linux/net/rxrpc/txbuf.c

135 lines
3.5 KiB
C
Raw Normal View History

// SPDX-License-Identifier: GPL-2.0-or-later
/* RxRPC Tx data buffering.
*
* Copyright (C) 2022 Red Hat, Inc. All Rights Reserved.
* Written by David Howells (dhowells@redhat.com)
*/
#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
#include <linux/slab.h>
#include "ar-internal.h"
static atomic_t rxrpc_txbuf_debug_ids;
atomic_t rxrpc_nr_txbuf;
/*
* Allocate and partially initialise an I/O request structure.
*/
struct rxrpc_txbuf *rxrpc_alloc_txbuf(struct rxrpc_call *call, u8 packet_type,
gfp_t gfp)
{
struct rxrpc_txbuf *txb;
txb = kmalloc(sizeof(*txb), gfp);
if (txb) {
INIT_LIST_HEAD(&txb->call_link);
INIT_LIST_HEAD(&txb->tx_link);
refcount_set(&txb->ref, 1);
txb->call_debug_id = call->debug_id;
txb->debug_id = atomic_inc_return(&rxrpc_txbuf_debug_ids);
txb->space = sizeof(txb->data);
txb->len = 0;
txb->offset = 0;
txb->flags = 0;
txb->ack_why = 0;
txb->seq = call->tx_prepared + 1;
txb->wire.epoch = htonl(call->conn->proto.epoch);
txb->wire.cid = htonl(call->cid);
txb->wire.callNumber = htonl(call->call_id);
txb->wire.seq = htonl(txb->seq);
txb->wire.type = packet_type;
txb->wire.flags = call->conn->out_clientflag;
txb->wire.userStatus = 0;
txb->wire.securityIndex = call->security_ix;
txb->wire._rsvd = 0;
txb->wire.serviceId = htons(call->dest_srx.srx_service);
trace_rxrpc_txbuf(txb->debug_id,
txb->call_debug_id, txb->seq, 1,
packet_type == RXRPC_PACKET_TYPE_DATA ?
rxrpc_txbuf_alloc_data :
rxrpc_txbuf_alloc_ack);
atomic_inc(&rxrpc_nr_txbuf);
}
return txb;
}
void rxrpc_get_txbuf(struct rxrpc_txbuf *txb, enum rxrpc_txbuf_trace what)
{
int r;
__refcount_inc(&txb->ref, &r);
trace_rxrpc_txbuf(txb->debug_id, txb->call_debug_id, txb->seq, r + 1, what);
}
void rxrpc_see_txbuf(struct rxrpc_txbuf *txb, enum rxrpc_txbuf_trace what)
{
int r = refcount_read(&txb->ref);
trace_rxrpc_txbuf(txb->debug_id, txb->call_debug_id, txb->seq, r, what);
}
static void rxrpc_free_txbuf(struct rcu_head *rcu)
{
struct rxrpc_txbuf *txb = container_of(rcu, struct rxrpc_txbuf, rcu);
trace_rxrpc_txbuf(txb->debug_id, txb->call_debug_id, txb->seq, 0,
rxrpc_txbuf_free);
kfree(txb);
atomic_dec(&rxrpc_nr_txbuf);
}
void rxrpc_put_txbuf(struct rxrpc_txbuf *txb, enum rxrpc_txbuf_trace what)
{
unsigned int debug_id, call_debug_id;
rxrpc_seq_t seq;
bool dead;
int r;
if (txb) {
debug_id = txb->debug_id;
call_debug_id = txb->call_debug_id;
seq = txb->seq;
dead = __refcount_dec_and_test(&txb->ref, &r);
trace_rxrpc_txbuf(debug_id, call_debug_id, seq, r - 1, what);
if (dead)
call_rcu(&txb->rcu, rxrpc_free_txbuf);
}
}
rxrpc: Don't use a ring buffer for call Tx queue Change the way the Tx queueing works to make the following ends easier to achieve: (1) The filling of packets, the encryption of packets and the transmission of packets can be handled in parallel by separate threads, rather than rxrpc_sendmsg() allocating, filling, encrypting and transmitting each packet before moving onto the next one. (2) Get rid of the fixed-size ring which sets a hard limit on the number of packets that can be retained in the ring. This allows the number of packets to increase without having to allocate a very large ring or having variable-sized rings. [Note: the downside of this is that it's then less efficient to locate a packet for retransmission as we then have to step through a list and examine each buffer in the list.] (3) Allow the filler/encrypter to run ahead of the transmission window. (4) Make it easier to do zero copy UDP from the packet buffers. (5) Make it easier to do zero copy from userspace to the packet buffers - and thence to UDP (only if for unauthenticated connections). To that end, the following changes are made: (1) Use the new rxrpc_txbuf struct instead of sk_buff for keeping packets to be transmitted in. This allows them to be placed on multiple queues simultaneously. An sk_buff isn't really necessary as it's never passed on to lower-level networking code. (2) Keep the transmissable packets in a linked list on the call struct rather than in a ring. As a consequence, the annotation buffer isn't used either; rather a flag is set on the packet to indicate