When an allocation in the txq_init path fails, the allocated buffers
end-up being freed twice: in the txq_init error path, and in txq_deinit.
This lead to issues as txq_deinit would work on already freed memory
regions:

    kernel BUG at mm/slub.c:3915!
    Internal error: Oops - BUG: 0 [#1] PREEMPT SMP

This patch fixes this by removing the txq_init own error path, as the
txq_deinit function is always called on errors. This was introduced by
TSO as way more buffers are allocated.

Fixes: 186cd4d4 ("net: mvpp2: software tso support")
Signed-off-by: NAntoine Tenart <antoine.tenart@free-electrons.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Jiri Pirko says:

====================
mlxsw: GRE offloading fixes

Petr says:

This patchset fixes a couple bugs in offloading GRE tunnels in mlxsw
driver.

Patch #1 fixes a problem that local routes pointing at a GRE tunnel
device are offloaded even if that netdevice is down.

Patch #2 detects that as a result of moving a GRE netdevice to a
different VRF, two tunnels now have a conflict of local addresses,
something that the mlxsw driver can't offload.

Patch #3 fixes a FIB abort caused by forming a route pointing at a
GRE tunnel that is eligible for offloading but already onloaded.

Patch #4 fixes a problem that next hops migrated to a new RIF kept the
old RIF reference, which went dangling shortly afterwards.
====================
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The function mlxsw_sp_nexthop_rif_update() walks the list of nexthops
associated with a RIF, and updates the corresponding entries in the
switch. It is used in particular when a tunnel underlay netdevice moves
to a different VRF, and all the nexthops are migrated over to a new RIF.
The problem is that each nexthop holds a reference to its RIF, and that
is not updated. So after the old RIF is gone, further activity on these
nexthops (such as downing the underlay netdevice) dereferences a
dangling pointer.

Fix the issue by updating rif of impacted nexthops before calling
mlxsw_sp_nexthop_rif_update().

Fixes: 0c5f1cd5 ("mlxsw: spectrum_router: Generalize __mlxsw_sp_ipip_entry_update_tunnel()")
Signed-off-by: NPetr Machata <petrm@mellanox.com>
Reviewed-by: NIdo Schimmel <idosch@mellanox.com>
Signed-off-by: NJiri Pirko <jiri@mellanox.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Some tunnels that are offloadable on their own can nonetheless be
demoted to slow path if their local address is in conflict with that of
another tunnel. When a route is formed for such a tunnel,
mlxsw_sp_nexthop_ipip_init() fails to find the corresponding IPIP entry,
and that triggers a FIB abort.

Resolve the problem by not assuming that a tunnel for which
mlxsw_sp_ipip_ops.can_offload() holds also automatically has an IPIP
entry.

Fixes: af641713 ("mlxsw: spectrum_router: Onload conflicting tunnels")
Signed-off-by: NPetr Machata <petrm@mellanox.com>
Reviewed-by: NIdo Schimmel <idosch@mellanox.com>
Signed-off-by: NJiri Pirko <jiri@mellanox.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The mlxsw driver currently doesn't offload GRE tunnels if they have the
same local address and use the same underlay VRF. When such a situation
arises, the tunnels in conflict are demoted to slow path.

However, the current code only verifies this condition on tunnel
creation and tunnel change, not when a tunnel is moved to a different
VRF. When the tunnel has no bound device, underlay and overlay are the
same. Thus moving a tunnel moves the underlay as well, and that can
cause local address conflict.

So modify mlxsw_sp_netdevice_ipip_ol_vrf_event() to check if there are
any conflicting tunnels, and demote them if yes.

Fixes: af641713 ("mlxsw: spectrum_router: Onload conflicting tunnels")
Signed-off-by: NPetr Machata <petrm@mellanox.com>
Reviewed-by: NIdo Schimmel <idosch@mellanox.com>
Signed-off-by: NJiri Pirko <jiri@mellanox.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

When a new local route is added, an IPIP entry is looked up to determine
whether the route should be offloaded as a tunnel decap or as a trap.
That decision should take into account whether the tunnel netdevice in
question is actually IFF_UP, and only install a decap offload if it is.

