For cases with dual stack (IPv4 + IPv6) connectivity, but where one
stack potentially is less reliable, strive to trying to connect over
both protocols in parallel, using whichever address connected first.
In cases with a hostname resolving to multiple IPv4 and IPv6
addresses, the current connection mechanism would try all addresses
in the order returned by getaddrinfo (with all IPv6 addresses ordered
before the IPv4 addresses normally). If connection attempts to the
IPv6 addresses return quickly with an error, this was no problem, but
if they were unsuccessful leading up to timeouts, the connection process
would have to wait for timeouts on all IPv6 target addresses before
attempting any IPv4 address.
Similar to what RFC 8305 suggests, reorder the list of addresses to
try connecting to, interleaving address families. After starting one
connection attempt, start another one in parallel after a small delay
(200 ms as suggested by the RFC).
For cases with unreliable IPv6 but reliable IPv4, this should make
connection attempts work as reliably as with plain IPv4, with only an
extra 200 ms of connection delay.
Signed-off-by: Martin Storsjö <martin@martin.st>
It was sort of optional before - if you didn't call it, networking was
initialized on demand, and an ugly warning was logged. Also, the doxygen
comments threatened that it would be made strictly required one day.
Make it explicitly optional. I would prefer to deprecate it fully, but
there might still be legitimate reasons to use this. But the average
user won't need it.
This is needed only for two reasons: to initialize TLS libraries like
OpenSSL and GnuTLS, and winsock.
OpenSSL and GnuTLS were already silently initialized on demand if the
global network init function was not called. They also have various
thread-safety acrobatics, which make concurrent initialization within
libavformat safe. In addition, the libraries are moving towards making
their global init functions safe, which removes all need for central
global init. In particular, GnuTLS 3.5.16 and OpenSSL 1.1.0g have been
found to have safe init functions. In all cases, they use internal
reference counters to avoid that the global uninit functions interfere
with concurrent uses of the library by other API users who called global
init.
winsock should be thread-safe as well, and maintains an internal
reference counter as well.
Since we still support ancient TLS libraries, which do not have this
fixed, and since it's unknown whether winsock and GnuTLS
reinitialization is costly in any way, don't deprecate the libavformat
functions yet.
It makes no sense to return an error after the first reconnect, and then
somehow resume the next time it's called. Usually this will lead to
demuxer errors. Make reconnecting block instead, until it has either
successfully reconnected, or given up.
Also make the wait reasonably interruptible. Since there is no mechanism
for this in the API, polling is the best we can do. This behaves roughly
the same as other interruptible network functions in libavformat.
(The original code would work if it returned AVERROR(EAGAIN) or so,
which would make retry_transfer_wrapper() repeat the read call. But I
think having an explicit loop for this is better anyway.)
I also snuck in a fix for reconnect_at_eof. It has to check for
AVERROR_EOF, not 0.
This can reduce latency and increase throughput, particularly on high
latency networks.
Signed-off-by: Aman Gupta <aman@tmm1.net>
Reviewed-by: Jeyapal, Karthick <kjeyapal@akamai.com>
If the remote end of a connection oriented socket hangs up, generating
an EPIPE error is preferable over an unhandled SIGPIPE signal.
Signed-off-by: Martin Storsjö <martin@martin.st>
This supports non-Linux systems (SOCK_CLOEXEC is non-standard) and
older Linux kernels to the extent possible.
Signed-off-by: Martin Storsjö <martin@martin.st>
This lowers the level of warnings printed if trying to connect
to a host name that provides both v6 and v4 addresses but the
service only is available on the v4 address (often occurring for
'localhost', with servers that aren't v6-aware).
Signed-off-by: Martin Storsjö <martin@martin.st>
Even if linking directly to getaddrinfo, use our version of
gai_strerror instead of the system's version. Microsoft explicitly
documents that their version of gai_strerror is thread-unsafe.
Signed-off-by: Martin Storsjö <martin@martin.st>
This avoids warnings if there already exists a definition.
This is the case on windows, where the getaddrinfo isn't available
and linked to (and we use our fallbacks instead, which actually
try to use the proper getaddrinfo version if found at runtime),
but gai_strerror still exists as a define.
Signed-off-by: Martin Storsjö <martin@martin.st>
This is useful if a proper getaddrinfo is loaded dynamically on
windows, while using the fallback implementation of gai_strerror.
Signed-off-by: Martin Storsjö <martin@martin.st>
Since the errno.h values don't match the error codes that winsock
returns, map the winsock error codes to the errno ones, to make
sure explicit checks against AVERROR(x) match.
Signed-off-by: Martin Storsjö <martin@martin.st>
This heaader is required for close() for sockets in network
code. For winsock, the equivalent function is defined in the
winsock2.h header.
This avoids having the HAVE_UNISTD_H in all files dealing with
raw sockets.
Signed-off-by: Martin Storsjö <martin@martin.st>
This is to make developers aware of the fact that they will
start using the new init function at some point.
Signed-off-by: Martin Storsjö <martin@martin.st>
Previously, the function would lead to an infinite wait (by
returning AVERROR(EAGAIN)) on sockets indicating an error
via either of these poll flags.
Signed-off-by: Martin Storsjö <martin@martin.st>
Support the URL_FLAG_NONBLOCK semantic and uniform the protocol.
The quick retry loop is already part of retry_transfer_wrapper.
The polling routine is common to the network protocols:
udp, tcp and, once merged, sctp.
Map EAGAIN and EINTR from ff_neterrno to the normal AVERROR()
error codes. Provide fallback definitions of other errno.h network
errors, mapping them to the corresponding winsock errors.
This eases catching these error codes in common code, without having
to distinguish between FF_NETERRNO(EAGAIN) and AVERROR(EAGAIN).
This fixes roundup issue 2614, unbreaking blocking network IO on
windows.
Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
(cherry picked from commit 28c4741a66)
Map EAGAIN and EINTR from ff_neterrno to the normal AVERROR()
error codes. Provide fallback definitions of other errno.h network
errors, mapping them to the corresponding winsock errors.
This eases catching these error codes in common code, without having
to distinguish between FF_NETERRNO(EAGAIN) and AVERROR(EAGAIN).
This fixes roundup issue 2614, unbreaking blocking network IO on
windows.
Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
Not all users of this function will have a full struct sockaddr_storage
available, and casting other sockaddrs to sockaddr_storage is wrong,
while any sockaddr can be cast to a base sockaddr.
Originally committed as revision 25388 to svn://svn.ffmpeg.org/ffmpeg/trunk
The header probably isn't the right permanent place for this function, but
it is quite small, and consensus seems to be that it can stay in the
header for now, instead of creating a new file network.c just for this one.
Originally committed as revision 25387 to svn://svn.ffmpeg.org/ffmpeg/trunk
for broken OSes. This is included in rtsp.h, as opposed to os_support.h.
Should fix OS/2 broken build on fate.
Originally committed as revision 25035 to svn://svn.ffmpeg.org/ffmpeg/trunk
jackarain
Anton Khirnov
Martin Storsjö
Jun Zhao
wm4
Rostislav Pehlivanov
Aman Gupta
Stephan Holljes
Michael Niedermayer
Rémi Denis-Courmont
Diego Biurrun
Luca Barbato
Andrey Utkin
Mans Rullgard
Ronald S. Bultje
Vincent Torri
Gil Pedersen