Reputation: 20350
Upgraded Ubuntu today from 14.10 to 15.04. Now seeing different behaviour either in boost::asio::async_read()
, boost::asio::posix::stream_descriptor
, or tap/tun interfaces. Calling async_read()
immediately returns boost::asio::error::eof
. If I ignore the error and loop back up to start a new async_read()
it does eventually read when bytes are available, and the application continues to work.
The problem with doing this workaround loop is the application now consumes 100% of a core as it sits in a tight loop continuously restarting the call to async_read()
.
This is how I'm setting things up:
fd = open("/dev/net/tun", O_RDWR);
....
boost::asio::posix::stream_descriptor my_stream( io_service);
my_stream.assign(fd, ec);
...
boost::asio::async_read(my_stream, my_buffer, boost::asio::transfer_at_least(16),
[=](const EC &error, std::size_t bytes_read)
{
if (error) // <- this triggers with EOF error
Anyone know what may have changed in the newer kernels (tun/tap), or boost 1.55, to cause this end-of-file error when doing asynchronous reads?
Upvotes: 1
Views: 1441
Reputation: 51891
Ubuntu 15.04 contains the 3.19 kernel, which has a reported regression in the TUN/TAP user API:
With kernel 3.19, a
read()
from a TUN/TAP file descriptor in non-blocking mode will return0
when no data is available, rather than fail withEAGAIN
.
Per the documentation, the return value from read()
should only be 0
when no message has been read and the the peer has performed an orderly shutdown. Hence, the Boost.Asio implementation treats a return of 0
as an indication that the peer has shutdown, and completes the async_read()
operation with an error code of boost::asio::error::eof
:
// Read some data.
signed_size_type bytes = socket_ops::recv(s, bufs, count, flags, ec);
// Check for end of stream.
if (is_stream && bytes == 0)
{
ec = boost::asio::error::eof;
return true;
}
Upvotes: 2