A ringbuffer-based implementation of loss interval history is easier to
maintain, allocate, and update.
Details:
* access to the Loss Interval Records via macro wrappers (with safety checks);
* simplified, on-demand allocation of entries (no extra memory consumption on
lossless links); cache allocation is local to the module / exported as
service;
* provision of RFC-compliant algorithm to re-compute average loss interval;
* provision of comprehensive, new loss detection algorithm
- support for all cases of loss, including re-ordered/duplicate packets;
- waiting for NDUPACK=3 packets to fill the hole;
- updating loss records when a late-arriving packet fills a hole.
Signed-off-by: Gerrit Renker <[EMAIL PROTECTED]>
Signed-off-by: Ian McDonald <[EMAIL PROTECTED]>
---
net/dccp/ccids/lib/loss_interval.c | 158 +++++++++++++++++++++++++++++-
net/dccp/ccids/lib/loss_interval.h | 58 +++++++++++-
net/dccp/ccids/lib/packet_history.c | 183 +++++++++++++++++++++++++++++++++++
net/dccp/ccids/lib/packet_history.h | 4 +
net/dccp/ccids/lib/tfrc.h | 3 +
5 files changed, 400 insertions(+), 6 deletions(-)
diff --git a/net/dccp/ccids/lib/loss_interval.c
b/net/dccp/ccids/lib/loss_interval.c
index 0ebdc67..cde5c7f 100644
--- a/net/dccp/ccids/lib/loss_interval.c
+++ b/net/dccp/ccids/lib/loss_interval.c
@@ -1,6 +1,7 @@
/*
* net/dccp/ccids/lib/loss_interval.c
*
+ * Copyright (c) 2007 The University of Aberdeen, Scotland, UK
* Copyright (c) 2005-7 The University of Waikato, Hamilton, New Zealand.
* Copyright (c) 2005-7 Ian McDonald <[EMAIL PROTECTED]>
* Copyright (c) 2005 Arnaldo Carvalho de Melo <[EMAIL PROTECTED]>
@@ -10,12 +11,7 @@
* the Free Software Foundation; either version 2 of the License, or
* (at your option) any later version.
*/
-
-#include <linux/module.h>
#include <net/sock.h>
-#include "../../dccp.h"
-#include "loss_interval.h"
-#include "packet_history.h"
#include "tfrc.h"
#define DCCP_LI_HIST_IVAL_F_LENGTH 8
@@ -27,6 +23,51 @@ struct dccp_li_hist_entry {
u32 dccplih_interval;
};
+static struct kmem_cache *tfrc_lh_slab __read_mostly;
+/* Loss Interval weights from [RFC 3448, 5.4], scaled by 10 */
+static const int tfrc_lh_weights[NINTERVAL] = { 10, 10, 10, 10, 8, 6, 4, 2 };
+
+/*
+ * Access macros: These require that at least one entry is present in lh,
+ * and implement array semantics (0 is first, n-1 is the last of n entries).