Fixes: 0063587d ("mlxsw: spectrum: Support decap-only IP-in-IP tunnels")
Signed-off-by: NPetr Machata <petrm@mellanox.com>
Reviewed-by: NIdo Schimmel <idosch@mellanox.com>
Signed-off-by: NJiri Pirko <jiri@mellanox.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Jeff Kirsher says:

====================
Intel Wired LAN Driver Updates 2017-11-27

This series contains updates to e1000, e1000e and i40e.

Gustavo A. R. Silva fixes a sizeof() issue where we were taking the size of
the pointer (which is always the size of the pointer).

Sasha does a follow up fix to a previous fix for buffer overrun, to resolve
community feedback from David Laight and the use of magic numbers.

Amritha fixes the reporting of error codes for when adding a cloud filter
fails.

Ahmad Fatoum brushes the dust off the e1000 driver to fix a code comment
and debug message which was incorrect about what the code was really doing.
====================
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Signed-off-by: NAhmad Fatoum <ahmad@a3f.at>
Tested-by: NAaron Brown <aaron.f.brown@intel.com>
Signed-off-by: NJeff Kirsher <jeffrey.t.kirsher@intel.com>

Adding cloud filters could fail for a number of reasons,
unsupported filter fields for example, which fails during
validation of fields itself. This will not result in admin
command errors and converting the admin queue status to posix
error code using i40e_aq_rc_to_posix would result in incorrect
error values. If the failure was due to AQ error itself,
reporting that correctly is handled in the inner function.
Signed-off-by: NAmritha Nambiar <amritha.nambiar@intel.com>
Tested-by: NAndrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: NJeff Kirsher <jeffrey.t.kirsher@intel.com>

This is a follow on to commit b10effb9 ("fix buffer overrun while the
 I219 is processing DMA transactions") to address David Laights concerns
about the use of "magic" numbers.  So define masks as well as add
additional code comments to give a better understanding of what needs to
be done to avoid a buffer overrun.
Signed-off-by: NSasha Neftin <sasha.neftin@intel.com>
Reviewed-by: NAlexander H Duyck <alexander.h.duyck@intel.com>
Reviewed-by: NDima Ruinskiy <dima.ruinskiy@intel.com>
Reviewed-by: NRaanan Avargil <raanan.avargil@intel.com>
Tested-by: NAaron Brown <aaron.f.brown@intel.com>
Signed-off-by: NJeff Kirsher <jeffrey.t.kirsher@intel.com>

sizeof when applied to a pointer typed expression gives the size of
the pointer.

The proper fix in this particular case is to code sizeof(*vfres)
instead of sizeof(vfres).

This issue was detected with the help of Coccinelle.
Signed-off-by: NGustavo A R Silva <garsilva@embeddedor.com>
Tested-by: NAndrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: NJeff Kirsher <jeffrey.t.kirsher@intel.com>

KASAN revealed another access after delete in group.c. This time
it found that we read the header of a received message after the
buffer has been released.
Signed-off-by: NJon Maloy <jon.maloy@ericsson.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

v2:
 * Replace busy wait with wait_event()/wake_up_all()
 * Cannot garantee that at the time xennet_remove is called, the
   xen_netback state will not be XenbusStateClosed, so added a
   condition for that
 * There's a small chance for the xen_netback state is
   XenbusStateUnknown by the time the xen_netfront switches to Closed,
   so added a condition for that.

When unloading module xen_netfront from guest, dmesg would output
warning messages like below:

  [  105.236836] xen:grant_table: WARNING: g.e. 0x903 still in use!
  [  105.236839] deferring g.e. 0x903 (pfn 0x35805)

This problem relies on netfront and netback being out of sync. By the time
netfront revokes the g.e.'s netback didn't have enough time to free all of
them, hence displaying the warnings on dmesg.

