Implement RxRPC slow-start, which is similar to RFC 5681 for TCP. A
tracepoint is added to log the state of the congestion management algorithm
and the decisions it makes.
Notes:
(1) Since we send fixed-size DATA packets (apart from the final packet in
each phase), counters and calculations are in terms of packets rather
than bytes.
(2) The ACK packet carries the equivalent of TCP SACK.
(3) The FLIGHT_SIZE calculation in RFC 5681 doesn't seem particularly
suited to SACK of a small number of packets. It seems that, almost
inevitably, by the time three 'duplicate' ACKs have been seen, we have
narrowed the loss down to one or two missing packets, and the
FLIGHT_SIZE calculation ends up as 2.
(4) In rxrpc_resend(), if there was no data that apparently needed
retransmission, we transmit a PING ACK to ask the peer to tell us what
its Rx window state is.
Signed-off-by: David Howells <dhowells@redhat.com>
Add a tracepoint to log transmission of DATA packets (including loss
injection).
Adjust the ACK transmission tracepoint to include the packet serial number
and to line this up with the DATA transmission display.
Signed-off-by: David Howells <dhowells@redhat.com>
Improve sk_buff tracing within AF_RXRPC by the following means:
(1) Use an enum to note the event type rather than plain integers and use
an array of event names rather than a big multi ?: list.
(2) Distinguish Rx from Tx packets and account them separately. This
requires the call phase to be tracked so that we know what we might
find in rxtx_buffer[].
(3) Add a parameter to rxrpc_{new,see,get,free}_skb() to indicate the
event type.
(4) A pair of 'rotate' events are added to indicate packets that are about
to be rotated out of the Rx and Tx windows.
(5) A pair of 'lost' events are added, along with rxrpc_lose_skb() for
packet loss injection recording.
Signed-off-by: David Howells <dhowells@redhat.com>
Add a pair of tracepoints, one to track rxrpc_connection struct ref
counting and the other to track the client connection cache state.
Signed-off-by: David Howells <dhowells@redhat.com>
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
Add a tracepoint for working out where local aborts happen. Each
tracepoint call is labelled with a 3-letter code so that they can be
distinguished - and the DATA sequence number is added too where available.
rxrpc_kernel_abort_call() also takes a 3-letter code so that AFS can
indicate the circumstances when it aborts a call.
Signed-off-by: David Howells <dhowells@redhat.com>
Don't expose skbs to in-kernel users, such as the AFS filesystem, but
instead provide a notification hook the indicates that a call needs
attention and another that indicates that there's a new call to be
collected.
This makes the following possibilities more achievable:
(1) Call refcounting can be made simpler if skbs don't hold refs to calls.
(2) skbs referring to non-data events will be able to be freed much sooner
rather than being queued for AFS to pick up as rxrpc_kernel_recv_data
will be able to consult the call state.
(3) We can shortcut the receive phase when a call is remotely aborted
because we don't have to go through all the packets to get to the one
cancelling the operation.
(4) It makes it easier to do encryption/decryption directly between AFS's
buffers and sk_buffs.
(5) Encryption/decryption can more easily be done in the AFS's thread
contexts - usually that of the userspace process that issued a syscall
- rather than in one of rxrpc's background threads on a workqueue.
(6) AFS will be able to wait synchronously on a call inside AF_RXRPC.
To make this work, the following interface function has been added:
int rxrpc_kernel_recv_data(
struct socket *sock, struct rxrpc_call *call,
void *buffer, size_t bufsize, size_t *_offset,
bool want_more, u32 *_abort_code);
This is the recvmsg equivalent. It allows the caller to find out about the
state of a specific call and to transfer received data into a buffer
piecemeal.
afs_extract_data() and rxrpc_kernel_recv_data() now do all the extraction
logic between them. They don't wait synchronously yet because the socket
lock needs to be dealt with.
Five interface functions have been removed:
rxrpc_kernel_is_data_last()
rxrpc_kernel_get_abort_code()
rxrpc_kernel_get_error_number()
rxrpc_kernel_free_skb()
rxrpc_kernel_data_consumed()
As a temporary hack, sk_buffs going to an in-kernel call are queued on the
rxrpc_call struct (->knlrecv_queue) rather than being handed over to the
in-kernel user. To process the queue internally, a temporary function,
temp_deliver_data() has been added. This will be replaced with common code
between the rxrpc_recvmsg() path and the kernel_rxrpc_recv_data() path in a
future patch.
Signed-off-by: David Howells <dhowells@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Condense the terminal states of a call state machine to a single state,
plus a separate completion type value. The value is then set, along with
error and abort code values, only when the call is transitioned to the
completion state.
Helpers are provided to simplify this.
Signed-off-by: David Howells <dhowells@redhat.com>
The call pointer in a channel on a connection will be NULL if there's no
active call on that channel. rxrpc_abort_calls() needs to check for this
before trying to take the call's state_lock.
