The peer->mtu_lock is only used to lock around writes to peer->max_data -
and nothing else; further, all such writes take place in the I/O thread and
the lock is only ever write-locked and never read-locked.
In a couple of places, the write_seqcount_begin() is wrapped in
preempt_disable/enable(), but not in all places. This can cause lockdep to
complain:
WARNING: CPU: 0 PID: 1549 at include/linux/seqlock.h:221 rxrpc_input_ack_trailer+0x305/0x430
...
RIP: 0010:rxrpc_input_ack_trailer+0x305/0x430
Fix this by just getting rid of the lock.
Fixes: eeaedc5449 ("rxrpc: Implement path-MTU probing using padded PING ACKs (RFC8899)")
Signed-off-by: David Howells <dhowells@redhat.com>
cc: Marc Dionne <marc.dionne@auristor.com>
cc: Simon Horman <horms@kernel.org>
cc: linux-afs@lists.infradead.org
Link: https://patch.msgid.link/20250218192250.296870-3-dhowells@redhat.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
There's a race in between the rxrpc I/O thread recording the end of the
receive phase of a call and recvmsg() examining the state of the call to
determine whether it has completed.
The problem is that call->_state records the I/O thread's view of the call,
not the application's view (which may lag), so that alone is not
sufficient. To this end, the application also checks whether there is
anything left in call->recvmsg_queue for it to pick up. The call must be
in state RXRPC_CALL_COMPLETE and the recvmsg_queue empty for the call to be
considered fully complete.
In rxrpc_input_queue_data(), the latest skbuff is added to the queue and
then, if it was marked as LAST_PACKET, the state is advanced... But this
is two separate operations with no locking around them.
As a consequence, the lack of locking means that sendmsg() can jump into
the gap on a service call and attempt to send the reply - but then get
rejected because the I/O thread hasn't advanced the state yet.
Simply flipping the order in which things are done isn't an option as that
impacts the client side, causing the checks in rxrpc_kernel_check_life() as
to whether the call is still alive to race instead.
Fix this by moving the update of call->_state inside the skb queue
spinlocked section where the packet is queued on the I/O thread side.
rxrpc's recvmsg() will then automatically sync against this because it has
to take the call->recvmsg_queue spinlock in order to dequeue the last
packet.
rxrpc's sendmsg() doesn't need amending as the app shouldn't be calling it
to send a reply until recvmsg() indicates it has returned all of the
request.
Fixes: 93368b6bd5 ("rxrpc: Move call state changes from recvmsg to I/O thread")
Signed-off-by: David Howells <dhowells@redhat.com>
cc: Marc Dionne <marc.dionne@auristor.com>
cc: Simon Horman <horms@kernel.org>
cc: linux-afs@lists.infradead.org
Link: https://patch.msgid.link/20250204230558.712536-3-dhowells@redhat.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The RXRPC_CALL_SERVER_SECURING state doesn't really belong with the other
states in the call's state set as the other states govern the call's Rx/Tx
phase transition and govern when packets can and can't be received or
transmitted. The "Securing" state doesn't actually govern the reception of
packets and would need to be split depending on whether or not we've
received the last packet yet (to mirror RECV_REQUEST/ACK_REQUEST).
The "Securing" state is more about whether or not we can start forwarding
packets to the application as recvmsg will need to decode them and the
decoding can't take place until the challenge/response exchange has
completed.
Fix this by removing the RXRPC_CALL_SERVER_SECURING state from the state
set and, instead, using a flag, RXRPC_CALL_CONN_CHALLENGING, to track
whether or not we can queue the call for reception by recvmsg() or notify
the kernel app that data is ready. In the event that we've already
received all the packets, the connection event handler will poke the app
layer in the appropriate manner.
Also there's a race whereby the app layer sees the last packet before rxrpc
has managed to end the rx phase and change the state to one amenable to
allowing a reply. Fix this by queuing the packet after calling
rxrpc_end_rx_phase().
Fixes: 17926a7932 ("[AF_RXRPC]: Provide secure RxRPC sockets for use by userspace and kernel both")
Signed-off-by: David Howells <dhowells@redhat.com>
cc: Marc Dionne <marc.dionne@auristor.com>
cc: Simon Horman <horms@kernel.org>
cc: linux-afs@lists.infradead.org
Link: https://patch.msgid.link/20250204230558.712536-2-dhowells@redhat.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
When an rxrpc call is in its transmission phase and is sending a lot of
packets, stalls occasionally occur that cause severe performance
degradation (eg. increasing the transmission time for a 256MiB payload from
0.7s to 2.5s over a 10G link).
rxrpc already implements TCP-style congestion control [RFC5681] and this
helps mitigate the effects, but occasionally we're missing a time event
that deals with a missing ACK, leading to a stall until the RTO expires.
Fix this by implementing RACK/TLP in rxrpc.
Signed-off-by: David Howells <dhowells@redhat.com>
cc: Marc Dionne <marc.dionne@auristor.com>
cc: linux-afs@lists.infradead.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Manage the determination of RTT on a per-call (ie. per-RPC op) basis rather
than on a per-peer basis, averaging across all calls going to that peer.