+ */
+#define __lih_index(lh, n) LIH_INDEX((lh)->counter - (n) - 1)
+#define __lih_entry(lh, n) (lh)->ring[__lih_index(lh, n)]
+#define __curr_entry(lh) (lh)->ring[LIH_INDEX((lh)->counter - 1)]
+#define __next_entry(lh) (lh)->ring[LIH_INDEX((lh)->counter)]
+
+/* given i with 0 <= i <= k, return I_i as per the rfc3448bis notation */
+static inline u32 tfrc_lh_get_interval(struct tfrc_loss_hist *lh, u8 i)
+{
+ BUG_ON(i >= lh->counter);
+ return __lih_entry(lh, i)->li_length;
+}
+
+static inline struct tfrc_loss_interval *tfrc_lh_peek(struct tfrc_loss_hist
*lh)
+{
+ return lh->counter ? __curr_entry(lh) : NULL;
+}
+
+/*
+ * On-demand allocation and de-allocation of entries
+ */
+static struct tfrc_loss_interval *tfrc_lh_demand_next(struct tfrc_loss_hist
*lh)
+{
+ if (__next_entry(lh) == NULL)
+ __next_entry(lh) = kmem_cache_alloc(tfrc_lh_slab, GFP_ATOMIC);
+
+ return __next_entry(lh);
+}
+
+void tfrc_lh_cleanup(struct tfrc_loss_hist *lh)
+{
+ if (tfrc_lh_is_initialised(lh))
+ for (lh->counter = 0; lh->counter < LIH_SIZE; lh->counter++)
+ if (__next_entry(lh) != NULL)
+ kmem_cache_free(tfrc_lh_slab, __next_entry(lh));
+}
+EXPORT_SYMBOL_GPL(tfrc_lh_cleanup);
+
static struct kmem_cache *dccp_li_cachep __read_mostly;
static inline struct dccp_li_hist_entry *dccp_li_hist_entry_new(const gfp_t
prio)
@@ -98,6 +139,65 @@ u32 dccp_li_hist_calc_i_mean(struct list_head *list)
EXPORT_SYMBOL_GPL(dccp_li_hist_calc_i_mean);
+static void tfrc_lh_calc_i_mean(struct tfrc_loss_hist *lh)
+{
+ u32 i_i, i_tot0 = 0, i_tot1 = 0, w_tot = 0;
+ int i, k = tfrc_lh_length(lh) - 1; /* k is as in rfc3448bis, 5.4 */
+
+ for (i=0; i <= k; i++) {
+ i_i = tfrc_lh_get_interval(lh, i);
+
+ if (i < k) {
+ i_tot0 += i_i * tfrc_lh_weights[i];
+ w_tot += tfrc_lh_weights[i];
+ }
+ if (i > 0)
+ i_tot1 += i_i * tfrc_lh_weights[i-1];
+ }
+
+ BUG_ON(w_tot == 0);
+ lh->i_mean = max(i_tot0, i_tot1) / w_tot;
+}
+
+/**
+ * tfrc_lh_update_i_mean - Update the `open' loss interval I_0
+ * For recomputing p: returns `true' if p > p_prev <=> 1/p < 1/p_prev
+ */
+u8 tfrc_lh_update_i_mean(struct tfrc_loss_hist *lh, struct sk_buff *skb)
+{
+ struct tfrc_loss_interval *cur = tfrc_lh_peek(lh);
+ u32 old_i_mean = lh->i_mean;
+ s64 length;
+
+ if (cur == NULL) /* not initialised */
+ return 0;
+
+ length = dccp_delta_seqno(cur->li_seqno, DCCP_SKB_CB(skb)->dccpd_seq);
+
+ if (length - cur->li_length <= 0) /* duplicate or reordered */
+ return 0;
+
+ if (SUB16(dccp_hdr(skb)->dccph_ccval, cur->li_ccval) > 4)
+ /*
+ * Implements RFC 4342, 10.2:
+ * If a packet S (skb) exists whose seqno comes `after' the one
+ * starting the current loss interval (cur) and if the modulo-16
+ * distance from C(cur) to C(S) is greater than 4, consider all
+ * subsequent packets as belonging to a new loss interval. This
+ * test is necessary since CCVal may wrap between intervals.
+ */
+ cur->li_is_closed = 1;
+
+ if (tfrc_lh_length(lh) == 1) /* due to RFC 3448, 6.3.1 */
+ return 0;
+
+ cur->li_length = length;
+ tfrc_lh_calc_i_mean(lh);
+
+ return (lh->i_mean < old_i_mean);
+}
+EXPORT_SYMBOL_GPL(tfrc_lh_update_i_mean);
+
static int dccp_li_hist_interval_new(struct list_head *list,
const u64 seq_loss, const u8 win_loss)
{
@@ -286,6 +386,54 @@ void dccp_li_update_li(struct sock *sk,
EXPORT_SYMBOL_GPL(dccp_li_update_li);
+/* Determine if `new_loss' does begin a new loss interval [RFC 4342, 10.2] */
+static inline u8 tfrc_lh_is_new_loss(struct tfrc_loss_interval *cur,
+ struct tfrc_rx_hist_entry *new_loss)
+{
+ return dccp_delta_seqno(cur->li_seqno, new_loss->seqno) > 0
+ && (cur->li_is_closed || SUB16(new_loss->ccval, cur->li_ccval) > 4);
+}
+
+/** tfrc_lh_interval_add - Insert new record into the Loss Interval database
+ * @lh: Loss Interval database
+ * @rh: Receive history containing a fresh loss event
+ * @calc_first_li: Caller-dependent routine to compute length of first interval
+ * @sk: Used by @calc_first_li in caller-specific way
(subtyping)
+ * Updates I_mean and returns 1 if a new interval has in fact been added to
@lh.
