Let's dive deep into Plug.Conn.send_file/5
Let’s dive into Plug.Conn.send_file/5, because we can find a lot of questioning online about this one. Is it efficient ? Does it use a linux syscall to send the file (yes) ?
⚠️ This not a fun post, but a giant string of jumping from file to file, first in a Phoenix project, then in Plug’s source, then in Cowboy, then in Ranch, then in Erlang’s source, then in C NIFs in Erlang’s source. Still it was interesting to encounter and read all of that code and see a bit more of the inner working of Erlang.
The call sites you’ll most frequently encounter look like this :
conn |> send_file(200, filename)
Digging into the source, we can see two optional parameters, offset (with the default value 0) and length (with the default value :all).
Our basic usage then looks like this with the defaults filled in :
conn |> send_file(200, filename, 0, :all)
Before the concrete implementation, the source contains a few guards :
🔗 source elixir-plug/plug/lib/plug/conn.ex:473
def send_file(%Conn{state: state}, status, _file, _offset, _length)
when not (state in @unsent) do
_ = Plug.Conn.Status.code(status)
raise AlreadySentError
end
If the Plug.Conn struct has been marked as sent, a Plug.Conn.AlreadySentError is raised. If the Plug.Conn struct has not been sent yet, we can move on to the next function head:
🔗 source elixir-plug/plug/lib/plug/conn.ex:479
def send_file(
%Conn{adapter: {adapter, payload}, owner: owner} = conn,
status,
file,
offset,
length
)
when is_binary(file)
A first check is made that the provided file name consists of a binary. The body then checks that the provided filename isn’t a null byte, and raises an ArgumentError if that was the case.
🔗 source elixir-plug/plug/lib/plug/conn.ex:487
if file =~ "\0" do
raise ArgumentError, "cannot send_file/5 with null byte"
end
The next line updates the Plug.Conn struct by calling the private function run_before_send/2 :
🔗 source elixir-plug/plug/lib/plug/conn.ex:491
conn =
run_before_send(%{conn | status: Plug.Conn.Status.code(status), resp_body: nil}, :set_file)
This function reduces a list of functions taking and returning a Plug.Conn, accumulated into the Plug.Conn struct’s :before_send key, over the Plug.Conn struct whose :state key has been updated to :new. In our case, the :state key has the value :set_file.
🔗 source elixir-plug/plug/lib/plug/conn.ex:1835
defp run_before_send(%Conn{before_send: before_send} = conn, new) do
conn = Enum.reduce(before_send, %{conn | state: new}, & &1.(&2))
if conn.state != new do
raise ArgumentError, "cannot send/change response from run_before_send callback"
end
%{conn | resp_headers: merge_headers(conn.resp_headers, conn.resp_cookies)}
end
If the resulting Plug.Conn struct :state key has been changed by one of the :before_send callbacks, an ArgumentError is raised because those :before_send callbacks aren’t allowed to perform this operation.
The response headers are then updated via another private function, merge_headers/2 which is out of scope for this post.
Back to the implementation of send_file/5 :
🔗 source elixir-plug/plug/lib/plug/conn.ex:494
{:ok, body, payload} =
adapter.send_file(payload, conn.status, conn.resp_headers, file, offset, length)
The actual implementation of send_file/5 is delegated to the Plug.Conn’s adapter’s send_file/6, of which we find the signature in Plug.Conn.Adapter :
🔗 source elixir-plug/plug/lib/plug/conn/adapter.ex:73
@doc """
Sends the given status, headers and file as a response
back to the client.
If the request has method `"HEAD"`, the adapter should
not send the response to the client.
Webservers are advised to return `nil` as the sent_body,
as the body can no longer be manipulated. However, the
test implementation returns the actual body so it can
be used during testing.
"""
@callback send_file(
payload,
status :: Conn.status(),
headers :: Conn.headers(),
file :: binary,
offset :: integer,
length :: integer | :all
) :: {:ok, sent_body :: binary | nil, payload}
In my case, Cowboy is used as a web server. We jump into Plug.Cowboy.Conn which implements the behaviour Plug.Conn.Adapter, and find the implementation of send_file/6 :
🔗 source elixir-plug/plug_cowboy/lib/plug/cowboy/conn.ex:40
@impl true
def send_file(req, status, headers, path, offset, length) do
%File.Stat{type: :regular, size: size} = File.stat!(path)
length =
cond do
length == :all -> size
is_integer(length) -> length
end
body = {:sendfile, offset, length, path}
headers = to_headers_map(headers)
req = :cowboy_req.reply(status, headers, body, req)
{:ok, nil, req}
end
It is first checked that the file is a regular file (as opposed to a directory, a symlink..). The length of the file to actually send is either the size of the file in case we passed :all as the length parameter, or the supplied length after we checked it was an integer.
:cowboy_req.reply is then called, the concrete call in the case of our response being sent with send_file(conn, 200, filename) becoming :cowboy_req.reply("200", headers, {:sendfile, offset, length, path}, req)
We then follow this track to the cowboy_req.erl file defining the module cowboy_req, to find the reply/4 function head :
If you are not used to read Erlang, a few things can maybe be off-putting :
Keeping in mind that variables are Uppercased and atoms are lowercase helps clear it up.
Here, Req is a variable and function_clause is an atom.
🔗 source ninenines/cowboy/src/cowboy_req.erl:793
-spec reply(cowboy:http_status(), cowboy:http_headers(), resp_body(), Req)
-> Req when Req::req().
reply(_, _, _, #{has_sent_resp := _}) ->
error(function_clause); %% @todo Better error message.
reply(Status, Headers, {sendfile, _, 0, _}, Req)
when is_integer(Status); is_binary(Status) ->
do_reply(Status, Headers#{
<<"content-length">> => <<"0">>
}, <<>>, Req);
reply(Status, Headers, SendFile = {sendfile, _, Len, _}, Req)
when is_integer(Status); is_binary(Status) ->
do_reply(Status, Headers#{
<<"content-length">> => integer_to_binary(Len)
}, SendFile, Req);
Going through the clauses one by one :
- if the response has already been sent, an error is raised
- if the length of the file is 0, the “Content-Length” header is set to 0 before calling
do_reply/4 - finally, if the file has a non-zero length, do_reply/4 is called, setting the “Content-Length” header to the length of the file
Following to do_reply/4 :
🔗 source ninenines/cowboy/src/cowboy_req.erl:832
do_reply(Status, Headers, Body, Req) ->
cast({response, Status, response_headers(Headers, Req), Body}, Req),
done_replying(Req, true).
done_replying(Req, HasSentResp) ->
maps:without([resp_cookies, resp_headers, resp_body], Req#{has_sent_resp => HasSentResp}).
A message of {response, "200", headers, {sendfile, Offset, Length, Path}} is sent via cast/2 to the Pid held in the Req argument :
🔗 source ninenines/cowboy/src/cowboy_req.erl:921
%% Stream handlers.
-spec cast(any(), req()) -> ok.
cast(Msg, #{pid := Pid, streamid := StreamID}) ->
Pid ! {{Pid, StreamID}, Msg},
ok.
Continuing my exploration, I find that this message is received by cowboy_http in a long receive block :
🔗 source ninenines/cowboy/src/cowboy_http.erl:255
%% ...
%% Messages pertaining to a stream.
{{Pid, StreamID}, Msg} when Pid =:= self() ->
loop(info(State, StreamID, Msg));
%% ...
Going to the definition of info in this module :
🔗 source ninenines/cowboy/src/cowboy_http.erl:934
info(State=#state{opts=Opts, streams=Streams0}, StreamID, Msg) ->
case lists:keyfind(StreamID, #stream.id, Streams0) of
Stream = #stream{state=StreamState0} ->
try cowboy_stream:info(StreamID, Msg, StreamState0) of
{Commands, StreamState} ->
Streams = lists:keyreplace(StreamID, #stream.id, Streams0,
Stream#stream{state=StreamState}),
commands(State#state{streams=Streams}, StreamID, Commands)
catch Class:Exception:Stacktrace ->
cowboy:log(cowboy_stream:make_error_log(info,
[StreamID, Msg, StreamState0],
Class, Exception, Stacktrace), Opts),
stream_terminate(State, StreamID, {internal_error, {Class, Exception},
'Unhandled exception in cowboy_stream:info/3.'})
end;
false ->
cowboy:log(warning, "Received message ~p for unknown stream ~p.~n",
[Msg, StreamID], Opts),
State
end.
We find a bit of error handling before delegating a call to commands/3 in the same module. Two definitions of commands/3 contain references to sendfile :
🔗 source ninenines/cowboy/src/cowboy_http.erl:1028
%% Send a full response.
%%
%% @todo Kill the stream if it sent a response when one has already been sent.
%% @todo Keep IsFin in the state.