ackedness. (3) Use the RXRPC_CALL_TX_LAST flag to indicate that the last packet to be transmitted has been queued. Add RXRPC_CALL_TX_ALL_ACKED to indicate that all packets up to and including the last got hard acked. (4) Wire headers are now stored in the txbuf rather than being concocted on the stack and they're stored immediately before the data, thereby allowing zerocopy of a single span. (5) Don't bother with instant-resend on transmission failure; rather, leave it for a timer or an ACK packet to trigger. Signed-off-by: David Howells <dhowells@redhat.com> cc: Marc Dionne <marc.dionne@auristor.com> cc: linux-afs@lists.infradead.org
2022-03-31 22:55:08 +00:00
/*
* Shrink the transmit buffer.
*/
void rxrpc_shrink_call_tx_buffer(struct rxrpc_call *call)
{
struct rxrpc_txbuf *txb;
rxrpc_seq_t hard_ack = smp_load_acquire(&call->acks_hard_ack);
bool wake = false;
rxrpc: Don't use a ring buffer for call Tx queue Change the way the Tx queueing works to make the following ends easier to achieve: (1) The filling of packets, the encryption of packets and the transmission of packets can be handled in parallel by separate threads, rather than rxrpc_sendmsg() allocating, filling, encrypting and transmitting each packet before moving onto the next one. (2) Get rid of the fixed-size ring which sets a hard limit on the number of packets that can be retained in the ring. This allows the number of packets to increase without having to allocate a very large ring or having variable-sized rings. [Note: the downside of this is that it's then less efficient to locate a packet for retransmission as we then have to step through a list and examine each buffer in the list.] (3) Allow the filler/encrypter to run ahead of the transmission window. (4) Make it easier to do zero copy UDP from the packet buffers. (5) Make it easier to do zero copy from userspace to the packet buffers - and thence to UDP (only if for unauthenticated connections). To that end, the following changes are made: (1) Use the new rxrpc_txbuf struct instead of sk_buff for keeping packets to be transmitted in. This allows them to be placed on multiple queues simultaneously. An sk_buff isn't really necessary as it's never passed on to lower-level networking code. (2) Keep the transmissable packets in a linked list on the call struct rather than in a ring. As a consequence, the annotation buffer isn't used either; rather a flag is set on the packet to indicate ackedness. (3) Use the RXRPC_CALL_TX_LAST flag to indicate that the last packet to be transmitted has been queued. Add RXRPC_CALL_TX_ALL_ACKED to indicate that all packets up to and including the last got hard acked. (4) Wire headers are now stored in the txbuf rather than being concocted on the stack and they're stored immediately before the data, thereby allowing zerocopy of a single span. (5) Don't bother with instant-resend on transmission failure; rather, leave it for a timer or an ACK packet to trigger. Signed-off-by: David Howells <dhowells@redhat.com> cc: Marc Dionne <marc.dionne@auristor.com> cc: linux-afs@lists.infradead.org
2022-03-31 22:55:08 +00:00
_enter("%x/%x/%x", call->tx_bottom, call->acks_hard_ack, call->tx_top);
while ((txb = list_first_entry_or_null(&call->tx_buffer,
struct rxrpc_txbuf, call_link))) {
rxrpc: Don't use a ring buffer for call Tx queue Change the way the Tx queueing works to make the following ends easier to achieve: (1) The filling of packets, the encryption of packets and the transmission of packets can be handled in parallel by separate threads, rather than rxrpc_sendmsg() allocating, filling, encrypting and transmitting each packet before moving onto the next one. (2) Get rid of the fixed-size ring which sets a hard limit on the number of packets that can be retained in the ring. This allows the number of packets to increase without having to allocate a very large ring or having variable-sized rings. [Note: the downside of this is that it's then less efficient to locate a packet for retransmission as we then have to step through a list and examine each buffer in the list.] (3) Allow the filler/encrypter to run ahead of the transmission window. (4) Make it easier