Skip to content

Latest commit

 

History

History
111 lines (74 loc) · 2.97 KB

README.rst

File metadata and controls

111 lines (74 loc) · 2.97 KB
Raw

Test for poll() and select() behavior

Date: 2017-11-17

What is it?

This is a test for poll() and select() created for analyzing the behavior of SO_SELECT_ERR_QUEUE when polling for transmit timestamps on a socket's error queue.

This program should be used with strace or another tracing tool to visualize the behavior of the system calls.

When the sender thread is enabled with --multi-thread option, -f option should be passed to strace.

In normal behavior, the main thread send UDP packets to localhost and poll in the same (unbound) socket. This has no effect. Thus, options can be specified such as:

  • Binding the socket (--bind-socket).
  • Enabling transmit timestamps (--tx-timestamp).
  • Requesting read-is-possible event (--in).

See all options with $ main --help. The order of the options does not matter.

Compiling

$ make

Behavior

Date:2019-08-01

Wake up and return:

  • For waking up a task, the events triggered by kernel (wake up key) must contain one or more of the events requested by user.
  • For returning events to the user (after the task is woken or before the task sleeps), the result events (which is generally the same as the triggered events) must contain one or more requested events. Only the matched events are returned.

Result events

The result event for an error condition is POLLERR.

Triggered events

The events triggered by kernel on an error condition are:

  • Before commit 6e5d58fdc9bedd0255a8 ("skbuff: Fix not waking applications when errors are enqueued") in Linux 4.16:
    • POLLIN and POLLPRI via sk_data_ready() (sock_def_readable()).
  • After commit 6e5d58fdc9bedd0255a8:
    • POLLERR via sk_error_report() (sock_queue_err_skb()).

Requested events

The request events for poll(): [1]

  • POLLERR | POLLHUP | <user_choice>

The request events for select(): [2][3]

  • readfds: POLLPRI | POLLERR | POLLIN | POLLHUP
  • writefds: POLLPRI | POLLERR | POLLOUT
  • exeptfds: POLLPRI

See [4] for where select() checks whether the result events were requested by user.

[1] https://git.kernel.org/pub/scm/linux/kernel/git/stable/
linux.git/tree/fs/select.c?id=6e5d58fdc9bedd0255a8#n820
[2] https://git.kernel.org/pub/scm/linux/kernel/git/stable/
linux.git/tree/fs/select.c?id=6e5d58fdc9bedd0255a8#n435
[3] https://git.kernel.org/pub/scm/linux/kernel/git/stable/
linux.git/tree/fs/select.c?id=6e5d58fdc9bedd0255a8#n443
[4] https://git.kernel.org/pub/scm/linux/kernel/git/stable/
linux.git/tree/fs/select.c?id=6e5d58fdc9bedd0255a8#n512

SO_SELECT_ERR_QUEUE

The socket option SO_SELECT_ERR_QUEUE was introduced in Linux 3.10 by the commit 7d4c04fc170087119727 ("net: add option to enable error queue packets waking select"). It simply adds POLLPRI to the result events (which already contains POLLERR) before it is checked for return.