Skip to content

Optional limit on message size #43

Open

Description

@GiuseppeSaluto

Feature request: optional limit on message size (including after permessage-deflate inflation)

Greetings!

Some context

I maintain aisstream, a small client for the AISStream WebSocket API built on top of mint_web_socket. Someone recently asked for an option to cap the size of incoming messages, so a misbehaving server can't make the client buffer unbounded data (aisstream_elixir#8). While looking into it I ran into two things I can't really solve from the outside.

The problem

Right now Mint.WebSocket doesn't have a size limit at any stage:

  1. Buffering: if a frame header declares a huge payload (say 4 GB) and the data keeps trickling in, decode/2 keeps buffering while waiting for the frame to complete.
  2. Decompression: with Mint.WebSocket.PerMessageDeflate, the whole message gets inflated in one go after the last fragment (:zlib.inflate/2 on the full payload), with no upper bound. So a small compressed message can blow up into a very large binary.

I can work around the first one by counting the bytes I pass to decode/2. Mint.WebSocket.t() is opaque and I'd rather not poke at its internals. The second one I can't do anything about: by the time I see the result, inflation has already allocated it.

This matters in practice because some services expect compression. AISStream, for example, applies bandwidth limits to uncompressed connections, so turning permessage-deflate off isn't really an option.

What I'm thinking

An optional limit, something like:

Mint.WebSocket.new(conn, ref, status, headers, max_message_size: 1_048_576)

Roughly:

  • It applies to the data buffered for an incomplete message (fragments included) and to the inflated size when permessage-deflate is on.
  • Inflation stops as soon as the limit is hit, instead of inflating everything first. Maybe :zlib.safeInflate/2 in a loop?
  • On overflow, decode/2 returns an error (e.g. {:error, websocket, %Mint.WebSocket.WebSocketError{reason: :message_too_large}}), so the caller can close with 1009 and reconnect.
  • The default stays :infinity (or nil), so nothing changes for existing users.

Totally open on naming and shape, of course.

Prior art

  • mint itself recently fixed the same pattern for HTTP/1 chunked bodies, where partial chunk data was buffered without a bound until the declared chunk length arrived (CVE-2026-56810, fixed in 1.9.1). WebSocket frame buffering and inflation seem to have the same shape.
  • Python's websockets has max_size (1 MiB by default), enforced on the decompressed size too.
  • websocket-driver (Ruby/Node) had the same gap, with the limit checked against compressed frame sizes, and fixed it by checking after extensions run (CVE-2026-54464).

Happy to help

If the approach sounds reasonable to you, I'd be glad to take a stab at a PR!

Thanks for Mint, by the way, it's been a pleasure to build on! 🌱

Activity

  1. the-mikedavis commented on Oct 27, 2026

    @the-mikedavis
    Collaborator

    This idea makes sense to me 👍

    I'm not sure if we would want to set a default (like 1 MiB as you mention the Python library does) since it might be a breaking change for some users passing a lot of data across a websocket connection. Feel free to open up a PR and we can discuss it further!

  2. ericmj commented on Oct 27, 2026

    @ericmj
    Member

    I would suggest setting a default even if it's a breaking change since no default is inherently unsafe. As an example Plug defaults to 8MB for Plug.Parsers.

    https://plug.hexdocs.pm/Plug.Parsers.html

  3. GiuseppeSaluto commented on Oct 28, 2026

    @GiuseppeSaluto
    Author

    Hello @the-mikedavis , @ericmj
    Thanks both!
    I'll open a PR with a :max_message_size option applied to buffered data and to the inflated size.

    On the default: how about a generous one like Plug's 8 MB, with :infinity to opt out? That keeps the safe-by-default behaviour while only affecting users with really large messages, who can opt out in one line.
    I'll make the default a single constant so it's easy to change whichever way you decide.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions