where id = 7", use(addr, "addr");
Object lifetime and immutability:
SOCI assumes that local variables provided as use
elements
live at least as long at it takes to execute the whole statement.
In short statement forms like above, the statement is executed sometime
at the end of the full expression and the whole process is driven by the invisible
temporary object handled by the library. If the data provided by user comes
from another temporary variable, it might be possible for the compiler to arrange
them in a way that the user data will be destroyed before the statement will
have its chance to execute, referencing objects that no longer exist:
// dangerous code: string getNameFromSomewhere(); sql << "insert into person(name) values(:n)", use(getNameFromSomewhere());
In the above example, the data passed to the database comes from the temporary
variable that is a result of call to getNameFromSomewhere
- this
should be avoided and named variables should be used to ensure safe lifetime relations:
// safe code: string getNameFromSomewhere(); string name = getNameFromSomewhere(); sql << "insert into person(name) values(:n)", use(name);
It is still possible to provide const
data for use elements.
Note that some database servers, like Oracle, allow PL/SQL procedures to modify their
in/out parameters - this is detected by the SOCI library and an error is reported
if the database attempts to modify the use
element that holds const
data.
The above example can be ultimately written in the following way:
// safe and efficient code: string getNameFromSomewhere(); const string & name = getNameFromSomewhere(); sql << "insert into person(name) values(:n)", use(name);
Portability notes:
The PostgreSQL backend allows to use the "native" PostgreSQL way of
naming parameters in the query, which is by numbers like $1
,
$2
, $3
, etc. In fact, the backend rewrites
the given query to the native form - and this is also one of the very few places
where SOCI intrudes into the SQL query. For portability reasons, it is
recommended to use named parameters, as shown in the examples above.
The query rewriting can be switched off by compiling the backend with
the SOCI_PGSQL_NOBINDBYNAME
name defined. Note that in this case
it is also necessary to define SOCI_PGSQL_NOPREPARE
, because
statement preparation relies on successful query rewriting.
In practice, both macros will be needed for PostgreSQL server older than 8.0.
In order to support null values and other conditions which are not real errors, the concept of indicator is provided.
For example, when the following SQL query is executed:
select name from person where id = 7
there are three possible outcomes:
Whereas the first alternative is easy to handle, the other two are more
complex. Moreover, they are not necessarily errors from the
application's point of view and what's more interesting, they are different
and the application may wish to detect which is the case.
The following example does this:
string name; indicator ind; sql << "select name from person where id = 7", into(name, ind); if (sql.got_data()) { switch (ind) { case i_ok: // the data was returned without problems break; case i_null: // there is a person, but he has no name (his name is null) break; case i_truncated: // the name was returned only in part, // because the provided buffer was too short // (not possible with std::string, but possible with char* and char[]) break; } } else { // no such person in the database }
The use of indicator variable is optional, but if it is not used and
the result would be i_null
,
then the exception is thrown. This means that you should use indicator
variables everywhere where the application logic (and database schema)
allow the "attribute not set" condition.
Indicator variables can be also used when binding input data, to control whether the data is to be used as provided, or explicitly overrided to be null:
int id = 7; string name; indicator ind = i_null; sql << "insert into person(id, name) values(:id, :name)", use(id), use(name, ind);
In the above example, the row is inserted with name
attribute set to null.
Indicator variables can also be used in conjunction with vector based insert, update, and select statements:
vector<string> names(100); vector<indicator> inds; sql << "select name from person where id = 7", into(names, inds);
The above example retrieves first 100 rows of data (or less). The
initial size of names
vector provides the (maximum)
number of rows that should be read. Both vectors will be
automatically resized according to the number of rows that were
actually read.
The following example inserts null for each value of name:
vector<int> ids; vector<string> names; vector<indicator> nameIndicators; for (int i = 0; i != 10; ++i) { ids.push_back(i); names.push_back(""); nameIndicators.push_back(i_null); } sql << "insert into person(id, name) values(:id, :name)", use(ids), use(name, nameIndicators);
See also Integration with Boost to learn how the Boost.Optional library can be used to handle null data conditions in a more natural way.
