6 How to implement a driver
This document was written a long time ago. A lot of it is still valid, but some things have changed since it was first written. Updates of this document are planned for the future. The reader is encouraged to also read the erl_driver, and the driver_entry documentation.
6.1 Introduction
This chapter tells you how to build your own driver for erlang.
A driver in Erlang is a library written in C, that is linked to the Erlang emulator and called from erlang. Drivers can be used when C is more suitable than Erlang, to speed things up, or to provide access to OS resources not directly accessible from Erlang.
A driver can be dynamically loaded, as a shared library (known as a DLL on windows), or statically loaded, linked with the emulator when it is compiled and linked. Only dynamically loaded drivers are described here, statically linked drivers are beyond the scope of this chapter.
When a driver is loaded it is executed in the context of the emulator, shares the same memory and the same thread. This means that all operations in the driver must be non-blocking, and that any crash in the driver will bring the whole emulator down. In short: you have to be extremely careful!
6.2 Sample driver
This is a simple driver for accessing a postgres database using the libpq C client library. Postgres is used because it's free and open source. For information on postgres, refer to the website www.postgres.org.
The driver is synchronous, it uses the synchronous calls of the client library. This is only for simplicity, and is generally not good, since it will halt the emulator while waiting for the database. This will be improved on below with an asynchronous sample driver.
The code is quite straight-forward: all communication between Erlang and the driver is done with port_control/3, and the driver returns data back using the rbuf.
An Erlang driver only exports one function: the driver entry function. This is defined with a macro, DRIVER_INIT, and returns a pointer to a C struct containing the entry points that are called from the emulator. The struct defines the entries that the emulator calls to call the driver, with a NULL pointer for entries that are not defined and used by the driver.
The start entry is called when the driver is opened as a port with open_port/2. Here we allocate memory for a user data structure. This user data will be passed every time the emulator calls us. First we store the driver handle, because it is needed in subsequent calls. We allocate memory for the connection handle that is used by LibPQ. We also set the port to return allocated driver binaries, by setting the flag PORT_CONTROL_FLAG_BINARY, calling set_port_control_flags. (This is because we don't know whether our data will fit in the result buffer of control, which has a default size set up by the emulator, currently 64 bytes.)
There is an entry init which is called when the driver is loaded, but we don't use this, since it is executed only once, and we want to have the possibility of several instances of the driver.
The stop entry is called when the port is closed.
The control entry is called from the emulator when the Erlang code calls port_control/3, to do the actual work. We have defined a simple set of commands: connect to login to the database, disconnect to log out and select to send a SQL-query and get the result. All results are returned through rbuf. The library ei in erl_interface is used to encode data in binary term format. The result is returned to the emulator as binary terms, so binary_to_term is called in Erlang to convert the result to term form.
The code is available in pg_sync.c in the sample directory of erts.
The driver entry contains the functions that will be called by the emulator. In our simple example, we only provide start, stop and control.
/* Driver interface declarations */
static ErlDrvData start(ErlDrvPort port, char *command);
static void stop(ErlDrvData drv_data);
static int control(ErlDrvData drv_data, unsigned int command, char *buf,
int len, char **rbuf, int rlen);
static ErlDrvEntry pq_driver_entry = {
NULL, /* init */
start,
stop,
NULL, /* output */
NULL, /* ready_input */
NULL, /* ready_output */
"pg_sync", /* the name of the driver */
NULL, /* finish */
NULL, /* handle */
control,
NULL, /* timeout */
NULL, /* outputv */
NULL, /* ready_async */
NULL, /* flush */
NULL, /* call */
NULL /* event */
};
We have a structure to store state needed by the driver, in this case we only need to keep the database connection.
typedef struct our_data_s {
PGconn* conn;
} our_data_t;
These are control codes we have defined.
