linux/net/dccp/ackvec.c
Gerrit Renker e28fe59f9c dccp ccid-2: Update code for the Ack Vector input/registration routine
This patch uupdates the code which registers new packets as received, using the
new circular buffer interface. It contributes a new algorithm which 
	* supports both tail/head pointers and buffer wrap-around and
	* deals with overflow (head/tail move in lock-step).

The updated code is also partioned differently, into
	1. dealing with the empty buffer,
	2. adding new packets into non-empty buffer,
	3. reserving space when encountering a `hole' in the sequence space,
	4. updating old state and deciding when old state is irrelevant.

Protection against large burst losses: With regard to (3), it is too costly to
reserve space when there are large bursts of losses. When bursts get too large,
the code does no longer reserve space and just fills in cells normally. This
measure reduces space consumption by a factor of 63.

The code reuses in part the previous implementation by Arnaldo de Melo.

Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
2008-09-04 07:45:37 +02:00

632 lines
19 KiB
C

/*
* net/dccp/ackvec.c
*
* An implementation of Ack Vectors for the DCCP protocol
* Copyright (c) 2007 University of Aberdeen, Scotland, UK
* Copyright (c) 2005 Arnaldo Carvalho de Melo <acme@ghostprotocols.net>
*
* This program is free software; you can redistribute it and/or modify it
* under the terms of the GNU General Public License as published by the
* Free Software Foundation; version 2 of the License;
*/
#include "ackvec.h"
#include "dccp.h"
#include <linux/init.h>
#include <linux/errno.h>
#include <linux/kernel.h>
#include <linux/skbuff.h>
#include <linux/slab.h>
#include <net/sock.h>
static struct kmem_cache *dccp_ackvec_slab;
static struct kmem_cache *dccp_ackvec_record_slab;
struct dccp_ackvec *dccp_ackvec_alloc(const gfp_t priority)
{
struct dccp_ackvec *av = kmem_cache_zalloc(dccp_ackvec_slab, priority);
if (av != NULL) {
av->av_buf_head = av->av_buf_tail = DCCPAV_MAX_ACKVEC_LEN - 1;
INIT_LIST_HEAD(&av->av_records);
}
return av;
}
static void dccp_ackvec_purge_records(struct dccp_ackvec *av)
{
struct dccp_ackvec_record *cur, *next;
list_for_each_entry_safe(cur, next, &av->av_records, avr_node)
kmem_cache_free(dccp_ackvec_record_slab, cur);
INIT_LIST_HEAD(&av->av_records);
}
void dccp_ackvec_free(struct dccp_ackvec *av)
{
if (likely(av != NULL)) {
dccp_ackvec_purge_records(av);
kmem_cache_free(dccp_ackvec_slab, av);
}
}
/**
* dccp_ackvec_update_records - Record information about sent Ack Vectors
* @av: Ack Vector records to update
* @seqno: Sequence number of the packet carrying the Ack Vector just sent
* @nonce_sum: The sum of all buffer nonces contained in the Ack Vector
*/
int dccp_ackvec_update_records(struct dccp_ackvec *av, u64 seqno, u8 nonce_sum)
{
struct dccp_ackvec_record *avr;
avr = kmem_cache_alloc(dccp_ackvec_record_slab, GFP_ATOMIC);
if (avr == NULL)
return -ENOBUFS;
avr->avr_ack_seqno = seqno;
avr->avr_ack_ptr = av->av_buf_head;
avr->avr_ack_ackno = av->av_buf_ackno;
avr->avr_ack_nonce = nonce_sum;
avr->avr_ack_runlen = dccp_ackvec_runlen(av->av_buf + av->av_buf_head);
/*
* When the buffer overflows, we keep no more than one record. This is
* the simplest way of disambiguating sender-Acks dating from before the
* overflow from sender-Acks which refer to after the overflow; a simple
* solution is preferable here since we are handling an exception.