Signed-off-by: David Howells <dhowells@redhat.com>
If a duplicate packet comes in for a call that has just completed on a
connection's channel then there will be an oops in the data_ready handler
because it tries to examine the connection struct via a call struct (which
we don't have - the pointer is unset).
Since the connection struct pointer is available to us, go direct instead.
Also, the ACK packet to be retransmitted needs three octets of padding
between the soft ack list and the ackinfo.
Fixes: 18bfeba50d ("rxrpc: Perform terminal call ACK/ABORT retransmission from conn processor")
Signed-off-by: David Howells <dhowells@redhat.com>
Perform terminal call ACK/ABORT retransmission in the connection processor
rather than in the call processor. With this change, once last_call is
set, no more incoming packets will be routed to the corresponding call or
any earlier calls on that channel (call IDs must only increase on a channel
on a connection).
Further, if a packet's callNumber is before the last_call ID or a packet is
aimed at successfully completed service call then that packet is discarded
and ignored.
Signed-off-by: David Howells <dhowells@redhat.com>
Each channel on a connection has a separate, independent number space from
which to allocate callNumber values. It is entirely possible, for example,
to have a connection with four active calls, each with call number 1.
Note that the callNumber values for any particular channel don't have to
start at 1, but they are supposed to increment monotonically for that
channel from a client's perspective and may not be reused once the call
number is transmitted (until the epoch cycles all the way back round).
Currently, however, call numbers are allocated on a per-connection basis
and, further, are held in an rb-tree. The rb-tree is redundant as the four
channel pointers in the rxrpc_connection struct are entirely capable of
pointing to all the calls currently in progress on a connection.
To this end, make the following changes:
(1) Handle call number allocation independently per channel.
(2) Get rid of the conn->calls rb-tree. This is overkill as a connection
may have a maximum of four calls in progress at any one time. Use the
pointers in the channels[] array instead, indexed by the channel
number from the packet.
(3) For each channel, save the result of the last call that was in
progress on that channel in conn->channels[] so that the final ACK or
ABORT packet can be replayed if necessary. Any call earlier than that
is just ignored. If we've seen the next call number in a packet, the
last one is most definitely defunct.
(4) When generating a RESPONSE packet for a connection, the call number
counter for each channel must be included in it.
(5) When parsing a RESPONSE packet for a connection, the call number
counters contained therein should be used to set the minimum expected
call numbers on each channel.
To do in future commits:
(1) Replay terminal packets based on the last call stored in
conn->channels[].
(2) Connections should be retired before the callNumber space on any
channel runs out.
(3) A server is expected to disregard or reject any new incoming call that
has a call number less than the current call number counter. The call
number counter for that channel must be advanced to the new call
number.
Note that the server cannot just require that the next call that it
sees on a channel be exactly the call number counter + 1 because then
there's a scenario that could cause a problem: The client transmits a
packet to initiate a connection, the network goes out, the server
sends an ACK (which gets lost), the client sends an ABORT (which also
gets lost); the network then reconnects, the client then reuses the
call number for the next call (it doesn't know the server already saw
the call number), but the server thinks it already has the first
packet of this call (it doesn't know that the client doesn't know that
it saw the call number the first time).
Signed-off-by: David Howells <dhowells@redhat.com>
Add RCU destruction for connections and calls as the RCU lookup from the
transport socket data_ready handler is going to come along shortly.
Whilst we're at it, move the cleanup workqueue flushing and RCU barrierage
into the destruction code for the objects that need it (locals and
connections) and add the extra RCU barrier required for connection cleanup.
Signed-off-by: David Howells <dhowells@redhat.com>
Rather than calling rxrpc_get_connection() manually before calling
rxrpc_queue_conn(), do it inside the queue wrapper.
This allows us to do some important fixes:
(1) If the usage count is 0, do nothing. This prevents connections from
being reanimated once they're dead.
(2) If rxrpc_queue_work() fails because the work item is already queued,
retract the usage count increment which would otherwise be lost.
(3) Don't take a ref on the connection in the work function. By passing
the ref through the work item, this is unnecessary. Doing it in the
work function is too late anyway. Previously, connection-directed
packets held a ref on the connection, but that's not really the best
idea.
And another useful changes:
(*) Don't need to take a refcount on the connection in the data_ready
handler unless we invoke the connection's work item. We're using RCU
there so that's otherwise redundant.
Signed-off-by: David Howells <dhowells@redhat.com>
Turn the connection event and state #define lists into enums and move
outside of the struct definition.
Whilst we're at it, change _SERVER to _SERVICE in those identifiers and add
EV_ into the event name to distinguish them from flags and states.
Also add a symbol indicating the number of states and use that in the state
text array.
Signed-off-by: David Howells <dhowells@redhat.com>
Provide queueing helper functions so that the queueing of local and
connection objects can be fixed later.
The issue is that a ref on the object needs to be passed to the work queue,
but the act of queueing the object may fail because the object is already
queued. Testing the queuedness of an object before hand doesn't work
because there can be a race with someone else trying to queue it. What
will have to be done is to adjust the refcount depending on the result of
the queue operation.