/* Keep the following definitions in alignment with the
* defines in erl_pq_sync.erl
*/
#define DRV_CONNECT 'C'
#define DRV_DISCONNECT 'D'
#define DRV_SELECT 'S'
This just returns the driver structure. The macro DRIVER_INIT defines the only exported function. All the other functions are static, and will not be exported from the library.
/* INITIALIZATION AFTER LOADING */
/*
* This is the init function called after this driver has been loaded.
* It must *not* be declared static. Must return the address to
* the driver entry.
*/
DRIVER_INIT(pq_drv)
{
return &pq_driver_entry;
}
Here we do some initialization, start is called from open_port. The data will be passed to control and stop.
/* DRIVER INTERFACE */
static ErlDrvData start(ErlDrvPort port, char *command)
{
our_data_t* data;
data = (our_data_t*)driver_alloc(sizeof(our_data_t));
data->conn = NULL;
set_port_control_flags(port, PORT_CONTROL_FLAG_BINARY);
return (ErlDrvData)data;
}
We call disconnect to log out from the database. (This should have been done from Erlang, but just in case.)
static int do_disconnect(our_data_t* data, ei_x_buff* x);
static void stop(ErlDrvData drv_data)
{
do_disconnect((our_data_t*)drv_data, NULL);
}
We use the binary format only to return data to the emulator; input data is a string paramater for connect and select. The returned data consists of Erlang terms.
The functions get_s and ei_x_to_new_binary are utilities that is used to make the code shorter. get_s duplicates the string and zero-terminates it, since the postgres client library wants that. ei_x_to_new_binary takes an ei_x_buff buffer and allocates a binary and copies the data there. This binary is returned in *rbuf. (Note that this binary is freed by the emulator, not by us.)
static char* get_s(const char* buf, int len);
static int do_connect(const char *s, our_data_t* data, ei_x_buff* x);
static int do_select(const char* s, our_data_t* data, ei_x_buff* x);
/* Since we are operating in binary mode, the return value from control
* is irrelevant, as long as it is not negative.
*/
static int control(ErlDrvData drv_data, unsigned int command, char *buf,
int len, char **rbuf, int rlen)
{
int r;
ei_x_buff x;
our_data_t* data = (our_data_t*)drv_data;
char* s = get_s(buf, len);
ei_x_new_with_version(&x);
switch (command) {
case DRV_CONNECT: r = do_connect(s, data, &x); break;
case DRV_DISCONNECT: r = do_disconnect(data, &x); break;
case DRV_SELECT: r = do_select(s, data, &x); break;
default: r = -1; break;
}
*rbuf = (char*)ei_x_to_new_binary(&x);
ei_x_free(&x);
driver_free(s);
return r;
}
In do_connect is where we log in to the database. If the connection was successful we store the connection handle in our driver data, and return ok. Otherwise, we return the error message from postgres, and store NULL in the driver data.
static int do_connect(const char *s, our_data_t* data, ei_x_buff* x)
{
PGconn* conn = PQconnectdb(s);
if (PQstatus(conn) != CONNECTION_OK) {
encode_error(x, conn);
PQfinish(conn);
conn = NULL;
} else {
encode_ok(x);
}
data->conn = conn;
return 0;
}
If we are connected (if the connection handle is not NULL), we log out from the database. We need to check if a we should encode an ok, since we might get here from the stop function, which doesn't return data to the emulator.
static int do_disconnect(our_data_t* data, ei_x_buff* x)
{
if (data->conn == NULL)
return 0;
PQfinish(data->conn);
data->conn = NULL;
if (x != NULL)
encode_ok(x);
return 0;
}
We execute a query and encodes the result. Encoding is done in another C module, pg_encode.c which is also provided as sample code.
static int do_select(const char* s, our_data_t* data, ei_x_buff* x)
{
PGresult* res = PQexec(data->conn, s);
encode_result(x, res, data->conn);
PQclear(res);
return 0;
}
Here we simply checks the result from postgres, and if it's data we encode it as lists of lists with column data. Everything from postgres is C strings, so we just use ei_x_encode_string to send the result as strings to Erlang. (The head of the list contains the column names.)