The problem is that the RTT measurements from the initial packets on a call
may be off because the server may do some setting up (such as getting a
lock on a file) before accepting the rest of the data in the RPC and,
further, the RTT may be affected by server-side file operations, for
instance if a large amount of data is being written or read.
Note: When handling the FS.StoreData-type RPCs, for example, the server
uses the userStatus field in the header of ACK packets as supplementary
flow control to aid in managing this. AF_RXRPC does not yet support this,
but it should be added.
Signed-off-by: David Howells <dhowells@redhat.com>
cc: Marc Dionne <marc.dionne@auristor.com>
cc: linux-afs@lists.infradead.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Tidy up the ACK parsing in the following ways:
(1) Put the serial number of the ACK packet into the rxrpc_ack_summary
struct and access it from there whilst parsing an ACK.
(2) Be consistent about using "if (summary.acked_serial)" rather than "if
(summary.acked_serial != 0)".
Signed-off-by: David Howells <dhowells@redhat.com>
cc: Marc Dionne <marc.dionne@auristor.com>
cc: linux-afs@lists.infradead.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Where a spinlock is used by both the application thread and the I/O thread,
use irq-disabling locking so that an interrupt taken on the app thread
doesn't also slow down the I/O thread.
Signed-off-by: David Howells <dhowells@redhat.com>
cc: Marc Dionne <marc.dionne@auristor.com>
cc: linux-afs@lists.infradead.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Send jumbo DATA packets if the path-MTU probing using padded PING ACK
packets shows up sufficient capacity to do so. This allows larger chunks
of data to be sent without reducing the retryability as the subpackets in a
jumbo packet can also be retransmitted individually.
Signed-off-by: David Howells <dhowells@redhat.com>
cc: Marc Dionne <marc.dionne@auristor.com>
cc: linux-afs@lists.infradead.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Adjust the rxrpc_rtt_rx tracepoint in the following ways:
(1) Display the collected RTT sample in the rxrpc_rtt_rx trace.
(2) Move the division of srtt by 8 to the TP_printk() rather doing it
before invoking the trace point.
(3) Display the min_rtt value.
Signed-off-by: David Howells <dhowells@redhat.com>
cc: Marc Dionne <marc.dionne@auristor.com>
cc: linux-afs@lists.infradead.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Store the serial number set on a DATA packet at the point of transmission
in the rxrpc_txqueue struct and when an ACK is received, match the
reference number in the ACK by trawling the txqueue rather than sharing an
RTT table with ACK RTT. This can be done as part of Tx queue rotation.
This means we have a lot more RTT samples available and is faster to search
with all the serial numbers packed together into a few cachelines rather
than being hung off different txbufs.
Signed-off-by: David Howells <dhowells@redhat.com>
cc: Marc Dionne <marc.dionne@auristor.com>
cc: linux-afs@lists.infradead.org
Link: https://patch.msgid.link/20241204074710.990092-25-dhowells@redhat.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Adjust some of the names of fields and constants to make them look a bit
more like the TCP congestion symbol names, such as flight_size -> in_flight
and congest_mode to ca_state.
Move the persistent congestion-related fields from the rxrpc_ack_summary
struct into the rxrpc_call struct rather than copying them out and back in
again. The rxrpc_congest tracepoint can fetch them from the call struct.
Rename the counters for soft acks and nacks to have an 's' on the front to
reflect the softness, e.g. nr_acks -> nr_sacks.
Make fields counting numbers of packets or numbers of acks u16 rather than
u8 to allow for windows of up to 8192 DATA packets in flight in future.
Signed-off-by: David Howells <dhowells@redhat.com>
cc: Marc Dionne <marc.dionne@auristor.com>
cc: linux-afs@lists.infradead.org
Link: https://patch.msgid.link/20241204074710.990092-23-dhowells@redhat.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Replace the call->acks_first_seq variable (which holds ack.firstPacket from
the latest ACK packet and indicates the sequence number of the first ack
slot in the SACK table) with call->acks_hard_ack which will hold the
highest sequence hard ACK'd. This is 1 less than call->acks_first_seq, but
it fits in the same schema as the other tracking variables which hold the
sequence of a packet, not one past it.
This will fix the rxrpc_congest tracepoint's calculation of SACK window
size which shows one fewer than it should - and will occasionally go to -1.
Signed-off-by: David Howells <dhowells@redhat.com>
cc: Marc Dionne <marc.dionne@auristor.com>
cc: linux-afs@lists.infradead.org
Link: https://patch.msgid.link/20241204074710.990092-21-dhowells@redhat.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
We need to scan the buffers in the transmission queue occasionally when
processing ACKs, but the transmission queue is currently a linked list of
transmission buffers which, when we eventually expand the Tx window to 8192
packets will be very slow to walk.
Instead, pull the fields we need to examine a lot (last sent time,
retransmitted flag) into a new struct rxrpc_txqueue and make each one hold
an array of 32 or 64 packets.
The transmission queue is then a list of these structs, each pointing to a
contiguous set of packets. Scanning is then a lot faster as the flags and
timestamps are concentrated in the CPU dcache.