+ */
+int tfrc_lh_interval_add(struct tfrc_loss_hist *lh, struct tfrc_rx_hist *rh,
+ u32 (*calc_first_li)(struct sock *), struct sock *sk)
+{
+ struct tfrc_loss_interval *cur = tfrc_lh_peek(lh), *new;
+
+ if (cur != NULL && !tfrc_lh_is_new_loss(cur, loss_prev(rh)))
+ return 0;
+
+ new = tfrc_lh_demand_next(lh);
+ if (unlikely(new == NULL)) {
+ DCCP_CRIT("Cannot allocate/add loss record.");
+ return 0;
+ }
+
+ new->li_seqno = loss_prev(rh)->seqno;
+ new->li_ccval = loss_prev(rh)->ccval;
+ new->li_is_closed = 0;
+
+ if (++lh->counter == 1)
+ lh->i_mean = new->li_length = (*calc_first_li)(sk);
+ else {
+ cur->li_length = dccp_delta_seqno(cur->li_seqno, new->li_seqno);
+ new->li_length = dccp_delta_seqno(new->li_seqno,
+ last_rcv(rh)->seqno);
+ if (lh->counter > (2*LIH_SIZE))
+ lh->counter -= LIH_SIZE;
+
+ tfrc_lh_calc_i_mean(lh);
+ }
+ return 1;
+}
+EXPORT_SYMBOL_GPL(tfrc_lh_interval_add);
+
int __init dccp_li_init(void)
{
dccp_li_cachep = kmem_cache_create("dccp_li_hist",
diff --git a/net/dccp/ccids/lib/loss_interval.h
b/net/dccp/ccids/lib/loss_interval.h
index 27bee92..16d37fb 100644
--- a/net/dccp/ccids/lib/loss_interval.h
+++ b/net/dccp/ccids/lib/loss_interval.h
@@ -3,6 +3,7 @@
/*
* net/dccp/ccids/lib/loss_interval.h
*
+ * Copyright (c) 2007 The University of Aberdeen, Scotland, UK
* Copyright (c) 2005-7 The University of Waikato, Hamilton, New Zealand.
* Copyright (c) 2005-7 Ian McDonald <[EMAIL PROTECTED]>
* Copyright (c) 2005 Arnaldo Carvalho de Melo <[EMAIL PROTECTED]>
@@ -12,11 +13,66 @@
* Software Foundation; either version 2 of the License, or (at your option)
* any later version.
*/
-
#include <linux/ktime.h>
#include <linux/list.h>
+#include <linux/slab.h>
+
+/*
+ * Number of loss intervals (RFC 4342, 8.6.1). The history size is one more
than
+ * NINTERVAL, since the `open' interval I_0 is always stored as the first
entry.
+ * LIH_INDEX is used to implement LIFO semantics on the array.