void encode_result(ei_x_buff* x, PGresult* res, PGconn* conn)
{
int row, n_rows, col, n_cols;
switch (PQresultStatus(res)) {
case PGRES_TUPLES_OK:
n_rows = PQntuples(res);
n_cols = PQnfields(res);
ei_x_encode_tuple_header(x, 2);
encode_ok(x);
ei_x_encode_list_header(x, n_rows+1);
ei_x_encode_list_header(x, n_cols);
for (col = 0; col < n_cols; ++col) {
ei_x_encode_string(x, PQfname(res, col));
}
ei_x_encode_empty_list(x);
for (row = 0; row < n_rows; ++row) {
ei_x_encode_list_header(x, n_cols);
for (col = 0; col < n_cols; ++col) {
ei_x_encode_string(x, PQgetvalue(res, row, col));
}
ei_x_encode_empty_list(x);
}
ei_x_encode_empty_list(x);
break;
case PGRES_COMMAND_OK:
ei_x_encode_tuple_header(x, 2);
encode_ok(x);
ei_x_encode_string(x, PQcmdTuples(res));
break;
default:
encode_error(x, conn);
break;
}
}
6.3 Compiling and linking the sample driver
The driver should be compiled and linked to a shared library (DLL on windows). With gcc this is done with the link flags -shared and -fpic. Since we use the ei library we should include it too. There are several versions of ei, compiled for debug or non-debug and multi-threaded or single-threaded. In the makefile for the samples the obj directory is used for the ei library, meaning that we use the non-debug, single-threaded version.
6.4 Calling a driver as a port in Erlang
Before a driver can be called from Erlang, it must be loaded and opened. Loading is done using the erl_ddll module (the erl_ddll driver that loads dynamic driver, is actually a driver itself). If loading is ok the port can be opened with open_port/2. The port name must match the name of the shared library and the name in the driver entry structure.
When the port has been opened, the driver can be called. In the pg_sync example, we don't have any data from the port, only the return value from the port_control.
The following code is the Erlang part of the synchronous postgres driver, pg_sync.erl.
-module(pg_sync).
-define(DRV_CONNECT, 1).
-define(DRV_DISCONNECT, 2).
-define(DRV_SELECT, 3).
-export([connect/1, disconnect/1, select/2]).
connect(ConnectStr) ->
case erl_ddll:load_driver(".", "pg_sync") of
ok -> ok;
{error, already_loaded} -> ok;
E -> exit({error, E})
end,
Port = open_port({spawn, ?MODULE}, []),
case binary_to_term(port_control(Port, ?DRV_CONNECT, ConnectStr)) of
ok -> {ok, Port};
Error -> Error
end.
disconnect(Port) ->
R = binary_to_term(port_control(Port, ?DRV_DISCONNECT, "")),
port_close(Port),
R.
select(Port, Query) ->
binary_to_term(port_control(Port, ?DRV_SELECT, Query)).
The api is simple: connect/1 loads the driver, opens it and logs on to the database, returning the Erlang port if successful, select/2 sends a query to the driver, and returns the result, disconnect/1 closes the database connection and the driver. (It does not unload it, however.) The connection string should be a connection string for postgres.
The driver is loaded with erl_ddll:load_driver/2, and if this is successful, or if it's already loaded, it is opened. This will call the start function in the driver.
We use the port_control/3 function for all calls into the driver, the result from the driver is returned immediately, and converted to terms by calling binary_to_term/1. (We trust that the terms returned from the driver are well-formed, otherwise the binary_to_term calls could be contained in a catch.)
6.5 Sample asynchronous driver
Sometimes database queries can take long time to complete, in our pg_sync driver, the emulator halts while the driver is doing it's job. This is often not acceptable, since no other Erlang processes gets a chance to do anything. To improve on our postgres driver, we reimplement it using the asynchronous calls in LibPQ.