The static binding for types is most useful when the types used in
the database are known at compile time - this was already presented
above with the help of into
and use
functions.
The following types are currently supported for use with into
and use
expressions:
char
(for character values)short
, int
, unsigned
long
, long long
, double
(for numeric values)char*
, char[]
, std::string
(for string values)std::tm
(for datetime
values)soci::statement
(for nested statements and PL/SQL
cursors)soci::blob
(for Binary Large OBjects)soci::row_id
(for row identifiers)See the test code that accompanies the library to see how each of these types is used.
Bulk inserts, updates, and selects are supported through the
following std::vector
based into and use types:
std::vector<char>
std::vector<short>
std::vector<int>
std::vector<unsigned long>
std::vector<long long>
std::vector<double>
std::vector<std::string>
std::vector<std::tm>
Use of the vector based types mirrors that of the standard types, with the size of the vector used to specify the number of records to process at a time. See below for examples.
Note that bulk operations are supported only for std::vector
s of the types listed above.
For certain applications it is desirable to be able to select data from
arbitrarily structured tables (e.g. via "select * from ...
") and format the
resulting data based upon its type.
SOCI supports this through the soci::row
and soci::column_properties
classes.
Data is selected into a row
object, which holds column_properties
objects describing
the attributes of data contained in each column. Once the data type for each
column is known, the data can be formatted appropriately.
For example, the code below creates an XML document from a selected row of data from an arbitrary table:
row r; sql << "select * from some_table", into(r); std::ostringstream doc; doc << "<row>" << std::endl; for(std::size_t i = 0; i != r.size(); ++i) { const column_properties & props = r.get_properties(i); doc << '<' << props.get_name() << '>'; switch(props.get_data_type()) { case dt_string: doc << r.get<std::string>(i); break; case dt_double: doc << r.get<double>(i); break; case dt_integer: doc << r.get<int>(i); break; case dt_unsigned_long: doc << r.get<unsigned long>(i); break; case dt_long_long: doc << r.get<long long>(i); break; case dt_date: std::tm when = r.get<std::tm>(i); doc << asctime(&when); break; } doc << "</" << props.get_name() << '>' << std::endl; } doc << "</row>";
The type T
parameter that should be passed to
row::get<T>()
depends on the SOCI data type that
is returned from
column_properties::get_data_type()
.
row::get<T>()
throws an exception of type
std::bad_cast
if an incorrect type T
is
requested.
SOCI Data Type | row::get<T> specialization |
---|---|
dt_double |
double |
dt_integer |
int |
dt_unsigned_long |
unsigned long |
dt_long_long |
long long |
dt_string |
std::string |
dt_date |
std::tm
|
The mapping of underlying database column types to SOCI datatypes is database specific. See the backend documentation for details.
It is also possible to extract data from the row
object using its stream-like
interface, where each extracted variable should have matching type respective to its position in the chain:
row r; sql << "select name, address, age from persons where id = 123", into(r); string name, address; int age; r >> name >> address >> age;
Note, however, that this interface is not compatible with the standard
std::istream
class and that it is only possible to extract a single row at a time
- for "safety" reasons the row boundary is preserved and it is necessary to perform the
fetch
operation explicitly for each consecutive row
(see next page).
SOCI can be easily extended with support for user-defined datatypes.
The extension mechanism relies on appropriate specialization of the type_conversion
struct that converts to and from one of the following SOCI base types:
double
int
unsigned long
long long
std::string
char
std::tm
There are three required class members for a valid type_conversion
specialization:
base_type
type definition, aliasing either one of the base types or another ser-defined typefrom_base()
static member function, converting from
the base typeto_base()
static member function, converting to the
base typeNote that no database-specific code is required to define user conversion.