*/
if (av->av_overflow)
dccp_ackvec_purge_records(av);
/*
* Since GSS is incremented for each packet, the list is automatically
* arranged in descending order of @ack_seqno.
*/
list_add(&avr->avr_node, &av->av_records);
dccp_pr_debug("Added Vector, ack_seqno=%llu, ack_ackno=%llu (rl=%u)\n",
(unsigned long long)avr->avr_ack_seqno,
(unsigned long long)avr->avr_ack_ackno,
avr->avr_ack_runlen);
return 0;
}
static struct dccp_ackvec_record *dccp_ackvec_lookup(struct list_head *av_list,
const u64 ackno)
{
struct dccp_ackvec_record *avr;
/*
* Exploit that records are inserted in descending order of sequence
* number, start with the oldest record first. If @ackno is `before'
* the earliest ack_ackno, the packet is too old to be considered.
*/
list_for_each_entry_reverse(avr, av_list, avr_node) {
if (avr->avr_ack_seqno == ackno)
return avr;
if (before48(ackno, avr->avr_ack_seqno))
break;
}
return NULL;
}
/*
* Buffer index and length computation using modulo-buffersize arithmetic.
* Note that, as pointers move from right to left, head is `before' tail.
*/
static inline u16 __ackvec_idx_add(const u16 a, const u16 b)
{
return (a + b) % DCCPAV_MAX_ACKVEC_LEN;
}
static inline u16 __ackvec_idx_sub(const u16 a, const u16 b)
{
return __ackvec_idx_add(a, DCCPAV_MAX_ACKVEC_LEN - b);
}
u16 dccp_ackvec_buflen(const struct dccp_ackvec *av)
{
if (unlikely(av->av_overflow))
return DCCPAV_MAX_ACKVEC_LEN;
return __ackvec_idx_sub(av->av_buf_tail, av->av_buf_head);
}
/**
* dccp_ackvec_update_old - Update previous state as per RFC 4340, 11.4.1
* @av: non-empty buffer to update
* @distance: negative or zero distance of @seqno from buf_ackno downward
* @seqno: the (old) sequence number whose record is to be updated
* @state: state in which packet carrying @seqno was received
*/
static void dccp_ackvec_update_old(struct dccp_ackvec *av, s64 distance,
u64 seqno, enum dccp_ackvec_states state)
{
u16 ptr = av->av_buf_head;
BUG_ON(distance > 0);
if (unlikely(dccp_ackvec_is_empty(av)))
return;
do {
u8 runlen = dccp_ackvec_runlen(av->av_buf + ptr);
if (distance + runlen >= 0) {
/*
* Only update the state if packet has not been received
* yet. This is OK as per the second table in RFC 4340,
* 11.4.1; i.e. here we are using the following table:
* RECEIVED
* 0 1 3
* S +---+---+---+
* T 0 | 0 | 0 | 0 |
* O +---+---+---+
* R 1 | 1 | 1 | 1 |
* E +---+---+---+
* D 3 | 0 | 1 | 3 |
* +---+---+---+
* The "Not Received" state was set by reserve_seats().