The asynchronous version of the driver is in the sample files pg_async.c and pg_asyng.erl.
/* Driver interface declarations */
static ErlDrvData start(ErlDrvPort port, char *command);
static void stop(ErlDrvData drv_data);
static int control(ErlDrvData drv_data, unsigned int command, char *buf,
int len, char **rbuf, int rlen);
static void ready_io(ErlDrvData drv_data, ErlDrvEvent event);
static ErlDrvEntry pq_driver_entry = {
NULL, /* init */
start,
stop,
NULL, /* output */
ready_io, /* ready_input */
ready_io, /* ready_output */
"pg_async", /* the name of the driver */
NULL, /* finish */
NULL, /* handle */
control,
NULL, /* timeout */
NULL, /* outputv */
NULL, /* ready_async */
NULL, /* flush */
NULL, /* call */
NULL /* event */
};
typedef struct our_data_t {
PGconn* conn;
ErlDrvPort port;
int socket;
int connecting;
} our_data_t;
Here some things have changed from pg_sync.c: we use the entry ready_io for ready_input and ready_output which will be called from the emulator only when there is input to be read from the socket. (Actually, the socket is used in a select function inside the emulator, and when the socket is signalled, indicating there is data to read, the ready_input entry is called. More on this below.)
Our driver data is also extended, we keep track of the socket used for communication with postgres, and also the port, which is needed when we send data to the port with driver_output. We have a flag connecting to tell whether the driver is waiting for a connection or waiting for the result of a query. (This is needed since the entry ready_io will be called both when connecting and when there is query result.)
static int do_connect(const char *s, our_data_t* data)
{
PGconn* conn = PQconnectStart(s);
if (PQstatus(conn) == CONNECTION_BAD) {
ei_x_buff x;
ei_x_new_with_version(&x);
encode_error(&x, conn);
PQfinish(conn);
conn = NULL;
driver_output(data->port, x.buff, x.index);
ei_x_free(&x);
}
PQconnectPoll(conn);
int socket = PQsocket(conn);
data->socket = socket;
driver_select(data->port, (ErlDrvEvent)socket, DO_READ, 1);
driver_select(data->port, (ErlDrvEvent)socket, DO_WRITE, 1);
data->conn = conn;
data->connecting = 1;
return 0;
}
The connect function looks a bit different too. We connect using the asynchronous PQconnectStart function. After the connection is started, we retrieve the socket for the connection with PQsocket. This socket is used with the driver_select function to wait for connection. When the socket is ready for input or for output, the ready_io function will be called.
Note that we only return data (with driver_output) if there is an error here, otherwise we wait for the connection to be completed, in which case our ready_io function will be called.
static int do_select(const char* s, our_data_t* data)
{
data->connecting = 0;
PGconn* conn = data->conn;
/* if there's an error return it now */
if (PQsendQuery(conn, s) == 0) {
ei_x_buff x;
ei_x_new_with_version(&x);
encode_error(&x, conn);
driver_output(data->port, x.buff, x.index);
ei_x_free(&x);
}
/* else wait for ready_output to get results */
return 0;
}
The do_select function initiates a select, and returns if there is no immediate error. The actual result will be returned when ready_io is called.