The trick here is to make netfront to wait until netback frees all the g.e.'s
and only then continue to cleanup for the module removal, and this is done by
manipulating both device states.
Signed-off-by: NEduardo Otubo <otubo@redhat.com>
Acked-by: NJuergen Gross <jgross@suse.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Merge tag 'mac80211-for-davem-2017-11-27' of git://git.kernel.org/pub/scm/linux/kernel/git/jberg/mac80211

Johannes Berg says:

====================
Four fixes:
 * CRYPTO_SHA256 is needed for regdb validation
 * mac80211: mesh path metric was wrong in some frames
 * mac80211: use QoS null-data packets on QoS connections
 * mac80211: tear down RX aggregation sessions first to
   drop fewer packets in HW restart scenarios
====================
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Xin Long says:

====================
sctp: a bunch of fixes for stream reconfig

This patchset is to make stream reset and asoc reset work more correctly
for stream reconfig.

Thank to Marcelo making them very clear.
====================
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

When doing asoc reset, if the sender of the response has already sent some
chunk and increased asoc->next_tsn before the duplicate request comes, the
response will use the old result with an incorrect sender next_tsn.

Better than asoc->next_tsn, asoc->ctsn_ack_point can't be changed after
the sender of the response has performed the asoc reset and before the
peer has confirmed it, and it's value is still asoc->next_tsn original
value minus 1.

This patch sets sender next_tsn for the old result with ctsn_ack_point
plus 1 when processing the duplicate request, to make sure the sender
next_tsn value peer gets will be always right.

Fixes: 692787ce ("sctp: implement receiver-side procedures for the SSN/TSN Reset Request Parameter")
Signed-off-by: NXin Long <lucien.xin@gmail.com>
Acked-by: NMarcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Now when doing asoc reset, it cleans up sacked and abandoned queues
by calling sctp_outq_free where it also cleans up unsent, retransmit
and transmitted queues.

It's safe for the sender of response, as these 3 queues are empty at
that time. But when the receiver of response is doing the reset, the
users may already enqueue some chunks into unsent during the time
waiting the response, and these chunks should not be flushed.

To void the chunks in it would be removed, it moves the queue into a
temp list, then gets it back after sctp_outq_free is done.

The patch also fixes some incorrect comments in
sctp_process_strreset_tsnreq.
Signed-off-by: NXin Long <lucien.xin@gmail.com>
Acked-by: NMarcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

As it says in rfc6525#section5.1.4, before sending the request,

   C2:  The sender has either no outstanding TSNs or considers all
        outstanding TSNs abandoned.

Prior to this patch, it tried to consider all outstanding TSNs abandoned
by dropping all chunks in all outqs with sctp_outq_free (even including
sacked, retransmit and transmitted queues) when doing this reset, which
is too aggressive.

To make it work gently, this patch will only allow the asoc reset when
the sender has no outstanding TSNs by checking if unsent, transmitted
and retransmit are all empty with sctp_outq_is_empty before sending
and processing the request.

Fixes: 692787ce ("sctp: implement receiver-side procedures for the SSN/TSN Reset Request Parameter")
Signed-off-by: NXin Long <lucien.xin@gmail.com>
Acked-by: NMarcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Now the out stream reset in sctp stream reconf could be done even if
the stream outq is not empty. It means that users can not be sure
since which msg the new ssn will be used.

To make this more synchronous, it shouldn't allow to do out stream
reset until these chunks in unsent outq all are sent out.

This patch checks the corresponding stream outqs when sending and
processing the request . If any of them has unsent chunks in outq,
it will return -EAGAIN instead or send SCTP_STRRESET_IN_PROGRESS
back to the sender.

Fixes: 7f9d68ac ("sctp: implement sender-side procedures for SSN Reset Request Parameter")
Suggested-by: NMarcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Signed-off-by: NXin Long <lucien.xin@gmail.com>
Acked-by: NMarcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Now in stream reconf part there are still some places using magic
number 2 for each stream number length. To make it more readable,
this patch is to replace them with sizeof(__u16).
Reported-by: NMarcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Signed-off-by: NXin Long <lucien.xin@gmail.com>
Acked-by: NMarcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

When doing HW restart we tear down aggregations.
Since at this point we are not TX'ing any aggregation, while
the peer is still sending RX aggregation over the air, it will
make sense to tear down the RX aggregations first.
Signed-off-by: NSara Sharon <sara.sharon@intel.com>
Signed-off-by: NLuca Coelho <luciano.coelho@intel.com>
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

The previous path metric update from RANN frame has not considered
the own link metric toward the transmitting mesh STA. Fix this.