*/
if (av->av_buf[ptr] == DCCPAV_NOT_RECEIVED)
av->av_buf[ptr] = state;
else
dccp_pr_debug("Not changing %llu state to %u\n",
(unsigned long long)seqno, state);
break;
}
distance += runlen + 1;
ptr = __ackvec_idx_add(ptr, 1);
} while (ptr != av->av_buf_tail);
}
/* Mark @num entries after buf_head as "Not yet received". */
static void dccp_ackvec_reserve_seats(struct dccp_ackvec *av, u16 num)
{
u16 start = __ackvec_idx_add(av->av_buf_head, 1),
len = DCCPAV_MAX_ACKVEC_LEN - start;
/* check for buffer wrap-around */
if (num > len) {
memset(av->av_buf + start, DCCPAV_NOT_RECEIVED, len);
start = 0;
num -= len;
}
if (num)
memset(av->av_buf + start, DCCPAV_NOT_RECEIVED, num);
}
/**
* dccp_ackvec_add_new - Record one or more new entries in Ack Vector buffer
* @av: container of buffer to update (can be empty or non-empty)
* @num_packets: number of packets to register (must be >= 1)
* @seqno: sequence number of the first packet in @num_packets
* @state: state in which packet carrying @seqno was received
*/
static void dccp_ackvec_add_new(struct dccp_ackvec *av, u32 num_packets,
u64 seqno, enum dccp_ackvec_states state)
{
u32 num_cells = num_packets;
if (num_packets > DCCPAV_BURST_THRESH) {
u32 lost_packets = num_packets - 1;
DCCP_WARN("Warning: large burst loss (%u)\n", lost_packets);
/*
* We received 1 packet and have a loss of size "num_packets-1"
* which we squeeze into num_cells-1 rather than reserving an
* entire byte for each lost packet.
* The reason is that the vector grows in O(burst_length); when
* it grows too large there will no room left for the payload.
* This is a trade-off: if a few packets out of the burst show
* up later, their state will not be changed; it is simply too
* costly to reshuffle/reallocate/copy the buffer each time.
* Should such problems persist, we will need to switch to a
* different underlying data structure.
*/
for (num_packets = num_cells = 1; lost_packets; ++num_cells) {
u8 len = min(lost_packets, (u32)DCCPAV_MAX_RUNLEN);
av->av_buf_head = __ackvec_idx_sub(av->av_buf_head, 1);
av->av_buf[av->av_buf_head] = DCCPAV_NOT_RECEIVED | len;
lost_packets -= len;
}
}
if (num_cells + dccp_ackvec_buflen(av) >= DCCPAV_MAX_ACKVEC_LEN) {
DCCP_CRIT("Ack Vector buffer overflow: dropping old entries\n");
av->av_overflow = true;
}
av->av_buf_head = __ackvec_idx_sub(av->av_buf_head, num_packets);
if (av->av_overflow)
av->av_buf_tail = av->av_buf_head;
av->av_buf[av->av_buf_head] = state;
av->av_buf_ackno = seqno;
if (num_packets > 1)
dccp_ackvec_reserve_seats(av, num_packets - 1);
}
/**
* dccp_ackvec_input - Register incoming packet in the buffer
*/
void dccp_ackvec_input(struct dccp_ackvec *av, struct sk_buff *skb)
{
u64 seqno = DCCP_SKB_CB(skb)->dccpd_seq;
enum dccp_ackvec_states state = DCCPAV_RECEIVED;
if (dccp_ackvec_is_empty(av)) {
dccp_ackvec_add_new(av, 1, seqno, state);
av->av_tail_ackno = seqno;
} else {
s64 num_packets = dccp_delta_seqno(av->av_buf_ackno, seqno);
u8 *current_head = av->av_buf + av->av_buf_head;
if (num_packets == 1 &&
dccp_ackvec_state(current_head) == state &&
dccp_ackvec_runlen(current_head) < DCCPAV_MAX_RUNLEN) {
*current_head += 1;
av->av_buf_ackno = seqno;
} else if (num_packets > 0) {
dccp_ackvec_add_new(av, num_packets, seqno, state);
} else {
dccp_ackvec_update_old(av, num_packets, seqno, state);
}
}
}
/*
* If several packets are missing, the HC-Receiver may prefer to enter multiple
* bytes with run length 0, rather than a single byte with a larger run length;
* this simplifies table updates if one of the missing packets arrives.