%% @todo Same two things above apply to DATA, possibly promise too.
commands(State0=#state{socket=Socket, transport=Transport, out_state=wait, streams=Streams}, StreamID,
[{response, StatusCode, Headers0, Body}|Tail]) ->
and
🔗 source ninenines/cowboy/src/cowboy_http.erl:1084
%% Send a response body chunk.
%% @todo We need to kill the stream if it tries to send data before headers.
commands(State0=#state{socket=Socket, transport=Transport, streams=Streams0, out_state=OutState},
StreamID, [{data, IsFin, Data}|Tail]) ->
In both cases, a call to sendfile/2 in the same module is issued :
🔗 source ninenines/cowboy/src/cowboy_http.erl:1237
%% We wrap the sendfile call into a try/catch because on OTP-20
%% and earlier a few different crashes could occur for sockets
%% that were closing or closed. For example a badarg in
%% erlang:port_get_data(#Port<...>) or a badmatch like
%% {{badmatch,{error,einval}},[{prim_file,sendfile,8,[]}...
%%
%% OTP-21 uses a NIF instead of a port so the implementation
%% and behavior has dramatically changed and it is unclear
%% whether it will be necessary in the future.
%%
%% This try/catch prevents some noisy logs to be written
%% when these errors occur.
sendfile(State=#state{socket=Socket, transport=Transport, opts=Opts},
{sendfile, Offset, Bytes, Path}) ->
try
%% When sendfile is disabled we explicitly use the fallback.
_ = case maps:get(sendfile, Opts, true) of
true -> Transport:sendfile(Socket, Path, Offset, Bytes);
false -> ranch_transport:sendfile(Transport, Socket, Path, Offset, Bytes, [])
end,
ok
catch _:_ ->
terminate(State, {socket_error, sendfile_crash,
'An error occurred when using the sendfile function.'})
end.
It eithers delegates the sending of the file to a transport-specific sendfile/4, or, if the ability to use the sendfile syscall isn’t available, delegates to ranch_transport:sendfile/6.
In ranch_transport’s documentation found here we can read :
Transports that manipulate TCP directly may use the file:sendfile/2,4,5 function, which calls the sendfile syscall where applicable (on Linux, for example). Other transports can use the sendfile/6 function exported from this module.
Let’s then follow to erlang’s file module, where we find two functions exported :
erlang/otp/lib/kernel/src/file.erl:60
%% Sendfile functions
-export([sendfile/2,sendfile/5]).
The first sends a whole file on a socket, and the second sends a chunk of a file.
From otp/lib/kernel/src/file.erl, we can see that the actual sending then goes through socket:sendfile/5, in a private sendfile/8 function defined in file :
🔗 source erlang/otp/lib/kernel/src/file.erl:1332
%% Internal sendfile functions
sendfile(#file_descriptor{ module = Mod } = Fd, Sock, Offset, Bytes,
ChunkSize, Headers, Trailers, Opts)
when is_integer(Offset), is_integer(Bytes) ->
case Sock of
?socket(_) when Headers =:= [], Trailers =:= [] ->
try socket:sendfile(Sock, Fd, Offset, Bytes, infinity)
catch error : notsup ->
sendfile_fallback(
Fd, socket_send(Sock), Offset, Bytes, ChunkSize,
Headers, Trailers)
end;
?socket(_) ->
sendfile_fallback(
Fd, socket_send(Sock), Offset, Bytes, ChunkSize,
Headers, Trailers);
Following to the socket module, we find sendfile/5, and learn that it handles sending the file via a NIF :
🔗 source erlang/otp/lib/kernel/src/socket.erl:3064
sendfile(
?socket(SockRef) = Socket, FileHandle_Cont, Offset, Count, Timeout)
when is_integer(Offset), is_integer(Count), 0 =< Count ->
%%
case FileHandle_Cont of
#file_descriptor{module = Module} = FileHandle ->
GetFRef = internal_get_nif_resource,
try Module:GetFRef(FileHandle) of
FRef ->
State = {FRef, Offset, Count},
sendfile_int(SockRef, State, Timeout)
catch
%% We could just crash here, since the caller
%% maybe broke the API and did not provide
%% a raw file as FileHandle, i.e GetFRef
%% is not implemented in Module;
%% but instead handle that nicely
After following a few calls in this function in socket.erl, we land on otp/erts/preloaded/src/prim_socket.erl and find further evidence of the use of a NIF :
🔗 source erlang/otp/erts/preloaded/src/prim_socket.erl:691
sendfile(SockRef, Offset, Count, SendRef) ->
nif_sendfile(SockRef, SendRef, Offset, Count).
sendfile(SockRef, FileRef, Offset, Count, SendRef) ->
nif_sendfile(SockRef, SendRef, Offset, Count, FileRef).
sendfile_deferred_close(SockRef) ->
nif_sendfile(SockRef).