Reported-by: Michael65535
Signed-off-by: NChun-Yeow Yeoh <yeohchunyeow@gmail.com>
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

When connected to a QoS/WMM AP, mac80211 should use a QoS NDP
for probing it, instead of a regular non-QoS one, fix this.

Change all the drivers to *not* allow QoS NDP for now, even
though it looks like most of them should be OK with that.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

If we want to add a datapath flow, which has more than 500 vxlan outputs'
action, we will get the following error reports:
  openvswitch: netlink: Flow action size 32832 bytes exceeds max
  openvswitch: netlink: Flow action size 32832 bytes exceeds max
  openvswitch: netlink: Actions may not be safe on all matching packets
  ... ...

It seems that we can simply enlarge the MAX_ACTIONS_BUFSIZE to fix it, but
this is not the root cause. For example, for a vxlan output action, we need
about 60 bytes for the nlattr, but after it is converted to the flow
action, it only occupies 24 bytes. This means that we can still support
more than 1000 vxlan output actions for a single datapath flow under the
the current 32k max limitation.

So even if the nla_len(attr) is larger than MAX_ACTIONS_BUFSIZE, we
shouldn't report EINVAL and keep it move on, as the judgement can be
done by the reserve_sfa_size.
Signed-off-by: Nzhangliping <zhangliping02@baidu.com>
Acked-by: NPravin B Shelar <pshelar@ovn.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

gso_type is being used in binary AND operations together with SKB_GSO_UDP.
The issue is that variable gso_type is of type unsigned short and
SKB_GSO_UDP expands to more than 16 bits:

SKB_GSO_UDP = 1 << 16

this makes any binary AND operation between gso_type and SKB_GSO_UDP to
be always zero, hence making some code unreachable and likely causing
undesired behavior.

Fix this by changing the data type of variable gso_type to unsigned int.

Addresses-Coverity-ID: 1462223
Fixes: 0c19f846 ("net: accept UFO datagrams from tuntap and packet")
Signed-off-by: NGustavo A. R. Silva <garsilva@embeddedor.com>
Acked-by: NWillem de Bruijn <willemb@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

New file seems to have missed the SPDX license scan and update.
Signed-off-by: NStephen Hemminger <sthemmin@microsoft.com>
Reviewed-by: NStefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Setting the refcount to 0 when allocating a tree to match the number of
switch devices it holds may cause an 'increment on 0; use-after-free',
if CONFIG_REFCOUNT_FULL is enabled.

To fix this, do not decrement the refcount of a newly allocated tree,
increment it when an already allocated tree is found, and decrement it
after the probing of a switch, as done with the previous behavior.

At the same time, make dsa_tree_get and dsa_tree_put accept a NULL
argument to simplify callers, and return the tree after incrementation,
as most kref users like of_node_get and of_node_put do.

Fixes: 8e5bf975 ("net: dsa: simplify tree reference counting")
Signed-off-by: NVivien Didelot <vivien.didelot@savoirfairelinux.com>
Reviewed-by: NFlorian Fainelli <f.fainelli@gmail.com>
Tested-by: NFlorian Fainelli <f.fainelli@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

When using the host personality, VMCI will grab a mutex for any
queue pair access. In the detach callback for the vmci vsock
transport, we call vsock_stream_has_data while holding a spinlock,
and vsock_stream_has_data will access a queue pair.

To avoid this, we can simply omit calling vsock_stream_has_data
for host side queue pairs, since the QPs are empty per default
when the guest has detached.

This bug affects users of VMware Workstation using kernel version
4.4 and later.

Testing: Ran vsock tests between guest and host, and verified that
with this change, the host isn't calling vsock_stream_has_data
during detach. Ran mixedTest between guest and host using both
guest and host as server.

v2: Rebased on top of recent change to sk_state values
Reviewed-by: NAdit Ranadive <aditr@vmware.com>
Reviewed-by: NAditya Sarwade <asarwade@vmware.com>
Reviewed-by: NStefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: NJorgen Hansen <jhansen@vmware.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

tcf_block_put_ext has assumed that all filters (and thus their goto
actions) are destroyed in RCU callback and thus can not race with our
list iteration. However, that is not true during netns cleanup (see
tcf_exts_get_net comment).