*/
static inline int dccp_ackvec_set_buf_head_state(struct dccp_ackvec *av,
const unsigned int packets,
const unsigned char state)
{
unsigned int gap;
long new_head;
if (av->av_vec_len + packets > DCCPAV_MAX_ACKVEC_LEN)
return -ENOBUFS;
gap = packets - 1;
new_head = av->av_buf_head - packets;
if (new_head < 0) {
if (gap > 0) {
memset(av->av_buf, DCCPAV_NOT_RECEIVED,
gap + new_head + 1);
gap = -new_head;
}
new_head += DCCPAV_MAX_ACKVEC_LEN;
}
av->av_buf_head = new_head;
if (gap > 0)
memset(av->av_buf + av->av_buf_head + 1,
DCCPAV_NOT_RECEIVED, gap);
av->av_buf[av->av_buf_head] = state;
av->av_vec_len += packets;
return 0;
}
/*
* Implements the RFC 4340, Appendix A
*/
int dccp_ackvec_add(struct dccp_ackvec *av, const struct sock *sk,
const u64 ackno, const u8 state)
{
u8 *cur_head = av->av_buf + av->av_buf_head,
*buf_end = av->av_buf + DCCPAV_MAX_ACKVEC_LEN;
/*
* Check at the right places if the buffer is full, if it is, tell the
* caller to start dropping packets till the HC-Sender acks our ACK
* vectors, when we will free up space in av_buf.
*
* We may well decide to do buffer compression, etc, but for now lets
* just drop.
*
* From Appendix A.1.1 (`New Packets'):
*
* Of course, the circular buffer may overflow, either when the
* HC-Sender is sending data at a very high rate, when the
* HC-Receiver's acknowledgements are not reaching the HC-Sender,
* or when the HC-Sender is forgetting to acknowledge those acks
* (so the HC-Receiver is unable to clean up old state). In this
* case, the HC-Receiver should either compress the buffer (by
* increasing run lengths when possible), transfer its state to
* a larger buffer, or, as a last resort, drop all received
* packets, without processing them whatsoever, until its buffer
* shrinks again.
*/
/* See if this is the first ackno being inserted */
if (av->av_vec_len == 0) {
*cur_head = state;
av->av_vec_len = 1;
} else if (after48(ackno, av->av_buf_ackno)) {
const u64 delta = dccp_delta_seqno(av->av_buf_ackno, ackno);
/*
* Look if the state of this packet is the same as the
* previous ackno and if so if we can bump the head len.
*/
if (delta == 1 && dccp_ackvec_state(cur_head) == state &&
dccp_ackvec_runlen(cur_head) < DCCPAV_MAX_RUNLEN)
*cur_head += 1;
else if (dccp_ackvec_set_buf_head_state(av, delta, state))
return -ENOBUFS;
} else {
/*
* A.1.2. Old Packets
*
* When a packet with Sequence Number S <= buf_ackno
* arrives, the HC-Receiver will scan the table for
* the byte corresponding to S. (Indexing structures
* could reduce the complexity of this scan.)
*/
u64 delta = dccp_delta_seqno(ackno, av->av_buf_ackno);
while (1) {
const u8 len = dccp_ackvec_runlen(cur_head);
/*
* valid packets not yet in av_buf have a reserved
* entry, with a len equal to 0.