That leads us to the file otp/erts/emulator/nifs/common/prim_socket_nif.c, that contains the definitions of nif_sendfile/1,4,5 :
🔗 source erlang/otp/erts/emulator/nifs/common/prim_socket_nif.c:5841
/* ----------------------------------------------------------------------
* nif_sendfile/1,4,5
*
* Description:
* Send a file on a socket
*
* Arguments:
* Socket (ref) - Points to the socket descriptor.
*
* SendRef - A unique id reference() for this (send) request.
*
* Offset - File offset to start from.
* Count - The number of bytes to send.
*
* InFileRef - A file NIF resource.
*/
Following a bit in the source, we can find these three macros :
🔗 source erlang/otp/erts/emulator/nifs/common/prim_socket_nif.c:5939
res = ESOCK_IO_SENDFILE_DC(env, descP);
// and
res = ESOCK_IO_SENDFILE_CONT(env, descP,
sockRef, sendRef,
offset, count);
// and
res = ESOCK_IO_SENDFILE_START(env, descP,
sockRef, sendRef,
offset, count, fRef);
they get rewritten as follows :
🔗 source erlang/otp/erts/emulator/nifs/common/prim_socket_nif.c:2581
#define ESOCK_IO_SENDFILE_START(ENV, D, \
SOR, SNR, \
O, CN, FR) \
((io_backend.sendfile_start != NULL) ? \
io_backend.sendfile_start((ENV), (D), \
(SOR), (SNR), \
(O), (CN), (FR)) : \
enif_raise_exception((ENV), MKA((ENV), "notsup")))
#define ESOCK_IO_SENDFILE_CONT(ENV, D, \
SOR, SNR, \
O, CP) \
((io_backend.sendfile_cont != NULL) ? \
io_backend.sendfile_cont((ENV), (D), \
(SOR), (SNR), \
(O), (CP)) : \
enif_raise_exception((ENV), MKA((ENV), "notsup")))
#define ESOCK_IO_SENDFILE_DC(ENV, D) \
((io_backend.sendfile_dc != NULL) ? \
io_backend.sendfile_dc((ENV), (D)) : \
enif_raise_exception((ENV), MKA((ENV), "notsup")))
Looking into erlang’s source for sendfile_start brings us to otp/erts/emulator/nifs/unix/unix_socket_syncio.c :
🔗 source erlang/otp/erts/emulator/nifs/unix/unix_socket_syncio.c:2286
/* ========================================================================
* Start a sendfile() operation
*/
extern
ERL_NIF_TERM essio_sendfile_start(ErlNifEnv* env,
ESockDescriptor* descP,
ERL_NIF_TERM sockRef,
ERL_NIF_TERM sendRef,
off_t offset,
size_t count,
ERL_NIF_TERM fRef)
{
This points us to essio_sendfile in this same file :
🔗 source erlang/otp/erts/emulator/nifs/unix/unix_socket_syncio.c:6103
/* *** Sendfile utility functions *** */
/* Platform independent sendfile() function
*
* Return < 0 for terminal error
* 0 for done
* > 0 for retry with select
*/
#if defined(HAVE_SENDFILE)
static
int essio_sendfile(ErlNifEnv* env,
ESockDescriptor* descP,
ERL_NIF_TERM sockRef,
off_t offset,
size_t* countP,
int* errP)
{
We finally find a call to sendfile as defined in sys/sendfile.h :
🔗 source erlang/otp/erts/emulator/nifs/unix/unix_socket_syncio.c:6144
#if defined (__linux__)
off_t prev_offset;
prev_offset = offset;
res =
sendfile(descP->sock, descP->sendfileHandle,
&offset, chunk_size);
error = (res < 0) ? sock_errno() : 0;
Sendfile is documented here : https://man7.org/linux/man-pages/man2/sendfile.2.html To quote this page :
sendfile() copies data between one file descriptor and another. Because this copying is done within the kernel, sendfile() is more efficient than the combination of read(2) and write(2), which would require transferring data to and from user space.
So.. yeah, sending a file with Plug.Conn.send_file is pretty efficient provided you’re running linux.
Sometimes it’s about the journey, even if the destination is well-known ;-) .