+ */
+#define NINTERVAL 8
+#define LIH_SIZE (NINTERVAL + 1)
+#define LIH_INDEX(ctr) (LIH_SIZE - 1 - ((ctr) % LIH_SIZE))
+
+/**
+ * tfrc_loss_interval - Loss history record for TFRC-based protocols
+ * @li_seqno: Highest received seqno before the start of loss
+ * @li_ccval: The CCVal belonging to @li_seqno
+ * @li_is_closed: Whether @li_seqno is older than 1 RTT
+ * @li_length: Loss interval sequence length
+ */
+struct tfrc_loss_interval {
+ u64 li_seqno:48,
+ li_ccval:4,
+ li_is_closed:1;
+ u32 li_length;
+};
+
+/**
+ * tfrc_loss_hist - Loss record database
+ * @ring: Circular queue managed in LIFO manner
+ * @counter: Current count of entries (can be more than %LIH_SIZE)
+ * @i_mean: Current Average Loss Interval [RFC 3448, 5.4]
+ */
+struct tfrc_loss_hist {
+ struct tfrc_loss_interval *ring[LIH_SIZE];
+ u8 counter;
+ u32 i_mean;
+};
+
+static inline void tfrc_lh_init(struct tfrc_loss_hist *lh)
+{
+ memset(lh, 0, sizeof(struct tfrc_loss_hist));
+}
+
+static inline u8 tfrc_lh_is_initialised(struct tfrc_loss_hist *lh)
+{
+ return lh->counter > 0;
+}
+
+static inline u8 tfrc_lh_length(struct tfrc_loss_hist *lh)
+{
+ return min(lh->counter, (u8)LIH_SIZE);
+}
extern void dccp_li_hist_purge(struct list_head *list);
+struct tfrc_rx_hist;
+extern int tfrc_lh_interval_add(struct tfrc_loss_hist *, struct tfrc_rx_hist
*,
+ u32 (*first_li)(struct sock *), struct sock *);
+extern u8 tfrc_lh_update_i_mean(struct tfrc_loss_hist *lh, struct sk_buff *);
+extern void tfrc_lh_cleanup(struct tfrc_loss_hist *lh);
extern u32 dccp_li_hist_calc_i_mean(struct list_head *list);
diff --git a/net/dccp/ccids/lib/packet_history.c
b/net/dccp/ccids/lib/packet_history.c
index e87e445..90b17da 100644
--- a/net/dccp/ccids/lib/packet_history.c
+++ b/net/dccp/ccids/lib/packet_history.c
@@ -172,6 +172,189 @@ void tfrc_rx_hist_cleanup(struct tfrc_rx_hist *h)
}
EXPORT_SYMBOL_GPL(tfrc_rx_hist_cleanup);
+/*
+ * Private helper functions for loss detection.
+ *
+ * In the descriptions, `Si' refers to the sequence number of entry number i,
+ * whose NDP count is `Ni' (lower case is used for variables).
+ * Note: All __after_loss functions expect that a test against duplicates has
+ * been performed already: the seqno of the skb must not be less than the
+ * seqno of loss_prev; and it must not equal that of any valid
hist_entry.
+ */
+static void __one_after_loss(struct tfrc_rx_hist *h, struct sk_buff *skb, u32
n2)
+{
+ u64 s0 = loss_prev(h)->seqno,
+ s1 = hist_entry(h, 1)->seqno,
+ s2 = DCCP_SKB_CB(skb)->dccpd_seq;
+ int n1 = hist_entry(h, 1)->ndp,
+ d12 = dccp_delta_seqno(s1, s2), d2;
+
+ if (d12 > 0) { /* S1 < S2 */
+ h->loss_count = 2;
+ tfrc_rx_hist_entry_from_skb(hist_entry(h, 2), skb, n2);
+ return;
+ }
+
+ /* S0 < S2 < S1 */
+ d2 = dccp_delta_seqno(s0, s2);
+
+ if (d2 == 1 || n2 >= d2) { /* S2 is direct successor of S0 */