*/
if (*cur_head == DCCPAV_NOT_RECEIVED && delta == 0) {
dccp_pr_debug("Found %llu reserved seat!\n",
(unsigned long long)ackno);
*cur_head = state;
goto out;
}
/* len == 0 means one packet */
if (delta < len + 1)
goto out_duplicate;
delta -= len + 1;
if (++cur_head == buf_end)
cur_head = av->av_buf;
}
}
av->av_buf_ackno = ackno;
out:
return 0;
out_duplicate:
/* Duplicate packet */
dccp_pr_debug("Received a dup or already considered lost "
"packet: %llu\n", (unsigned long long)ackno);
return -EILSEQ;
}
static void dccp_ackvec_throw_record(struct dccp_ackvec *av,
struct dccp_ackvec_record *avr)
{
struct dccp_ackvec_record *next;
/* sort out vector length */
if (av->av_buf_head <= avr->avr_ack_ptr)
av->av_vec_len = avr->avr_ack_ptr - av->av_buf_head;
else
av->av_vec_len = DCCPAV_MAX_ACKVEC_LEN - 1 -
av->av_buf_head + avr->avr_ack_ptr;
/* free records */
list_for_each_entry_safe_from(avr, next, &av->av_records, avr_node) {
list_del(&avr->avr_node);
kmem_cache_free(dccp_ackvec_record_slab, avr);
}
}
void dccp_ackvec_check_rcv_ackno(struct dccp_ackvec *av, struct sock *sk,
const u64 ackno)
{
struct dccp_ackvec_record *avr;
/*
* If we traverse backwards, it should be faster when we have large
* windows. We will be receiving ACKs for stuff we sent a while back
* -sorbo.
*/
list_for_each_entry_reverse(avr, &av->av_records, avr_node) {
if (ackno == avr->avr_ack_seqno) {
dccp_pr_debug("%s ACK packet 0, len=%d, ack_seqno=%llu, "
"ack_ackno=%llu, ACKED!\n",
dccp_role(sk), avr->avr_ack_runlen,
(unsigned long long)avr->avr_ack_seqno,
(unsigned long long)avr->avr_ack_ackno);
dccp_ackvec_throw_record(av, avr);
break;
} else if (avr->avr_ack_seqno > ackno)
break; /* old news */
}
}
static void dccp_ackvec_check_rcv_ackvector(struct dccp_ackvec *av,
struct sock *sk, u64 *ackno,
const unsigned char len,
const unsigned char *vector)
{
unsigned char i;
struct dccp_ackvec_record *avr;
/* Check if we actually sent an ACK vector */
if (list_empty(&av->av_records))
return;
i = len;
/*
* XXX
* I think it might be more efficient to work backwards. See comment on
* rcv_ackno. -sorbo.
*/
avr = list_entry(av->av_records.next, struct dccp_ackvec_record, avr_node);
while (i--) {
const u8 rl = dccp_ackvec_runlen(vector);
u64 ackno_end_rl;
dccp_set_seqno(&ackno_end_rl, *ackno - rl);
/*
* If our AVR sequence number is greater than the ack, go
* forward in the AVR list until it is not so.
*/
list_for_each_entry_from(avr, &av->av_records, avr_node) {
if (!after48(avr->avr_ack_seqno, *ackno))
goto found;
}
/* End of the av_records list, not found, exit */
break;
found:
if (between48(avr->avr_ack_seqno, ackno_end_rl, *ackno)) {
if (dccp_ackvec_state(vector) != DCCPAV_NOT_RECEIVED) {
dccp_pr_debug("%s ACK vector 0, len=%d, "
"ack_seqno=%llu, ack_ackno=%llu, "
"ACKED!\n",
dccp_role(sk), len,
(unsigned long long)
avr->avr_ack_seqno,
(unsigned long long)
avr->avr_ack_ackno);
dccp_ackvec_throw_record(av, avr);
break;
}
/*
* If it wasn't received, continue scanning... we might
* find another one.
*/
}
dccp_set_seqno(ackno, ackno_end_rl - 1);
++vector;
}
}
int dccp_ackvec_parse(struct sock *sk, const struct sk_buff *skb,
u64 *ackno, const u8 opt, const u8 *value, const u8 len)
{
if (len > DCCP_SINGLE_OPT_MAXLEN)
return -1;
/* dccp_ackvector_print(DCCP_SKB_CB(skb)->dccpd_ack_seq, value, len); */
dccp_ackvec_check_rcv_ackvector(dccp_sk(sk)->dccps_hc_rx_ackvec, sk,
ackno, len, value);
return 0;
}
/**
* dccp_ackvec_clear_state - Perform house-keeping / garbage-collection
* This routine is called when the peer acknowledges the receipt of Ack Vectors
* up to and including @ackno. While based on on section A.3 of RFC 4340, here
* are additional precautions to prevent corrupted buffer state. In particular,
* we use tail_ackno to identify outdated records; it always marks the earliest
* packet of group (2) in 11.4.2.
