[tcp] Send TCP keepalives on idle established connections
In some circumstances, intermediate devices may lose state in a way
that temporarily prevents the successful delivery of packets from a
TCP peer. For example, a firewall may drop a NAT forwarding table
entry.
Since iPXE spends most of its time downloading files (and hence purely
receiving data, sending only TCP ACKs), this can easily happen in a
situation in which there is no reason for iPXE's TCP stack to generate
any retransmissions. The temporary loss of connectivity can therefore
effectively become permanent.
Work around this problem by sending TCP keepalives after a period of
inactivity on an established connection.
TCP keepalives usually send a single garbage byte in sequence number
space that has already been ACKed by the peer. Since we do not need
to elicit a response from the peer, we instead send pure ACKs (with no
garbage data) in order to keep the transmit code path simple.
Originally-implemented-by: Ladi Prosek <lprosek@redhat.com>
Debugged-by: Ladi Prosek <lprosek@redhat.com>
Signed-off-by: Michael Brown <mcb30@ipxe.org>
[tcp] Gracefully close connections during shutdown
We currently do not wait for a received FIN before exiting to boot a
loaded OS. In the common case of booting from an HTTP server, this
means that the TCP connection is left consuming resources on the
server side: the server will retransmit the FIN several times before
giving up.
Fix by initiating a graceful close of all TCP connections and waiting
(for up to one second) for all connections to finish closing
gracefully (i.e. for the outgoing FIN to have been sent and ACKed, and
for the incoming FIN to have been received and ACKed at least once).
Signed-off-by: Michael Brown <mcb30@ipxe.org>
[tcp] Implement support for TCP Selective Acknowledgements (SACK)
The TCP Selective Acknowledgement option (specified in RFC2018)
provides a mechanism for the receiver to indicate packets that have
been received out of order (e.g. due to earlier dropped packets).
iPXE often operates in environments in which there is a high
probability of packet loss. For example, the legacy USB keyboard
emulation in some BIOSes involves polling the USB bus from within a
system management interrupt: this introduces an invisible delay of
around 500us which is long enough for around 40 full-length packets to
be dropped. Similarly, almost all 1Gbps USB2 devices will eventually
end up dropping packets because the USB2 bus does not provide enough
bandwidth to sustain a 1Gbps stream, and most devices will not provide
enough internal buffering to hold a full TCP window's worth of
received packets.
Add support for sending TCP Selective Acknowledgements. This provides
the sender with more detailed information about which packets have
been lost, and so allows for a more efficient retransmission strategy.
We include a SACK-permitted option in our SYN packet, since
experimentation shows that at least Linux peers will not include a
SACK-permitted option in the SYN-ACK packet if one was not present in
the initial SYN. (RFC2018 does not seem to mandate this behaviour,
but it is consistent with the approach taken in RFC1323.) We ignore
any received SACK options; this is safe to do since SACK is only ever
advisory and we never have to send non-trivial amounts of data.
Since our TCP receive queue is a candidate for cache discarding under
low memory conditions, we may end up discarding data that has been
reported as received via a SACK option. This is permitted by RFC2018.
We follow the stricture that SACK blocks must not report data which is
no longer held by the receiver: previously-reported blocks are
validated against the current receive queue before being included
within the current SACK block list.
Experiments in a qemu VM using forced packet drops (by setting
NETDEV_DISCARD_RATE to 32) show that implementing SACK improves
throughput by around 400%.
Experiments with a USB2 NIC (an SMSC7500) show that implementing SACK
improves throughput by around 700%, increasing the download rate from
35Mbps up to 250Mbps (which is approximately the usable bandwidth
limit for USB2).
Signed-off-by: Michael Brown <mcb30@ipxe.org>
[legal] Relicense files under GPL2_OR_LATER_OR_UBDL
These files cannot be automatically relicensed by util/relicense.pl
since they either contain unusual but trivial contributions (such as
the addition of __nonnull function attributes), or contain lines
dating back to the initial git revision (and so require manual
knowledge of the code's origin).
Signed-off-by: Michael Brown <mcb30@ipxe.org>
iPXE currently advertises a fixed MSS of 1460, which is correct only
for IPv4 over Ethernet. For IPv6 over Ethernet, the value should be
1440 (allowing for the larger IPv6 header). For non-Ethernet link
layers, the value should reflect the MTU of the underlying network
device.
Use tcpip_mtu() to calculate the transport-layer MTU associated with
the peer address, and calculate the MSS to allow for an optionless TCP
header as per RFC 6691.
As a side benefit, we can now fail a connection immediately with a
meaningful error message if we have no route to the destination
address.