+ int d21 = -d12;
+
+ if (d21 == 1 || n1 >= d21) {
+ /* hole is filled: S0, S2, and S1 are consecutive */
+ h->loss_count = 0;
+ h->loss_start = hist_index(h, 1);
+ } else
+ /* gap between S2 and S1: just update loss_prev */
+ tfrc_rx_hist_entry_from_skb(loss_prev(h), skb, n2);
+
+ } else { /* hole between S0 and S2 */
+ /*
+ * Reorder history to insert S2 between S0 and s1
+ */
+ tfrc_rx_hist_swap(&hist_entry(h, 0), &hist_entry(h, 3));
+ h->loss_start = hist_index(h, 3);
+ tfrc_rx_hist_entry_from_skb(hist_entry(h, 1), skb, n2);
+ h->loss_count = 2;
+ }
+}
+
+/* return 1 if a new loss event has been identified */
+static int __two_after_loss(struct tfrc_rx_hist *h, struct sk_buff *skb, u32
n3)
+{
+ u64 s0 = loss_prev(h)->seqno,
+ s1 = hist_entry(h, 1)->seqno,
+ s2 = hist_entry(h, 2)->seqno,
+ s3 = DCCP_SKB_CB(skb)->dccpd_seq;
+ int n1 = hist_entry(h, 1)->ndp,
+ d23 = dccp_delta_seqno(s2, s3), d13, d3, d31;
+
+ if (d23 > 0) { /* S2 < S3 */
+ h->loss_count = 3;
+ tfrc_rx_hist_entry_from_skb(hist_entry(h, 3), skb, n3);
+ return 1;
+ }
+
+ /* S3 < S2 */
+ d13 = dccp_delta_seqno(s1, s3);
+
+ if (d13 > 0) {
+ /*
+ * The sequence number order is S1, S3, S2
+ * Reorder history to insert entry between S1 and S2
+ */
+ tfrc_rx_hist_swap(&hist_entry(h, 2), &hist_entry(h, 3));
+ tfrc_rx_hist_entry_from_skb(hist_entry(h, 2), skb, n3);
+ h->loss_count = 3;
+ return 1;
+ }
+
+ /* S0 < S3 < S1 */
+ d31 = -d13;
+ d3 = dccp_delta_seqno(s0, s3);
+
+ if (d3 == 1 || n3 >= d3) { /* S3 is a successor of S0 */
+
+ if (d31 == 1 || n1 >= d31) {
+ /* hole between S0 and S1 filled by S3 */
+ int d2 = dccp_delta_seqno(s1, s2),
+ n2 = hist_entry(h, 2)->ndp;
+
+ if (d2 == 1 || n2 >= d2) {
+ /* entire hole filled by S0, S3, S1, S2 */
+ h->loss_start = hist_index(h, 2);
+ h->loss_count = 0;
+ } else {
+ /* gap remains between S1 and S2 */
+ h->loss_start = hist_index(h, 1);
+ h->loss_count = 1;
+ }
+
+ } else /* gap exists between S3 and S1, loss_count stays at 2 */
+ tfrc_rx_hist_entry_from_skb(loss_prev(h), skb, n3);
+
+ return 0;
+ }
+
+ /*
+ * The remaining case: S3 is not a successor of S0.
+ * Sequence order is S0, S3, S1, S2; reorder to insert between S0 and S1
+ */
+ tfrc_rx_hist_swap(&hist_entry(h, 0), &hist_entry(h, 3));
+ h->loss_start = hist_index(h, 3);
+ tfrc_rx_hist_entry_from_skb(hist_entry(h, 1), skb, n3);
+ h->loss_count = 3;
+
+ return 1;
+}
+
+/* recycle RX history records to continue loss detection if necessary */
+static void __three_after_loss(struct tfrc_rx_hist *h)
+{
+ /*
+ * The distance between S0 and S1 is always greater than 1 and the NDP
+ * count of S1 is smaller than this distance. Otherwise there would
+ * have been no loss. Hence it is only necessary to see whether there
+ * are further missing data packets between S1/S2 and S2/S3.