*/
void dccp_ackvec_clear_state(struct dccp_ackvec *av, const u64 ackno)
{
struct dccp_ackvec_record *avr, *next;
u8 runlen_now, eff_runlen;
s64 delta;
avr = dccp_ackvec_lookup(&av->av_records, ackno);
if (avr == NULL)
return;
/*
* Deal with outdated acknowledgments: this arises when e.g. there are
* several old records and the acks from the peer come in slowly. In
* that case we may still have records that pre-date tail_ackno.
*/
delta = dccp_delta_seqno(av->av_tail_ackno, avr->avr_ack_ackno);
if (delta < 0)
goto free_records;
/*
* Deal with overlapping Ack Vectors: don't subtract more than the
* number of packets between tail_ackno and ack_ackno.
*/
eff_runlen = delta < avr->avr_ack_runlen ? delta : avr->avr_ack_runlen;
runlen_now = dccp_ackvec_runlen(av->av_buf + avr->avr_ack_ptr);
/*
* The run length of Ack Vector cells does not decrease over time. If
* the run length is the same as at the time the Ack Vector was sent, we
* free the ack_ptr cell. That cell can however not be freed if the run
* length has increased: in this case we need to move the tail pointer
* backwards (towards higher indices), to its next-oldest neighbour.
*/
if (runlen_now > eff_runlen) {
av->av_buf[avr->avr_ack_ptr] -= eff_runlen + 1;
av->av_buf_tail = __ackvec_idx_add(avr->avr_ack_ptr, 1);
/* This move may not have cleared the overflow flag. */
if (av->av_overflow)
av->av_overflow = (av->av_buf_head == av->av_buf_tail);
} else {
av->av_buf_tail = avr->avr_ack_ptr;
/*
* We have made sure that avr points to a valid cell within the
* buffer. This cell is either older than head, or equals head
* (empty buffer): in both cases we no longer have any overflow.
*/
av->av_overflow = 0;
}
/*
* The peer has acknowledged up to and including ack_ackno. Hence the
* first packet in group (2) of 11.4.2 is the successor of ack_ackno.
*/
av->av_tail_ackno = ADD48(avr->avr_ack_ackno, 1);
free_records:
list_for_each_entry_safe_from(avr, next, &av->av_records, avr_node) {
list_del(&avr->avr_node);
kmem_cache_free(dccp_ackvec_record_slab, avr);
}
}
int __init dccp_ackvec_init(void)
{
dccp_ackvec_slab = kmem_cache_create("dccp_ackvec",
sizeof(struct dccp_ackvec), 0,
SLAB_HWCACHE_ALIGN, NULL);
if (dccp_ackvec_slab == NULL)
goto out_err;
dccp_ackvec_record_slab = kmem_cache_create("dccp_ackvec_record",
sizeof(struct dccp_ackvec_record),
0, SLAB_HWCACHE_ALIGN, NULL);
if (dccp_ackvec_record_slab == NULL)
goto out_destroy_slab;
return 0;
out_destroy_slab:
kmem_cache_destroy(dccp_ackvec_slab);
dccp_ackvec_slab = NULL;
out_err:
DCCP_CRIT("Unable to create Ack Vector slab cache");
return -ENOBUFS;
}
void dccp_ackvec_exit(void)
{
if (dccp_ackvec_slab != NULL) {
kmem_cache_destroy(dccp_ackvec_slab);
dccp_ackvec_slab = NULL;
}
if (dccp_ackvec_record_slab != NULL) {
kmem_cache_destroy(dccp_ackvec_record_slab);
dccp_ackvec_record_slab = NULL;
}
}