Reported-by: Anton D. Kachalov <mouse@yandex-team.ru>
Signed-off-by: Michael Brown <mcb30@ipxe.org>
The path MTU is currently hardcoded to 1460 bytes, which fails to
allow space for TCP options. Sending a maximum-sized datagram (which
is viable when using HTTP POST) will therefore fail since the Ethernet
MTU will be exceeded.
Reduce the hardcoded path MTU to produce a maximum datagram of 1280
bytes, which is the size required of data link layers by IPv6. It is
a reasonable assumption that all intermediary data link layers will be
able to convey this packet without fragmentation, even for IPv4.
Note that this reduction has a minimal impact upon download
throughput, since it affects only the transmit data path.
Originally-fixed-by: Suresh Sundriyal <ssundriy@vmware.com>
Signed-off-by: Michael Brown <mcb30@ipxe.org>
A window size of 256kB should be sufficient to allow for
full-bandwidth transfers over a Gigabit LAN, and for acceptable
transfer speeds over other typical links.
Signed-off-by: Michael Brown <mcb30@ipxe.org>
The maximum unscaled TCP window (64kB) implies a maximum bandwidth of
around 300kB/s on a WAN link with an RTT of 200ms. Add support for
the TCP window scaling option to remove this upper limit.
Signed-off-by: Michael Brown <mcb30@ipxe.org>
TCP currently neglects to allow sufficient space for its own headers
when allocating I/O buffers. This problem is masked by the fact that
the maximum link-layer header size (802.11) is substantially larger
than the common Ethernet link-layer header.
Fix by allowing sufficient space for any TCP headers, as well as the
network-layer and link-layer headers.
Reported-by: Scott K Logan <logans@cottsay.net>
Debugged-by: Scott K Logan <logans@cottsay.net>
Tested-by: Scott K Logan <logans@cottsay.net>
Signed-off-by: Michael Brown <mcb30@ipxe.org>
[build] Fix misaligned table entries when using gcc 4.5
Declarations without the accompanying __table_entry cause misalignment
of the table entries when using gcc 4.5. Fix by adding the
appropriate __table_entry macro or (where possible) by removing
unnecessary forward declarations.
Signed-off-by: Piotr Jaroszyński <p.jaroszynski@gmail.com>
Signed-off-by: Michael Brown <mcb30@ipxe.org>
Maintain a queue of received packets, so that lost packets need not
result in retransmission of the entire TCP window.
Increase the TCP window to 8kB, in order that we can potentially
transmit enough duplicate ACKs to trigger Fast Retransmission at the
sender.
Using a 10MB HTTP download in qemu-kvm with an artificial drop rate of
1 in 64 packets, this reduces the download time from around 26s to
around 4s.
Signed-off-by: Michael Brown <mcb30@ipxe.org>
Access to the gpxe.org and etherboot.org domains and associated
resources has been revoked by the registrant of the domain. Work
around this problem by renaming project from gPXE to iPXE, and
updating URLs to match.
Also update README, LOG and COPYRIGHTS to remove obsolete information.
Signed-off-by: Michael Brown <mcb30@ipxe.org>
[tcp] Improve robustness in the presence of duplicated received packets
gPXE responds to duplicated ACKs with an immediate retransmission,
which can lead to a sorceror's apprentice syndrome. It also responds
to out-of-range (or old duplicate) ACKs with a RST, which can cause
valid connections to be dropped.
Fix the sorceror's apprentice syndrome by leaving the retransmission
timer running (and so inhibiting the immediate retransmission) when we
receive a potential duplicate ACK. This seems to match the behaviour
of Linux observed via wireshark traces.
Fix the RST issue by sending RST only on out-of-range ACKs that occur
before the connection is fully established, as per RFC 793.
These problems were exposed during development of the 802.11 wireless
link layer; the 802.11 protocol has a failure mode that can easily
cause duplicated packets. The fixes were tested in a controlled way
by faking large numbers of duplicated packets in the rtl8139 driver.
Originally-fixed-by: Joshua Oreman <oremanj@rwcr.net>
Redefine TCP state to include "flags that have been sent" rather than
"flags that are currently being sent". This allows at least one special
case (checking that we haven't already sent a FIN in tcp_rx_fin()) to be
collapsed.
I want to get to the point where any header in include/ reflects a
standard user-level header (e.g. a POSIX header), while everything that's
specific to gPXE lives in include/gpxe/. Headers that reflect a Linux
header (e.g. if_ether.h) should also be in include/gpxe/, with the same
name as the Linux header and, preferably, the same names used for the
definitions.
Michael Brown