+ */
+ int d2 = tfrc_rx_hist_delta_seqno(h, 1, 2),
+ d3 = tfrc_rx_hist_delta_seqno(h, 2, 3),
+ n2 = hist_entry(h, 2)->ndp,
+ n3 = hist_entry(h, 3)->ndp;
+
+ if (d2 == 1 || n2 >= d2) { /* S2 is successor to S1 */
+
+ if (d3 == 1 || n3 >= d3) {
+ /* S3 is successor of S2: entire hole is filled */
+ h->loss_start = hist_index(h, 3);
+ h->loss_count = 0;
+ } else {
+ /* gap between S2 and S3 */
+ h->loss_start = hist_index(h, 2);
+ h->loss_count = 1;
+ }
+
+ } else { /* gap between S1 and S2 */
+ h->loss_start = hist_index(h, 1);
+ h->loss_count = 2;
+ }
+}
+
+/**
+ * tfrc_rx_handle_loss - Loss detection and further processing
+ * @h: The non-empty RX history object
+ * @lh: Loss Intervals database to update
+ * @skb: Currently received packet
+ * @ndp: The NDP count belonging to @skb
+ * @calc_first_li: Caller-dependent computation of first loss interval in @lh
+ * @sk: Used by @calc_first_li (see tfrc_lh_interval_add)
+ * Chooses action according to pending loss, updates LI database when a new
+ * loss was detected, and does required post-processing. Returns 1 when caller
+ * should send feedback, 0 otherwise.
+ */
+int tfrc_rx_handle_loss(struct tfrc_rx_hist *h,
+ struct tfrc_loss_hist *lh,
+ struct sk_buff *skb, u32 ndp,
+ u32 (*calc_first_li)(struct sock *), struct sock *sk)
+{
+ int is_new_loss = 0;
+
+ if (h->loss_count == 1) {
+ __one_after_loss(h, skb, ndp);
+ } else if (h->loss_count != 2) {
+ DCCP_BUG("invalid loss_count %d", h->loss_count);
+ } else if (__two_after_loss(h, skb, ndp)) {
+ /*
+ * Update Loss Interval database and recycle RX records
+ */
+ is_new_loss = tfrc_lh_interval_add(lh, h, calc_first_li, sk);
+ __three_after_loss(h);
+ }
+ return is_new_loss;
+}
+EXPORT_SYMBOL_GPL(tfrc_rx_handle_loss);
+
/**
* tfrc_rx_sample_rtt - Sample RTT from timestamp / CCVal
* Based on ideas presented in RFC 4342, 8.1. Returns 0 if it was not able
diff --git a/net/dccp/ccids/lib/packet_history.h
b/net/dccp/ccids/lib/packet_history.h
index 706f871..ae7c9be 100644
--- a/net/dccp/ccids/lib/packet_history.h
+++ b/net/dccp/ccids/lib/packet_history.h
@@ -184,6 +184,10 @@ static inline void tfrc_rx_hist_update(struct tfrc_rx_hist
*h,
tfrc_rx_hist_entry_from_skb(last_rcv(h), skb, ndp);
}
+struct tfrc_loss_hist;
+extern int tfrc_rx_handle_loss(struct tfrc_rx_hist *, struct tfrc_loss_hist *,
+ struct sk_buff *skb, u32 ndp,
+ u32 (*first_li)(struct sock *), struct sock *);
extern u32 tfrc_rx_sample_rtt(struct tfrc_rx_hist *, struct sk_buff *);
extern int tfrc_rx_hist_init(struct tfrc_rx_hist *);
extern void tfrc_rx_hist_cleanup(struct tfrc_rx_hist *);
diff --git a/net/dccp/ccids/lib/tfrc.h b/net/dccp/ccids/lib/tfrc.h
index ab8848c..1fb1187 100644
--- a/net/dccp/ccids/lib/tfrc.h
+++ b/net/dccp/ccids/lib/tfrc.h
@@ -17,6 +17,9 @@
#include <linux/types.h>
#include <asm/div64.h>
#include "../../dccp.h"
+/* internal includes that this module exports: */
+#include "loss_interval.h"
+#include "packet_history.h"
#ifdef CONFIG_IP_DCCP_TFRC_DEBUG
extern int tfrc_debug;
--
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to [EMAIL PROTECTED]
More majordomo info at http://vger.kernel.org/majordomo-info.html