static void ready_io(ErlDrvData drv_data, ErlDrvEvent event)
{
PGresult* res = NULL;
our_data_t* data = (our_data_t*)drv_data;
PGconn* conn = data->conn;
ei_x_buff x;
ei_x_new_with_version(&x);
if (data->connecting) {
ConnStatusType status;
PQconnectPoll(conn);
status = PQstatus(conn);
if (status == CONNECTION_OK)
encode_ok(&x);
else if (status == CONNECTION_BAD)
encode_error(&x, conn);
} else {
PQconsumeInput(conn);
if (PQisBusy(conn))
return;
res = PQgetResult(conn);
encode_result(&x, res, conn);
PQclear(res);
for (;;) {
res = PQgetResult(conn);
if (res == NULL)
break;
PQclear(res);
}
}
if (x.index > 1) {
driver_output(data->port, x.buff, x.index);
if (data->connecting)
driver_select(data->port, (ErlDrvEvent)data->socket, DO_WRITE, 0);
}
ei_x_free(&x);
}
The ready_io function will be called when the socket we got from postgres is ready for input or output. Here we first check if we are connecting to the database. In that case we check connection status and return ok if the connection is successful, or error if it's not. If the connection is not yet established, we simply return; ready_io will be called again.
If we have result from a connect, indicated that we have data in the x buffer, we no longer need to select on output (ready_output), so we remove this by calling driver_select.
If we're not connecting, we're waiting for results from a PQsendQuery, so we get the result and return it. The encoding is done with the same functions as in the earlier example.
We should add error handling here, for instance checking that the socket is still open, but this is just a simple example.
The Erlang part of the asynchronous driver consists of the sample file pg_async.erl.
-module(pg_async).
-define(DRV_CONNECT, $C).
-define(DRV_DISCONNECT, $D).
-define(DRV_SELECT, $S).
-export([connect/1, disconnect/1, select/2]).
connect(ConnectStr) ->
case erl_ddll:load_driver(".", "pg_async") of
ok -> ok;
{error, already_loaded} -> ok;
_ -> exit({error, could_not_load_driver})
end,
Port = open_port({spawn, ?MODULE}, [binary]),
port_control(Port, ?DRV_CONNECT, ConnectStr),
case return_port_data(Port) of
ok ->
{ok, Port};
Error ->
Error
end.
disconnect(Port) ->
port_control(Port, ?DRV_DISCONNECT, ""),
R = return_port_data(Port),
port_close(Port),
R.
select(Port, Query) ->
port_control(Port, ?DRV_SELECT, Query),
return_port_data(Port).
return_port_data(Port) ->
receive
{Port, {data, Data}} ->
binary_to_term(Data)
end.
The Erlang code is slightly different, this is because we don't return the result synchronously from port_control, instead we get it from driver_output as data in the message queue. The function return_port_data above receives data from the port. Since the data is in binary format, we use binary_to_term/1 to convert it to Erlang term. Note that the driver is opened in binary mode, open_port/2 is called with the option [binary]. This means that data sent from the driver to the emulator is sent as binaries. Without the binary option, they would have been lists of integers.
6.6 An asynchronous driver using driver_async
As a final example we demonstrate the use of driver_async. We also use the driver term interface. The driver is written in C++. This enables us to use an algorithm from STL. We will use the next_permutation algorithm to get the next permutation of a list of integers. For large lists (more than 100000 elements), this will take some time, so we will perform this as an asynchronous task.
The asynchronous api for drivers are quite complicated. First of all, the work must be prepared. In our example we do this in output. We could have used control just as well, but we want some variation in our examples. In our driver, we allocate a structure that contains all needed for the asynchronous task to do the work. This is done in the main emulator thread. Then the asynchronous function is called from a driver thread, separate from the main emulator thread. Note that the driver- functions are not reentrant, so they shouldn't be used. Finally, after the function is completed, the driver callback ready_async is called from the main emulator thread, this is where we return the result to Erlang. (We can't return the result from within the asynchronous function, since we can't call the driver-functions.)
The code below is from the sample file next_perm.cc.