Prevent the user after free by holding all chains (except 0, that one is
already held). foreach_safe is not enough in this case.

To reproduce, run the following in a netns and then delete the ns:
    ip link add dtest type dummy
    tc qdisc add dev dtest ingress
    tc filter add dev dtest chain 1 parent ffff: handle 1 prio 1 flower action goto chain 2

Fixes: 822e86d9 ("net_sched: remove tcf_block_put_deferred()")
Signed-off-by: NRoman Kapl <code@rkapl.cz>
Acked-by: NJiri Pirko <jiri@mellanox.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Commit 86dabda4 ("net: thunderbolt: Clear finished Tx frame bus
address in tbnet_tx_callback()") fixed a DMA-API violation where the
driver called dma_unmap_page() in tbnet_free_buffers() for a bus address
that might already be unmapped. The fix was to zero out the bus address
of a frame in tbnet_tx_callback().

However, as pointed out by David Miller, zero might well be valid
mapping (at least in theory) so it is not good idea to use it here.

It turns out that we don't need the whole map/unmap dance for Tx buffers
at all. Instead we can map the buffers when they are initially allocated
and unmap them when the interface is brought down. In between we just
DMA sync the buffers for the CPU or device as needed.
Signed-off-by: NMika Westerberg <mika.westerberg@linux.intel.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Don't offload IP header checksum to NIC.

This fixes a previous patch which enabled checksum offloading
for both IPv4 and IPv6 packets.  So L3 checksum offload was
getting enabled for IPv6 pkts.  And HW is dropping these pkts
as it assumes the pkt is IPv4 when IP csum offload is set
in the SQ descriptor.

Fixes:  3a9024f5 ("net: thunderx: Enable TSO and checksum offloads for ipv6")
Signed-off-by: NSunil Goutham <sgoutham@cavium.com>
Signed-off-by: NAleksey Makarov <aleksey.makarov@auriga.com>
Reviewed-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

When regulatory database certificates are built-in, they're
currently using the SHA256 digest algorithm, so add that to
the build in that case.

Also add a note that for custom certificates, one may need
to add the right algorithms.
Reported-by: NFlorian Fainelli <f.fainelli@gmail.com>
Tested-by: NFlorian Fainelli <f.fainelli@gmail.com>
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

The function pci_unmap_page is obsolete. So it is replaced with
the function dma_unmap_page.

CC: Srinivas Eeda <srinivas.eeda@oracle.com>
CC: Joe Jin <joe.jin@oracle.com>
CC: Junxiao Bi <junxiao.bi@oracle.com>
Signed-off-by: NZhu Yanjun <yanjun.zhu@oracle.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

David Howells says:

====================
rxrpc: Fixes and improvements

Here's a set of patches that fix and improve some stuff in the AF_RXRPC
protocol:

The patches are:

 (1) Unlock mutex returned by rxrpc_accept_call().

 (2) Don't set connection upgrade by default.

 (3) Differentiate the call->user_mutex used by the kernel from that used
     by userspace calling sendmsg() to avoid lockdep warnings.

 (4) Delay terminal ACK transmission to a work queue so that it can be
     replaced by the next call if there is one.

 (5) Split the call parameters from the connection parameters so that more
     call-specific parameters can be passed through.

 (6) Fix the call timeouts to work the same as for other RxRPC/AFS
     implementations.

 (7) Don't transmit DELAY ACKs immediately, but instead delay them slightly
     so that can be discarded or can represent more packets.

 (8) Use RTT to calculate certain protocol timeouts.

 (9) Add a timeout to detect lost ACK/DATA packets.

(10) Add a keepalive function so that we ping the peer if we haven't
     transmitted for a short while, thereby keeping intervening firewall
     routes open.

(11) Make service endpoints expire like they're supposed to so that the UDP
     port can be reused.

(12) Fix connection expiry timers to make cleanup happen in a more timely
     fashion.
====================
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Fix the rxrpc connection expiry timers so that connections for closed
AF_RXRPC sockets get deleted in a more timely fashion, freeing up the
transport UDP port much more quickly.

