How does HTTP file upload work?

When I submit a simple form like this with a file attached:

<form enctype="multipart/form-data" action="http://localhost:3000/upload?upload_progress_id=12344" method="POST">
<input type="hidden" name="MAX_FILE_SIZE" value="100000" />
Choose a file to upload: <input name="uploadedfile" type="file" /><br />
<input type="submit" value="Upload File" />
</form>

How does it send the file internally? Is the file sent as part of the HTTP body as data? In the headers of this request, I don't see anything related to the name of the file.

I just would like the know the internal workings of the HTTP when sending a file.

Solution 1:

Let's take a look at what happens when you select a file and submit your form (I've truncated the headers for brevity):

POST /upload?upload_progress_id=12344 HTTP/1.1
Host: localhost:3000
Content-Length: 1325
Origin: http://localhost:3000
... other headers ...
Content-Type: multipart/form-data; boundary=----WebKitFormBoundaryePkpFF7tjBAqx29L

------WebKitFormBoundaryePkpFF7tjBAqx29L
Content-Disposition: form-data; name="MAX_FILE_SIZE"

100000
------WebKitFormBoundaryePkpFF7tjBAqx29L
Content-Disposition: form-data; name="uploadedfile"; filename="hello.o"
Content-Type: application/x-object

... contents of file goes here ...
------WebKitFormBoundaryePkpFF7tjBAqx29L--

NOTE: each boundary string must be prefixed with an extra --, just like in the end of the last boundary string. The example above already includes this, but it can be easy to miss. See comment by @Andreas below.

Instead of URL encoding the form parameters, the form parameters (including the file data) are sent as sections in a multipart document in the body of the request.

In the example above, you can see the input MAX_FILE_SIZE with the value set in the form, as well as a section containing the file data. The file name is part of the Content-Disposition header.

The full details are here.

Solution 2:

How does it send the file internally?

The format is called multipart/form-data, as asked at: What does enctype='multipart/form-data' mean?

I'm going to:

add some more HTML5 references
explain why he is right with a form submit example

HTML5 references

There are three possibilities for enctype:

x-www-urlencoded
multipart/form-data (spec points to RFC2388)
text-plain. This is "not reliably interpretable by computer", so it should never be used in production, and we will not look further into it.

How to generate the examples

Once you see an example of each method, it becomes obvious how they work, and when you should use each one.

You can produce examples using:

nc -l or an ECHO server: HTTP test server accepting GET/POST requests
an user agent like a browser or cURL

Save the form to a minimal .html file:

<!DOCTYPE html>
<html lang="en">
<head>
  <meta charset="utf-8"/>
  <title>upload</title>
</head>
<body>
  <form action="http://localhost:8000" method="post" enctype="multipart/form-data">
  <p><input type="text" name="text1" value="text default">
  <p><input type="text" name="text2" value="a&#x03C9;b">
  <p><input type="file" name="file1">
  <p><input type="file" name="file2">
  <p><input type="file" name="file3">
  <p><button type="submit">Submit</button>
</form>
</body>
</html>

We set the default text value to aωb, which means aωb because ω is U+03C9, which are the bytes 61 CF 89 62 in UTF-8.

Create files to upload:

echo 'Content of a.txt.' > a.txt

echo '<!DOCTYPE html><title>Content of a.html.</title>' > a.html

# Binary file containing 4 bytes: 'a', 1, 2 and 'b'.
printf 'a\xCF\x89b' > binary

Run our little echo server:

while true; do printf '' | nc -l 8000 localhost; done

Open the HTML on your browser, select the files and click on submit and check the terminal.

nc prints the request received.

Tested on: Ubuntu 14.04.3, nc BSD 1.105, Firefox 40.

multipart/form-data

Firefox sent:

POST / HTTP/1.1
[[ Less interesting headers ... ]]
Content-Type: multipart/form-data; boundary=---------------------------735323031399963166993862150
Content-Length: 834

-----------------------------735323031399963166993862150
Content-Disposition: form-data; name="text1"

text default
-----------------------------735323031399963166993862150
Content-Disposition: form-data; name="text2"

aωb
-----------------------------735323031399963166993862150
Content-Disposition: form-data; name="file1"; filename="a.txt"
Content-Type: text/plain

Content of a.txt.

-----------------------------735323031399963166993862150
Content-Disposition: form-data; name="file2"; filename="a.html"
Content-Type: text/html

<!DOCTYPE html><title>Content of a.html.</title>

-----------------------------735323031399963166993862150
Content-Disposition: form-data; name="file3"; filename="binary"
Content-Type: application/octet-stream

aωb
-----------------------------735323031399963166993862150--

For the binary file and text field, the bytes 61 CF 89 62 (aωb in UTF-8) are sent literally. You could verify that with nc -l localhost 8000 | hd, which says that the bytes:

61 CF 89 62

were sent (61 == 'a' and 62 == 'b').

Therefore it is clear that:

Content-Type: multipart/form-data; boundary=---------------------------735323031399963166993862150 sets the content type to multipart/form-data and says that the fields are separated by the given boundary string.

But note that the:
```
boundary=---------------------------735323031399963166993862150
```
has two less dadhes -- than the actual barrier
```
-----------------------------735323031399963166993862150
```
This is because the standard requires the boundary to start with two dashes --. The other dashes appear to be just how Firefox chose to implement the arbitrary boundary. RFC 7578 clearly mentions that those two leading dashes -- are required:

4.1. "Boundary" Parameter of multipart/form-data

As with other multipart types, the parts are delimited with a boundary delimiter, constructed using CRLF, "--", and the value of the "boundary" parameter.

every field gets some sub headers before its data: Content-Disposition: form-data;, the field name, the filename, followed by the data.

The server reads the data until the next boundary string. The browser must choose a boundary that will not appear in any of the fields, so this is why the boundary may vary between requests.

Because we have the unique boundary, no encoding of the data is necessary: binary data is sent as is.

TODO: what is the optimal boundary size (log(N) I bet), and name / running time of the algorithm that finds it? Asked at: https://cs.stackexchange.com/questions/39687/find-the-shortest-sequence-that-is-not-a-sub-sequence-of-a-set-of-sequences
Content-Type is automatically determined by the browser.

How it is determined exactly was asked at: How is mime type of an uploaded file determined by browser?

application/x-www-form-urlencoded

Now change the enctype to application/x-www-form-urlencoded, reload the browser, and resubmit.

Firefox sent:

POST / HTTP/1.1
[[ Less interesting headers ... ]]
Content-Type: application/x-www-form-urlencoded
Content-Length: 51

text1=text+default&text2=a%CF%89b&file1=a.txt&file2=a.html&file3=binary

Clearly the file data was not sent, only the basenames. So this cannot be used for files.

As for the text field, we see that usual printable characters like a and b were sent in one byte, while non-printable ones like 0xCF and 0x89 took up 3 bytes each: %CF%89!

Comparison

File uploads often contain lots of non-printable characters (e.g. images), while text forms almost never do.

From the examples we have seen that:

multipart/form-data: adds a few bytes of boundary overhead to the message, and must spend some time calculating it, but sends each byte in one byte.
application/x-www-form-urlencoded: has a single byte boundary per field (&), but adds a linear overhead factor of 3x for every non-printable character.

Therefore, even if we could send files with application/x-www-form-urlencoded, we wouldn't want to, because it is so inefficient.

But for printable characters found in text fields, it does not matter and generates less overhead, so we just use it.