The transmission timestamps are stored as a number of microseconds from a
base ktime to reduce memory requirements. This should be fine provided we
manage to transmit an entire buffer within an hour.
This will make implementing RACK-TLP [RFC8985] easier as it will be less
costly to scan the transmission buffers.
Signed-off-by: David Howells <dhowells@redhat.com>
cc: Marc Dionne <marc.dionne@auristor.com>
cc: linux-afs@lists.infradead.org
Link: https://patch.msgid.link/20241204074710.990092-19-dhowells@redhat.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Starvation can happen in the rxrpc I/O thread because it goes back to the
top of the I/O loop after it does any one thing without trying to give any
other connection or call CPU time. Also, because it processes one call
packet at a time, it tries to do the retransmission loop after each ACK
without checking to see if there are other ACKs already in the queue that
can update the SACK state.
Fix this by:
(1) Add a received-packet queue on each call.
(2) Distribute packets from the master Rx queue to the individual call,
conn and error queues and 'poking' calls to add them to the attend
queue first thing in the I/O thread.
(3) Go through all the attention-seeking connections and calls before
going back to the top of the I/O thread. Each queue is extracted as a
whole and then gone through so that new additions to insert themselves
into the queue.
(4) Make the call event handler go through all the packets currently on
the call's rx_queue before transmitting and retransmitting DATA
packets.
(5) Drop the skb argument from the call event handler as this is now
replaced with the rx_queue. Instead, keep track of whether we
received a packet or an ACK for the tests that used to rely on that.
Signed-off-by: David Howells <dhowells@redhat.com>
cc: Marc Dionne <marc.dionne@auristor.com>
cc: linux-afs@lists.infradead.org
Link: https://patch.msgid.link/20241204074710.990092-14-dhowells@redhat.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Implement path-MTU probing (along the lines of RFC8899) by padding some of
the PING ACKs we send. PING ACKs get their own individual responses quite
apart from the acking of data (though, as ACKs, they fulfil that role
also).
The probing concentrates on packet sizes that correspond how many
subpackets can be stuffed inside a jumbo packet as jumbo DATA packets are
just aggregations of individual DATA packets and can be split easily for
retransmission purposes.
If we want to perform probing, we advertise this by setting the maximum
number of jumbo subpackets to 0 in the ack trailer when we send an ACK and
see if the peer is also advertising the service. This is interpreted by
non-supporting Rx stacks as an indication that jumbo packets aren't
supported.
The MTU sizes advertised in the ACK trailer AF_RXRPC transmits are pegged
at a maximum of 1444 unless pmtud is supported by both sides.
Signed-off-by: David Howells <dhowells@redhat.com>
cc: Marc Dionne <marc.dionne@auristor.com>
cc: linux-afs@lists.infradead.org
Link: https://patch.msgid.link/20241204074710.990092-10-dhowells@redhat.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Only generate one ACK packet for all the subpackets in a jumbo packet. If
we would like to generate more than one ACK, we prioritise them base on
their reason code, in the order, highest first:
OutOfSeq > NoSpace > ExceedsWin > Duplicate > Requested > Delay > Idle
For the first four, we reference the lowest offending subpacket; for the
last three, the highest.
This reduces the number of ACKs we end up transmitting to one per UDP
packet transmitted to reduce network loading and packet parsing.
Fixes: 5d7edbc923 ("rxrpc: Get rid of the Rx ring")
Signed-off-by: David Howells <dhowells@redhat.com>
cc: Marc Dionne <marc.dionne@auristor.com>
cc: linux-afs@lists.infradead.org
Reviewed-by: Jeffrey Altman <jaltman@auristor.com <mailto:jaltman@auristor.com>>
Link: https://lore.kernel.org/r/20240503150749.1001323-3-dhowells@redhat.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Make the following fixes to the congestion control algorithm:
(1) Don't vary the cwnd starting value by the size of RXRPC_TX_SMSS since
that's currently held constant - set to the size of a jumbo subpacket
payload so that we can create jumbo packets on the fly. The current
code invariably picks 3 as the starting value.
Further, the starting cwnd needs to be an even number because we ack
every other packet, so set it to 4.
(2) Don't cut ssthresh when we see an ACK come from the peer with a
receive window (rwind) less than ssthresh. ssthresh keeps track of
characteristics of the connection whereas rwind may be reduced by the
peer for any reason - and may be reduced to 0.
Fixes: 1fc4fa2ac9 ("rxrpc: Fix congestion management")
Fixes: 0851115090 ("rxrpc: Reduce ssthresh to peer's receive window")
Signed-off-by: David Howells <dhowells@redhat.com>
Suggested-by: Simon Wilkinson <sxw@auristor.com>
cc: Marc Dionne <marc.dionne@auristor.com>
cc: linux-afs@lists.infradead.org
Reviewed-by: Jeffrey Altman <jaltman@auristor.com <mailto:jaltman@auristor.com>>
Link: https://lore.kernel.org/r/20240503150749.1001323-2-dhowells@redhat.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