to do zero copy UDP from the packet buffers. (5) Make it easier to do zero copy from userspace to the packet buffers - and thence to UDP (only if for unauthenticated connections). To that end, the following changes are made: (1) Use the new rxrpc_txbuf struct instead of sk_buff for keeping packets to be transmitted in. This allows them to be placed on multiple queues simultaneously. An sk_buff isn't really necessary as it's never passed on to lower-level networking code. (2) Keep the transmissable packets in a linked list on the call struct rather than in a ring. As a consequence, the annotation buffer isn't used either; rather a flag is set on the packet to indicate ackedness. (3) Use the RXRPC_CALL_TX_LAST flag to indicate that the last packet to be transmitted has been queued. Add RXRPC_CALL_TX_ALL_ACKED to indicate that all packets up to and including the last got hard acked. (4) Wire headers are now stored in the txbuf rather than being concocted on the stack and they're stored immediately before the data, thereby allowing zerocopy of a single span. (5) Don't bother with instant-resend on transmission failure; rather, leave it for a timer or an ACK packet to trigger. Signed-off-by: David Howells <dhowells@redhat.com> cc: Marc Dionne <marc.dionne@auristor.com> cc: linux-afs@lists.infradead.org
2022-03-31 22:55:08 +00:00
hard_ack = smp_load_acquire(&call->acks_hard_ack);
if (before(hard_ack, txb->seq))
break;
rxrpc: Don't hold a ref for call timer or workqueue Currently, rxrpc gives the call timer a ref on the call when it starts it and this is passed along to the workqueue by the timer expiration function. The problem comes when queue_work() fails (ie. the work item is already queued): the timer routine must put the ref - but this may cause the cleanup code to run. This has the unfortunate effect that the cleanup code may then be run in softirq context - which means that any spinlocks it might need to touch have to be guarded to disable softirqs (ie. they need a "_bh" suffix). Fix this by: (1) Don't give a ref to the timer. (2) Making the expiration function not do anything if the refcount is 0. Note that this is more of an optimisation. (3) Make sure that the cleanup routine waits for timer to complete. However, this has a consequence that timer cannot give a ref to the work item. Therefore the following fixes are also necessary: (4) Don't give a ref to the work item. (5) Make the work item return asap if it sees the ref count is 0. (6) Make sure that the cleanup routine waits for the work item to complete. Unfortunately, neither the timer nor the work item can simply get around the problem by just using refcount_inc_not_zero() as the waits would still have to be done, and there would still be the possibility of having to put the ref in the expiration function. Note the call work item is going to go away with the work being transferred to the I/O thread, so the wait in (6) will become obsolete. Signed-off-by: David Howells <dhowells@redhat.com> cc: Marc Dionne <marc.dionne@auristor.com> cc: linux-afs@lists.infradead.org
2022-11-25 09:00:55 +00:00
if (txb->seq != call->tx_bottom + 1)
rxrpc_see_txbuf(txb, rxrpc_txbuf_see_out_of_step);
rxrpc: Don't use a ring buffer for call Tx queue Change the way the Tx queueing works to make the following ends easier to achieve: (1) The filling of packets, the encryption of packets and the transmission of packets can be handled in parallel by separate threads, rather than rxrpc_sendmsg() allocating, filling, encrypting and transmitting each packet before moving onto the next one. (2) Get rid of the fixed-size ring which sets a hard limit on the number of packets that can be retained in the ring. This allows the number of packets to increase without having to allocate a very large ring or having variable-sized rings. [Note: the downside of this is that it's then less efficient to locate a packet for retransmission as we then have to step through a list and examine each buffer in the list.] (3) Allow the filler/encrypter to run ahead of the transmission window. (4) Make it easier to do zero copy UDP from the packet buffers. (5) Make it easier to do zero