The following example shows how the user can extend SOCI to
support his own type MyInt
, which here is some wrapper
for the fundamental int
type:
class MyInt { public: MyInt() {} MyInt(int i) : i_(i) {} void set(int i) { i_ = i; } int get() const { return i_; } private: int i_; }; namespace soci { template <> struct type_conversion<MyInt> { typedef int base_type; static void from_base(int i, indicator ind, MyInt & mi) { if (ind == i_null) { throw soci_error("Null value not allowed for this type"); } mi.set(i); } static void to_base(const MyInt & mi, int & i, indicator & ind) { i = mi.get(); ind = i_ok; } }; }
The above specialization for soci::type_conversion<MyInt>
is enough
to enable the following:
MyInt i; sql << "select count(*) from person", into(i); cout << "We have " << i.get() << " persons in the database.\n";
Note that there is a number of types from the Boost library integrated with SOCI out of the box, see Integration with Boost for complete description. Use these as examples of conversions for more complext data types.
Note also that user-defined datatypes are not supported with bulk data transfer.
Another possibility to extend SOCI with custom data types is to use
the into_type<T>
and use_type<T>
class templates, which specializations can be user-provided. These
specializations need to implement the interface defined by,
respectively, the into_type_base
and use_type_base
classes.
Note that when specializing these template classes the only convention is that when the indicator variable is used (see below), it should appear in the second position. Please refer to the library source code to see how this is done for the standard types.
SOCI provides a class called values
specifically to
enable object-relational mapping via type_conversion
specializations.
For example, the following code maps a Person
object to
and from a
database table containing columns "ID"
, "FIRST_NAME"
, "LAST_NAME"
, and
"GENDER"
.
Note that the mapping is non-invasive - the Person
object
itself does not contain any SOCI-specific code:
struct Person { int id; std::string firstName; std::string lastName; std::string gender; }; namespace soci { template<> struct TypeConversion<Person> { typedef values base_type; static void from_base(values const & v, indicator /* ind */, Person & p) { p.id = v.get<int>("ID"); p.firstName = v.get<std::string>("FIRST_NAME"); p.lastName = v.get<std::string>("LAST_NAME"); // p.gender will be set to the default value "unknown" // when the column is null: p.gender = v.get<std::string>("GENDER", "unknown"); // alternatively, the indicator can be tested directly: // if (v.indicator("GENDER") == i_null) // { // p.gender = "unknown"; // } // else // { // p.gender = v.get<std::string>("GENDER"); // } } static void to_base(const Person & p, values & v, indicator & ind) { v.set("ID", p.id); v.set("FIRST_NAME", p.firstName); v.set("LAST_NAME", p.lastName); v.set("GENDER", p.gender, p.gender.empty() ? i_null : i_ok); ind = i_ok; } }; }
With the above type_conversion
specialization in place, it
is possible to use Person
directly with SOCI:
session sql(oracle, "service=db1 user=scott password=tiger"); Person p; p.id = 1; p.lastName = "Smith"; p.firstName = "Pat"; sql << "insert into person(id, first_name, last_name) " "values(:ID, :FIRST_NAME, :LAST_NAME)", use(p); Person p1; sql << "select * from person", into(p1); assert(p1.id == 1); assert(p1.firstName + p.lastName == "PatSmith"); assert(p1.gender == "unknown"); p.firstName = "Patricia"; sql << "update person set first_name = :FIRST_NAME " "where id = :ID", use(p);
Note: The values
class is currently not suited for use outside of type_conversion
specializations. It is specially designed to facilitate
object-relational mapping when used as shown above.
The SOCI library provides also an interface for basic operations on large objects (BLOBs - Binary Large OBjects).
blob b(sql); // sql is a session object sql << "select mp3 from mymusic where id = 123", into(b);
The following functions are provided in the blob
interface, mimicking the file-like operations:
std::size_t get_len();
std::size_t read(std::size_t offset, char *buf, std::size_t
toRead);
std::size_t write(std::size_t offset, char const *buf,
std::size_t toWrite);
std::size_t append(char const *buf, std::size_t toWrite);
void trim(std::size_t newLen);
The offset
parameter is always counted from the beginning
of the BLOB's data.
Portability notes:
trim
function is not currently available for
the PostgreSQL backend.Previous (Connections and simple queries) | Next (Statements, procedures and transactions) |
Copyright © 2004-2008 Maciej Sobczak, Stephen Hutton
Generated on Sun Oct 3 2010 17:42:17 for EXTRAS-SOCI by Doxygen 1.7.1