The driver entry looks like before, but also contains the call-back ready_async.
static ErlDrvEntry next_perm_driver_entry = {
NULL, /* init */
start,
NULL, /* stop */
output,
NULL, /* ready_input */
NULL, /* ready_output */
"next_perm", /* the name of the driver */
NULL, /* finish */
NULL, /* handle */
NULL, /* control */
NULL, /* timeout */
NULL, /* outputv */
ready_async,
NULL, /* flush */
NULL, /* call */
NULL /* event */
};
The output function allocates the work-area of the asynchronous function. Since we use C++, we use a struct, and stuff the data in it. We have to copy the original data, it is not valid after we have returned from the output function, and the do_perm function will be called later, and from another thread. We return no data here, instead it will be sent later from the ready_async call-back.
The async_data will be passed to the do_perm function. We do not use a async_free function (the last argument to driver_async, it's only used if the task is cancelled programmatically.
struct our_async_data {
bool prev;
vector<int> data;
our_async_data(ErlDrvPort p, int command, const char* buf, int len);
};
our_async_data::our_async_data(ErlDrvPort p, int command,
const char* buf, int len)
: prev(command == 2),
data((int*)buf, (int*)buf + len / sizeof(int))
{
}
static void do_perm(void* async_data);
static void output(ErlDrvData drv_data, char *buf, int len)
{
if (*buf < 1 || *buf > 2) return;
ErlDrvPort port = reinterpret_cast<ErlDrvPort>(drv_data);
void* async_data = new our_async_data(port, *buf, buf+1, len);
driver_async(port, NULL, do_perm, async_data, do_free);
}
In the do_perm we simply do the work, operating on the structure that was allocated in output.
static void do_perm(void* async_data)
{
our_async_data* d = reinterpret_cast<our_async_data*>(async_data);
if (d->prev)
prev_permutation(d->data.begin(), d->data.end());
else
next_permutation(d->data.begin(), d->data.end());
}
In the ready_async function, the output is sent back to the emulator. We use the driver term format instead of ei. This is the only way to send Erlang terms directly to a driver, without having the Erlang code to call binary_to_term/1. In our simple example this works well, and we don't need to use ei to handle the binary term format.
When the data is returned we deallocate our data.
static void ready_async(ErlDrvData drv_data, ErlDrvThreadData async_data)
{
ErlDrvPort port = reinterpret_cast<ErlDrvPort>(drv_data);
our_async_data* d = reinterpret_cast<our_async_data*>(async_data);
int n = d->data.size(), result_n = n*2 + 3;
ErlDrvTermData* result = new ErlDrvTermData[result_n], * rp = result;
for (vector<int>::iterator i = d->data.begin();
i != d->data.end(); ++i) {
*rp++ = ERL_DRV_INT;
*rp++ = *i;
}
*rp++ = ERL_DRV_NIL;
*rp++ = ERL_DRV_LIST;
*rp++ = n+1;
driver_output_term(port, result, result_n);
delete[] result;
delete d;
}
This driver is called like the others from Erlang, however, since we use driver_output_term, there is no need to call binary_to_term. The Erlang code is in the sample file next_perm.erl.
The input is changed into a list of integers and sent to the driver.
-module(next_perm).
-export([next_perm/1, prev_perm/1, load/0, all_perm/1]).
load() ->
case whereis(next_perm) of
undefined ->
case erl_ddll:load_driver(".", "next_perm") of
ok -> ok;
{error, already_loaded} -> ok;
E -> exit(E)
end,
Port = open_port({spawn, "next_perm"}, []),
register(next_perm, Port);
_ ->
ok
end.
list_to_integer_binaries(L) ->
[<<I:32/integer-native>> || I <- L].
next_perm(L) ->
next_perm(L, 1).
prev_perm(L) ->
next_perm(L, 2).
next_perm(L, Nxt) ->
load(),
B = list_to_integer_binaries(L),
port_control(next_perm, Nxt, B),
receive
Result ->
Result
end.
all_perm(L) ->
New = prev_perm(L),
all_perm(New, L, [New]).
all_perm(L, L, Acc) ->
Acc;
all_perm(L, Orig, Acc) ->
New = prev_perm(L),
all_perm(New, Orig, [New | Acc]).