copy from userspace to the packet buffers - and thence to UDP (only if for unauthenticated connections). To that end, the following changes are made: (1) Use the new rxrpc_txbuf struct instead of sk_buff for keeping packets to be transmitted in. This allows them to be placed on multiple queues simultaneously. An sk_buff isn't really necessary as it's never passed on to lower-level networking code. (2) Keep the transmissable packets in a linked list on the call struct rather than in a ring. As a consequence, the annotation buffer isn't used either; rather a flag is set on the packet to indicate ackedness. (3) Use the RXRPC_CALL_TX_LAST flag to indicate that the last packet to be transmitted has been queued. Add RXRPC_CALL_TX_ALL_ACKED to indicate that all packets up to and including the last got hard acked. (4) Wire headers are now stored in the txbuf rather than being concocted on the stack and they're stored immediately before the data, thereby allowing zerocopy of a single span. (5) Don't bother with instant-resend on transmission failure; rather, leave it for a timer or an ACK packet to trigger. Signed-off-by: David Howells <dhowells@redhat.com> cc: Marc Dionne <marc.dionne@auristor.com> cc: linux-afs@lists.infradead.org
2022-03-31 22:55:08 +00:00
ASSERTCMP(txb->seq, ==, call->tx_bottom + 1);
smp_store_release(&call->tx_bottom, call->tx_bottom + 1);
rxrpc: Don't use a ring buffer for call Tx queue Change the way the Tx queueing works to make the following ends easier to achieve: (1) The filling of packets, the encryption of packets and the transmission of packets can be handled in parallel by separate threads, rather than rxrpc_sendmsg() allocating, filling, encrypting and transmitting each packet before moving onto the next one. (2) Get rid of the fixed-size ring which sets a hard limit on the number of packets that can be retained in the ring. This allows the number of packets to increase without having to allocate a very large ring or having variable-sized rings. [Note: the downside of this is that it's then less efficient to locate a packet for retransmission as we then have to step through a list and examine each buffer in the list.] (3) Allow the filler/encrypter to run ahead of the transmission window. (4) Make it easier to do zero copy UDP from the packet buffers. (5) Make it easier to do zero copy from userspace to the packet buffers - and thence to UDP (only if for unauthenticated connections). To that end, the following changes are made: (1) Use the new rxrpc_txbuf struct instead of sk_buff for keeping packets to be transmitted in. This allows them to be placed on multiple queues simultaneously. An sk_buff isn't really necessary as it's never passed on to lower-level networking code. (2) Keep the transmissable packets in a linked list on the call struct rather than in a ring. As a consequence, the annotation buffer isn't used either; rather a flag is set on the packet to indicate ackedness. (3) Use the RXRPC_CALL_TX_LAST flag to indicate that the last packet to be transmitted has been queued. Add RXRPC_CALL_TX_ALL_ACKED to indicate that all packets up to and including the last got hard acked. (4) Wire headers are now stored in the txbuf rather than being concocted on the stack and they're stored immediately before the data, thereby allowing zerocopy of a single span. (5) Don't bother with instant-resend on transmission failure; rather, leave it for a timer or an ACK packet to trigger. Signed-off-by: David Howells <dhowells@redhat.com> cc: Marc Dionne <marc.dionne@auristor.com> cc: linux-afs@lists.infradead.org
2022-03-31 22:55:08 +00:00
list_del_rcu(&txb->call_link);
trace_rxrpc_txqueue(call, rxrpc_txqueue_dequeue);
rxrpc_put_txbuf(txb, rxrpc_txbuf_put_rotated);
if (after(call->acks_hard_ack, call->tx_bottom + 128))
wake = true;