Signed-off-by: David Howells <dhowells@redhat.com>
rxkad uses stack memory in SG lists which would not work if stacks were
allocated from vmalloc memory. In fact, in most cases this isn't even
necessary as the stack memory ends up getting copied over to kmalloc
memory.
This patch eliminates all the unnecessary stack memory uses by supplying
the final destination directly to the crypto API. In two instances where a
temporary buffer is actually needed we also switch use a scratch area in
the rxrpc_call struct (only one DATA packet will be being secured or
verified at a time).
Finally there is no need to split a split-page buffer into two SG entries
so code dealing with that has been removed.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Andy Lutomirski <luto@kernel.org>
Signed-off-by: David Howells <dhowells@redhat.com>
Provide refcount helper functions for connections so that the code doesn't
touch local or connection usage counts directly.
Also make it such that local and peer put functions can take a NULL
pointer.
Signed-off-by: David Howells <dhowells@redhat.com>
Replace accesses of conn->trans->{local,peer} with
conn->params.{local,peer} thus making it easier for a future commit to
remove the rxrpc_transport struct.
This also reduces the number of memory accesses involved.
Signed-off-by: David Howells <dhowells@redhat.com>
Define and use a structure to hold connection parameters. This makes it
easier to pass multiple connection parameters around.
Define and use a structure to hold protocol information used to hash a
connection for lookup on incoming packet. Most of these fields will be
disposed of eventually, including the duplicate local pointer.
Whilst we're at it rename "proto" to "family" when referring to a protocol
family.
Signed-off-by: David Howells <dhowells@redhat.com>
Rework the local RxRPC endpoint management.
Local endpoint objects are maintained in a flat list as before. This
should be okay as there shouldn't be more than one per open AF_RXRPC socket
(there can be fewer as local endpoints can be shared if their local service
ID is 0 and they share the same local transport parameters).
Changes:
(1) Local endpoints may now only be shared if they have local service ID 0
(ie. they're not being used for listening).
This prevents a scenario where process A is listening of the Cache
Manager port and process B contacts a fileserver - which may then
attempt to send CM requests back to B. But if A and B are sharing a
local endpoint, A will get the CM requests meant for B.
(2) We use a mutex to handle lookups and don't provide RCU-only lookups
since we only expect to access the list when opening a socket or
destroying an endpoint.
The local endpoint object is pointed to by the transport socket's
sk_user_data for the life of the transport socket - allowing us to
refer to it directly from the sk_data_ready and sk_error_report
callbacks.
(3) atomic_inc_not_zero() now exists and can be used to only share a local
endpoint if the last reference hasn't yet gone.
(4) We can remove rxrpc_local_lock - a spinlock that had to be taken with
BH processing disabled given that we assume sk_user_data won't change
under us.
(5) The transport socket is shut down before we clear the sk_user_data
pointer so that we can be sure that the transport socket's callbacks
won't be invoked once the RCU destruction is scheduled.
(6) Local endpoints have a work item that handles both destruction and
event processing. The means that destruction doesn't then need to
wait for event processing. The event queues can then be cleared after
the transport socket is shut down.
(7) Local endpoints are no longer available for resurrection beyond the
life of the sockets that had them open. As soon as their last ref
goes, they are scheduled for destruction and may not have their usage
count moved from 0.
Signed-off-by: David Howells <dhowells@redhat.com>
Rename files matching net/rxrpc/ar-*.c to get rid of the "ar-" prefix.
This will aid splitting those files by making easier to come up with new
names.
Note that the not all files are simply renamed from ar-X.c to X.c. The
following exceptions are made:
(*) ar-call.c -> call_object.c
ar-ack.c -> call_event.c
call_object.c is going to contain the core of the call object
handling. Call event handling is all going to be in call_event.c.
(*) ar-accept.c -> call_accept.c
Incoming call handling is going to be here.
(*) ar-connection.c -> conn_object.c
ar-connevent.c -> conn_event.c
The former file is going to have the basic connection object handling,
but there will likely be some differentiation between client
connections and service connections in additional files later. The
latter file will have all the connection-level event handling.
(*) ar-local.c -> local_object.c
This will have the local endpoint object handling code. The local
endpoint event handling code will later be split out into
local_event.c.
(*) ar-peer.c -> peer_object.c
This will have the peer endpoint object handling code. Peer event
handling code will be placed in peer_event.c (for the moment, there is
none).
(*) ar-error.c -> peer_event.c
This will become the peer event handling code, though for the moment
it's actually driven from the local endpoint's perspective.
Note that I haven't renamed ar-transport.c to transport_object.c as the
intention is to delete it when the rxrpc_transport struct is excised.
The only file that actually has its contents changed is net/rxrpc/Makefile.
net/rxrpc/ar-internal.h will need its section marker comments updating, but
I'll do that in a separate patch to make it easier for git to follow the
history across the rename. I may also want to rename ar-internal.h at some
point - but that would mean updating all the #includes and I'd rather do
that in a separate step.
Signed-off-by: David Howells <dhowells@redhat.com.
David Howells