 (1) Replace the delayed work items with work items plus timers so that
     timer_reduce() can be used to shorten them and so that the timer
     doesn't requeue the work item if the net namespace is dead.

 (2) Don't use queue_delayed_work() as that won't alter the timeout if the
     timer is already running.

 (3) Don't rearm the timers if the network namespace is dead.
Signed-off-by: NDavid Howells <dhowells@redhat.com>

RxRPC service endpoints expire like they're supposed to by the following
means:

 (1) Mark dead rxrpc_net structs (with ->live) rather than twiddling the
     global service conn timeout, otherwise the first rxrpc_net struct to
     die will cause connections on all others to expire immediately from
     then on.

 (2) Mark local service endpoints for which the socket has been closed
     (->service_closed) so that the expiration timeout can be much
     shortened for service and client connections going through that
     endpoint.

 (3) rxrpc_put_service_conn() needs to schedule the reaper when the usage
     count reaches 1, not 0, as idle conns have a 1 count.

 (4) The accumulator for the earliest time we might want to schedule for
     should be initialised to jiffies + MAX_JIFFY_OFFSET, not ULONG_MAX as
     the comparison functions use signed arithmetic.

 (5) Simplify the expiration handling, adding the expiration value to the
     idle timestamp each time rather than keeping track of the time in the
     past before which the idle timestamp must go to be expired.  This is
     much easier to read.

 (6) Ignore the timeouts if the net namespace is dead.

 (7) Restart the service reaper work item rather the client reaper.
Signed-off-by: NDavid Howells <dhowells@redhat.com>

We need to transmit a packet every so often to act as a keepalive for the
peer (which has a timeout from the last time it received a packet) and also
to prevent any intervening firewalls from closing the route.

Do this by resetting a timer every time we transmit a packet.  If the timer
ever expires, we transmit a PING ACK packet and thereby also elicit a PING
RESPONSE ACK from the other side - which prevents our last-rx timeout from
expiring.

The timer is set to 1/6 of the last-rx timeout so that we can detect the
other side going away if it misses 6 replies in a row.

This is particularly necessary for servers where the processing of the
service function may take a significant amount of time.
Signed-off-by: NDavid Howells <dhowells@redhat.com>

Add an extra timeout that is set/updated when we send a DATA packet that
has the request-ack flag set.  This allows us to detect if we don't get an
ACK in response to the latest flagged packet.

The ACK packet is adjudged to have been lost if it doesn't turn up within
2*RTT of the transmission.

If the timeout occurs, we schedule the sending of a PING ACK to find out
the state of the other side.  If a new DATA packet is ready to go sooner,
we cancel the sending of the ping and set the request-ack flag on that
instead.

If we get back a PING-RESPONSE ACK that indicates a lower tx_top than what
we had at the time of the ping transmission, we adjudge all the DATA
packets sent between the response tx_top and the ping-time tx_top to have
been lost and retransmit immediately.

Rather than sending a PING ACK, we could just pick a DATA packet and
speculatively retransmit that with request-ack set.  It should result in
either a REQUESTED ACK or a DUPLICATE ACK which we can then use in lieu the
a PING-RESPONSE ACK mentioned above.
Signed-off-by: NDavid Howells <dhowells@redhat.com>

Express protocol timeouts for data retransmission and deferred ack
generation in terms on RTT rather than specified timeouts once we have
sufficient RTT samples.

For the moment, this requires just one RTT sample to be able to use this
for ack deferral and two for data retransmission.

The data retransmission timeout is set at RTT*1.5 and the ACK deferral
timeout is set at RTT.

Note that the calculated timeout is limited to a minimum of 4ns to make
sure it doesn't happen too quickly.
Signed-off-by: NDavid Howells <dhowells@redhat.com>

Don't transmit a DELAY ACK immediately on proposal when the Rx window is
rotated, but rather defer it to the work function.  This means that we have
a chance to queue/consume more received packets before we actually send the
DELAY ACK, or even cancel it entirely, thereby reducing the number of
packets transmitted.

We do, however, want to continue sending other types of packet immediately,
particularly REQUESTED ACKs, as they may be used for RTT calculation by the
other side.
Signed-off-by: NDavid Howells <dhowells@redhat.com>