rxrpc: Don't use a ring buffer for call Tx queue Change the way the Tx queueing works to make the following ends easier to achieve: (1) The filling of packets, the encryption of packets and the transmission of packets can be handled in parallel by separate threads, rather than rxrpc_sendmsg() allocating, filling, encrypting and transmitting each packet before moving onto the next one. (2) Get rid of the fixed-size ring which sets a hard limit on the number of packets that can be retained in the ring. This allows the number of packets to increase without having to allocate a very large ring or having variable-sized rings. [Note: the downside of this is that it's then less efficient to locate a packet for retransmission as we then have to step through a list and examine each buffer in the list.] (3) Allow the filler/encrypter to run ahead of the transmission window. (4) Make it easier to do zero copy UDP from the packet buffers. (5) Make it easier to do zero copy from userspace to the packet buffers - and thence to UDP (only if for unauthenticated connections). To that end, the following changes are made: (1) Use the new rxrpc_txbuf struct instead of sk_buff for keeping packets to be transmitted in. This allows them to be placed on multiple queues simultaneously. An sk_buff isn't really necessary as it's never passed on to lower-level networking code. (2) Keep the transmissable packets in a linked list on the call struct rather than in a ring. As a consequence, the annotation buffer isn't used either; rather a flag is set on the packet to indicate ackedness. (3) Use the RXRPC_CALL_TX_LAST flag to indicate that the last packet to be transmitted has been queued. Add RXRPC_CALL_TX_ALL_ACKED to indicate that all packets up to and including the last got hard acked. (4) Wire headers are now stored in the txbuf rather than being concocted on the stack and they're stored immediately before the data, thereby allowing zerocopy of a single span. (5) Don't bother with instant-resend on transmission failure; rather, leave it for a timer or an ACK packet to trigger. Signed-off-by: David Howells <dhowells@redhat.com> cc: Marc Dionne <marc.dionne@auristor.com> cc: linux-afs@lists.infradead.org
2022-03-31 22:55:08 +00:00
}
if (wake)
wake_up(&call->waitq);
rxrpc: Don't use a ring buffer for call Tx queue Change the way the Tx queueing works to make the following ends easier to achieve: (1) The filling of packets, the encryption of packets and the transmission of packets can be handled in parallel by separate threads, rather than rxrpc_sendmsg() allocating, filling, encrypting and transmitting each packet before moving onto the next one. (2) Get rid of the fixed-size ring which sets a hard limit on the number of packets that can be retained in the ring. This allows the number of packets to increase without having to allocate a very large ring or having variable-sized rings. [Note: the downside of this is that it's then less efficient to locate a packet for retransmission as we then have to step through a list and examine each buffer in the list.] (3) Allow the filler/encrypter to run ahead of the transmission window. (4) Make it easier to do zero copy UDP from the packet buffers. (5) Make it easier to do zero copy from userspace to the packet buffers - and thence to UDP (only if for unauthenticated connections). To that end, the following changes are made: (1) Use the new rxrpc_txbuf struct instead of sk_buff for keeping packets to be transmitted in. This allows them to be placed on multiple queues simultaneously. An sk_buff isn't really necessary as it's never passed on to lower-level networking code. (2) Keep the transmissable packets in a linked list on the call struct rather than in a ring. As a consequence, the annotation buffer isn't used either; rather a flag is set on the packet to indicate ackedness. (3) Use the RXRPC_CALL_TX_LAST flag to indicate that the last packet to be transmitted has been queued. Add RXRPC_CALL_TX_ALL_ACKED to indicate that all packets up to and including the last got hard acked. (4) Wire headers are now stored in the txbuf rather than being concocted on the stack and they're stored immediately before the data, thereby allowing zerocopy of a single span. (5) Don't bother with instant-resend on transmission failure; rather, leave it for a timer or an ACK packet to trigger. Signed-off-by: David Howells <dhowells@redhat.com> cc: Marc Dionne <marc.dionne@auristor.com> cc: linux-afs@lists.infradead.org
2022-03-31 22:55:08 +00:00
}